xref: /openbmc/linux/tools/perf/builtin-stat.c (revision 930beb5a)
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 bool			forever				= false;
142 static struct timespec		ref_time;
143 static struct cpu_map		*aggr_map;
144 static int			(*aggr_get_id)(struct cpu_map *m, int cpu);
145 
146 static volatile int done = 0;
147 
148 struct perf_stat {
149 	struct stats	  res_stats[3];
150 };
151 
152 static inline void diff_timespec(struct timespec *r, struct timespec *a,
153 				 struct timespec *b)
154 {
155 	r->tv_sec = a->tv_sec - b->tv_sec;
156 	if (a->tv_nsec < b->tv_nsec) {
157 		r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
158 		r->tv_sec--;
159 	} else {
160 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
161 	}
162 }
163 
164 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
165 {
166 	return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
167 }
168 
169 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
170 {
171 	return perf_evsel__cpus(evsel)->nr;
172 }
173 
174 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
175 {
176 	memset(evsel->priv, 0, sizeof(struct perf_stat));
177 }
178 
179 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
180 {
181 	evsel->priv = zalloc(sizeof(struct perf_stat));
182 	return evsel->priv == NULL ? -ENOMEM : 0;
183 }
184 
185 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
186 {
187 	free(evsel->priv);
188 	evsel->priv = NULL;
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 	free(evsel->prev_raw_counts);
211 	evsel->prev_raw_counts = NULL;
212 }
213 
214 static void perf_evlist__free_stats(struct perf_evlist *evlist)
215 {
216 	struct perf_evsel *evsel;
217 
218 	list_for_each_entry(evsel, &evlist->entries, node) {
219 		perf_evsel__free_stat_priv(evsel);
220 		perf_evsel__free_counts(evsel);
221 		perf_evsel__free_prev_raw_counts(evsel);
222 	}
223 }
224 
225 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
226 {
227 	struct perf_evsel *evsel;
228 
229 	list_for_each_entry(evsel, &evlist->entries, node) {
230 		if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
231 		    perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
232 		    (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
233 			goto out_free;
234 	}
235 
236 	return 0;
237 
238 out_free:
239 	perf_evlist__free_stats(evlist);
240 	return -1;
241 }
242 
243 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
244 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
245 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
246 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
247 static struct stats runtime_branches_stats[MAX_NR_CPUS];
248 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
249 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
250 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
251 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
252 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
253 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
254 static struct stats runtime_cycles_in_tx_stats[MAX_NR_CPUS];
255 static struct stats walltime_nsecs_stats;
256 static struct stats runtime_transaction_stats[MAX_NR_CPUS];
257 static struct stats runtime_elision_stats[MAX_NR_CPUS];
258 
259 static void perf_stat__reset_stats(struct perf_evlist *evlist)
260 {
261 	struct perf_evsel *evsel;
262 
263 	list_for_each_entry(evsel, &evlist->entries, node) {
264 		perf_evsel__reset_stat_priv(evsel);
265 		perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
266 	}
267 
268 	memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
269 	memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
270 	memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
271 	memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
272 	memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
273 	memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
274 	memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
275 	memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
276 	memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
277 	memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
278 	memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
279 	memset(runtime_cycles_in_tx_stats, 0,
280 			sizeof(runtime_cycles_in_tx_stats));
281 	memset(runtime_transaction_stats, 0,
282 		sizeof(runtime_transaction_stats));
283 	memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
284 	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
285 }
286 
287 static int create_perf_stat_counter(struct perf_evsel *evsel)
288 {
289 	struct perf_event_attr *attr = &evsel->attr;
290 
291 	if (scale)
292 		attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
293 				    PERF_FORMAT_TOTAL_TIME_RUNNING;
294 
295 	attr->inherit = !no_inherit;
296 
297 	if (target__has_cpu(&target))
298 		return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
299 
300 	if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
301 		attr->disabled = 1;
302 		if (!initial_delay)
303 			attr->enable_on_exec = 1;
304 	}
305 
306 	return perf_evsel__open_per_thread(evsel, evsel_list->threads);
307 }
308 
309 /*
310  * Does the counter have nsecs as a unit?
311  */
312 static inline int nsec_counter(struct perf_evsel *evsel)
313 {
314 	if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
315 	    perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
316 		return 1;
317 
318 	return 0;
319 }
320 
321 static struct perf_evsel *nth_evsel(int n)
322 {
323 	static struct perf_evsel **array;
324 	static int array_len;
325 	struct perf_evsel *ev;
326 	int j;
327 
328 	/* Assumes this only called when evsel_list does not change anymore. */
329 	if (!array) {
330 		list_for_each_entry(ev, &evsel_list->entries, node)
331 			array_len++;
332 		array = malloc(array_len * sizeof(void *));
333 		if (!array)
334 			exit(ENOMEM);
335 		j = 0;
336 		list_for_each_entry(ev, &evsel_list->entries, node)
337 			array[j++] = ev;
338 	}
339 	if (n < array_len)
340 		return array[n];
341 	return NULL;
342 }
343 
344 /*
345  * Update various tracking values we maintain to print
346  * more semantic information such as miss/hit ratios,
347  * instruction rates, etc:
348  */
349 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
350 {
351 	if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
352 		update_stats(&runtime_nsecs_stats[0], count[0]);
353 	else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
354 		update_stats(&runtime_cycles_stats[0], count[0]);
355 	else if (transaction_run &&
356 		 perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX)))
357 		update_stats(&runtime_cycles_in_tx_stats[0], count[0]);
358 	else if (transaction_run &&
359 		 perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START)))
360 		update_stats(&runtime_transaction_stats[0], count[0]);
361 	else if (transaction_run &&
362 		 perf_evsel__cmp(counter, nth_evsel(T_ELISION_START)))
363 		update_stats(&runtime_elision_stats[0], count[0]);
364 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
365 		update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
366 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
367 		update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
368 	else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
369 		update_stats(&runtime_branches_stats[0], count[0]);
370 	else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
371 		update_stats(&runtime_cacherefs_stats[0], count[0]);
372 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
373 		update_stats(&runtime_l1_dcache_stats[0], count[0]);
374 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
375 		update_stats(&runtime_l1_icache_stats[0], count[0]);
376 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
377 		update_stats(&runtime_ll_cache_stats[0], count[0]);
378 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
379 		update_stats(&runtime_dtlb_cache_stats[0], count[0]);
380 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
381 		update_stats(&runtime_itlb_cache_stats[0], count[0]);
382 }
383 
384 /*
385  * Read out the results of a single counter:
386  * aggregate counts across CPUs in system-wide mode
387  */
388 static int read_counter_aggr(struct perf_evsel *counter)
389 {
390 	struct perf_stat *ps = counter->priv;
391 	u64 *count = counter->counts->aggr.values;
392 	int i;
393 
394 	if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
395 			       thread_map__nr(evsel_list->threads), scale) < 0)
396 		return -1;
397 
398 	for (i = 0; i < 3; i++)
399 		update_stats(&ps->res_stats[i], count[i]);
400 
401 	if (verbose) {
402 		fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
403 			perf_evsel__name(counter), count[0], count[1], count[2]);
404 	}
405 
406 	/*
407 	 * Save the full runtime - to allow normalization during printout:
408 	 */
409 	update_shadow_stats(counter, count);
410 
411 	return 0;
412 }
413 
414 /*
415  * Read out the results of a single counter:
416  * do not aggregate counts across CPUs in system-wide mode
417  */
418 static int read_counter(struct perf_evsel *counter)
419 {
420 	u64 *count;
421 	int cpu;
422 
423 	for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
424 		if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
425 			return -1;
426 
427 		count = counter->counts->cpu[cpu].values;
428 
429 		update_shadow_stats(counter, count);
430 	}
431 
432 	return 0;
433 }
434 
435 static void print_interval(void)
436 {
437 	static int num_print_interval;
438 	struct perf_evsel *counter;
439 	struct perf_stat *ps;
440 	struct timespec ts, rs;
441 	char prefix[64];
442 
443 	if (aggr_mode == AGGR_GLOBAL) {
444 		list_for_each_entry(counter, &evsel_list->entries, node) {
445 			ps = counter->priv;
446 			memset(ps->res_stats, 0, sizeof(ps->res_stats));
447 			read_counter_aggr(counter);
448 		}
449 	} else	{
450 		list_for_each_entry(counter, &evsel_list->entries, node) {
451 			ps = counter->priv;
452 			memset(ps->res_stats, 0, sizeof(ps->res_stats));
453 			read_counter(counter);
454 		}
455 	}
456 
457 	clock_gettime(CLOCK_MONOTONIC, &ts);
458 	diff_timespec(&rs, &ts, &ref_time);
459 	sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
460 
461 	if (num_print_interval == 0 && !csv_output) {
462 		switch (aggr_mode) {
463 		case AGGR_SOCKET:
464 			fprintf(output, "#           time socket cpus             counts events\n");
465 			break;
466 		case AGGR_CORE:
467 			fprintf(output, "#           time core         cpus             counts events\n");
468 			break;
469 		case AGGR_NONE:
470 			fprintf(output, "#           time CPU                 counts events\n");
471 			break;
472 		case AGGR_GLOBAL:
473 		default:
474 			fprintf(output, "#           time             counts events\n");
475 		}
476 	}
477 
478 	if (++num_print_interval == 25)
479 		num_print_interval = 0;
480 
481 	switch (aggr_mode) {
482 	case AGGR_CORE:
483 	case AGGR_SOCKET:
484 		print_aggr(prefix);
485 		break;
486 	case AGGR_NONE:
487 		list_for_each_entry(counter, &evsel_list->entries, node)
488 			print_counter(counter, prefix);
489 		break;
490 	case AGGR_GLOBAL:
491 	default:
492 		list_for_each_entry(counter, &evsel_list->entries, node)
493 			print_counter_aggr(counter, prefix);
494 	}
495 
496 	fflush(output);
497 }
498 
499 static void handle_initial_delay(void)
500 {
501 	struct perf_evsel *counter;
502 
503 	if (initial_delay) {
504 		const int ncpus = cpu_map__nr(evsel_list->cpus),
505 			nthreads = thread_map__nr(evsel_list->threads);
506 
507 		usleep(initial_delay * 1000);
508 		list_for_each_entry(counter, &evsel_list->entries, node)
509 			perf_evsel__enable(counter, ncpus, nthreads);
510 	}
511 }
512 
513 static int __run_perf_stat(int argc, const char **argv)
514 {
515 	char msg[512];
516 	unsigned long long t0, t1;
517 	struct perf_evsel *counter;
518 	struct timespec ts;
519 	int status = 0;
520 	const bool forks = (argc > 0);
521 
522 	if (interval) {
523 		ts.tv_sec  = interval / 1000;
524 		ts.tv_nsec = (interval % 1000) * 1000000;
525 	} else {
526 		ts.tv_sec  = 1;
527 		ts.tv_nsec = 0;
528 	}
529 
530 	if (forks) {
531 		if (perf_evlist__prepare_workload(evsel_list, &target, argv,
532 						  false, false) < 0) {
533 			perror("failed to prepare workload");
534 			return -1;
535 		}
536 		child_pid = evsel_list->workload.pid;
537 	}
538 
539 	if (group)
540 		perf_evlist__set_leader(evsel_list);
541 
542 	list_for_each_entry(counter, &evsel_list->entries, node) {
543 		if (create_perf_stat_counter(counter) < 0) {
544 			/*
545 			 * PPC returns ENXIO for HW counters until 2.6.37
546 			 * (behavior changed with commit b0a873e).
547 			 */
548 			if (errno == EINVAL || errno == ENOSYS ||
549 			    errno == ENOENT || errno == EOPNOTSUPP ||
550 			    errno == ENXIO) {
551 				if (verbose)
552 					ui__warning("%s event is not supported by the kernel.\n",
553 						    perf_evsel__name(counter));
554 				counter->supported = false;
555 				continue;
556 			}
557 
558 			perf_evsel__open_strerror(counter, &target,
559 						  errno, msg, sizeof(msg));
560 			ui__error("%s\n", msg);
561 
562 			if (child_pid != -1)
563 				kill(child_pid, SIGTERM);
564 
565 			return -1;
566 		}
567 		counter->supported = true;
568 	}
569 
570 	if (perf_evlist__apply_filters(evsel_list)) {
571 		error("failed to set filter with %d (%s)\n", errno,
572 			strerror(errno));
573 		return -1;
574 	}
575 
576 	/*
577 	 * Enable counters and exec the command:
578 	 */
579 	t0 = rdclock();
580 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
581 
582 	if (forks) {
583 		perf_evlist__start_workload(evsel_list);
584 		handle_initial_delay();
585 
586 		if (interval) {
587 			while (!waitpid(child_pid, &status, WNOHANG)) {
588 				nanosleep(&ts, NULL);
589 				print_interval();
590 			}
591 		}
592 		wait(&status);
593 		if (WIFSIGNALED(status))
594 			psignal(WTERMSIG(status), argv[0]);
595 	} else {
596 		handle_initial_delay();
597 		while (!done) {
598 			nanosleep(&ts, NULL);
599 			if (interval)
600 				print_interval();
601 		}
602 	}
603 
604 	t1 = rdclock();
605 
606 	update_stats(&walltime_nsecs_stats, t1 - t0);
607 
608 	if (aggr_mode == AGGR_GLOBAL) {
609 		list_for_each_entry(counter, &evsel_list->entries, node) {
610 			read_counter_aggr(counter);
611 			perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
612 					     thread_map__nr(evsel_list->threads));
613 		}
614 	} else {
615 		list_for_each_entry(counter, &evsel_list->entries, node) {
616 			read_counter(counter);
617 			perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
618 		}
619 	}
620 
621 	return WEXITSTATUS(status);
622 }
623 
624 static int run_perf_stat(int argc __maybe_unused, const char **argv)
625 {
626 	int ret;
627 
628 	if (pre_cmd) {
629 		ret = system(pre_cmd);
630 		if (ret)
631 			return ret;
632 	}
633 
634 	if (sync_run)
635 		sync();
636 
637 	ret = __run_perf_stat(argc, argv);
638 	if (ret)
639 		return ret;
640 
641 	if (post_cmd) {
642 		ret = system(post_cmd);
643 		if (ret)
644 			return ret;
645 	}
646 
647 	return ret;
648 }
649 
650 static void print_noise_pct(double total, double avg)
651 {
652 	double pct = rel_stddev_stats(total, avg);
653 
654 	if (csv_output)
655 		fprintf(output, "%s%.2f%%", csv_sep, pct);
656 	else if (pct)
657 		fprintf(output, "  ( +-%6.2f%% )", pct);
658 }
659 
660 static void print_noise(struct perf_evsel *evsel, double avg)
661 {
662 	struct perf_stat *ps;
663 
664 	if (run_count == 1)
665 		return;
666 
667 	ps = evsel->priv;
668 	print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
669 }
670 
671 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
672 {
673 	switch (aggr_mode) {
674 	case AGGR_CORE:
675 		fprintf(output, "S%d-C%*d%s%*d%s",
676 			cpu_map__id_to_socket(id),
677 			csv_output ? 0 : -8,
678 			cpu_map__id_to_cpu(id),
679 			csv_sep,
680 			csv_output ? 0 : 4,
681 			nr,
682 			csv_sep);
683 		break;
684 	case AGGR_SOCKET:
685 		fprintf(output, "S%*d%s%*d%s",
686 			csv_output ? 0 : -5,
687 			id,
688 			csv_sep,
689 			csv_output ? 0 : 4,
690 			nr,
691 			csv_sep);
692 			break;
693 	case AGGR_NONE:
694 		fprintf(output, "CPU%*d%s",
695 			csv_output ? 0 : -4,
696 			perf_evsel__cpus(evsel)->map[id], csv_sep);
697 		break;
698 	case AGGR_GLOBAL:
699 	default:
700 		break;
701 	}
702 }
703 
704 static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
705 {
706 	double msecs = avg / 1e6;
707 	const char *fmt = csv_output ? "%.6f%s%s" : "%18.6f%s%-25s";
708 	char name[25];
709 
710 	aggr_printout(evsel, cpu, nr);
711 
712 	scnprintf(name, sizeof(name), "%s%s",
713 		  perf_evsel__name(evsel), csv_output ? "" : " (msec)");
714 	fprintf(output, fmt, msecs, csv_sep, name);
715 
716 	if (evsel->cgrp)
717 		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
718 
719 	if (csv_output || interval)
720 		return;
721 
722 	if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
723 		fprintf(output, " # %8.3f CPUs utilized          ",
724 			avg / avg_stats(&walltime_nsecs_stats));
725 	else
726 		fprintf(output, "                                   ");
727 }
728 
729 /* used for get_ratio_color() */
730 enum grc_type {
731 	GRC_STALLED_CYCLES_FE,
732 	GRC_STALLED_CYCLES_BE,
733 	GRC_CACHE_MISSES,
734 	GRC_MAX_NR
735 };
736 
737 static const char *get_ratio_color(enum grc_type type, double ratio)
738 {
739 	static const double grc_table[GRC_MAX_NR][3] = {
740 		[GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
741 		[GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
742 		[GRC_CACHE_MISSES] 	= { 20.0, 10.0, 5.0 },
743 	};
744 	const char *color = PERF_COLOR_NORMAL;
745 
746 	if (ratio > grc_table[type][0])
747 		color = PERF_COLOR_RED;
748 	else if (ratio > grc_table[type][1])
749 		color = PERF_COLOR_MAGENTA;
750 	else if (ratio > grc_table[type][2])
751 		color = PERF_COLOR_YELLOW;
752 
753 	return color;
754 }
755 
756 static void print_stalled_cycles_frontend(int cpu,
757 					  struct perf_evsel *evsel
758 					  __maybe_unused, double avg)
759 {
760 	double total, ratio = 0.0;
761 	const char *color;
762 
763 	total = avg_stats(&runtime_cycles_stats[cpu]);
764 
765 	if (total)
766 		ratio = avg / total * 100.0;
767 
768 	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
769 
770 	fprintf(output, " #  ");
771 	color_fprintf(output, color, "%6.2f%%", ratio);
772 	fprintf(output, " frontend cycles idle   ");
773 }
774 
775 static void print_stalled_cycles_backend(int cpu,
776 					 struct perf_evsel *evsel
777 					 __maybe_unused, double avg)
778 {
779 	double total, ratio = 0.0;
780 	const char *color;
781 
782 	total = avg_stats(&runtime_cycles_stats[cpu]);
783 
784 	if (total)
785 		ratio = avg / total * 100.0;
786 
787 	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
788 
789 	fprintf(output, " #  ");
790 	color_fprintf(output, color, "%6.2f%%", ratio);
791 	fprintf(output, " backend  cycles idle   ");
792 }
793 
794 static void print_branch_misses(int cpu,
795 				struct perf_evsel *evsel __maybe_unused,
796 				double avg)
797 {
798 	double total, ratio = 0.0;
799 	const char *color;
800 
801 	total = avg_stats(&runtime_branches_stats[cpu]);
802 
803 	if (total)
804 		ratio = avg / total * 100.0;
805 
806 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
807 
808 	fprintf(output, " #  ");
809 	color_fprintf(output, color, "%6.2f%%", ratio);
810 	fprintf(output, " of all branches        ");
811 }
812 
813 static void print_l1_dcache_misses(int cpu,
814 				   struct perf_evsel *evsel __maybe_unused,
815 				   double avg)
816 {
817 	double total, ratio = 0.0;
818 	const char *color;
819 
820 	total = avg_stats(&runtime_l1_dcache_stats[cpu]);
821 
822 	if (total)
823 		ratio = avg / total * 100.0;
824 
825 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
826 
827 	fprintf(output, " #  ");
828 	color_fprintf(output, color, "%6.2f%%", ratio);
829 	fprintf(output, " of all L1-dcache hits  ");
830 }
831 
832 static void print_l1_icache_misses(int cpu,
833 				   struct perf_evsel *evsel __maybe_unused,
834 				   double avg)
835 {
836 	double total, ratio = 0.0;
837 	const char *color;
838 
839 	total = avg_stats(&runtime_l1_icache_stats[cpu]);
840 
841 	if (total)
842 		ratio = avg / total * 100.0;
843 
844 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
845 
846 	fprintf(output, " #  ");
847 	color_fprintf(output, color, "%6.2f%%", ratio);
848 	fprintf(output, " of all L1-icache hits  ");
849 }
850 
851 static void print_dtlb_cache_misses(int cpu,
852 				    struct perf_evsel *evsel __maybe_unused,
853 				    double avg)
854 {
855 	double total, ratio = 0.0;
856 	const char *color;
857 
858 	total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
859 
860 	if (total)
861 		ratio = avg / total * 100.0;
862 
863 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
864 
865 	fprintf(output, " #  ");
866 	color_fprintf(output, color, "%6.2f%%", ratio);
867 	fprintf(output, " of all dTLB cache hits ");
868 }
869 
870 static void print_itlb_cache_misses(int cpu,
871 				    struct perf_evsel *evsel __maybe_unused,
872 				    double avg)
873 {
874 	double total, ratio = 0.0;
875 	const char *color;
876 
877 	total = avg_stats(&runtime_itlb_cache_stats[cpu]);
878 
879 	if (total)
880 		ratio = avg / total * 100.0;
881 
882 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
883 
884 	fprintf(output, " #  ");
885 	color_fprintf(output, color, "%6.2f%%", ratio);
886 	fprintf(output, " of all iTLB cache hits ");
887 }
888 
889 static void print_ll_cache_misses(int cpu,
890 				  struct perf_evsel *evsel __maybe_unused,
891 				  double avg)
892 {
893 	double total, ratio = 0.0;
894 	const char *color;
895 
896 	total = avg_stats(&runtime_ll_cache_stats[cpu]);
897 
898 	if (total)
899 		ratio = avg / total * 100.0;
900 
901 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
902 
903 	fprintf(output, " #  ");
904 	color_fprintf(output, color, "%6.2f%%", ratio);
905 	fprintf(output, " of all LL-cache hits   ");
906 }
907 
908 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
909 {
910 	double total, ratio = 0.0, total2;
911 	const char *fmt;
912 
913 	if (csv_output)
914 		fmt = "%.0f%s%s";
915 	else if (big_num)
916 		fmt = "%'18.0f%s%-25s";
917 	else
918 		fmt = "%18.0f%s%-25s";
919 
920 	aggr_printout(evsel, cpu, nr);
921 
922 	if (aggr_mode == AGGR_GLOBAL)
923 		cpu = 0;
924 
925 	fprintf(output, fmt, avg, csv_sep, perf_evsel__name(evsel));
926 
927 	if (evsel->cgrp)
928 		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
929 
930 	if (csv_output || interval)
931 		return;
932 
933 	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
934 		total = avg_stats(&runtime_cycles_stats[cpu]);
935 		if (total) {
936 			ratio = avg / total;
937 			fprintf(output, " #   %5.2f  insns per cycle        ", ratio);
938 		}
939 		total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
940 		total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
941 
942 		if (total && avg) {
943 			ratio = total / avg;
944 			fprintf(output, "\n                                             #   %5.2f  stalled cycles per insn", ratio);
945 		}
946 
947 	} else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
948 			runtime_branches_stats[cpu].n != 0) {
949 		print_branch_misses(cpu, evsel, avg);
950 	} else if (
951 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
952 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
953 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
954 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
955 			runtime_l1_dcache_stats[cpu].n != 0) {
956 		print_l1_dcache_misses(cpu, evsel, avg);
957 	} else if (
958 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
959 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
960 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
961 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
962 			runtime_l1_icache_stats[cpu].n != 0) {
963 		print_l1_icache_misses(cpu, evsel, avg);
964 	} else if (
965 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
966 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
967 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
968 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
969 			runtime_dtlb_cache_stats[cpu].n != 0) {
970 		print_dtlb_cache_misses(cpu, evsel, avg);
971 	} else if (
972 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
973 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
974 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
975 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
976 			runtime_itlb_cache_stats[cpu].n != 0) {
977 		print_itlb_cache_misses(cpu, evsel, avg);
978 	} else if (
979 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
980 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
981 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
982 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
983 			runtime_ll_cache_stats[cpu].n != 0) {
984 		print_ll_cache_misses(cpu, evsel, avg);
985 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
986 			runtime_cacherefs_stats[cpu].n != 0) {
987 		total = avg_stats(&runtime_cacherefs_stats[cpu]);
988 
989 		if (total)
990 			ratio = avg * 100 / total;
991 
992 		fprintf(output, " # %8.3f %% of all cache refs    ", ratio);
993 
994 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
995 		print_stalled_cycles_frontend(cpu, evsel, avg);
996 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
997 		print_stalled_cycles_backend(cpu, evsel, avg);
998 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
999 		total = avg_stats(&runtime_nsecs_stats[cpu]);
1000 
1001 		if (total) {
1002 			ratio = avg / total;
1003 			fprintf(output, " # %8.3f GHz                    ", ratio);
1004 		}
1005 	} else if (transaction_run &&
1006 		   perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1007 		total = avg_stats(&runtime_cycles_stats[cpu]);
1008 		if (total)
1009 			fprintf(output,
1010 				" #   %5.2f%% transactional cycles   ",
1011 				100.0 * (avg / total));
1012 	} else if (transaction_run &&
1013 		   perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1014 		total = avg_stats(&runtime_cycles_stats[cpu]);
1015 		total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1016 		if (total2 < avg)
1017 			total2 = avg;
1018 		if (total)
1019 			fprintf(output,
1020 				" #   %5.2f%% aborted cycles         ",
1021 				100.0 * ((total2-avg) / total));
1022 	} else if (transaction_run &&
1023 		   perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1024 		   avg > 0 &&
1025 		   runtime_cycles_in_tx_stats[cpu].n != 0) {
1026 		total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1027 
1028 		if (total)
1029 			ratio = total / avg;
1030 
1031 		fprintf(output, " # %8.0f cycles / transaction   ", ratio);
1032 	} else if (transaction_run &&
1033 		   perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1034 		   avg > 0 &&
1035 		   runtime_cycles_in_tx_stats[cpu].n != 0) {
1036 		total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1037 
1038 		if (total)
1039 			ratio = total / avg;
1040 
1041 		fprintf(output, " # %8.0f cycles / elision       ", ratio);
1042 	} else if (runtime_nsecs_stats[cpu].n != 0) {
1043 		char unit = 'M';
1044 
1045 		total = avg_stats(&runtime_nsecs_stats[cpu]);
1046 
1047 		if (total)
1048 			ratio = 1000.0 * avg / total;
1049 		if (ratio < 0.001) {
1050 			ratio *= 1000;
1051 			unit = 'K';
1052 		}
1053 
1054 		fprintf(output, " # %8.3f %c/sec                  ", ratio, unit);
1055 	} else {
1056 		fprintf(output, "                                   ");
1057 	}
1058 }
1059 
1060 static void print_aggr(char *prefix)
1061 {
1062 	struct perf_evsel *counter;
1063 	int cpu, cpu2, s, s2, id, nr;
1064 	u64 ena, run, val;
1065 
1066 	if (!(aggr_map || aggr_get_id))
1067 		return;
1068 
1069 	for (s = 0; s < aggr_map->nr; s++) {
1070 		id = aggr_map->map[s];
1071 		list_for_each_entry(counter, &evsel_list->entries, node) {
1072 			val = ena = run = 0;
1073 			nr = 0;
1074 			for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1075 				cpu2 = perf_evsel__cpus(counter)->map[cpu];
1076 				s2 = aggr_get_id(evsel_list->cpus, cpu2);
1077 				if (s2 != id)
1078 					continue;
1079 				val += counter->counts->cpu[cpu].val;
1080 				ena += counter->counts->cpu[cpu].ena;
1081 				run += counter->counts->cpu[cpu].run;
1082 				nr++;
1083 			}
1084 			if (prefix)
1085 				fprintf(output, "%s", prefix);
1086 
1087 			if (run == 0 || ena == 0) {
1088 				aggr_printout(counter, id, nr);
1089 
1090 				fprintf(output, "%*s%s%*s",
1091 					csv_output ? 0 : 18,
1092 					counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1093 					csv_sep,
1094 					csv_output ? 0 : -24,
1095 					perf_evsel__name(counter));
1096 
1097 				if (counter->cgrp)
1098 					fprintf(output, "%s%s",
1099 						csv_sep, counter->cgrp->name);
1100 
1101 				fputc('\n', output);
1102 				continue;
1103 			}
1104 
1105 			if (nsec_counter(counter))
1106 				nsec_printout(id, nr, counter, val);
1107 			else
1108 				abs_printout(id, nr, counter, val);
1109 
1110 			if (!csv_output) {
1111 				print_noise(counter, 1.0);
1112 
1113 				if (run != ena)
1114 					fprintf(output, "  (%.2f%%)",
1115 						100.0 * run / ena);
1116 			}
1117 			fputc('\n', output);
1118 		}
1119 	}
1120 }
1121 
1122 /*
1123  * Print out the results of a single counter:
1124  * aggregated counts in system-wide mode
1125  */
1126 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1127 {
1128 	struct perf_stat *ps = counter->priv;
1129 	double avg = avg_stats(&ps->res_stats[0]);
1130 	int scaled = counter->counts->scaled;
1131 
1132 	if (prefix)
1133 		fprintf(output, "%s", prefix);
1134 
1135 	if (scaled == -1) {
1136 		fprintf(output, "%*s%s%*s",
1137 			csv_output ? 0 : 18,
1138 			counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1139 			csv_sep,
1140 			csv_output ? 0 : -24,
1141 			perf_evsel__name(counter));
1142 
1143 		if (counter->cgrp)
1144 			fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1145 
1146 		fputc('\n', output);
1147 		return;
1148 	}
1149 
1150 	if (nsec_counter(counter))
1151 		nsec_printout(-1, 0, counter, avg);
1152 	else
1153 		abs_printout(-1, 0, counter, avg);
1154 
1155 	print_noise(counter, avg);
1156 
1157 	if (csv_output) {
1158 		fputc('\n', output);
1159 		return;
1160 	}
1161 
1162 	if (scaled) {
1163 		double avg_enabled, avg_running;
1164 
1165 		avg_enabled = avg_stats(&ps->res_stats[1]);
1166 		avg_running = avg_stats(&ps->res_stats[2]);
1167 
1168 		fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1169 	}
1170 	fprintf(output, "\n");
1171 }
1172 
1173 /*
1174  * Print out the results of a single counter:
1175  * does not use aggregated count in system-wide
1176  */
1177 static void print_counter(struct perf_evsel *counter, char *prefix)
1178 {
1179 	u64 ena, run, val;
1180 	int cpu;
1181 
1182 	for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1183 		val = counter->counts->cpu[cpu].val;
1184 		ena = counter->counts->cpu[cpu].ena;
1185 		run = counter->counts->cpu[cpu].run;
1186 
1187 		if (prefix)
1188 			fprintf(output, "%s", prefix);
1189 
1190 		if (run == 0 || ena == 0) {
1191 			fprintf(output, "CPU%*d%s%*s%s%*s",
1192 				csv_output ? 0 : -4,
1193 				perf_evsel__cpus(counter)->map[cpu], csv_sep,
1194 				csv_output ? 0 : 18,
1195 				counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1196 				csv_sep,
1197 				csv_output ? 0 : -24,
1198 				perf_evsel__name(counter));
1199 
1200 			if (counter->cgrp)
1201 				fprintf(output, "%s%s",
1202 					csv_sep, counter->cgrp->name);
1203 
1204 			fputc('\n', output);
1205 			continue;
1206 		}
1207 
1208 		if (nsec_counter(counter))
1209 			nsec_printout(cpu, 0, counter, val);
1210 		else
1211 			abs_printout(cpu, 0, counter, val);
1212 
1213 		if (!csv_output) {
1214 			print_noise(counter, 1.0);
1215 
1216 			if (run != ena)
1217 				fprintf(output, "  (%.2f%%)",
1218 					100.0 * run / ena);
1219 		}
1220 		fputc('\n', output);
1221 	}
1222 }
1223 
1224 static void print_stat(int argc, const char **argv)
1225 {
1226 	struct perf_evsel *counter;
1227 	int i;
1228 
1229 	fflush(stdout);
1230 
1231 	if (!csv_output) {
1232 		fprintf(output, "\n");
1233 		fprintf(output, " Performance counter stats for ");
1234 		if (target.system_wide)
1235 			fprintf(output, "\'system wide");
1236 		else if (target.cpu_list)
1237 			fprintf(output, "\'CPU(s) %s", target.cpu_list);
1238 		else if (!target__has_task(&target)) {
1239 			fprintf(output, "\'%s", argv[0]);
1240 			for (i = 1; i < argc; i++)
1241 				fprintf(output, " %s", argv[i]);
1242 		} else if (target.pid)
1243 			fprintf(output, "process id \'%s", target.pid);
1244 		else
1245 			fprintf(output, "thread id \'%s", target.tid);
1246 
1247 		fprintf(output, "\'");
1248 		if (run_count > 1)
1249 			fprintf(output, " (%d runs)", run_count);
1250 		fprintf(output, ":\n\n");
1251 	}
1252 
1253 	switch (aggr_mode) {
1254 	case AGGR_CORE:
1255 	case AGGR_SOCKET:
1256 		print_aggr(NULL);
1257 		break;
1258 	case AGGR_GLOBAL:
1259 		list_for_each_entry(counter, &evsel_list->entries, node)
1260 			print_counter_aggr(counter, NULL);
1261 		break;
1262 	case AGGR_NONE:
1263 		list_for_each_entry(counter, &evsel_list->entries, node)
1264 			print_counter(counter, NULL);
1265 		break;
1266 	default:
1267 		break;
1268 	}
1269 
1270 	if (!csv_output) {
1271 		if (!null_run)
1272 			fprintf(output, "\n");
1273 		fprintf(output, " %17.9f seconds time elapsed",
1274 				avg_stats(&walltime_nsecs_stats)/1e9);
1275 		if (run_count > 1) {
1276 			fprintf(output, "                                        ");
1277 			print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1278 					avg_stats(&walltime_nsecs_stats));
1279 		}
1280 		fprintf(output, "\n\n");
1281 	}
1282 }
1283 
1284 static volatile int signr = -1;
1285 
1286 static void skip_signal(int signo)
1287 {
1288 	if ((child_pid == -1) || interval)
1289 		done = 1;
1290 
1291 	signr = signo;
1292 	/*
1293 	 * render child_pid harmless
1294 	 * won't send SIGTERM to a random
1295 	 * process in case of race condition
1296 	 * and fast PID recycling
1297 	 */
1298 	child_pid = -1;
1299 }
1300 
1301 static void sig_atexit(void)
1302 {
1303 	sigset_t set, oset;
1304 
1305 	/*
1306 	 * avoid race condition with SIGCHLD handler
1307 	 * in skip_signal() which is modifying child_pid
1308 	 * goal is to avoid send SIGTERM to a random
1309 	 * process
1310 	 */
1311 	sigemptyset(&set);
1312 	sigaddset(&set, SIGCHLD);
1313 	sigprocmask(SIG_BLOCK, &set, &oset);
1314 
1315 	if (child_pid != -1)
1316 		kill(child_pid, SIGTERM);
1317 
1318 	sigprocmask(SIG_SETMASK, &oset, NULL);
1319 
1320 	if (signr == -1)
1321 		return;
1322 
1323 	signal(signr, SIG_DFL);
1324 	kill(getpid(), signr);
1325 }
1326 
1327 static int stat__set_big_num(const struct option *opt __maybe_unused,
1328 			     const char *s __maybe_unused, int unset)
1329 {
1330 	big_num_opt = unset ? 0 : 1;
1331 	return 0;
1332 }
1333 
1334 static int perf_stat_init_aggr_mode(void)
1335 {
1336 	switch (aggr_mode) {
1337 	case AGGR_SOCKET:
1338 		if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1339 			perror("cannot build socket map");
1340 			return -1;
1341 		}
1342 		aggr_get_id = cpu_map__get_socket;
1343 		break;
1344 	case AGGR_CORE:
1345 		if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1346 			perror("cannot build core map");
1347 			return -1;
1348 		}
1349 		aggr_get_id = cpu_map__get_core;
1350 		break;
1351 	case AGGR_NONE:
1352 	case AGGR_GLOBAL:
1353 	default:
1354 		break;
1355 	}
1356 	return 0;
1357 }
1358 
1359 static int setup_events(const char * const *attrs, unsigned len)
1360 {
1361 	unsigned i;
1362 
1363 	for (i = 0; i < len; i++) {
1364 		if (parse_events(evsel_list, attrs[i]))
1365 			return -1;
1366 	}
1367 	return 0;
1368 }
1369 
1370 /*
1371  * Add default attributes, if there were no attributes specified or
1372  * if -d/--detailed, -d -d or -d -d -d is used:
1373  */
1374 static int add_default_attributes(void)
1375 {
1376 	struct perf_event_attr default_attrs[] = {
1377 
1378   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1379   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1380   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1381   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1382 
1383   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1384   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1385   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1386   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1387   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1388   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1389 
1390 };
1391 
1392 /*
1393  * Detailed stats (-d), covering the L1 and last level data caches:
1394  */
1395 	struct perf_event_attr detailed_attrs[] = {
1396 
1397   { .type = PERF_TYPE_HW_CACHE,
1398     .config =
1399 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1400 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1401 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1402 
1403   { .type = PERF_TYPE_HW_CACHE,
1404     .config =
1405 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1406 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1407 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1408 
1409   { .type = PERF_TYPE_HW_CACHE,
1410     .config =
1411 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1412 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1413 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1414 
1415   { .type = PERF_TYPE_HW_CACHE,
1416     .config =
1417 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1418 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1419 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1420 };
1421 
1422 /*
1423  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1424  */
1425 	struct perf_event_attr very_detailed_attrs[] = {
1426 
1427   { .type = PERF_TYPE_HW_CACHE,
1428     .config =
1429 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1430 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1431 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1432 
1433   { .type = PERF_TYPE_HW_CACHE,
1434     .config =
1435 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1436 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1437 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1438 
1439   { .type = PERF_TYPE_HW_CACHE,
1440     .config =
1441 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1442 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1443 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1444 
1445   { .type = PERF_TYPE_HW_CACHE,
1446     .config =
1447 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1448 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1449 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1450 
1451   { .type = PERF_TYPE_HW_CACHE,
1452     .config =
1453 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1454 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1455 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1456 
1457   { .type = PERF_TYPE_HW_CACHE,
1458     .config =
1459 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1460 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1461 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1462 
1463 };
1464 
1465 /*
1466  * Very, very detailed stats (-d -d -d), adding prefetch events:
1467  */
1468 	struct perf_event_attr very_very_detailed_attrs[] = {
1469 
1470   { .type = PERF_TYPE_HW_CACHE,
1471     .config =
1472 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1473 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1474 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1475 
1476   { .type = PERF_TYPE_HW_CACHE,
1477     .config =
1478 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1479 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1480 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1481 };
1482 
1483 	/* Set attrs if no event is selected and !null_run: */
1484 	if (null_run)
1485 		return 0;
1486 
1487 	if (transaction_run) {
1488 		int err;
1489 		if (pmu_have_event("cpu", "cycles-ct") &&
1490 		    pmu_have_event("cpu", "el-start"))
1491 			err = setup_events(transaction_attrs,
1492 					ARRAY_SIZE(transaction_attrs));
1493 		else
1494 			err = setup_events(transaction_limited_attrs,
1495 				 ARRAY_SIZE(transaction_limited_attrs));
1496 		if (err < 0) {
1497 			fprintf(stderr, "Cannot set up transaction events\n");
1498 			return -1;
1499 		}
1500 		return 0;
1501 	}
1502 
1503 	if (!evsel_list->nr_entries) {
1504 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1505 			return -1;
1506 	}
1507 
1508 	/* Detailed events get appended to the event list: */
1509 
1510 	if (detailed_run <  1)
1511 		return 0;
1512 
1513 	/* Append detailed run extra attributes: */
1514 	if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1515 		return -1;
1516 
1517 	if (detailed_run < 2)
1518 		return 0;
1519 
1520 	/* Append very detailed run extra attributes: */
1521 	if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1522 		return -1;
1523 
1524 	if (detailed_run < 3)
1525 		return 0;
1526 
1527 	/* Append very, very detailed run extra attributes: */
1528 	return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1529 }
1530 
1531 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1532 {
1533 	bool append_file = false;
1534 	int output_fd = 0;
1535 	const char *output_name	= NULL;
1536 	const struct option options[] = {
1537 	OPT_BOOLEAN('T', "transaction", &transaction_run,
1538 		    "hardware transaction statistics"),
1539 	OPT_CALLBACK('e', "event", &evsel_list, "event",
1540 		     "event selector. use 'perf list' to list available events",
1541 		     parse_events_option),
1542 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1543 		     "event filter", parse_filter),
1544 	OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1545 		    "child tasks do not inherit counters"),
1546 	OPT_STRING('p', "pid", &target.pid, "pid",
1547 		   "stat events on existing process id"),
1548 	OPT_STRING('t', "tid", &target.tid, "tid",
1549 		   "stat events on existing thread id"),
1550 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1551 		    "system-wide collection from all CPUs"),
1552 	OPT_BOOLEAN('g', "group", &group,
1553 		    "put the counters into a counter group"),
1554 	OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1555 	OPT_INCR('v', "verbose", &verbose,
1556 		    "be more verbose (show counter open errors, etc)"),
1557 	OPT_INTEGER('r', "repeat", &run_count,
1558 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1559 	OPT_BOOLEAN('n', "null", &null_run,
1560 		    "null run - dont start any counters"),
1561 	OPT_INCR('d', "detailed", &detailed_run,
1562 		    "detailed run - start a lot of events"),
1563 	OPT_BOOLEAN('S', "sync", &sync_run,
1564 		    "call sync() before starting a run"),
1565 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1566 			   "print large numbers with thousands\' separators",
1567 			   stat__set_big_num),
1568 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1569 		    "list of cpus to monitor in system-wide"),
1570 	OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1571 		    "disable CPU count aggregation", AGGR_NONE),
1572 	OPT_STRING('x', "field-separator", &csv_sep, "separator",
1573 		   "print counts with custom separator"),
1574 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1575 		     "monitor event in cgroup name only", parse_cgroups),
1576 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1577 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1578 	OPT_INTEGER(0, "log-fd", &output_fd,
1579 		    "log output to fd, instead of stderr"),
1580 	OPT_STRING(0, "pre", &pre_cmd, "command",
1581 			"command to run prior to the measured command"),
1582 	OPT_STRING(0, "post", &post_cmd, "command",
1583 			"command to run after to the measured command"),
1584 	OPT_UINTEGER('I', "interval-print", &interval,
1585 		    "print counts at regular interval in ms (>= 100)"),
1586 	OPT_SET_UINT(0, "per-socket", &aggr_mode,
1587 		     "aggregate counts per processor socket", AGGR_SOCKET),
1588 	OPT_SET_UINT(0, "per-core", &aggr_mode,
1589 		     "aggregate counts per physical processor core", AGGR_CORE),
1590 	OPT_UINTEGER('D', "delay", &initial_delay,
1591 		     "ms to wait before starting measurement after program start"),
1592 	OPT_END()
1593 	};
1594 	const char * const stat_usage[] = {
1595 		"perf stat [<options>] [<command>]",
1596 		NULL
1597 	};
1598 	int status = -EINVAL, run_idx;
1599 	const char *mode;
1600 
1601 	setlocale(LC_ALL, "");
1602 
1603 	evsel_list = perf_evlist__new();
1604 	if (evsel_list == NULL)
1605 		return -ENOMEM;
1606 
1607 	argc = parse_options(argc, argv, options, stat_usage,
1608 		PARSE_OPT_STOP_AT_NON_OPTION);
1609 
1610 	output = stderr;
1611 	if (output_name && strcmp(output_name, "-"))
1612 		output = NULL;
1613 
1614 	if (output_name && output_fd) {
1615 		fprintf(stderr, "cannot use both --output and --log-fd\n");
1616 		parse_options_usage(stat_usage, options, "o", 1);
1617 		parse_options_usage(NULL, options, "log-fd", 0);
1618 		goto out;
1619 	}
1620 
1621 	if (output_fd < 0) {
1622 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
1623 		parse_options_usage(stat_usage, options, "log-fd", 0);
1624 		goto out;
1625 	}
1626 
1627 	if (!output) {
1628 		struct timespec tm;
1629 		mode = append_file ? "a" : "w";
1630 
1631 		output = fopen(output_name, mode);
1632 		if (!output) {
1633 			perror("failed to create output file");
1634 			return -1;
1635 		}
1636 		clock_gettime(CLOCK_REALTIME, &tm);
1637 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1638 	} else if (output_fd > 0) {
1639 		mode = append_file ? "a" : "w";
1640 		output = fdopen(output_fd, mode);
1641 		if (!output) {
1642 			perror("Failed opening logfd");
1643 			return -errno;
1644 		}
1645 	}
1646 
1647 	if (csv_sep) {
1648 		csv_output = true;
1649 		if (!strcmp(csv_sep, "\\t"))
1650 			csv_sep = "\t";
1651 	} else
1652 		csv_sep = DEFAULT_SEPARATOR;
1653 
1654 	/*
1655 	 * let the spreadsheet do the pretty-printing
1656 	 */
1657 	if (csv_output) {
1658 		/* User explicitly passed -B? */
1659 		if (big_num_opt == 1) {
1660 			fprintf(stderr, "-B option not supported with -x\n");
1661 			parse_options_usage(stat_usage, options, "B", 1);
1662 			parse_options_usage(NULL, options, "x", 1);
1663 			goto out;
1664 		} else /* Nope, so disable big number formatting */
1665 			big_num = false;
1666 	} else if (big_num_opt == 0) /* User passed --no-big-num */
1667 		big_num = false;
1668 
1669 	if (!argc && target__none(&target))
1670 		usage_with_options(stat_usage, options);
1671 
1672 	if (run_count < 0) {
1673 		pr_err("Run count must be a positive number\n");
1674 		parse_options_usage(stat_usage, options, "r", 1);
1675 		goto out;
1676 	} else if (run_count == 0) {
1677 		forever = true;
1678 		run_count = 1;
1679 	}
1680 
1681 	/* no_aggr, cgroup are for system-wide only */
1682 	if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1683 	    !target__has_cpu(&target)) {
1684 		fprintf(stderr, "both cgroup and no-aggregation "
1685 			"modes only available in system-wide mode\n");
1686 
1687 		parse_options_usage(stat_usage, options, "G", 1);
1688 		parse_options_usage(NULL, options, "A", 1);
1689 		parse_options_usage(NULL, options, "a", 1);
1690 		goto out;
1691 	}
1692 
1693 	if (add_default_attributes())
1694 		goto out;
1695 
1696 	target__validate(&target);
1697 
1698 	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1699 		if (target__has_task(&target)) {
1700 			pr_err("Problems finding threads of monitor\n");
1701 			parse_options_usage(stat_usage, options, "p", 1);
1702 			parse_options_usage(NULL, options, "t", 1);
1703 		} else if (target__has_cpu(&target)) {
1704 			perror("failed to parse CPUs map");
1705 			parse_options_usage(stat_usage, options, "C", 1);
1706 			parse_options_usage(NULL, options, "a", 1);
1707 		}
1708 		goto out;
1709 	}
1710 	if (interval && interval < 100) {
1711 		pr_err("print interval must be >= 100ms\n");
1712 		parse_options_usage(stat_usage, options, "I", 1);
1713 		goto out_free_maps;
1714 	}
1715 
1716 	if (perf_evlist__alloc_stats(evsel_list, interval))
1717 		goto out_free_maps;
1718 
1719 	if (perf_stat_init_aggr_mode())
1720 		goto out_free_maps;
1721 
1722 	/*
1723 	 * We dont want to block the signals - that would cause
1724 	 * child tasks to inherit that and Ctrl-C would not work.
1725 	 * What we want is for Ctrl-C to work in the exec()-ed
1726 	 * task, but being ignored by perf stat itself:
1727 	 */
1728 	atexit(sig_atexit);
1729 	if (!forever)
1730 		signal(SIGINT,  skip_signal);
1731 	signal(SIGCHLD, skip_signal);
1732 	signal(SIGALRM, skip_signal);
1733 	signal(SIGABRT, skip_signal);
1734 
1735 	status = 0;
1736 	for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1737 		if (run_count != 1 && verbose)
1738 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1739 				run_idx + 1);
1740 
1741 		status = run_perf_stat(argc, argv);
1742 		if (forever && status != -1) {
1743 			print_stat(argc, argv);
1744 			perf_stat__reset_stats(evsel_list);
1745 		}
1746 	}
1747 
1748 	if (!forever && status != -1 && !interval)
1749 		print_stat(argc, argv);
1750 
1751 	perf_evlist__free_stats(evsel_list);
1752 out_free_maps:
1753 	perf_evlist__delete_maps(evsel_list);
1754 out:
1755 	perf_evlist__delete(evsel_list);
1756 	return status;
1757 }
1758