xref: /openbmc/linux/tools/perf/builtin-stat.c (revision cd5d5810)
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(char *prefix);
72 
73 static struct perf_evlist	*evsel_list;
74 
75 static struct perf_target	target = {
76 	.uid	= UINT_MAX,
77 };
78 
79 enum aggr_mode {
80 	AGGR_NONE,
81 	AGGR_GLOBAL,
82 	AGGR_SOCKET,
83 	AGGR_CORE,
84 };
85 
86 static int			run_count			=  1;
87 static bool			no_inherit			= false;
88 static bool			scale				=  true;
89 static enum aggr_mode		aggr_mode			= AGGR_GLOBAL;
90 static volatile pid_t		child_pid			= -1;
91 static bool			null_run			=  false;
92 static int			detailed_run			=  0;
93 static bool			big_num				=  true;
94 static int			big_num_opt			=  -1;
95 static const char		*csv_sep			= NULL;
96 static bool			csv_output			= false;
97 static bool			group				= false;
98 static FILE			*output				= NULL;
99 static const char		*pre_cmd			= NULL;
100 static const char		*post_cmd			= NULL;
101 static bool			sync_run			= false;
102 static unsigned int		interval			= 0;
103 static unsigned int		initial_delay			= 0;
104 static bool			forever				= false;
105 static struct timespec		ref_time;
106 static struct cpu_map		*aggr_map;
107 static int			(*aggr_get_id)(struct cpu_map *m, int cpu);
108 
109 static volatile int done = 0;
110 
111 struct perf_stat {
112 	struct stats	  res_stats[3];
113 };
114 
115 static inline void diff_timespec(struct timespec *r, struct timespec *a,
116 				 struct timespec *b)
117 {
118 	r->tv_sec = a->tv_sec - b->tv_sec;
119 	if (a->tv_nsec < b->tv_nsec) {
120 		r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
121 		r->tv_sec--;
122 	} else {
123 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
124 	}
125 }
126 
127 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
128 {
129 	return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
130 }
131 
132 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
133 {
134 	return perf_evsel__cpus(evsel)->nr;
135 }
136 
137 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
138 {
139 	memset(evsel->priv, 0, sizeof(struct perf_stat));
140 }
141 
142 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
143 {
144 	evsel->priv = zalloc(sizeof(struct perf_stat));
145 	return evsel->priv == NULL ? -ENOMEM : 0;
146 }
147 
148 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
149 {
150 	free(evsel->priv);
151 	evsel->priv = NULL;
152 }
153 
154 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
155 {
156 	void *addr;
157 	size_t sz;
158 
159 	sz = sizeof(*evsel->counts) +
160 	     (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
161 
162 	addr = zalloc(sz);
163 	if (!addr)
164 		return -ENOMEM;
165 
166 	evsel->prev_raw_counts =  addr;
167 
168 	return 0;
169 }
170 
171 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
172 {
173 	free(evsel->prev_raw_counts);
174 	evsel->prev_raw_counts = NULL;
175 }
176 
177 static void perf_evlist__free_stats(struct perf_evlist *evlist)
178 {
179 	struct perf_evsel *evsel;
180 
181 	list_for_each_entry(evsel, &evlist->entries, node) {
182 		perf_evsel__free_stat_priv(evsel);
183 		perf_evsel__free_counts(evsel);
184 		perf_evsel__free_prev_raw_counts(evsel);
185 	}
186 }
187 
188 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
189 {
190 	struct perf_evsel *evsel;
191 
192 	list_for_each_entry(evsel, &evlist->entries, node) {
193 		if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
194 		    perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
195 		    (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
196 			goto out_free;
197 	}
198 
199 	return 0;
200 
201 out_free:
202 	perf_evlist__free_stats(evlist);
203 	return -1;
204 }
205 
206 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
207 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
208 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
209 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
210 static struct stats runtime_branches_stats[MAX_NR_CPUS];
211 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
212 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
213 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
214 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
215 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
216 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
217 static struct stats walltime_nsecs_stats;
218 
219 static void perf_stat__reset_stats(struct perf_evlist *evlist)
220 {
221 	struct perf_evsel *evsel;
222 
223 	list_for_each_entry(evsel, &evlist->entries, node) {
224 		perf_evsel__reset_stat_priv(evsel);
225 		perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
226 	}
227 
228 	memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
229 	memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
230 	memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
231 	memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
232 	memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
233 	memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
234 	memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
235 	memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
236 	memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
237 	memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
238 	memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
239 	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
240 }
241 
242 static int create_perf_stat_counter(struct perf_evsel *evsel)
243 {
244 	struct perf_event_attr *attr = &evsel->attr;
245 
246 	if (scale)
247 		attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
248 				    PERF_FORMAT_TOTAL_TIME_RUNNING;
249 
250 	attr->inherit = !no_inherit;
251 
252 	if (perf_target__has_cpu(&target))
253 		return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
254 
255 	if (!perf_target__has_task(&target) &&
256 	    perf_evsel__is_group_leader(evsel)) {
257 		attr->disabled = 1;
258 		if (!initial_delay)
259 			attr->enable_on_exec = 1;
260 	}
261 
262 	return perf_evsel__open_per_thread(evsel, evsel_list->threads);
263 }
264 
265 /*
266  * Does the counter have nsecs as a unit?
267  */
268 static inline int nsec_counter(struct perf_evsel *evsel)
269 {
270 	if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
271 	    perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
272 		return 1;
273 
274 	return 0;
275 }
276 
277 /*
278  * Update various tracking values we maintain to print
279  * more semantic information such as miss/hit ratios,
280  * instruction rates, etc:
281  */
282 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
283 {
284 	if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
285 		update_stats(&runtime_nsecs_stats[0], count[0]);
286 	else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
287 		update_stats(&runtime_cycles_stats[0], count[0]);
288 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
289 		update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
290 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
291 		update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
292 	else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
293 		update_stats(&runtime_branches_stats[0], count[0]);
294 	else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
295 		update_stats(&runtime_cacherefs_stats[0], count[0]);
296 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
297 		update_stats(&runtime_l1_dcache_stats[0], count[0]);
298 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
299 		update_stats(&runtime_l1_icache_stats[0], count[0]);
300 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
301 		update_stats(&runtime_ll_cache_stats[0], count[0]);
302 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
303 		update_stats(&runtime_dtlb_cache_stats[0], count[0]);
304 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
305 		update_stats(&runtime_itlb_cache_stats[0], count[0]);
306 }
307 
308 /*
309  * Read out the results of a single counter:
310  * aggregate counts across CPUs in system-wide mode
311  */
312 static int read_counter_aggr(struct perf_evsel *counter)
313 {
314 	struct perf_stat *ps = counter->priv;
315 	u64 *count = counter->counts->aggr.values;
316 	int i;
317 
318 	if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
319 			       thread_map__nr(evsel_list->threads), scale) < 0)
320 		return -1;
321 
322 	for (i = 0; i < 3; i++)
323 		update_stats(&ps->res_stats[i], count[i]);
324 
325 	if (verbose) {
326 		fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
327 			perf_evsel__name(counter), count[0], count[1], count[2]);
328 	}
329 
330 	/*
331 	 * Save the full runtime - to allow normalization during printout:
332 	 */
333 	update_shadow_stats(counter, count);
334 
335 	return 0;
336 }
337 
338 /*
339  * Read out the results of a single counter:
340  * do not aggregate counts across CPUs in system-wide mode
341  */
342 static int read_counter(struct perf_evsel *counter)
343 {
344 	u64 *count;
345 	int cpu;
346 
347 	for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
348 		if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
349 			return -1;
350 
351 		count = counter->counts->cpu[cpu].values;
352 
353 		update_shadow_stats(counter, count);
354 	}
355 
356 	return 0;
357 }
358 
359 static void print_interval(void)
360 {
361 	static int num_print_interval;
362 	struct perf_evsel *counter;
363 	struct perf_stat *ps;
364 	struct timespec ts, rs;
365 	char prefix[64];
366 
367 	if (aggr_mode == AGGR_GLOBAL) {
368 		list_for_each_entry(counter, &evsel_list->entries, node) {
369 			ps = counter->priv;
370 			memset(ps->res_stats, 0, sizeof(ps->res_stats));
371 			read_counter_aggr(counter);
372 		}
373 	} else	{
374 		list_for_each_entry(counter, &evsel_list->entries, node) {
375 			ps = counter->priv;
376 			memset(ps->res_stats, 0, sizeof(ps->res_stats));
377 			read_counter(counter);
378 		}
379 	}
380 
381 	clock_gettime(CLOCK_MONOTONIC, &ts);
382 	diff_timespec(&rs, &ts, &ref_time);
383 	sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
384 
385 	if (num_print_interval == 0 && !csv_output) {
386 		switch (aggr_mode) {
387 		case AGGR_SOCKET:
388 			fprintf(output, "#           time socket cpus             counts events\n");
389 			break;
390 		case AGGR_CORE:
391 			fprintf(output, "#           time core         cpus             counts events\n");
392 			break;
393 		case AGGR_NONE:
394 			fprintf(output, "#           time CPU                 counts events\n");
395 			break;
396 		case AGGR_GLOBAL:
397 		default:
398 			fprintf(output, "#           time             counts events\n");
399 		}
400 	}
401 
402 	if (++num_print_interval == 25)
403 		num_print_interval = 0;
404 
405 	switch (aggr_mode) {
406 	case AGGR_CORE:
407 	case AGGR_SOCKET:
408 		print_aggr(prefix);
409 		break;
410 	case AGGR_NONE:
411 		list_for_each_entry(counter, &evsel_list->entries, node)
412 			print_counter(counter, prefix);
413 		break;
414 	case AGGR_GLOBAL:
415 	default:
416 		list_for_each_entry(counter, &evsel_list->entries, node)
417 			print_counter_aggr(counter, prefix);
418 	}
419 
420 	fflush(output);
421 }
422 
423 static void handle_initial_delay(void)
424 {
425 	struct perf_evsel *counter;
426 
427 	if (initial_delay) {
428 		const int ncpus = cpu_map__nr(evsel_list->cpus),
429 			nthreads = thread_map__nr(evsel_list->threads);
430 
431 		usleep(initial_delay * 1000);
432 		list_for_each_entry(counter, &evsel_list->entries, node)
433 			perf_evsel__enable(counter, ncpus, nthreads);
434 	}
435 }
436 
437 static int __run_perf_stat(int argc, const char **argv)
438 {
439 	char msg[512];
440 	unsigned long long t0, t1;
441 	struct perf_evsel *counter;
442 	struct timespec ts;
443 	int status = 0;
444 	const bool forks = (argc > 0);
445 
446 	if (interval) {
447 		ts.tv_sec  = interval / 1000;
448 		ts.tv_nsec = (interval % 1000) * 1000000;
449 	} else {
450 		ts.tv_sec  = 1;
451 		ts.tv_nsec = 0;
452 	}
453 
454 	if (forks) {
455 		if (perf_evlist__prepare_workload(evsel_list, &target, argv,
456 						  false, false) < 0) {
457 			perror("failed to prepare workload");
458 			return -1;
459 		}
460 		child_pid = evsel_list->workload.pid;
461 	}
462 
463 	if (group)
464 		perf_evlist__set_leader(evsel_list);
465 
466 	list_for_each_entry(counter, &evsel_list->entries, node) {
467 		if (create_perf_stat_counter(counter) < 0) {
468 			/*
469 			 * PPC returns ENXIO for HW counters until 2.6.37
470 			 * (behavior changed with commit b0a873e).
471 			 */
472 			if (errno == EINVAL || errno == ENOSYS ||
473 			    errno == ENOENT || errno == EOPNOTSUPP ||
474 			    errno == ENXIO) {
475 				if (verbose)
476 					ui__warning("%s event is not supported by the kernel.\n",
477 						    perf_evsel__name(counter));
478 				counter->supported = false;
479 				continue;
480 			}
481 
482 			perf_evsel__open_strerror(counter, &target,
483 						  errno, msg, sizeof(msg));
484 			ui__error("%s\n", msg);
485 
486 			if (child_pid != -1)
487 				kill(child_pid, SIGTERM);
488 
489 			return -1;
490 		}
491 		counter->supported = true;
492 	}
493 
494 	if (perf_evlist__apply_filters(evsel_list)) {
495 		error("failed to set filter with %d (%s)\n", errno,
496 			strerror(errno));
497 		return -1;
498 	}
499 
500 	/*
501 	 * Enable counters and exec the command:
502 	 */
503 	t0 = rdclock();
504 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
505 
506 	if (forks) {
507 		perf_evlist__start_workload(evsel_list);
508 		handle_initial_delay();
509 
510 		if (interval) {
511 			while (!waitpid(child_pid, &status, WNOHANG)) {
512 				nanosleep(&ts, NULL);
513 				print_interval();
514 			}
515 		}
516 		wait(&status);
517 		if (WIFSIGNALED(status))
518 			psignal(WTERMSIG(status), argv[0]);
519 	} else {
520 		handle_initial_delay();
521 		while (!done) {
522 			nanosleep(&ts, NULL);
523 			if (interval)
524 				print_interval();
525 		}
526 	}
527 
528 	t1 = rdclock();
529 
530 	update_stats(&walltime_nsecs_stats, t1 - t0);
531 
532 	if (aggr_mode == AGGR_GLOBAL) {
533 		list_for_each_entry(counter, &evsel_list->entries, node) {
534 			read_counter_aggr(counter);
535 			perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
536 					     thread_map__nr(evsel_list->threads));
537 		}
538 	} else {
539 		list_for_each_entry(counter, &evsel_list->entries, node) {
540 			read_counter(counter);
541 			perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
542 		}
543 	}
544 
545 	return WEXITSTATUS(status);
546 }
547 
548 static int run_perf_stat(int argc __maybe_unused, const char **argv)
549 {
550 	int ret;
551 
552 	if (pre_cmd) {
553 		ret = system(pre_cmd);
554 		if (ret)
555 			return ret;
556 	}
557 
558 	if (sync_run)
559 		sync();
560 
561 	ret = __run_perf_stat(argc, argv);
562 	if (ret)
563 		return ret;
564 
565 	if (post_cmd) {
566 		ret = system(post_cmd);
567 		if (ret)
568 			return ret;
569 	}
570 
571 	return ret;
572 }
573 
574 static void print_noise_pct(double total, double avg)
575 {
576 	double pct = rel_stddev_stats(total, avg);
577 
578 	if (csv_output)
579 		fprintf(output, "%s%.2f%%", csv_sep, pct);
580 	else if (pct)
581 		fprintf(output, "  ( +-%6.2f%% )", pct);
582 }
583 
584 static void print_noise(struct perf_evsel *evsel, double avg)
585 {
586 	struct perf_stat *ps;
587 
588 	if (run_count == 1)
589 		return;
590 
591 	ps = evsel->priv;
592 	print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
593 }
594 
595 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
596 {
597 	switch (aggr_mode) {
598 	case AGGR_CORE:
599 		fprintf(output, "S%d-C%*d%s%*d%s",
600 			cpu_map__id_to_socket(id),
601 			csv_output ? 0 : -8,
602 			cpu_map__id_to_cpu(id),
603 			csv_sep,
604 			csv_output ? 0 : 4,
605 			nr,
606 			csv_sep);
607 		break;
608 	case AGGR_SOCKET:
609 		fprintf(output, "S%*d%s%*d%s",
610 			csv_output ? 0 : -5,
611 			id,
612 			csv_sep,
613 			csv_output ? 0 : 4,
614 			nr,
615 			csv_sep);
616 			break;
617 	case AGGR_NONE:
618 		fprintf(output, "CPU%*d%s",
619 			csv_output ? 0 : -4,
620 			perf_evsel__cpus(evsel)->map[id], csv_sep);
621 		break;
622 	case AGGR_GLOBAL:
623 	default:
624 		break;
625 	}
626 }
627 
628 static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
629 {
630 	double msecs = avg / 1e6;
631 	const char *fmt = csv_output ? "%.6f%s%s" : "%18.6f%s%-25s";
632 
633 	aggr_printout(evsel, cpu, nr);
634 
635 	fprintf(output, fmt, msecs, csv_sep, perf_evsel__name(evsel));
636 
637 	if (evsel->cgrp)
638 		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
639 
640 	if (csv_output || interval)
641 		return;
642 
643 	if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
644 		fprintf(output, " # %8.3f CPUs utilized          ",
645 			avg / avg_stats(&walltime_nsecs_stats));
646 	else
647 		fprintf(output, "                                   ");
648 }
649 
650 /* used for get_ratio_color() */
651 enum grc_type {
652 	GRC_STALLED_CYCLES_FE,
653 	GRC_STALLED_CYCLES_BE,
654 	GRC_CACHE_MISSES,
655 	GRC_MAX_NR
656 };
657 
658 static const char *get_ratio_color(enum grc_type type, double ratio)
659 {
660 	static const double grc_table[GRC_MAX_NR][3] = {
661 		[GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
662 		[GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
663 		[GRC_CACHE_MISSES] 	= { 20.0, 10.0, 5.0 },
664 	};
665 	const char *color = PERF_COLOR_NORMAL;
666 
667 	if (ratio > grc_table[type][0])
668 		color = PERF_COLOR_RED;
669 	else if (ratio > grc_table[type][1])
670 		color = PERF_COLOR_MAGENTA;
671 	else if (ratio > grc_table[type][2])
672 		color = PERF_COLOR_YELLOW;
673 
674 	return color;
675 }
676 
677 static void print_stalled_cycles_frontend(int cpu,
678 					  struct perf_evsel *evsel
679 					  __maybe_unused, double avg)
680 {
681 	double total, ratio = 0.0;
682 	const char *color;
683 
684 	total = avg_stats(&runtime_cycles_stats[cpu]);
685 
686 	if (total)
687 		ratio = avg / total * 100.0;
688 
689 	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
690 
691 	fprintf(output, " #  ");
692 	color_fprintf(output, color, "%6.2f%%", ratio);
693 	fprintf(output, " frontend cycles idle   ");
694 }
695 
696 static void print_stalled_cycles_backend(int cpu,
697 					 struct perf_evsel *evsel
698 					 __maybe_unused, double avg)
699 {
700 	double total, ratio = 0.0;
701 	const char *color;
702 
703 	total = avg_stats(&runtime_cycles_stats[cpu]);
704 
705 	if (total)
706 		ratio = avg / total * 100.0;
707 
708 	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
709 
710 	fprintf(output, " #  ");
711 	color_fprintf(output, color, "%6.2f%%", ratio);
712 	fprintf(output, " backend  cycles idle   ");
713 }
714 
715 static void print_branch_misses(int cpu,
716 				struct perf_evsel *evsel __maybe_unused,
717 				double avg)
718 {
719 	double total, ratio = 0.0;
720 	const char *color;
721 
722 	total = avg_stats(&runtime_branches_stats[cpu]);
723 
724 	if (total)
725 		ratio = avg / total * 100.0;
726 
727 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
728 
729 	fprintf(output, " #  ");
730 	color_fprintf(output, color, "%6.2f%%", ratio);
731 	fprintf(output, " of all branches        ");
732 }
733 
734 static void print_l1_dcache_misses(int cpu,
735 				   struct perf_evsel *evsel __maybe_unused,
736 				   double avg)
737 {
738 	double total, ratio = 0.0;
739 	const char *color;
740 
741 	total = avg_stats(&runtime_l1_dcache_stats[cpu]);
742 
743 	if (total)
744 		ratio = avg / total * 100.0;
745 
746 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
747 
748 	fprintf(output, " #  ");
749 	color_fprintf(output, color, "%6.2f%%", ratio);
750 	fprintf(output, " of all L1-dcache hits  ");
751 }
752 
753 static void print_l1_icache_misses(int cpu,
754 				   struct perf_evsel *evsel __maybe_unused,
755 				   double avg)
756 {
757 	double total, ratio = 0.0;
758 	const char *color;
759 
760 	total = avg_stats(&runtime_l1_icache_stats[cpu]);
761 
762 	if (total)
763 		ratio = avg / total * 100.0;
764 
765 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
766 
767 	fprintf(output, " #  ");
768 	color_fprintf(output, color, "%6.2f%%", ratio);
769 	fprintf(output, " of all L1-icache hits  ");
770 }
771 
772 static void print_dtlb_cache_misses(int cpu,
773 				    struct perf_evsel *evsel __maybe_unused,
774 				    double avg)
775 {
776 	double total, ratio = 0.0;
777 	const char *color;
778 
779 	total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
780 
781 	if (total)
782 		ratio = avg / total * 100.0;
783 
784 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
785 
786 	fprintf(output, " #  ");
787 	color_fprintf(output, color, "%6.2f%%", ratio);
788 	fprintf(output, " of all dTLB cache hits ");
789 }
790 
791 static void print_itlb_cache_misses(int cpu,
792 				    struct perf_evsel *evsel __maybe_unused,
793 				    double avg)
794 {
795 	double total, ratio = 0.0;
796 	const char *color;
797 
798 	total = avg_stats(&runtime_itlb_cache_stats[cpu]);
799 
800 	if (total)
801 		ratio = avg / total * 100.0;
802 
803 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
804 
805 	fprintf(output, " #  ");
806 	color_fprintf(output, color, "%6.2f%%", ratio);
807 	fprintf(output, " of all iTLB cache hits ");
808 }
809 
810 static void print_ll_cache_misses(int cpu,
811 				  struct perf_evsel *evsel __maybe_unused,
812 				  double avg)
813 {
814 	double total, ratio = 0.0;
815 	const char *color;
816 
817 	total = avg_stats(&runtime_ll_cache_stats[cpu]);
818 
819 	if (total)
820 		ratio = avg / total * 100.0;
821 
822 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
823 
824 	fprintf(output, " #  ");
825 	color_fprintf(output, color, "%6.2f%%", ratio);
826 	fprintf(output, " of all LL-cache hits   ");
827 }
828 
829 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
830 {
831 	double total, ratio = 0.0;
832 	const char *fmt;
833 
834 	if (csv_output)
835 		fmt = "%.0f%s%s";
836 	else if (big_num)
837 		fmt = "%'18.0f%s%-25s";
838 	else
839 		fmt = "%18.0f%s%-25s";
840 
841 	aggr_printout(evsel, cpu, nr);
842 
843 	if (aggr_mode == AGGR_GLOBAL)
844 		cpu = 0;
845 
846 	fprintf(output, fmt, avg, csv_sep, perf_evsel__name(evsel));
847 
848 	if (evsel->cgrp)
849 		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
850 
851 	if (csv_output || interval)
852 		return;
853 
854 	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
855 		total = avg_stats(&runtime_cycles_stats[cpu]);
856 		if (total)
857 			ratio = avg / total;
858 
859 		fprintf(output, " #   %5.2f  insns per cycle        ", ratio);
860 
861 		total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
862 		total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
863 
864 		if (total && avg) {
865 			ratio = total / avg;
866 			fprintf(output, "\n                                             #   %5.2f  stalled cycles per insn", ratio);
867 		}
868 
869 	} else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
870 			runtime_branches_stats[cpu].n != 0) {
871 		print_branch_misses(cpu, evsel, avg);
872 	} else if (
873 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
874 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
875 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
876 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
877 			runtime_l1_dcache_stats[cpu].n != 0) {
878 		print_l1_dcache_misses(cpu, evsel, avg);
879 	} else if (
880 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
881 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
882 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
883 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
884 			runtime_l1_icache_stats[cpu].n != 0) {
885 		print_l1_icache_misses(cpu, evsel, avg);
886 	} else if (
887 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
888 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
889 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
890 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
891 			runtime_dtlb_cache_stats[cpu].n != 0) {
892 		print_dtlb_cache_misses(cpu, evsel, avg);
893 	} else if (
894 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
895 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
896 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
897 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
898 			runtime_itlb_cache_stats[cpu].n != 0) {
899 		print_itlb_cache_misses(cpu, evsel, avg);
900 	} else if (
901 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
902 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
903 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
904 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
905 			runtime_ll_cache_stats[cpu].n != 0) {
906 		print_ll_cache_misses(cpu, evsel, avg);
907 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
908 			runtime_cacherefs_stats[cpu].n != 0) {
909 		total = avg_stats(&runtime_cacherefs_stats[cpu]);
910 
911 		if (total)
912 			ratio = avg * 100 / total;
913 
914 		fprintf(output, " # %8.3f %% of all cache refs    ", ratio);
915 
916 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
917 		print_stalled_cycles_frontend(cpu, evsel, avg);
918 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
919 		print_stalled_cycles_backend(cpu, evsel, avg);
920 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
921 		total = avg_stats(&runtime_nsecs_stats[cpu]);
922 
923 		if (total)
924 			ratio = 1.0 * avg / total;
925 
926 		fprintf(output, " # %8.3f GHz                    ", ratio);
927 	} else if (runtime_nsecs_stats[cpu].n != 0) {
928 		char unit = 'M';
929 
930 		total = avg_stats(&runtime_nsecs_stats[cpu]);
931 
932 		if (total)
933 			ratio = 1000.0 * avg / total;
934 		if (ratio < 0.001) {
935 			ratio *= 1000;
936 			unit = 'K';
937 		}
938 
939 		fprintf(output, " # %8.3f %c/sec                  ", ratio, unit);
940 	} else {
941 		fprintf(output, "                                   ");
942 	}
943 }
944 
945 static void print_aggr(char *prefix)
946 {
947 	struct perf_evsel *counter;
948 	int cpu, cpu2, s, s2, id, nr;
949 	u64 ena, run, val;
950 
951 	if (!(aggr_map || aggr_get_id))
952 		return;
953 
954 	for (s = 0; s < aggr_map->nr; s++) {
955 		id = aggr_map->map[s];
956 		list_for_each_entry(counter, &evsel_list->entries, node) {
957 			val = ena = run = 0;
958 			nr = 0;
959 			for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
960 				cpu2 = perf_evsel__cpus(counter)->map[cpu];
961 				s2 = aggr_get_id(evsel_list->cpus, cpu2);
962 				if (s2 != id)
963 					continue;
964 				val += counter->counts->cpu[cpu].val;
965 				ena += counter->counts->cpu[cpu].ena;
966 				run += counter->counts->cpu[cpu].run;
967 				nr++;
968 			}
969 			if (prefix)
970 				fprintf(output, "%s", prefix);
971 
972 			if (run == 0 || ena == 0) {
973 				aggr_printout(counter, id, nr);
974 
975 				fprintf(output, "%*s%s%*s",
976 					csv_output ? 0 : 18,
977 					counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
978 					csv_sep,
979 					csv_output ? 0 : -24,
980 					perf_evsel__name(counter));
981 
982 				if (counter->cgrp)
983 					fprintf(output, "%s%s",
984 						csv_sep, counter->cgrp->name);
985 
986 				fputc('\n', output);
987 				continue;
988 			}
989 
990 			if (nsec_counter(counter))
991 				nsec_printout(id, nr, counter, val);
992 			else
993 				abs_printout(id, nr, counter, val);
994 
995 			if (!csv_output) {
996 				print_noise(counter, 1.0);
997 
998 				if (run != ena)
999 					fprintf(output, "  (%.2f%%)",
1000 						100.0 * run / ena);
1001 			}
1002 			fputc('\n', output);
1003 		}
1004 	}
1005 }
1006 
1007 /*
1008  * Print out the results of a single counter:
1009  * aggregated counts in system-wide mode
1010  */
1011 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1012 {
1013 	struct perf_stat *ps = counter->priv;
1014 	double avg = avg_stats(&ps->res_stats[0]);
1015 	int scaled = counter->counts->scaled;
1016 
1017 	if (prefix)
1018 		fprintf(output, "%s", prefix);
1019 
1020 	if (scaled == -1) {
1021 		fprintf(output, "%*s%s%*s",
1022 			csv_output ? 0 : 18,
1023 			counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1024 			csv_sep,
1025 			csv_output ? 0 : -24,
1026 			perf_evsel__name(counter));
1027 
1028 		if (counter->cgrp)
1029 			fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1030 
1031 		fputc('\n', output);
1032 		return;
1033 	}
1034 
1035 	if (nsec_counter(counter))
1036 		nsec_printout(-1, 0, counter, avg);
1037 	else
1038 		abs_printout(-1, 0, counter, avg);
1039 
1040 	print_noise(counter, avg);
1041 
1042 	if (csv_output) {
1043 		fputc('\n', output);
1044 		return;
1045 	}
1046 
1047 	if (scaled) {
1048 		double avg_enabled, avg_running;
1049 
1050 		avg_enabled = avg_stats(&ps->res_stats[1]);
1051 		avg_running = avg_stats(&ps->res_stats[2]);
1052 
1053 		fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1054 	}
1055 	fprintf(output, "\n");
1056 }
1057 
1058 /*
1059  * Print out the results of a single counter:
1060  * does not use aggregated count in system-wide
1061  */
1062 static void print_counter(struct perf_evsel *counter, char *prefix)
1063 {
1064 	u64 ena, run, val;
1065 	int cpu;
1066 
1067 	for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1068 		val = counter->counts->cpu[cpu].val;
1069 		ena = counter->counts->cpu[cpu].ena;
1070 		run = counter->counts->cpu[cpu].run;
1071 
1072 		if (prefix)
1073 			fprintf(output, "%s", prefix);
1074 
1075 		if (run == 0 || ena == 0) {
1076 			fprintf(output, "CPU%*d%s%*s%s%*s",
1077 				csv_output ? 0 : -4,
1078 				perf_evsel__cpus(counter)->map[cpu], csv_sep,
1079 				csv_output ? 0 : 18,
1080 				counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1081 				csv_sep,
1082 				csv_output ? 0 : -24,
1083 				perf_evsel__name(counter));
1084 
1085 			if (counter->cgrp)
1086 				fprintf(output, "%s%s",
1087 					csv_sep, counter->cgrp->name);
1088 
1089 			fputc('\n', output);
1090 			continue;
1091 		}
1092 
1093 		if (nsec_counter(counter))
1094 			nsec_printout(cpu, 0, counter, val);
1095 		else
1096 			abs_printout(cpu, 0, counter, val);
1097 
1098 		if (!csv_output) {
1099 			print_noise(counter, 1.0);
1100 
1101 			if (run != ena)
1102 				fprintf(output, "  (%.2f%%)",
1103 					100.0 * run / ena);
1104 		}
1105 		fputc('\n', output);
1106 	}
1107 }
1108 
1109 static void print_stat(int argc, const char **argv)
1110 {
1111 	struct perf_evsel *counter;
1112 	int i;
1113 
1114 	fflush(stdout);
1115 
1116 	if (!csv_output) {
1117 		fprintf(output, "\n");
1118 		fprintf(output, " Performance counter stats for ");
1119 		if (!perf_target__has_task(&target)) {
1120 			fprintf(output, "\'%s", argv[0]);
1121 			for (i = 1; i < argc; i++)
1122 				fprintf(output, " %s", argv[i]);
1123 		} else if (target.pid)
1124 			fprintf(output, "process id \'%s", target.pid);
1125 		else
1126 			fprintf(output, "thread id \'%s", target.tid);
1127 
1128 		fprintf(output, "\'");
1129 		if (run_count > 1)
1130 			fprintf(output, " (%d runs)", run_count);
1131 		fprintf(output, ":\n\n");
1132 	}
1133 
1134 	switch (aggr_mode) {
1135 	case AGGR_CORE:
1136 	case AGGR_SOCKET:
1137 		print_aggr(NULL);
1138 		break;
1139 	case AGGR_GLOBAL:
1140 		list_for_each_entry(counter, &evsel_list->entries, node)
1141 			print_counter_aggr(counter, NULL);
1142 		break;
1143 	case AGGR_NONE:
1144 		list_for_each_entry(counter, &evsel_list->entries, node)
1145 			print_counter(counter, NULL);
1146 		break;
1147 	default:
1148 		break;
1149 	}
1150 
1151 	if (!csv_output) {
1152 		if (!null_run)
1153 			fprintf(output, "\n");
1154 		fprintf(output, " %17.9f seconds time elapsed",
1155 				avg_stats(&walltime_nsecs_stats)/1e9);
1156 		if (run_count > 1) {
1157 			fprintf(output, "                                        ");
1158 			print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1159 					avg_stats(&walltime_nsecs_stats));
1160 		}
1161 		fprintf(output, "\n\n");
1162 	}
1163 }
1164 
1165 static volatile int signr = -1;
1166 
1167 static void skip_signal(int signo)
1168 {
1169 	if ((child_pid == -1) || interval)
1170 		done = 1;
1171 
1172 	signr = signo;
1173 	/*
1174 	 * render child_pid harmless
1175 	 * won't send SIGTERM to a random
1176 	 * process in case of race condition
1177 	 * and fast PID recycling
1178 	 */
1179 	child_pid = -1;
1180 }
1181 
1182 static void sig_atexit(void)
1183 {
1184 	sigset_t set, oset;
1185 
1186 	/*
1187 	 * avoid race condition with SIGCHLD handler
1188 	 * in skip_signal() which is modifying child_pid
1189 	 * goal is to avoid send SIGTERM to a random
1190 	 * process
1191 	 */
1192 	sigemptyset(&set);
1193 	sigaddset(&set, SIGCHLD);
1194 	sigprocmask(SIG_BLOCK, &set, &oset);
1195 
1196 	if (child_pid != -1)
1197 		kill(child_pid, SIGTERM);
1198 
1199 	sigprocmask(SIG_SETMASK, &oset, NULL);
1200 
1201 	if (signr == -1)
1202 		return;
1203 
1204 	signal(signr, SIG_DFL);
1205 	kill(getpid(), signr);
1206 }
1207 
1208 static int stat__set_big_num(const struct option *opt __maybe_unused,
1209 			     const char *s __maybe_unused, int unset)
1210 {
1211 	big_num_opt = unset ? 0 : 1;
1212 	return 0;
1213 }
1214 
1215 static int perf_stat_init_aggr_mode(void)
1216 {
1217 	switch (aggr_mode) {
1218 	case AGGR_SOCKET:
1219 		if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1220 			perror("cannot build socket map");
1221 			return -1;
1222 		}
1223 		aggr_get_id = cpu_map__get_socket;
1224 		break;
1225 	case AGGR_CORE:
1226 		if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1227 			perror("cannot build core map");
1228 			return -1;
1229 		}
1230 		aggr_get_id = cpu_map__get_core;
1231 		break;
1232 	case AGGR_NONE:
1233 	case AGGR_GLOBAL:
1234 	default:
1235 		break;
1236 	}
1237 	return 0;
1238 }
1239 
1240 
1241 /*
1242  * Add default attributes, if there were no attributes specified or
1243  * if -d/--detailed, -d -d or -d -d -d is used:
1244  */
1245 static int add_default_attributes(void)
1246 {
1247 	struct perf_event_attr default_attrs[] = {
1248 
1249   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1250   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1251   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1252   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1253 
1254   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1255   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1256   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1257   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1258   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1259   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1260 
1261 };
1262 
1263 /*
1264  * Detailed stats (-d), covering the L1 and last level data caches:
1265  */
1266 	struct perf_event_attr detailed_attrs[] = {
1267 
1268   { .type = PERF_TYPE_HW_CACHE,
1269     .config =
1270 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1271 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1272 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1273 
1274   { .type = PERF_TYPE_HW_CACHE,
1275     .config =
1276 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1277 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1278 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1279 
1280   { .type = PERF_TYPE_HW_CACHE,
1281     .config =
1282 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1283 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1284 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1285 
1286   { .type = PERF_TYPE_HW_CACHE,
1287     .config =
1288 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1289 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1290 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1291 };
1292 
1293 /*
1294  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1295  */
1296 	struct perf_event_attr very_detailed_attrs[] = {
1297 
1298   { .type = PERF_TYPE_HW_CACHE,
1299     .config =
1300 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1301 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1302 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1303 
1304   { .type = PERF_TYPE_HW_CACHE,
1305     .config =
1306 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1307 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1308 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1309 
1310   { .type = PERF_TYPE_HW_CACHE,
1311     .config =
1312 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1313 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1314 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1315 
1316   { .type = PERF_TYPE_HW_CACHE,
1317     .config =
1318 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1319 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1320 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1321 
1322   { .type = PERF_TYPE_HW_CACHE,
1323     .config =
1324 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1325 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1326 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1327 
1328   { .type = PERF_TYPE_HW_CACHE,
1329     .config =
1330 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1331 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1332 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1333 
1334 };
1335 
1336 /*
1337  * Very, very detailed stats (-d -d -d), adding prefetch events:
1338  */
1339 	struct perf_event_attr very_very_detailed_attrs[] = {
1340 
1341   { .type = PERF_TYPE_HW_CACHE,
1342     .config =
1343 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1344 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1345 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1346 
1347   { .type = PERF_TYPE_HW_CACHE,
1348     .config =
1349 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1350 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1351 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1352 };
1353 
1354 	/* Set attrs if no event is selected and !null_run: */
1355 	if (null_run)
1356 		return 0;
1357 
1358 	if (!evsel_list->nr_entries) {
1359 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1360 			return -1;
1361 	}
1362 
1363 	/* Detailed events get appended to the event list: */
1364 
1365 	if (detailed_run <  1)
1366 		return 0;
1367 
1368 	/* Append detailed run extra attributes: */
1369 	if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1370 		return -1;
1371 
1372 	if (detailed_run < 2)
1373 		return 0;
1374 
1375 	/* Append very detailed run extra attributes: */
1376 	if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1377 		return -1;
1378 
1379 	if (detailed_run < 3)
1380 		return 0;
1381 
1382 	/* Append very, very detailed run extra attributes: */
1383 	return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1384 }
1385 
1386 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1387 {
1388 	bool append_file = false;
1389 	int output_fd = 0;
1390 	const char *output_name	= NULL;
1391 	const struct option options[] = {
1392 	OPT_CALLBACK('e', "event", &evsel_list, "event",
1393 		     "event selector. use 'perf list' to list available events",
1394 		     parse_events_option),
1395 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1396 		     "event filter", parse_filter),
1397 	OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1398 		    "child tasks do not inherit counters"),
1399 	OPT_STRING('p', "pid", &target.pid, "pid",
1400 		   "stat events on existing process id"),
1401 	OPT_STRING('t', "tid", &target.tid, "tid",
1402 		   "stat events on existing thread id"),
1403 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1404 		    "system-wide collection from all CPUs"),
1405 	OPT_BOOLEAN('g', "group", &group,
1406 		    "put the counters into a counter group"),
1407 	OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1408 	OPT_INCR('v', "verbose", &verbose,
1409 		    "be more verbose (show counter open errors, etc)"),
1410 	OPT_INTEGER('r', "repeat", &run_count,
1411 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1412 	OPT_BOOLEAN('n', "null", &null_run,
1413 		    "null run - dont start any counters"),
1414 	OPT_INCR('d', "detailed", &detailed_run,
1415 		    "detailed run - start a lot of events"),
1416 	OPT_BOOLEAN('S', "sync", &sync_run,
1417 		    "call sync() before starting a run"),
1418 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1419 			   "print large numbers with thousands\' separators",
1420 			   stat__set_big_num),
1421 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1422 		    "list of cpus to monitor in system-wide"),
1423 	OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1424 		    "disable CPU count aggregation", AGGR_NONE),
1425 	OPT_STRING('x', "field-separator", &csv_sep, "separator",
1426 		   "print counts with custom separator"),
1427 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1428 		     "monitor event in cgroup name only", parse_cgroups),
1429 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1430 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1431 	OPT_INTEGER(0, "log-fd", &output_fd,
1432 		    "log output to fd, instead of stderr"),
1433 	OPT_STRING(0, "pre", &pre_cmd, "command",
1434 			"command to run prior to the measured command"),
1435 	OPT_STRING(0, "post", &post_cmd, "command",
1436 			"command to run after to the measured command"),
1437 	OPT_UINTEGER('I', "interval-print", &interval,
1438 		    "print counts at regular interval in ms (>= 100)"),
1439 	OPT_SET_UINT(0, "per-socket", &aggr_mode,
1440 		     "aggregate counts per processor socket", AGGR_SOCKET),
1441 	OPT_SET_UINT(0, "per-core", &aggr_mode,
1442 		     "aggregate counts per physical processor core", AGGR_CORE),
1443 	OPT_UINTEGER('D', "delay", &initial_delay,
1444 		     "ms to wait before starting measurement after program start"),
1445 	OPT_END()
1446 	};
1447 	const char * const stat_usage[] = {
1448 		"perf stat [<options>] [<command>]",
1449 		NULL
1450 	};
1451 	int status = -ENOMEM, run_idx;
1452 	const char *mode;
1453 
1454 	setlocale(LC_ALL, "");
1455 
1456 	evsel_list = perf_evlist__new();
1457 	if (evsel_list == NULL)
1458 		return -ENOMEM;
1459 
1460 	argc = parse_options(argc, argv, options, stat_usage,
1461 		PARSE_OPT_STOP_AT_NON_OPTION);
1462 
1463 	output = stderr;
1464 	if (output_name && strcmp(output_name, "-"))
1465 		output = NULL;
1466 
1467 	if (output_name && output_fd) {
1468 		fprintf(stderr, "cannot use both --output and --log-fd\n");
1469 		usage_with_options(stat_usage, options);
1470 	}
1471 
1472 	if (output_fd < 0) {
1473 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
1474 		usage_with_options(stat_usage, options);
1475 	}
1476 
1477 	if (!output) {
1478 		struct timespec tm;
1479 		mode = append_file ? "a" : "w";
1480 
1481 		output = fopen(output_name, mode);
1482 		if (!output) {
1483 			perror("failed to create output file");
1484 			return -1;
1485 		}
1486 		clock_gettime(CLOCK_REALTIME, &tm);
1487 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1488 	} else if (output_fd > 0) {
1489 		mode = append_file ? "a" : "w";
1490 		output = fdopen(output_fd, mode);
1491 		if (!output) {
1492 			perror("Failed opening logfd");
1493 			return -errno;
1494 		}
1495 	}
1496 
1497 	if (csv_sep) {
1498 		csv_output = true;
1499 		if (!strcmp(csv_sep, "\\t"))
1500 			csv_sep = "\t";
1501 	} else
1502 		csv_sep = DEFAULT_SEPARATOR;
1503 
1504 	/*
1505 	 * let the spreadsheet do the pretty-printing
1506 	 */
1507 	if (csv_output) {
1508 		/* User explicitly passed -B? */
1509 		if (big_num_opt == 1) {
1510 			fprintf(stderr, "-B option not supported with -x\n");
1511 			usage_with_options(stat_usage, options);
1512 		} else /* Nope, so disable big number formatting */
1513 			big_num = false;
1514 	} else if (big_num_opt == 0) /* User passed --no-big-num */
1515 		big_num = false;
1516 
1517 	if (!argc && !perf_target__has_task(&target))
1518 		usage_with_options(stat_usage, options);
1519 	if (run_count < 0) {
1520 		usage_with_options(stat_usage, options);
1521 	} else if (run_count == 0) {
1522 		forever = true;
1523 		run_count = 1;
1524 	}
1525 
1526 	/* no_aggr, cgroup are for system-wide only */
1527 	if ((aggr_mode != AGGR_GLOBAL || nr_cgroups)
1528 	     && !perf_target__has_cpu(&target)) {
1529 		fprintf(stderr, "both cgroup and no-aggregation "
1530 			"modes only available in system-wide mode\n");
1531 
1532 		usage_with_options(stat_usage, options);
1533 		return -1;
1534 	}
1535 
1536 	if (add_default_attributes())
1537 		goto out;
1538 
1539 	perf_target__validate(&target);
1540 
1541 	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1542 		if (perf_target__has_task(&target))
1543 			pr_err("Problems finding threads of monitor\n");
1544 		if (perf_target__has_cpu(&target))
1545 			perror("failed to parse CPUs map");
1546 
1547 		usage_with_options(stat_usage, options);
1548 		return -1;
1549 	}
1550 	if (interval && interval < 100) {
1551 		pr_err("print interval must be >= 100ms\n");
1552 		usage_with_options(stat_usage, options);
1553 		return -1;
1554 	}
1555 
1556 	if (perf_evlist__alloc_stats(evsel_list, interval))
1557 		goto out_free_maps;
1558 
1559 	if (perf_stat_init_aggr_mode())
1560 		goto out;
1561 
1562 	/*
1563 	 * We dont want to block the signals - that would cause
1564 	 * child tasks to inherit that and Ctrl-C would not work.
1565 	 * What we want is for Ctrl-C to work in the exec()-ed
1566 	 * task, but being ignored by perf stat itself:
1567 	 */
1568 	atexit(sig_atexit);
1569 	if (!forever)
1570 		signal(SIGINT,  skip_signal);
1571 	signal(SIGCHLD, skip_signal);
1572 	signal(SIGALRM, skip_signal);
1573 	signal(SIGABRT, skip_signal);
1574 
1575 	status = 0;
1576 	for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1577 		if (run_count != 1 && verbose)
1578 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1579 				run_idx + 1);
1580 
1581 		status = run_perf_stat(argc, argv);
1582 		if (forever && status != -1) {
1583 			print_stat(argc, argv);
1584 			perf_stat__reset_stats(evsel_list);
1585 		}
1586 	}
1587 
1588 	if (!forever && status != -1 && !interval)
1589 		print_stat(argc, argv);
1590 
1591 	perf_evlist__free_stats(evsel_list);
1592 out_free_maps:
1593 	perf_evlist__delete_maps(evsel_list);
1594 out:
1595 	perf_evlist__delete(evsel_list);
1596 	return status;
1597 }
1598