xref: /openbmc/linux/tools/perf/builtin-stat.c (revision 8a10bc9d)
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/pmu.h"
50 #include "util/event.h"
51 #include "util/evlist.h"
52 #include "util/evsel.h"
53 #include "util/debug.h"
54 #include "util/color.h"
55 #include "util/stat.h"
56 #include "util/header.h"
57 #include "util/cpumap.h"
58 #include "util/thread.h"
59 #include "util/thread_map.h"
60 
61 #include <stdlib.h>
62 #include <sys/prctl.h>
63 #include <locale.h>
64 
65 #define DEFAULT_SEPARATOR	" "
66 #define CNTR_NOT_SUPPORTED	"<not supported>"
67 #define CNTR_NOT_COUNTED	"<not counted>"
68 
69 static void print_stat(int argc, const char **argv);
70 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
71 static void print_counter(struct perf_evsel *counter, char *prefix);
72 static void print_aggr(char *prefix);
73 
74 /* Default events used for perf stat -T */
75 static const char * const transaction_attrs[] = {
76 	"task-clock",
77 	"{"
78 	"instructions,"
79 	"cycles,"
80 	"cpu/cycles-t/,"
81 	"cpu/tx-start/,"
82 	"cpu/el-start/,"
83 	"cpu/cycles-ct/"
84 	"}"
85 };
86 
87 /* More limited version when the CPU does not have all events. */
88 static const char * const transaction_limited_attrs[] = {
89 	"task-clock",
90 	"{"
91 	"instructions,"
92 	"cycles,"
93 	"cpu/cycles-t/,"
94 	"cpu/tx-start/"
95 	"}"
96 };
97 
98 /* must match transaction_attrs and the beginning limited_attrs */
99 enum {
100 	T_TASK_CLOCK,
101 	T_INSTRUCTIONS,
102 	T_CYCLES,
103 	T_CYCLES_IN_TX,
104 	T_TRANSACTION_START,
105 	T_ELISION_START,
106 	T_CYCLES_IN_TX_CP,
107 };
108 
109 static struct perf_evlist	*evsel_list;
110 
111 static struct target target = {
112 	.uid	= UINT_MAX,
113 };
114 
115 enum aggr_mode {
116 	AGGR_NONE,
117 	AGGR_GLOBAL,
118 	AGGR_SOCKET,
119 	AGGR_CORE,
120 };
121 
122 static int			run_count			=  1;
123 static bool			no_inherit			= false;
124 static bool			scale				=  true;
125 static enum aggr_mode		aggr_mode			= AGGR_GLOBAL;
126 static volatile pid_t		child_pid			= -1;
127 static bool			null_run			=  false;
128 static int			detailed_run			=  0;
129 static bool			transaction_run;
130 static bool			big_num				=  true;
131 static int			big_num_opt			=  -1;
132 static const char		*csv_sep			= NULL;
133 static bool			csv_output			= false;
134 static bool			group				= false;
135 static FILE			*output				= NULL;
136 static const char		*pre_cmd			= NULL;
137 static const char		*post_cmd			= NULL;
138 static bool			sync_run			= false;
139 static unsigned int		interval			= 0;
140 static unsigned int		initial_delay			= 0;
141 static unsigned int		unit_width			= 4; /* strlen("unit") */
142 static bool			forever				= false;
143 static struct timespec		ref_time;
144 static struct cpu_map		*aggr_map;
145 static int			(*aggr_get_id)(struct cpu_map *m, int cpu);
146 
147 static volatile int done = 0;
148 
149 struct perf_stat {
150 	struct stats	  res_stats[3];
151 };
152 
153 static inline void diff_timespec(struct timespec *r, struct timespec *a,
154 				 struct timespec *b)
155 {
156 	r->tv_sec = a->tv_sec - b->tv_sec;
157 	if (a->tv_nsec < b->tv_nsec) {
158 		r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
159 		r->tv_sec--;
160 	} else {
161 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
162 	}
163 }
164 
165 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
166 {
167 	return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
168 }
169 
170 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
171 {
172 	return perf_evsel__cpus(evsel)->nr;
173 }
174 
175 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
176 {
177 	memset(evsel->priv, 0, sizeof(struct perf_stat));
178 }
179 
180 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
181 {
182 	evsel->priv = zalloc(sizeof(struct perf_stat));
183 	return evsel->priv == NULL ? -ENOMEM : 0;
184 }
185 
186 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
187 {
188 	zfree(&evsel->priv);
189 }
190 
191 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
192 {
193 	void *addr;
194 	size_t sz;
195 
196 	sz = sizeof(*evsel->counts) +
197 	     (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
198 
199 	addr = zalloc(sz);
200 	if (!addr)
201 		return -ENOMEM;
202 
203 	evsel->prev_raw_counts =  addr;
204 
205 	return 0;
206 }
207 
208 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
209 {
210 	zfree(&evsel->prev_raw_counts);
211 }
212 
213 static void perf_evlist__free_stats(struct perf_evlist *evlist)
214 {
215 	struct perf_evsel *evsel;
216 
217 	evlist__for_each(evlist, evsel) {
218 		perf_evsel__free_stat_priv(evsel);
219 		perf_evsel__free_counts(evsel);
220 		perf_evsel__free_prev_raw_counts(evsel);
221 	}
222 }
223 
224 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
225 {
226 	struct perf_evsel *evsel;
227 
228 	evlist__for_each(evlist, evsel) {
229 		if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
230 		    perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
231 		    (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
232 			goto out_free;
233 	}
234 
235 	return 0;
236 
237 out_free:
238 	perf_evlist__free_stats(evlist);
239 	return -1;
240 }
241 
242 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
243 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
244 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
245 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
246 static struct stats runtime_branches_stats[MAX_NR_CPUS];
247 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
248 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
249 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
250 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
251 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
252 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
253 static struct stats runtime_cycles_in_tx_stats[MAX_NR_CPUS];
254 static struct stats walltime_nsecs_stats;
255 static struct stats runtime_transaction_stats[MAX_NR_CPUS];
256 static struct stats runtime_elision_stats[MAX_NR_CPUS];
257 
258 static void perf_stat__reset_stats(struct perf_evlist *evlist)
259 {
260 	struct perf_evsel *evsel;
261 
262 	evlist__for_each(evlist, evsel) {
263 		perf_evsel__reset_stat_priv(evsel);
264 		perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
265 	}
266 
267 	memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
268 	memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
269 	memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
270 	memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
271 	memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
272 	memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
273 	memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
274 	memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
275 	memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
276 	memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
277 	memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
278 	memset(runtime_cycles_in_tx_stats, 0,
279 			sizeof(runtime_cycles_in_tx_stats));
280 	memset(runtime_transaction_stats, 0,
281 		sizeof(runtime_transaction_stats));
282 	memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
283 	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
284 }
285 
286 static int create_perf_stat_counter(struct perf_evsel *evsel)
287 {
288 	struct perf_event_attr *attr = &evsel->attr;
289 
290 	if (scale)
291 		attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
292 				    PERF_FORMAT_TOTAL_TIME_RUNNING;
293 
294 	attr->inherit = !no_inherit;
295 
296 	if (target__has_cpu(&target))
297 		return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
298 
299 	if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
300 		attr->disabled = 1;
301 		if (!initial_delay)
302 			attr->enable_on_exec = 1;
303 	}
304 
305 	return perf_evsel__open_per_thread(evsel, evsel_list->threads);
306 }
307 
308 /*
309  * Does the counter have nsecs as a unit?
310  */
311 static inline int nsec_counter(struct perf_evsel *evsel)
312 {
313 	if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
314 	    perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
315 		return 1;
316 
317 	return 0;
318 }
319 
320 static struct perf_evsel *nth_evsel(int n)
321 {
322 	static struct perf_evsel **array;
323 	static int array_len;
324 	struct perf_evsel *ev;
325 	int j;
326 
327 	/* Assumes this only called when evsel_list does not change anymore. */
328 	if (!array) {
329 		evlist__for_each(evsel_list, ev)
330 			array_len++;
331 		array = malloc(array_len * sizeof(void *));
332 		if (!array)
333 			exit(ENOMEM);
334 		j = 0;
335 		evlist__for_each(evsel_list, ev)
336 			array[j++] = ev;
337 	}
338 	if (n < array_len)
339 		return array[n];
340 	return NULL;
341 }
342 
343 /*
344  * Update various tracking values we maintain to print
345  * more semantic information such as miss/hit ratios,
346  * instruction rates, etc:
347  */
348 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
349 {
350 	if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
351 		update_stats(&runtime_nsecs_stats[0], count[0]);
352 	else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
353 		update_stats(&runtime_cycles_stats[0], count[0]);
354 	else if (transaction_run &&
355 		 perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX)))
356 		update_stats(&runtime_cycles_in_tx_stats[0], count[0]);
357 	else if (transaction_run &&
358 		 perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START)))
359 		update_stats(&runtime_transaction_stats[0], count[0]);
360 	else if (transaction_run &&
361 		 perf_evsel__cmp(counter, nth_evsel(T_ELISION_START)))
362 		update_stats(&runtime_elision_stats[0], count[0]);
363 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
364 		update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
365 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
366 		update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
367 	else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
368 		update_stats(&runtime_branches_stats[0], count[0]);
369 	else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
370 		update_stats(&runtime_cacherefs_stats[0], count[0]);
371 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
372 		update_stats(&runtime_l1_dcache_stats[0], count[0]);
373 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
374 		update_stats(&runtime_l1_icache_stats[0], count[0]);
375 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
376 		update_stats(&runtime_ll_cache_stats[0], count[0]);
377 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
378 		update_stats(&runtime_dtlb_cache_stats[0], count[0]);
379 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
380 		update_stats(&runtime_itlb_cache_stats[0], count[0]);
381 }
382 
383 /*
384  * Read out the results of a single counter:
385  * aggregate counts across CPUs in system-wide mode
386  */
387 static int read_counter_aggr(struct perf_evsel *counter)
388 {
389 	struct perf_stat *ps = counter->priv;
390 	u64 *count = counter->counts->aggr.values;
391 	int i;
392 
393 	if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
394 			       thread_map__nr(evsel_list->threads), scale) < 0)
395 		return -1;
396 
397 	for (i = 0; i < 3; i++)
398 		update_stats(&ps->res_stats[i], count[i]);
399 
400 	if (verbose) {
401 		fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
402 			perf_evsel__name(counter), count[0], count[1], count[2]);
403 	}
404 
405 	/*
406 	 * Save the full runtime - to allow normalization during printout:
407 	 */
408 	update_shadow_stats(counter, count);
409 
410 	return 0;
411 }
412 
413 /*
414  * Read out the results of a single counter:
415  * do not aggregate counts across CPUs in system-wide mode
416  */
417 static int read_counter(struct perf_evsel *counter)
418 {
419 	u64 *count;
420 	int cpu;
421 
422 	for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
423 		if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
424 			return -1;
425 
426 		count = counter->counts->cpu[cpu].values;
427 
428 		update_shadow_stats(counter, count);
429 	}
430 
431 	return 0;
432 }
433 
434 static void print_interval(void)
435 {
436 	static int num_print_interval;
437 	struct perf_evsel *counter;
438 	struct perf_stat *ps;
439 	struct timespec ts, rs;
440 	char prefix[64];
441 
442 	if (aggr_mode == AGGR_GLOBAL) {
443 		evlist__for_each(evsel_list, counter) {
444 			ps = counter->priv;
445 			memset(ps->res_stats, 0, sizeof(ps->res_stats));
446 			read_counter_aggr(counter);
447 		}
448 	} else	{
449 		evlist__for_each(evsel_list, counter) {
450 			ps = counter->priv;
451 			memset(ps->res_stats, 0, sizeof(ps->res_stats));
452 			read_counter(counter);
453 		}
454 	}
455 
456 	clock_gettime(CLOCK_MONOTONIC, &ts);
457 	diff_timespec(&rs, &ts, &ref_time);
458 	sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
459 
460 	if (num_print_interval == 0 && !csv_output) {
461 		switch (aggr_mode) {
462 		case AGGR_SOCKET:
463 			fprintf(output, "#           time socket cpus             counts %*s events\n", unit_width, "unit");
464 			break;
465 		case AGGR_CORE:
466 			fprintf(output, "#           time core         cpus             counts %*s events\n", unit_width, "unit");
467 			break;
468 		case AGGR_NONE:
469 			fprintf(output, "#           time CPU                counts %*s events\n", unit_width, "unit");
470 			break;
471 		case AGGR_GLOBAL:
472 		default:
473 			fprintf(output, "#           time             counts %*s events\n", unit_width, "unit");
474 		}
475 	}
476 
477 	if (++num_print_interval == 25)
478 		num_print_interval = 0;
479 
480 	switch (aggr_mode) {
481 	case AGGR_CORE:
482 	case AGGR_SOCKET:
483 		print_aggr(prefix);
484 		break;
485 	case AGGR_NONE:
486 		evlist__for_each(evsel_list, counter)
487 			print_counter(counter, prefix);
488 		break;
489 	case AGGR_GLOBAL:
490 	default:
491 		evlist__for_each(evsel_list, counter)
492 			print_counter_aggr(counter, prefix);
493 	}
494 
495 	fflush(output);
496 }
497 
498 static void handle_initial_delay(void)
499 {
500 	struct perf_evsel *counter;
501 
502 	if (initial_delay) {
503 		const int ncpus = cpu_map__nr(evsel_list->cpus),
504 			nthreads = thread_map__nr(evsel_list->threads);
505 
506 		usleep(initial_delay * 1000);
507 		evlist__for_each(evsel_list, counter)
508 			perf_evsel__enable(counter, ncpus, nthreads);
509 	}
510 }
511 
512 static volatile int workload_exec_errno;
513 
514 /*
515  * perf_evlist__prepare_workload will send a SIGUSR1
516  * if the fork fails, since we asked by setting its
517  * want_signal to true.
518  */
519 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
520 					void *ucontext __maybe_unused)
521 {
522 	workload_exec_errno = info->si_value.sival_int;
523 }
524 
525 static int __run_perf_stat(int argc, const char **argv)
526 {
527 	char msg[512];
528 	unsigned long long t0, t1;
529 	struct perf_evsel *counter;
530 	struct timespec ts;
531 	size_t l;
532 	int status = 0;
533 	const bool forks = (argc > 0);
534 
535 	if (interval) {
536 		ts.tv_sec  = interval / 1000;
537 		ts.tv_nsec = (interval % 1000) * 1000000;
538 	} else {
539 		ts.tv_sec  = 1;
540 		ts.tv_nsec = 0;
541 	}
542 
543 	if (forks) {
544 		if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
545 						  workload_exec_failed_signal) < 0) {
546 			perror("failed to prepare workload");
547 			return -1;
548 		}
549 		child_pid = evsel_list->workload.pid;
550 	}
551 
552 	if (group)
553 		perf_evlist__set_leader(evsel_list);
554 
555 	evlist__for_each(evsel_list, counter) {
556 		if (create_perf_stat_counter(counter) < 0) {
557 			/*
558 			 * PPC returns ENXIO for HW counters until 2.6.37
559 			 * (behavior changed with commit b0a873e).
560 			 */
561 			if (errno == EINVAL || errno == ENOSYS ||
562 			    errno == ENOENT || errno == EOPNOTSUPP ||
563 			    errno == ENXIO) {
564 				if (verbose)
565 					ui__warning("%s event is not supported by the kernel.\n",
566 						    perf_evsel__name(counter));
567 				counter->supported = false;
568 				continue;
569 			}
570 
571 			perf_evsel__open_strerror(counter, &target,
572 						  errno, msg, sizeof(msg));
573 			ui__error("%s\n", msg);
574 
575 			if (child_pid != -1)
576 				kill(child_pid, SIGTERM);
577 
578 			return -1;
579 		}
580 		counter->supported = true;
581 
582 		l = strlen(counter->unit);
583 		if (l > unit_width)
584 			unit_width = l;
585 	}
586 
587 	if (perf_evlist__apply_filters(evsel_list)) {
588 		error("failed to set filter with %d (%s)\n", errno,
589 			strerror(errno));
590 		return -1;
591 	}
592 
593 	/*
594 	 * Enable counters and exec the command:
595 	 */
596 	t0 = rdclock();
597 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
598 
599 	if (forks) {
600 		perf_evlist__start_workload(evsel_list);
601 		handle_initial_delay();
602 
603 		if (interval) {
604 			while (!waitpid(child_pid, &status, WNOHANG)) {
605 				nanosleep(&ts, NULL);
606 				print_interval();
607 			}
608 		}
609 		wait(&status);
610 
611 		if (workload_exec_errno) {
612 			const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
613 			pr_err("Workload failed: %s\n", emsg);
614 			return -1;
615 		}
616 
617 		if (WIFSIGNALED(status))
618 			psignal(WTERMSIG(status), argv[0]);
619 	} else {
620 		handle_initial_delay();
621 		while (!done) {
622 			nanosleep(&ts, NULL);
623 			if (interval)
624 				print_interval();
625 		}
626 	}
627 
628 	t1 = rdclock();
629 
630 	update_stats(&walltime_nsecs_stats, t1 - t0);
631 
632 	if (aggr_mode == AGGR_GLOBAL) {
633 		evlist__for_each(evsel_list, counter) {
634 			read_counter_aggr(counter);
635 			perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
636 					     thread_map__nr(evsel_list->threads));
637 		}
638 	} else {
639 		evlist__for_each(evsel_list, counter) {
640 			read_counter(counter);
641 			perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
642 		}
643 	}
644 
645 	return WEXITSTATUS(status);
646 }
647 
648 static int run_perf_stat(int argc, const char **argv)
649 {
650 	int ret;
651 
652 	if (pre_cmd) {
653 		ret = system(pre_cmd);
654 		if (ret)
655 			return ret;
656 	}
657 
658 	if (sync_run)
659 		sync();
660 
661 	ret = __run_perf_stat(argc, argv);
662 	if (ret)
663 		return ret;
664 
665 	if (post_cmd) {
666 		ret = system(post_cmd);
667 		if (ret)
668 			return ret;
669 	}
670 
671 	return ret;
672 }
673 
674 static void print_noise_pct(double total, double avg)
675 {
676 	double pct = rel_stddev_stats(total, avg);
677 
678 	if (csv_output)
679 		fprintf(output, "%s%.2f%%", csv_sep, pct);
680 	else if (pct)
681 		fprintf(output, "  ( +-%6.2f%% )", pct);
682 }
683 
684 static void print_noise(struct perf_evsel *evsel, double avg)
685 {
686 	struct perf_stat *ps;
687 
688 	if (run_count == 1)
689 		return;
690 
691 	ps = evsel->priv;
692 	print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
693 }
694 
695 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
696 {
697 	switch (aggr_mode) {
698 	case AGGR_CORE:
699 		fprintf(output, "S%d-C%*d%s%*d%s",
700 			cpu_map__id_to_socket(id),
701 			csv_output ? 0 : -8,
702 			cpu_map__id_to_cpu(id),
703 			csv_sep,
704 			csv_output ? 0 : 4,
705 			nr,
706 			csv_sep);
707 		break;
708 	case AGGR_SOCKET:
709 		fprintf(output, "S%*d%s%*d%s",
710 			csv_output ? 0 : -5,
711 			id,
712 			csv_sep,
713 			csv_output ? 0 : 4,
714 			nr,
715 			csv_sep);
716 			break;
717 	case AGGR_NONE:
718 		fprintf(output, "CPU%*d%s",
719 			csv_output ? 0 : -4,
720 			perf_evsel__cpus(evsel)->map[id], csv_sep);
721 		break;
722 	case AGGR_GLOBAL:
723 	default:
724 		break;
725 	}
726 }
727 
728 static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
729 {
730 	double msecs = avg / 1e6;
731 	const char *fmt_v, *fmt_n;
732 	char name[25];
733 
734 	fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
735 	fmt_n = csv_output ? "%s" : "%-25s";
736 
737 	aggr_printout(evsel, cpu, nr);
738 
739 	scnprintf(name, sizeof(name), "%s%s",
740 		  perf_evsel__name(evsel), csv_output ? "" : " (msec)");
741 
742 	fprintf(output, fmt_v, msecs, csv_sep);
743 
744 	if (csv_output)
745 		fprintf(output, "%s%s", evsel->unit, csv_sep);
746 	else
747 		fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
748 
749 	fprintf(output, fmt_n, name);
750 
751 	if (evsel->cgrp)
752 		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
753 
754 	if (csv_output || interval)
755 		return;
756 
757 	if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
758 		fprintf(output, " # %8.3f CPUs utilized          ",
759 			avg / avg_stats(&walltime_nsecs_stats));
760 	else
761 		fprintf(output, "                                   ");
762 }
763 
764 /* used for get_ratio_color() */
765 enum grc_type {
766 	GRC_STALLED_CYCLES_FE,
767 	GRC_STALLED_CYCLES_BE,
768 	GRC_CACHE_MISSES,
769 	GRC_MAX_NR
770 };
771 
772 static const char *get_ratio_color(enum grc_type type, double ratio)
773 {
774 	static const double grc_table[GRC_MAX_NR][3] = {
775 		[GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
776 		[GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
777 		[GRC_CACHE_MISSES] 	= { 20.0, 10.0, 5.0 },
778 	};
779 	const char *color = PERF_COLOR_NORMAL;
780 
781 	if (ratio > grc_table[type][0])
782 		color = PERF_COLOR_RED;
783 	else if (ratio > grc_table[type][1])
784 		color = PERF_COLOR_MAGENTA;
785 	else if (ratio > grc_table[type][2])
786 		color = PERF_COLOR_YELLOW;
787 
788 	return color;
789 }
790 
791 static void print_stalled_cycles_frontend(int cpu,
792 					  struct perf_evsel *evsel
793 					  __maybe_unused, double avg)
794 {
795 	double total, ratio = 0.0;
796 	const char *color;
797 
798 	total = avg_stats(&runtime_cycles_stats[cpu]);
799 
800 	if (total)
801 		ratio = avg / total * 100.0;
802 
803 	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
804 
805 	fprintf(output, " #  ");
806 	color_fprintf(output, color, "%6.2f%%", ratio);
807 	fprintf(output, " frontend cycles idle   ");
808 }
809 
810 static void print_stalled_cycles_backend(int cpu,
811 					 struct perf_evsel *evsel
812 					 __maybe_unused, double avg)
813 {
814 	double total, ratio = 0.0;
815 	const char *color;
816 
817 	total = avg_stats(&runtime_cycles_stats[cpu]);
818 
819 	if (total)
820 		ratio = avg / total * 100.0;
821 
822 	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
823 
824 	fprintf(output, " #  ");
825 	color_fprintf(output, color, "%6.2f%%", ratio);
826 	fprintf(output, " backend  cycles idle   ");
827 }
828 
829 static void print_branch_misses(int cpu,
830 				struct perf_evsel *evsel __maybe_unused,
831 				double avg)
832 {
833 	double total, ratio = 0.0;
834 	const char *color;
835 
836 	total = avg_stats(&runtime_branches_stats[cpu]);
837 
838 	if (total)
839 		ratio = avg / total * 100.0;
840 
841 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
842 
843 	fprintf(output, " #  ");
844 	color_fprintf(output, color, "%6.2f%%", ratio);
845 	fprintf(output, " of all branches        ");
846 }
847 
848 static void print_l1_dcache_misses(int cpu,
849 				   struct perf_evsel *evsel __maybe_unused,
850 				   double avg)
851 {
852 	double total, ratio = 0.0;
853 	const char *color;
854 
855 	total = avg_stats(&runtime_l1_dcache_stats[cpu]);
856 
857 	if (total)
858 		ratio = avg / total * 100.0;
859 
860 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
861 
862 	fprintf(output, " #  ");
863 	color_fprintf(output, color, "%6.2f%%", ratio);
864 	fprintf(output, " of all L1-dcache hits  ");
865 }
866 
867 static void print_l1_icache_misses(int cpu,
868 				   struct perf_evsel *evsel __maybe_unused,
869 				   double avg)
870 {
871 	double total, ratio = 0.0;
872 	const char *color;
873 
874 	total = avg_stats(&runtime_l1_icache_stats[cpu]);
875 
876 	if (total)
877 		ratio = avg / total * 100.0;
878 
879 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
880 
881 	fprintf(output, " #  ");
882 	color_fprintf(output, color, "%6.2f%%", ratio);
883 	fprintf(output, " of all L1-icache hits  ");
884 }
885 
886 static void print_dtlb_cache_misses(int cpu,
887 				    struct perf_evsel *evsel __maybe_unused,
888 				    double avg)
889 {
890 	double total, ratio = 0.0;
891 	const char *color;
892 
893 	total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
894 
895 	if (total)
896 		ratio = avg / total * 100.0;
897 
898 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
899 
900 	fprintf(output, " #  ");
901 	color_fprintf(output, color, "%6.2f%%", ratio);
902 	fprintf(output, " of all dTLB cache hits ");
903 }
904 
905 static void print_itlb_cache_misses(int cpu,
906 				    struct perf_evsel *evsel __maybe_unused,
907 				    double avg)
908 {
909 	double total, ratio = 0.0;
910 	const char *color;
911 
912 	total = avg_stats(&runtime_itlb_cache_stats[cpu]);
913 
914 	if (total)
915 		ratio = avg / total * 100.0;
916 
917 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
918 
919 	fprintf(output, " #  ");
920 	color_fprintf(output, color, "%6.2f%%", ratio);
921 	fprintf(output, " of all iTLB cache hits ");
922 }
923 
924 static void print_ll_cache_misses(int cpu,
925 				  struct perf_evsel *evsel __maybe_unused,
926 				  double avg)
927 {
928 	double total, ratio = 0.0;
929 	const char *color;
930 
931 	total = avg_stats(&runtime_ll_cache_stats[cpu]);
932 
933 	if (total)
934 		ratio = avg / total * 100.0;
935 
936 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
937 
938 	fprintf(output, " #  ");
939 	color_fprintf(output, color, "%6.2f%%", ratio);
940 	fprintf(output, " of all LL-cache hits   ");
941 }
942 
943 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
944 {
945 	double total, ratio = 0.0, total2;
946 	double sc =  evsel->scale;
947 	const char *fmt;
948 
949 	if (csv_output) {
950 		fmt = sc != 1.0 ?  "%.2f%s" : "%.0f%s";
951 	} else {
952 		if (big_num)
953 			fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
954 		else
955 			fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
956 	}
957 
958 	aggr_printout(evsel, cpu, nr);
959 
960 	if (aggr_mode == AGGR_GLOBAL)
961 		cpu = 0;
962 
963 	fprintf(output, fmt, avg, csv_sep);
964 
965 	if (evsel->unit)
966 		fprintf(output, "%-*s%s",
967 			csv_output ? 0 : unit_width,
968 			evsel->unit, csv_sep);
969 
970 	fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
971 
972 	if (evsel->cgrp)
973 		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
974 
975 	if (csv_output || interval)
976 		return;
977 
978 	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
979 		total = avg_stats(&runtime_cycles_stats[cpu]);
980 		if (total) {
981 			ratio = avg / total;
982 			fprintf(output, " #   %5.2f  insns per cycle        ", ratio);
983 		}
984 		total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
985 		total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
986 
987 		if (total && avg) {
988 			ratio = total / avg;
989 			fprintf(output, "\n");
990 			if (aggr_mode == AGGR_NONE)
991 				fprintf(output, "        ");
992 			fprintf(output, "                                                  #   %5.2f  stalled cycles per insn", ratio);
993 		}
994 
995 	} else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
996 			runtime_branches_stats[cpu].n != 0) {
997 		print_branch_misses(cpu, evsel, avg);
998 	} else if (
999 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
1000 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
1001 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1002 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1003 			runtime_l1_dcache_stats[cpu].n != 0) {
1004 		print_l1_dcache_misses(cpu, evsel, avg);
1005 	} else if (
1006 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
1007 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
1008 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1009 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1010 			runtime_l1_icache_stats[cpu].n != 0) {
1011 		print_l1_icache_misses(cpu, evsel, avg);
1012 	} else if (
1013 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
1014 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
1015 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1016 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1017 			runtime_dtlb_cache_stats[cpu].n != 0) {
1018 		print_dtlb_cache_misses(cpu, evsel, avg);
1019 	} else if (
1020 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
1021 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
1022 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1023 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1024 			runtime_itlb_cache_stats[cpu].n != 0) {
1025 		print_itlb_cache_misses(cpu, evsel, avg);
1026 	} else if (
1027 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
1028 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
1029 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1030 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1031 			runtime_ll_cache_stats[cpu].n != 0) {
1032 		print_ll_cache_misses(cpu, evsel, avg);
1033 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
1034 			runtime_cacherefs_stats[cpu].n != 0) {
1035 		total = avg_stats(&runtime_cacherefs_stats[cpu]);
1036 
1037 		if (total)
1038 			ratio = avg * 100 / total;
1039 
1040 		fprintf(output, " # %8.3f %% of all cache refs    ", ratio);
1041 
1042 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1043 		print_stalled_cycles_frontend(cpu, evsel, avg);
1044 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1045 		print_stalled_cycles_backend(cpu, evsel, avg);
1046 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1047 		total = avg_stats(&runtime_nsecs_stats[cpu]);
1048 
1049 		if (total) {
1050 			ratio = avg / total;
1051 			fprintf(output, " # %8.3f GHz                    ", ratio);
1052 		}
1053 	} else if (transaction_run &&
1054 		   perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1055 		total = avg_stats(&runtime_cycles_stats[cpu]);
1056 		if (total)
1057 			fprintf(output,
1058 				" #   %5.2f%% transactional cycles   ",
1059 				100.0 * (avg / total));
1060 	} else if (transaction_run &&
1061 		   perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1062 		total = avg_stats(&runtime_cycles_stats[cpu]);
1063 		total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1064 		if (total2 < avg)
1065 			total2 = avg;
1066 		if (total)
1067 			fprintf(output,
1068 				" #   %5.2f%% aborted cycles         ",
1069 				100.0 * ((total2-avg) / total));
1070 	} else if (transaction_run &&
1071 		   perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1072 		   avg > 0 &&
1073 		   runtime_cycles_in_tx_stats[cpu].n != 0) {
1074 		total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1075 
1076 		if (total)
1077 			ratio = total / avg;
1078 
1079 		fprintf(output, " # %8.0f cycles / transaction   ", ratio);
1080 	} else if (transaction_run &&
1081 		   perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1082 		   avg > 0 &&
1083 		   runtime_cycles_in_tx_stats[cpu].n != 0) {
1084 		total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1085 
1086 		if (total)
1087 			ratio = total / avg;
1088 
1089 		fprintf(output, " # %8.0f cycles / elision       ", ratio);
1090 	} else if (runtime_nsecs_stats[cpu].n != 0) {
1091 		char unit = 'M';
1092 
1093 		total = avg_stats(&runtime_nsecs_stats[cpu]);
1094 
1095 		if (total)
1096 			ratio = 1000.0 * avg / total;
1097 		if (ratio < 0.001) {
1098 			ratio *= 1000;
1099 			unit = 'K';
1100 		}
1101 
1102 		fprintf(output, " # %8.3f %c/sec                  ", ratio, unit);
1103 	} else {
1104 		fprintf(output, "                                   ");
1105 	}
1106 }
1107 
1108 static void print_aggr(char *prefix)
1109 {
1110 	struct perf_evsel *counter;
1111 	int cpu, cpu2, s, s2, id, nr;
1112 	double uval;
1113 	u64 ena, run, val;
1114 
1115 	if (!(aggr_map || aggr_get_id))
1116 		return;
1117 
1118 	for (s = 0; s < aggr_map->nr; s++) {
1119 		id = aggr_map->map[s];
1120 		evlist__for_each(evsel_list, counter) {
1121 			val = ena = run = 0;
1122 			nr = 0;
1123 			for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1124 				cpu2 = perf_evsel__cpus(counter)->map[cpu];
1125 				s2 = aggr_get_id(evsel_list->cpus, cpu2);
1126 				if (s2 != id)
1127 					continue;
1128 				val += counter->counts->cpu[cpu].val;
1129 				ena += counter->counts->cpu[cpu].ena;
1130 				run += counter->counts->cpu[cpu].run;
1131 				nr++;
1132 			}
1133 			if (prefix)
1134 				fprintf(output, "%s", prefix);
1135 
1136 			if (run == 0 || ena == 0) {
1137 				aggr_printout(counter, id, nr);
1138 
1139 				fprintf(output, "%*s%s",
1140 					csv_output ? 0 : 18,
1141 					counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1142 					csv_sep);
1143 
1144 				fprintf(output, "%-*s%s",
1145 					csv_output ? 0 : unit_width,
1146 					counter->unit, csv_sep);
1147 
1148 				fprintf(output, "%*s",
1149 					csv_output ? 0 : -25,
1150 					perf_evsel__name(counter));
1151 
1152 				if (counter->cgrp)
1153 					fprintf(output, "%s%s",
1154 						csv_sep, counter->cgrp->name);
1155 
1156 				fputc('\n', output);
1157 				continue;
1158 			}
1159 			uval = val * counter->scale;
1160 
1161 			if (nsec_counter(counter))
1162 				nsec_printout(id, nr, counter, uval);
1163 			else
1164 				abs_printout(id, nr, counter, uval);
1165 
1166 			if (!csv_output) {
1167 				print_noise(counter, 1.0);
1168 
1169 				if (run != ena)
1170 					fprintf(output, "  (%.2f%%)",
1171 						100.0 * run / ena);
1172 			}
1173 			fputc('\n', output);
1174 		}
1175 	}
1176 }
1177 
1178 /*
1179  * Print out the results of a single counter:
1180  * aggregated counts in system-wide mode
1181  */
1182 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1183 {
1184 	struct perf_stat *ps = counter->priv;
1185 	double avg = avg_stats(&ps->res_stats[0]);
1186 	int scaled = counter->counts->scaled;
1187 	double uval;
1188 
1189 	if (prefix)
1190 		fprintf(output, "%s", prefix);
1191 
1192 	if (scaled == -1) {
1193 		fprintf(output, "%*s%s",
1194 			csv_output ? 0 : 18,
1195 			counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1196 			csv_sep);
1197 		fprintf(output, "%-*s%s",
1198 			csv_output ? 0 : unit_width,
1199 			counter->unit, csv_sep);
1200 		fprintf(output, "%*s",
1201 			csv_output ? 0 : -25,
1202 			perf_evsel__name(counter));
1203 
1204 		if (counter->cgrp)
1205 			fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1206 
1207 		fputc('\n', output);
1208 		return;
1209 	}
1210 
1211 	uval = avg * counter->scale;
1212 
1213 	if (nsec_counter(counter))
1214 		nsec_printout(-1, 0, counter, uval);
1215 	else
1216 		abs_printout(-1, 0, counter, uval);
1217 
1218 	print_noise(counter, avg);
1219 
1220 	if (csv_output) {
1221 		fputc('\n', output);
1222 		return;
1223 	}
1224 
1225 	if (scaled) {
1226 		double avg_enabled, avg_running;
1227 
1228 		avg_enabled = avg_stats(&ps->res_stats[1]);
1229 		avg_running = avg_stats(&ps->res_stats[2]);
1230 
1231 		fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1232 	}
1233 	fprintf(output, "\n");
1234 }
1235 
1236 /*
1237  * Print out the results of a single counter:
1238  * does not use aggregated count in system-wide
1239  */
1240 static void print_counter(struct perf_evsel *counter, char *prefix)
1241 {
1242 	u64 ena, run, val;
1243 	double uval;
1244 	int cpu;
1245 
1246 	for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1247 		val = counter->counts->cpu[cpu].val;
1248 		ena = counter->counts->cpu[cpu].ena;
1249 		run = counter->counts->cpu[cpu].run;
1250 
1251 		if (prefix)
1252 			fprintf(output, "%s", prefix);
1253 
1254 		if (run == 0 || ena == 0) {
1255 			fprintf(output, "CPU%*d%s%*s%s",
1256 				csv_output ? 0 : -4,
1257 				perf_evsel__cpus(counter)->map[cpu], csv_sep,
1258 				csv_output ? 0 : 18,
1259 				counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1260 				csv_sep);
1261 
1262 				fprintf(output, "%-*s%s",
1263 					csv_output ? 0 : unit_width,
1264 					counter->unit, csv_sep);
1265 
1266 				fprintf(output, "%*s",
1267 					csv_output ? 0 : -25,
1268 					perf_evsel__name(counter));
1269 
1270 			if (counter->cgrp)
1271 				fprintf(output, "%s%s",
1272 					csv_sep, counter->cgrp->name);
1273 
1274 			fputc('\n', output);
1275 			continue;
1276 		}
1277 
1278 		uval = val * counter->scale;
1279 
1280 		if (nsec_counter(counter))
1281 			nsec_printout(cpu, 0, counter, uval);
1282 		else
1283 			abs_printout(cpu, 0, counter, uval);
1284 
1285 		if (!csv_output) {
1286 			print_noise(counter, 1.0);
1287 
1288 			if (run != ena)
1289 				fprintf(output, "  (%.2f%%)",
1290 					100.0 * run / ena);
1291 		}
1292 		fputc('\n', output);
1293 	}
1294 }
1295 
1296 static void print_stat(int argc, const char **argv)
1297 {
1298 	struct perf_evsel *counter;
1299 	int i;
1300 
1301 	fflush(stdout);
1302 
1303 	if (!csv_output) {
1304 		fprintf(output, "\n");
1305 		fprintf(output, " Performance counter stats for ");
1306 		if (target.system_wide)
1307 			fprintf(output, "\'system wide");
1308 		else if (target.cpu_list)
1309 			fprintf(output, "\'CPU(s) %s", target.cpu_list);
1310 		else if (!target__has_task(&target)) {
1311 			fprintf(output, "\'%s", argv[0]);
1312 			for (i = 1; i < argc; i++)
1313 				fprintf(output, " %s", argv[i]);
1314 		} else if (target.pid)
1315 			fprintf(output, "process id \'%s", target.pid);
1316 		else
1317 			fprintf(output, "thread id \'%s", target.tid);
1318 
1319 		fprintf(output, "\'");
1320 		if (run_count > 1)
1321 			fprintf(output, " (%d runs)", run_count);
1322 		fprintf(output, ":\n\n");
1323 	}
1324 
1325 	switch (aggr_mode) {
1326 	case AGGR_CORE:
1327 	case AGGR_SOCKET:
1328 		print_aggr(NULL);
1329 		break;
1330 	case AGGR_GLOBAL:
1331 		evlist__for_each(evsel_list, counter)
1332 			print_counter_aggr(counter, NULL);
1333 		break;
1334 	case AGGR_NONE:
1335 		evlist__for_each(evsel_list, counter)
1336 			print_counter(counter, NULL);
1337 		break;
1338 	default:
1339 		break;
1340 	}
1341 
1342 	if (!csv_output) {
1343 		if (!null_run)
1344 			fprintf(output, "\n");
1345 		fprintf(output, " %17.9f seconds time elapsed",
1346 				avg_stats(&walltime_nsecs_stats)/1e9);
1347 		if (run_count > 1) {
1348 			fprintf(output, "                                        ");
1349 			print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1350 					avg_stats(&walltime_nsecs_stats));
1351 		}
1352 		fprintf(output, "\n\n");
1353 	}
1354 }
1355 
1356 static volatile int signr = -1;
1357 
1358 static void skip_signal(int signo)
1359 {
1360 	if ((child_pid == -1) || interval)
1361 		done = 1;
1362 
1363 	signr = signo;
1364 	/*
1365 	 * render child_pid harmless
1366 	 * won't send SIGTERM to a random
1367 	 * process in case of race condition
1368 	 * and fast PID recycling
1369 	 */
1370 	child_pid = -1;
1371 }
1372 
1373 static void sig_atexit(void)
1374 {
1375 	sigset_t set, oset;
1376 
1377 	/*
1378 	 * avoid race condition with SIGCHLD handler
1379 	 * in skip_signal() which is modifying child_pid
1380 	 * goal is to avoid send SIGTERM to a random
1381 	 * process
1382 	 */
1383 	sigemptyset(&set);
1384 	sigaddset(&set, SIGCHLD);
1385 	sigprocmask(SIG_BLOCK, &set, &oset);
1386 
1387 	if (child_pid != -1)
1388 		kill(child_pid, SIGTERM);
1389 
1390 	sigprocmask(SIG_SETMASK, &oset, NULL);
1391 
1392 	if (signr == -1)
1393 		return;
1394 
1395 	signal(signr, SIG_DFL);
1396 	kill(getpid(), signr);
1397 }
1398 
1399 static int stat__set_big_num(const struct option *opt __maybe_unused,
1400 			     const char *s __maybe_unused, int unset)
1401 {
1402 	big_num_opt = unset ? 0 : 1;
1403 	return 0;
1404 }
1405 
1406 static int perf_stat_init_aggr_mode(void)
1407 {
1408 	switch (aggr_mode) {
1409 	case AGGR_SOCKET:
1410 		if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1411 			perror("cannot build socket map");
1412 			return -1;
1413 		}
1414 		aggr_get_id = cpu_map__get_socket;
1415 		break;
1416 	case AGGR_CORE:
1417 		if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1418 			perror("cannot build core map");
1419 			return -1;
1420 		}
1421 		aggr_get_id = cpu_map__get_core;
1422 		break;
1423 	case AGGR_NONE:
1424 	case AGGR_GLOBAL:
1425 	default:
1426 		break;
1427 	}
1428 	return 0;
1429 }
1430 
1431 static int setup_events(const char * const *attrs, unsigned len)
1432 {
1433 	unsigned i;
1434 
1435 	for (i = 0; i < len; i++) {
1436 		if (parse_events(evsel_list, attrs[i]))
1437 			return -1;
1438 	}
1439 	return 0;
1440 }
1441 
1442 /*
1443  * Add default attributes, if there were no attributes specified or
1444  * if -d/--detailed, -d -d or -d -d -d is used:
1445  */
1446 static int add_default_attributes(void)
1447 {
1448 	struct perf_event_attr default_attrs[] = {
1449 
1450   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1451   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1452   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1453   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1454 
1455   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1456   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1457   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1458   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1459   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1460   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1461 
1462 };
1463 
1464 /*
1465  * Detailed stats (-d), covering the L1 and last level data caches:
1466  */
1467 	struct perf_event_attr detailed_attrs[] = {
1468 
1469   { .type = PERF_TYPE_HW_CACHE,
1470     .config =
1471 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1472 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1473 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1474 
1475   { .type = PERF_TYPE_HW_CACHE,
1476     .config =
1477 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1478 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1479 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1480 
1481   { .type = PERF_TYPE_HW_CACHE,
1482     .config =
1483 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1484 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1485 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1486 
1487   { .type = PERF_TYPE_HW_CACHE,
1488     .config =
1489 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1490 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1491 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1492 };
1493 
1494 /*
1495  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1496  */
1497 	struct perf_event_attr very_detailed_attrs[] = {
1498 
1499   { .type = PERF_TYPE_HW_CACHE,
1500     .config =
1501 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1502 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1503 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1504 
1505   { .type = PERF_TYPE_HW_CACHE,
1506     .config =
1507 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1508 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1509 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1510 
1511   { .type = PERF_TYPE_HW_CACHE,
1512     .config =
1513 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1514 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1515 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1516 
1517   { .type = PERF_TYPE_HW_CACHE,
1518     .config =
1519 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1520 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1521 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1522 
1523   { .type = PERF_TYPE_HW_CACHE,
1524     .config =
1525 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1526 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1527 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1528 
1529   { .type = PERF_TYPE_HW_CACHE,
1530     .config =
1531 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1532 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1533 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1534 
1535 };
1536 
1537 /*
1538  * Very, very detailed stats (-d -d -d), adding prefetch events:
1539  */
1540 	struct perf_event_attr very_very_detailed_attrs[] = {
1541 
1542   { .type = PERF_TYPE_HW_CACHE,
1543     .config =
1544 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1545 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1546 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1547 
1548   { .type = PERF_TYPE_HW_CACHE,
1549     .config =
1550 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1551 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1552 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1553 };
1554 
1555 	/* Set attrs if no event is selected and !null_run: */
1556 	if (null_run)
1557 		return 0;
1558 
1559 	if (transaction_run) {
1560 		int err;
1561 		if (pmu_have_event("cpu", "cycles-ct") &&
1562 		    pmu_have_event("cpu", "el-start"))
1563 			err = setup_events(transaction_attrs,
1564 					ARRAY_SIZE(transaction_attrs));
1565 		else
1566 			err = setup_events(transaction_limited_attrs,
1567 				 ARRAY_SIZE(transaction_limited_attrs));
1568 		if (err < 0) {
1569 			fprintf(stderr, "Cannot set up transaction events\n");
1570 			return -1;
1571 		}
1572 		return 0;
1573 	}
1574 
1575 	if (!evsel_list->nr_entries) {
1576 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1577 			return -1;
1578 	}
1579 
1580 	/* Detailed events get appended to the event list: */
1581 
1582 	if (detailed_run <  1)
1583 		return 0;
1584 
1585 	/* Append detailed run extra attributes: */
1586 	if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1587 		return -1;
1588 
1589 	if (detailed_run < 2)
1590 		return 0;
1591 
1592 	/* Append very detailed run extra attributes: */
1593 	if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1594 		return -1;
1595 
1596 	if (detailed_run < 3)
1597 		return 0;
1598 
1599 	/* Append very, very detailed run extra attributes: */
1600 	return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1601 }
1602 
1603 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1604 {
1605 	bool append_file = false;
1606 	int output_fd = 0;
1607 	const char *output_name	= NULL;
1608 	const struct option options[] = {
1609 	OPT_BOOLEAN('T', "transaction", &transaction_run,
1610 		    "hardware transaction statistics"),
1611 	OPT_CALLBACK('e', "event", &evsel_list, "event",
1612 		     "event selector. use 'perf list' to list available events",
1613 		     parse_events_option),
1614 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1615 		     "event filter", parse_filter),
1616 	OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1617 		    "child tasks do not inherit counters"),
1618 	OPT_STRING('p', "pid", &target.pid, "pid",
1619 		   "stat events on existing process id"),
1620 	OPT_STRING('t', "tid", &target.tid, "tid",
1621 		   "stat events on existing thread id"),
1622 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1623 		    "system-wide collection from all CPUs"),
1624 	OPT_BOOLEAN('g', "group", &group,
1625 		    "put the counters into a counter group"),
1626 	OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1627 	OPT_INCR('v', "verbose", &verbose,
1628 		    "be more verbose (show counter open errors, etc)"),
1629 	OPT_INTEGER('r', "repeat", &run_count,
1630 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1631 	OPT_BOOLEAN('n', "null", &null_run,
1632 		    "null run - dont start any counters"),
1633 	OPT_INCR('d', "detailed", &detailed_run,
1634 		    "detailed run - start a lot of events"),
1635 	OPT_BOOLEAN('S', "sync", &sync_run,
1636 		    "call sync() before starting a run"),
1637 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1638 			   "print large numbers with thousands\' separators",
1639 			   stat__set_big_num),
1640 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1641 		    "list of cpus to monitor in system-wide"),
1642 	OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1643 		    "disable CPU count aggregation", AGGR_NONE),
1644 	OPT_STRING('x', "field-separator", &csv_sep, "separator",
1645 		   "print counts with custom separator"),
1646 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1647 		     "monitor event in cgroup name only", parse_cgroups),
1648 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1649 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1650 	OPT_INTEGER(0, "log-fd", &output_fd,
1651 		    "log output to fd, instead of stderr"),
1652 	OPT_STRING(0, "pre", &pre_cmd, "command",
1653 			"command to run prior to the measured command"),
1654 	OPT_STRING(0, "post", &post_cmd, "command",
1655 			"command to run after to the measured command"),
1656 	OPT_UINTEGER('I', "interval-print", &interval,
1657 		    "print counts at regular interval in ms (>= 100)"),
1658 	OPT_SET_UINT(0, "per-socket", &aggr_mode,
1659 		     "aggregate counts per processor socket", AGGR_SOCKET),
1660 	OPT_SET_UINT(0, "per-core", &aggr_mode,
1661 		     "aggregate counts per physical processor core", AGGR_CORE),
1662 	OPT_UINTEGER('D', "delay", &initial_delay,
1663 		     "ms to wait before starting measurement after program start"),
1664 	OPT_END()
1665 	};
1666 	const char * const stat_usage[] = {
1667 		"perf stat [<options>] [<command>]",
1668 		NULL
1669 	};
1670 	int status = -EINVAL, run_idx;
1671 	const char *mode;
1672 
1673 	setlocale(LC_ALL, "");
1674 
1675 	evsel_list = perf_evlist__new();
1676 	if (evsel_list == NULL)
1677 		return -ENOMEM;
1678 
1679 	argc = parse_options(argc, argv, options, stat_usage,
1680 		PARSE_OPT_STOP_AT_NON_OPTION);
1681 
1682 	output = stderr;
1683 	if (output_name && strcmp(output_name, "-"))
1684 		output = NULL;
1685 
1686 	if (output_name && output_fd) {
1687 		fprintf(stderr, "cannot use both --output and --log-fd\n");
1688 		parse_options_usage(stat_usage, options, "o", 1);
1689 		parse_options_usage(NULL, options, "log-fd", 0);
1690 		goto out;
1691 	}
1692 
1693 	if (output_fd < 0) {
1694 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
1695 		parse_options_usage(stat_usage, options, "log-fd", 0);
1696 		goto out;
1697 	}
1698 
1699 	if (!output) {
1700 		struct timespec tm;
1701 		mode = append_file ? "a" : "w";
1702 
1703 		output = fopen(output_name, mode);
1704 		if (!output) {
1705 			perror("failed to create output file");
1706 			return -1;
1707 		}
1708 		clock_gettime(CLOCK_REALTIME, &tm);
1709 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1710 	} else if (output_fd > 0) {
1711 		mode = append_file ? "a" : "w";
1712 		output = fdopen(output_fd, mode);
1713 		if (!output) {
1714 			perror("Failed opening logfd");
1715 			return -errno;
1716 		}
1717 	}
1718 
1719 	if (csv_sep) {
1720 		csv_output = true;
1721 		if (!strcmp(csv_sep, "\\t"))
1722 			csv_sep = "\t";
1723 	} else
1724 		csv_sep = DEFAULT_SEPARATOR;
1725 
1726 	/*
1727 	 * let the spreadsheet do the pretty-printing
1728 	 */
1729 	if (csv_output) {
1730 		/* User explicitly passed -B? */
1731 		if (big_num_opt == 1) {
1732 			fprintf(stderr, "-B option not supported with -x\n");
1733 			parse_options_usage(stat_usage, options, "B", 1);
1734 			parse_options_usage(NULL, options, "x", 1);
1735 			goto out;
1736 		} else /* Nope, so disable big number formatting */
1737 			big_num = false;
1738 	} else if (big_num_opt == 0) /* User passed --no-big-num */
1739 		big_num = false;
1740 
1741 	if (!argc && target__none(&target))
1742 		usage_with_options(stat_usage, options);
1743 
1744 	if (run_count < 0) {
1745 		pr_err("Run count must be a positive number\n");
1746 		parse_options_usage(stat_usage, options, "r", 1);
1747 		goto out;
1748 	} else if (run_count == 0) {
1749 		forever = true;
1750 		run_count = 1;
1751 	}
1752 
1753 	/* no_aggr, cgroup are for system-wide only */
1754 	if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1755 	    !target__has_cpu(&target)) {
1756 		fprintf(stderr, "both cgroup and no-aggregation "
1757 			"modes only available in system-wide mode\n");
1758 
1759 		parse_options_usage(stat_usage, options, "G", 1);
1760 		parse_options_usage(NULL, options, "A", 1);
1761 		parse_options_usage(NULL, options, "a", 1);
1762 		goto out;
1763 	}
1764 
1765 	if (add_default_attributes())
1766 		goto out;
1767 
1768 	target__validate(&target);
1769 
1770 	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1771 		if (target__has_task(&target)) {
1772 			pr_err("Problems finding threads of monitor\n");
1773 			parse_options_usage(stat_usage, options, "p", 1);
1774 			parse_options_usage(NULL, options, "t", 1);
1775 		} else if (target__has_cpu(&target)) {
1776 			perror("failed to parse CPUs map");
1777 			parse_options_usage(stat_usage, options, "C", 1);
1778 			parse_options_usage(NULL, options, "a", 1);
1779 		}
1780 		goto out;
1781 	}
1782 	if (interval && interval < 100) {
1783 		pr_err("print interval must be >= 100ms\n");
1784 		parse_options_usage(stat_usage, options, "I", 1);
1785 		goto out;
1786 	}
1787 
1788 	if (perf_evlist__alloc_stats(evsel_list, interval))
1789 		goto out;
1790 
1791 	if (perf_stat_init_aggr_mode())
1792 		goto out;
1793 
1794 	/*
1795 	 * We dont want to block the signals - that would cause
1796 	 * child tasks to inherit that and Ctrl-C would not work.
1797 	 * What we want is for Ctrl-C to work in the exec()-ed
1798 	 * task, but being ignored by perf stat itself:
1799 	 */
1800 	atexit(sig_atexit);
1801 	if (!forever)
1802 		signal(SIGINT,  skip_signal);
1803 	signal(SIGCHLD, skip_signal);
1804 	signal(SIGALRM, skip_signal);
1805 	signal(SIGABRT, skip_signal);
1806 
1807 	status = 0;
1808 	for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1809 		if (run_count != 1 && verbose)
1810 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1811 				run_idx + 1);
1812 
1813 		status = run_perf_stat(argc, argv);
1814 		if (forever && status != -1) {
1815 			print_stat(argc, argv);
1816 			perf_stat__reset_stats(evsel_list);
1817 		}
1818 	}
1819 
1820 	if (!forever && status != -1 && !interval)
1821 		print_stat(argc, argv);
1822 
1823 	perf_evlist__free_stats(evsel_list);
1824 out:
1825 	perf_evlist__delete(evsel_list);
1826 	return status;
1827 }
1828