xref: /openbmc/linux/tools/perf/builtin-stat.c (revision e6e8c6c2)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * builtin-stat.c
4  *
5  * Builtin stat command: Give a precise performance counters summary
6  * overview about any workload, CPU or specific PID.
7  *
8  * Sample output:
9 
10    $ perf stat ./hackbench 10
11 
12   Time: 0.118
13 
14   Performance counter stats for './hackbench 10':
15 
16        1708.761321 task-clock                #   11.037 CPUs utilized
17             41,190 context-switches          #    0.024 M/sec
18              6,735 CPU-migrations            #    0.004 M/sec
19             17,318 page-faults               #    0.010 M/sec
20      5,205,202,243 cycles                    #    3.046 GHz
21      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
22      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
23      2,603,501,247 instructions              #    0.50  insns per cycle
24                                              #    1.48  stalled cycles per insn
25        484,357,498 branches                  #  283.455 M/sec
26          6,388,934 branch-misses             #    1.32% of all branches
27 
28         0.154822978  seconds time elapsed
29 
30  *
31  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32  *
33  * Improvements and fixes by:
34  *
35  *   Arjan van de Ven <arjan@linux.intel.com>
36  *   Yanmin Zhang <yanmin.zhang@intel.com>
37  *   Wu Fengguang <fengguang.wu@intel.com>
38  *   Mike Galbraith <efault@gmx.de>
39  *   Paul Mackerras <paulus@samba.org>
40  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
41  */
42 
43 #include "builtin.h"
44 #include "perf.h"
45 #include "util/cgroup.h"
46 #include <subcmd/parse-options.h>
47 #include "util/parse-events.h"
48 #include "util/pmu.h"
49 #include "util/event.h"
50 #include "util/evlist.h"
51 #include "util/evlist-hybrid.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_map.h"
59 #include "util/counts.h"
60 #include "util/topdown.h"
61 #include "util/session.h"
62 #include "util/tool.h"
63 #include "util/string2.h"
64 #include "util/metricgroup.h"
65 #include "util/synthetic-events.h"
66 #include "util/target.h"
67 #include "util/time-utils.h"
68 #include "util/top.h"
69 #include "util/affinity.h"
70 #include "util/pfm.h"
71 #include "util/bpf_counter.h"
72 #include "util/iostat.h"
73 #include "util/pmu-hybrid.h"
74 #include "asm/bug.h"
75 
76 #include <linux/time64.h>
77 #include <linux/zalloc.h>
78 #include <api/fs/fs.h>
79 #include <errno.h>
80 #include <signal.h>
81 #include <stdlib.h>
82 #include <sys/prctl.h>
83 #include <inttypes.h>
84 #include <locale.h>
85 #include <math.h>
86 #include <sys/types.h>
87 #include <sys/stat.h>
88 #include <sys/wait.h>
89 #include <unistd.h>
90 #include <sys/time.h>
91 #include <sys/resource.h>
92 #include <linux/err.h>
93 
94 #include <linux/ctype.h>
95 #include <perf/evlist.h>
96 
97 #define DEFAULT_SEPARATOR	" "
98 #define FREEZE_ON_SMI_PATH	"devices/cpu/freeze_on_smi"
99 
100 static void print_counters(struct timespec *ts, int argc, const char **argv);
101 
102 /* Default events used for perf stat -T */
103 static const char *transaction_attrs = {
104 	"task-clock,"
105 	"{"
106 	"instructions,"
107 	"cycles,"
108 	"cpu/cycles-t/,"
109 	"cpu/tx-start/,"
110 	"cpu/el-start/,"
111 	"cpu/cycles-ct/"
112 	"}"
113 };
114 
115 /* More limited version when the CPU does not have all events. */
116 static const char * transaction_limited_attrs = {
117 	"task-clock,"
118 	"{"
119 	"instructions,"
120 	"cycles,"
121 	"cpu/cycles-t/,"
122 	"cpu/tx-start/"
123 	"}"
124 };
125 
126 static const char * topdown_attrs[] = {
127 	"topdown-total-slots",
128 	"topdown-slots-retired",
129 	"topdown-recovery-bubbles",
130 	"topdown-fetch-bubbles",
131 	"topdown-slots-issued",
132 	NULL,
133 };
134 
135 static const char *topdown_metric_attrs[] = {
136 	"slots",
137 	"topdown-retiring",
138 	"topdown-bad-spec",
139 	"topdown-fe-bound",
140 	"topdown-be-bound",
141 	NULL,
142 };
143 
144 static const char *topdown_metric_L2_attrs[] = {
145 	"slots",
146 	"topdown-retiring",
147 	"topdown-bad-spec",
148 	"topdown-fe-bound",
149 	"topdown-be-bound",
150 	"topdown-heavy-ops",
151 	"topdown-br-mispredict",
152 	"topdown-fetch-lat",
153 	"topdown-mem-bound",
154 	NULL,
155 };
156 
157 #define TOPDOWN_MAX_LEVEL			2
158 
159 static const char *smi_cost_attrs = {
160 	"{"
161 	"msr/aperf/,"
162 	"msr/smi/,"
163 	"cycles"
164 	"}"
165 };
166 
167 static struct evlist	*evsel_list;
168 static bool all_counters_use_bpf = true;
169 
170 static struct target target = {
171 	.uid	= UINT_MAX,
172 };
173 
174 #define METRIC_ONLY_LEN 20
175 
176 static volatile pid_t		child_pid			= -1;
177 static int			detailed_run			=  0;
178 static bool			transaction_run;
179 static bool			topdown_run			= false;
180 static bool			smi_cost			= false;
181 static bool			smi_reset			= false;
182 static int			big_num_opt			=  -1;
183 static bool			group				= false;
184 static const char		*pre_cmd			= NULL;
185 static const char		*post_cmd			= NULL;
186 static bool			sync_run			= false;
187 static bool			forever				= false;
188 static bool			force_metric_only		= false;
189 static struct timespec		ref_time;
190 static bool			append_file;
191 static bool			interval_count;
192 static const char		*output_name;
193 static int			output_fd;
194 
195 struct perf_stat {
196 	bool			 record;
197 	struct perf_data	 data;
198 	struct perf_session	*session;
199 	u64			 bytes_written;
200 	struct perf_tool	 tool;
201 	bool			 maps_allocated;
202 	struct perf_cpu_map	*cpus;
203 	struct perf_thread_map *threads;
204 	enum aggr_mode		 aggr_mode;
205 };
206 
207 static struct perf_stat		perf_stat;
208 #define STAT_RECORD		perf_stat.record
209 
210 static volatile int done = 0;
211 
212 static struct perf_stat_config stat_config = {
213 	.aggr_mode		= AGGR_GLOBAL,
214 	.scale			= true,
215 	.unit_width		= 4, /* strlen("unit") */
216 	.run_count		= 1,
217 	.metric_only_len	= METRIC_ONLY_LEN,
218 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
219 	.ru_stats		= &ru_stats,
220 	.big_num		= true,
221 	.ctl_fd			= -1,
222 	.ctl_fd_ack		= -1,
223 	.iostat_run		= false,
224 };
225 
226 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
227 {
228 	if (!a->core.cpus && !b->core.cpus)
229 		return true;
230 
231 	if (!a->core.cpus || !b->core.cpus)
232 		return false;
233 
234 	if (perf_cpu_map__nr(a->core.cpus) != perf_cpu_map__nr(b->core.cpus))
235 		return false;
236 
237 	for (int i = 0; i < perf_cpu_map__nr(a->core.cpus); i++) {
238 		if (perf_cpu_map__cpu(a->core.cpus, i).cpu !=
239 		    perf_cpu_map__cpu(b->core.cpus, i).cpu)
240 			return false;
241 	}
242 
243 	return true;
244 }
245 
246 static void evlist__check_cpu_maps(struct evlist *evlist)
247 {
248 	struct evsel *evsel, *pos, *leader;
249 	char buf[1024];
250 
251 	if (evlist__has_hybrid(evlist))
252 		evlist__warn_hybrid_group(evlist);
253 
254 	evlist__for_each_entry(evlist, evsel) {
255 		leader = evsel__leader(evsel);
256 
257 		/* Check that leader matches cpus with each member. */
258 		if (leader == evsel)
259 			continue;
260 		if (cpus_map_matched(leader, evsel))
261 			continue;
262 
263 		/* If there's mismatch disable the group and warn user. */
264 		WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
265 		evsel__group_desc(leader, buf, sizeof(buf));
266 		pr_warning("  %s\n", buf);
267 
268 		if (verbose) {
269 			cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
270 			pr_warning("     %s: %s\n", leader->name, buf);
271 			cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
272 			pr_warning("     %s: %s\n", evsel->name, buf);
273 		}
274 
275 		for_each_group_evsel(pos, leader)
276 			evsel__remove_from_group(pos, leader);
277 	}
278 }
279 
280 static inline void diff_timespec(struct timespec *r, struct timespec *a,
281 				 struct timespec *b)
282 {
283 	r->tv_sec = a->tv_sec - b->tv_sec;
284 	if (a->tv_nsec < b->tv_nsec) {
285 		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
286 		r->tv_sec--;
287 	} else {
288 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
289 	}
290 }
291 
292 static void perf_stat__reset_stats(void)
293 {
294 	int i;
295 
296 	evlist__reset_stats(evsel_list);
297 	perf_stat__reset_shadow_stats();
298 
299 	for (i = 0; i < stat_config.stats_num; i++)
300 		perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
301 }
302 
303 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
304 				     union perf_event *event,
305 				     struct perf_sample *sample __maybe_unused,
306 				     struct machine *machine __maybe_unused)
307 {
308 	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
309 		pr_err("failed to write perf data, error: %m\n");
310 		return -1;
311 	}
312 
313 	perf_stat.bytes_written += event->header.size;
314 	return 0;
315 }
316 
317 static int write_stat_round_event(u64 tm, u64 type)
318 {
319 	return perf_event__synthesize_stat_round(NULL, tm, type,
320 						 process_synthesized_event,
321 						 NULL);
322 }
323 
324 #define WRITE_STAT_ROUND_EVENT(time, interval) \
325 	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
326 
327 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
328 
329 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
330 				   struct perf_counts_values *count)
331 {
332 	struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
333 	struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
334 
335 	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
336 					   process_synthesized_event, NULL);
337 }
338 
339 static int read_single_counter(struct evsel *counter, int cpu_map_idx,
340 			       int thread, struct timespec *rs)
341 {
342 	switch(counter->tool_event) {
343 		case PERF_TOOL_DURATION_TIME: {
344 			u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
345 			struct perf_counts_values *count =
346 				perf_counts(counter->counts, cpu_map_idx, thread);
347 			count->ena = count->run = val;
348 			count->val = val;
349 			return 0;
350 		}
351 		case PERF_TOOL_USER_TIME:
352 		case PERF_TOOL_SYSTEM_TIME: {
353 			u64 val;
354 			struct perf_counts_values *count =
355 				perf_counts(counter->counts, cpu_map_idx, thread);
356 			if (counter->tool_event == PERF_TOOL_USER_TIME)
357 				val = ru_stats.ru_utime_usec_stat.mean;
358 			else
359 				val = ru_stats.ru_stime_usec_stat.mean;
360 			count->ena = count->run = val;
361 			count->val = val;
362 			return 0;
363 		}
364 		default:
365 		case PERF_TOOL_NONE:
366 			return evsel__read_counter(counter, cpu_map_idx, thread);
367 		case PERF_TOOL_MAX:
368 			/* This should never be reached */
369 			return 0;
370 	}
371 }
372 
373 /*
374  * Read out the results of a single counter:
375  * do not aggregate counts across CPUs in system-wide mode
376  */
377 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx)
378 {
379 	int nthreads = perf_thread_map__nr(evsel_list->core.threads);
380 	int thread;
381 
382 	if (!counter->supported)
383 		return -ENOENT;
384 
385 	for (thread = 0; thread < nthreads; thread++) {
386 		struct perf_counts_values *count;
387 
388 		count = perf_counts(counter->counts, cpu_map_idx, thread);
389 
390 		/*
391 		 * The leader's group read loads data into its group members
392 		 * (via evsel__read_counter()) and sets their count->loaded.
393 		 */
394 		if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
395 		    read_single_counter(counter, cpu_map_idx, thread, rs)) {
396 			counter->counts->scaled = -1;
397 			perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
398 			perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
399 			return -1;
400 		}
401 
402 		perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
403 
404 		if (STAT_RECORD) {
405 			if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
406 				pr_err("failed to write stat event\n");
407 				return -1;
408 			}
409 		}
410 
411 		if (verbose > 1) {
412 			fprintf(stat_config.output,
413 				"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
414 					evsel__name(counter),
415 					perf_cpu_map__cpu(evsel__cpus(counter),
416 							  cpu_map_idx).cpu,
417 					count->val, count->ena, count->run);
418 		}
419 	}
420 
421 	return 0;
422 }
423 
424 static int read_affinity_counters(struct timespec *rs)
425 {
426 	struct evlist_cpu_iterator evlist_cpu_itr;
427 	struct affinity saved_affinity, *affinity;
428 
429 	if (all_counters_use_bpf)
430 		return 0;
431 
432 	if (!target__has_cpu(&target) || target__has_per_thread(&target))
433 		affinity = NULL;
434 	else if (affinity__setup(&saved_affinity) < 0)
435 		return -1;
436 	else
437 		affinity = &saved_affinity;
438 
439 	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
440 		struct evsel *counter = evlist_cpu_itr.evsel;
441 
442 		if (evsel__is_bpf(counter))
443 			continue;
444 
445 		if (!counter->err) {
446 			counter->err = read_counter_cpu(counter, rs,
447 							evlist_cpu_itr.cpu_map_idx);
448 		}
449 	}
450 	if (affinity)
451 		affinity__cleanup(&saved_affinity);
452 
453 	return 0;
454 }
455 
456 static int read_bpf_map_counters(void)
457 {
458 	struct evsel *counter;
459 	int err;
460 
461 	evlist__for_each_entry(evsel_list, counter) {
462 		if (!evsel__is_bpf(counter))
463 			continue;
464 
465 		err = bpf_counter__read(counter);
466 		if (err)
467 			return err;
468 	}
469 	return 0;
470 }
471 
472 static void read_counters(struct timespec *rs)
473 {
474 	struct evsel *counter;
475 
476 	if (!stat_config.stop_read_counter) {
477 		if (read_bpf_map_counters() ||
478 		    read_affinity_counters(rs))
479 			return;
480 	}
481 
482 	evlist__for_each_entry(evsel_list, counter) {
483 		if (counter->err)
484 			pr_debug("failed to read counter %s\n", counter->name);
485 		if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
486 			pr_warning("failed to process counter %s\n", counter->name);
487 		counter->err = 0;
488 	}
489 }
490 
491 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
492 {
493 	int i;
494 
495 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
496 	if (!config->stats)
497 		return -1;
498 
499 	config->stats_num = nthreads;
500 
501 	for (i = 0; i < nthreads; i++)
502 		runtime_stat__init(&config->stats[i]);
503 
504 	return 0;
505 }
506 
507 static void runtime_stat_delete(struct perf_stat_config *config)
508 {
509 	int i;
510 
511 	if (!config->stats)
512 		return;
513 
514 	for (i = 0; i < config->stats_num; i++)
515 		runtime_stat__exit(&config->stats[i]);
516 
517 	zfree(&config->stats);
518 }
519 
520 static void runtime_stat_reset(struct perf_stat_config *config)
521 {
522 	int i;
523 
524 	if (!config->stats)
525 		return;
526 
527 	for (i = 0; i < config->stats_num; i++)
528 		perf_stat__reset_shadow_per_stat(&config->stats[i]);
529 }
530 
531 static void process_interval(void)
532 {
533 	struct timespec ts, rs;
534 
535 	clock_gettime(CLOCK_MONOTONIC, &ts);
536 	diff_timespec(&rs, &ts, &ref_time);
537 
538 	perf_stat__reset_shadow_per_stat(&rt_stat);
539 	runtime_stat_reset(&stat_config);
540 	read_counters(&rs);
541 
542 	if (STAT_RECORD) {
543 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
544 			pr_err("failed to write stat round event\n");
545 	}
546 
547 	init_stats(&walltime_nsecs_stats);
548 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
549 	print_counters(&rs, 0, NULL);
550 }
551 
552 static bool handle_interval(unsigned int interval, int *times)
553 {
554 	if (interval) {
555 		process_interval();
556 		if (interval_count && !(--(*times)))
557 			return true;
558 	}
559 	return false;
560 }
561 
562 static int enable_counters(void)
563 {
564 	struct evsel *evsel;
565 	int err;
566 
567 	evlist__for_each_entry(evsel_list, evsel) {
568 		if (!evsel__is_bpf(evsel))
569 			continue;
570 
571 		err = bpf_counter__enable(evsel);
572 		if (err)
573 			return err;
574 	}
575 
576 	if (stat_config.initial_delay < 0) {
577 		pr_info(EVLIST_DISABLED_MSG);
578 		return 0;
579 	}
580 
581 	if (stat_config.initial_delay > 0) {
582 		pr_info(EVLIST_DISABLED_MSG);
583 		usleep(stat_config.initial_delay * USEC_PER_MSEC);
584 	}
585 
586 	/*
587 	 * We need to enable counters only if:
588 	 * - we don't have tracee (attaching to task or cpu)
589 	 * - we have initial delay configured
590 	 */
591 	if (!target__none(&target) || stat_config.initial_delay) {
592 		if (!all_counters_use_bpf)
593 			evlist__enable(evsel_list);
594 		if (stat_config.initial_delay > 0)
595 			pr_info(EVLIST_ENABLED_MSG);
596 	}
597 	return 0;
598 }
599 
600 static void disable_counters(void)
601 {
602 	struct evsel *counter;
603 
604 	/*
605 	 * If we don't have tracee (attaching to task or cpu), counters may
606 	 * still be running. To get accurate group ratios, we must stop groups
607 	 * from counting before reading their constituent counters.
608 	 */
609 	if (!target__none(&target)) {
610 		evlist__for_each_entry(evsel_list, counter)
611 			bpf_counter__disable(counter);
612 		if (!all_counters_use_bpf)
613 			evlist__disable(evsel_list);
614 	}
615 }
616 
617 static volatile int workload_exec_errno;
618 
619 /*
620  * evlist__prepare_workload will send a SIGUSR1
621  * if the fork fails, since we asked by setting its
622  * want_signal to true.
623  */
624 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
625 					void *ucontext __maybe_unused)
626 {
627 	workload_exec_errno = info->si_value.sival_int;
628 }
629 
630 static bool evsel__should_store_id(struct evsel *counter)
631 {
632 	return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
633 }
634 
635 static bool is_target_alive(struct target *_target,
636 			    struct perf_thread_map *threads)
637 {
638 	struct stat st;
639 	int i;
640 
641 	if (!target__has_task(_target))
642 		return true;
643 
644 	for (i = 0; i < threads->nr; i++) {
645 		char path[PATH_MAX];
646 
647 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
648 			  threads->map[i].pid);
649 
650 		if (!stat(path, &st))
651 			return true;
652 	}
653 
654 	return false;
655 }
656 
657 static void process_evlist(struct evlist *evlist, unsigned int interval)
658 {
659 	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
660 
661 	if (evlist__ctlfd_process(evlist, &cmd) > 0) {
662 		switch (cmd) {
663 		case EVLIST_CTL_CMD_ENABLE:
664 			if (interval)
665 				process_interval();
666 			break;
667 		case EVLIST_CTL_CMD_DISABLE:
668 			if (interval)
669 				process_interval();
670 			break;
671 		case EVLIST_CTL_CMD_SNAPSHOT:
672 		case EVLIST_CTL_CMD_ACK:
673 		case EVLIST_CTL_CMD_UNSUPPORTED:
674 		case EVLIST_CTL_CMD_EVLIST:
675 		case EVLIST_CTL_CMD_STOP:
676 		case EVLIST_CTL_CMD_PING:
677 		default:
678 			break;
679 		}
680 	}
681 }
682 
683 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
684 			int *time_to_sleep)
685 {
686 	int tts = *time_to_sleep;
687 	struct timespec time_diff;
688 
689 	diff_timespec(&time_diff, time_stop, time_start);
690 
691 	tts -= time_diff.tv_sec * MSEC_PER_SEC +
692 	       time_diff.tv_nsec / NSEC_PER_MSEC;
693 
694 	if (tts < 0)
695 		tts = 0;
696 
697 	*time_to_sleep = tts;
698 }
699 
700 static int dispatch_events(bool forks, int timeout, int interval, int *times)
701 {
702 	int child_exited = 0, status = 0;
703 	int time_to_sleep, sleep_time;
704 	struct timespec time_start, time_stop;
705 
706 	if (interval)
707 		sleep_time = interval;
708 	else if (timeout)
709 		sleep_time = timeout;
710 	else
711 		sleep_time = 1000;
712 
713 	time_to_sleep = sleep_time;
714 
715 	while (!done) {
716 		if (forks)
717 			child_exited = waitpid(child_pid, &status, WNOHANG);
718 		else
719 			child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
720 
721 		if (child_exited)
722 			break;
723 
724 		clock_gettime(CLOCK_MONOTONIC, &time_start);
725 		if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
726 			if (timeout || handle_interval(interval, times))
727 				break;
728 			time_to_sleep = sleep_time;
729 		} else { /* fd revent */
730 			process_evlist(evsel_list, interval);
731 			clock_gettime(CLOCK_MONOTONIC, &time_stop);
732 			compute_tts(&time_start, &time_stop, &time_to_sleep);
733 		}
734 	}
735 
736 	return status;
737 }
738 
739 enum counter_recovery {
740 	COUNTER_SKIP,
741 	COUNTER_RETRY,
742 	COUNTER_FATAL,
743 };
744 
745 static enum counter_recovery stat_handle_error(struct evsel *counter)
746 {
747 	char msg[BUFSIZ];
748 	/*
749 	 * PPC returns ENXIO for HW counters until 2.6.37
750 	 * (behavior changed with commit b0a873e).
751 	 */
752 	if (errno == EINVAL || errno == ENOSYS ||
753 	    errno == ENOENT || errno == EOPNOTSUPP ||
754 	    errno == ENXIO) {
755 		if (verbose > 0)
756 			ui__warning("%s event is not supported by the kernel.\n",
757 				    evsel__name(counter));
758 		counter->supported = false;
759 		/*
760 		 * errored is a sticky flag that means one of the counter's
761 		 * cpu event had a problem and needs to be reexamined.
762 		 */
763 		counter->errored = true;
764 
765 		if ((evsel__leader(counter) != counter) ||
766 		    !(counter->core.leader->nr_members > 1))
767 			return COUNTER_SKIP;
768 	} else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
769 		if (verbose > 0)
770 			ui__warning("%s\n", msg);
771 		return COUNTER_RETRY;
772 	} else if (target__has_per_thread(&target) &&
773 		   evsel_list->core.threads &&
774 		   evsel_list->core.threads->err_thread != -1) {
775 		/*
776 		 * For global --per-thread case, skip current
777 		 * error thread.
778 		 */
779 		if (!thread_map__remove(evsel_list->core.threads,
780 					evsel_list->core.threads->err_thread)) {
781 			evsel_list->core.threads->err_thread = -1;
782 			return COUNTER_RETRY;
783 		}
784 	}
785 
786 	evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
787 	ui__error("%s\n", msg);
788 
789 	if (child_pid != -1)
790 		kill(child_pid, SIGTERM);
791 	return COUNTER_FATAL;
792 }
793 
794 static int __run_perf_stat(int argc, const char **argv, int run_idx)
795 {
796 	int interval = stat_config.interval;
797 	int times = stat_config.times;
798 	int timeout = stat_config.timeout;
799 	char msg[BUFSIZ];
800 	unsigned long long t0, t1;
801 	struct evsel *counter;
802 	size_t l;
803 	int status = 0;
804 	const bool forks = (argc > 0);
805 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
806 	struct evlist_cpu_iterator evlist_cpu_itr;
807 	struct affinity saved_affinity, *affinity = NULL;
808 	int err;
809 	bool second_pass = false;
810 
811 	if (forks) {
812 		if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
813 			perror("failed to prepare workload");
814 			return -1;
815 		}
816 		child_pid = evsel_list->workload.pid;
817 	}
818 
819 	if (group)
820 		evlist__set_leader(evsel_list);
821 
822 	if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
823 		if (affinity__setup(&saved_affinity) < 0)
824 			return -1;
825 		affinity = &saved_affinity;
826 	}
827 
828 	evlist__for_each_entry(evsel_list, counter) {
829 		counter->reset_group = false;
830 		if (bpf_counter__load(counter, &target))
831 			return -1;
832 		if (!evsel__is_bpf(counter))
833 			all_counters_use_bpf = false;
834 	}
835 
836 	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
837 		counter = evlist_cpu_itr.evsel;
838 
839 		/*
840 		 * bperf calls evsel__open_per_cpu() in bperf__load(), so
841 		 * no need to call it again here.
842 		 */
843 		if (target.use_bpf)
844 			break;
845 
846 		if (counter->reset_group || counter->errored)
847 			continue;
848 		if (evsel__is_bpf(counter))
849 			continue;
850 try_again:
851 		if (create_perf_stat_counter(counter, &stat_config, &target,
852 					     evlist_cpu_itr.cpu_map_idx) < 0) {
853 
854 			/*
855 			 * Weak group failed. We cannot just undo this here
856 			 * because earlier CPUs might be in group mode, and the kernel
857 			 * doesn't support mixing group and non group reads. Defer
858 			 * it to later.
859 			 * Don't close here because we're in the wrong affinity.
860 			 */
861 			if ((errno == EINVAL || errno == EBADF) &&
862 				evsel__leader(counter) != counter &&
863 				counter->weak_group) {
864 				evlist__reset_weak_group(evsel_list, counter, false);
865 				assert(counter->reset_group);
866 				second_pass = true;
867 				continue;
868 			}
869 
870 			switch (stat_handle_error(counter)) {
871 			case COUNTER_FATAL:
872 				return -1;
873 			case COUNTER_RETRY:
874 				goto try_again;
875 			case COUNTER_SKIP:
876 				continue;
877 			default:
878 				break;
879 			}
880 
881 		}
882 		counter->supported = true;
883 	}
884 
885 	if (second_pass) {
886 		/*
887 		 * Now redo all the weak group after closing them,
888 		 * and also close errored counters.
889 		 */
890 
891 		/* First close errored or weak retry */
892 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
893 			counter = evlist_cpu_itr.evsel;
894 
895 			if (!counter->reset_group && !counter->errored)
896 				continue;
897 
898 			perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
899 		}
900 		/* Now reopen weak */
901 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
902 			counter = evlist_cpu_itr.evsel;
903 
904 			if (!counter->reset_group && !counter->errored)
905 				continue;
906 			if (!counter->reset_group)
907 				continue;
908 try_again_reset:
909 			pr_debug2("reopening weak %s\n", evsel__name(counter));
910 			if (create_perf_stat_counter(counter, &stat_config, &target,
911 						     evlist_cpu_itr.cpu_map_idx) < 0) {
912 
913 				switch (stat_handle_error(counter)) {
914 				case COUNTER_FATAL:
915 					return -1;
916 				case COUNTER_RETRY:
917 					goto try_again_reset;
918 				case COUNTER_SKIP:
919 					continue;
920 				default:
921 					break;
922 				}
923 			}
924 			counter->supported = true;
925 		}
926 	}
927 	affinity__cleanup(affinity);
928 
929 	evlist__for_each_entry(evsel_list, counter) {
930 		if (!counter->supported) {
931 			perf_evsel__free_fd(&counter->core);
932 			continue;
933 		}
934 
935 		l = strlen(counter->unit);
936 		if (l > stat_config.unit_width)
937 			stat_config.unit_width = l;
938 
939 		if (evsel__should_store_id(counter) &&
940 		    evsel__store_ids(counter, evsel_list))
941 			return -1;
942 	}
943 
944 	if (evlist__apply_filters(evsel_list, &counter)) {
945 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
946 			counter->filter, evsel__name(counter), errno,
947 			str_error_r(errno, msg, sizeof(msg)));
948 		return -1;
949 	}
950 
951 	if (STAT_RECORD) {
952 		int fd = perf_data__fd(&perf_stat.data);
953 
954 		if (is_pipe) {
955 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
956 		} else {
957 			err = perf_session__write_header(perf_stat.session, evsel_list,
958 							 fd, false);
959 		}
960 
961 		if (err < 0)
962 			return err;
963 
964 		err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
965 							 process_synthesized_event, is_pipe);
966 		if (err < 0)
967 			return err;
968 	}
969 
970 	err = enable_counters();
971 	if (err)
972 		return -1;
973 
974 	/* Exec the command, if any */
975 	if (forks)
976 		evlist__start_workload(evsel_list);
977 
978 	t0 = rdclock();
979 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
980 
981 	if (forks) {
982 		if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
983 			status = dispatch_events(forks, timeout, interval, &times);
984 		if (child_pid != -1) {
985 			if (timeout)
986 				kill(child_pid, SIGTERM);
987 			wait4(child_pid, &status, 0, &stat_config.ru_data);
988 		}
989 
990 		if (workload_exec_errno) {
991 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
992 			pr_err("Workload failed: %s\n", emsg);
993 			return -1;
994 		}
995 
996 		if (WIFSIGNALED(status))
997 			psignal(WTERMSIG(status), argv[0]);
998 	} else {
999 		status = dispatch_events(forks, timeout, interval, &times);
1000 	}
1001 
1002 	disable_counters();
1003 
1004 	t1 = rdclock();
1005 
1006 	if (stat_config.walltime_run_table)
1007 		stat_config.walltime_run[run_idx] = t1 - t0;
1008 
1009 	if (interval && stat_config.summary) {
1010 		stat_config.interval = 0;
1011 		stat_config.stop_read_counter = true;
1012 		init_stats(&walltime_nsecs_stats);
1013 		update_stats(&walltime_nsecs_stats, t1 - t0);
1014 
1015 		if (stat_config.aggr_mode == AGGR_GLOBAL)
1016 			evlist__save_aggr_prev_raw_counts(evsel_list);
1017 
1018 		evlist__copy_prev_raw_counts(evsel_list);
1019 		evlist__reset_prev_raw_counts(evsel_list);
1020 		runtime_stat_reset(&stat_config);
1021 		perf_stat__reset_shadow_per_stat(&rt_stat);
1022 	} else {
1023 		update_stats(&walltime_nsecs_stats, t1 - t0);
1024 		update_rusage_stats(&ru_stats, &stat_config.ru_data);
1025 	}
1026 
1027 	/*
1028 	 * Closing a group leader splits the group, and as we only disable
1029 	 * group leaders, results in remaining events becoming enabled. To
1030 	 * avoid arbitrary skew, we must read all counters before closing any
1031 	 * group leaders.
1032 	 */
1033 	read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
1034 
1035 	/*
1036 	 * We need to keep evsel_list alive, because it's processed
1037 	 * later the evsel_list will be closed after.
1038 	 */
1039 	if (!STAT_RECORD)
1040 		evlist__close(evsel_list);
1041 
1042 	return WEXITSTATUS(status);
1043 }
1044 
1045 static int run_perf_stat(int argc, const char **argv, int run_idx)
1046 {
1047 	int ret;
1048 
1049 	if (pre_cmd) {
1050 		ret = system(pre_cmd);
1051 		if (ret)
1052 			return ret;
1053 	}
1054 
1055 	if (sync_run)
1056 		sync();
1057 
1058 	ret = __run_perf_stat(argc, argv, run_idx);
1059 	if (ret)
1060 		return ret;
1061 
1062 	if (post_cmd) {
1063 		ret = system(post_cmd);
1064 		if (ret)
1065 			return ret;
1066 	}
1067 
1068 	return ret;
1069 }
1070 
1071 static void print_counters(struct timespec *ts, int argc, const char **argv)
1072 {
1073 	/* Do not print anything if we record to the pipe. */
1074 	if (STAT_RECORD && perf_stat.data.is_pipe)
1075 		return;
1076 	if (stat_config.quiet)
1077 		return;
1078 
1079 	evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1080 }
1081 
1082 static volatile int signr = -1;
1083 
1084 static void skip_signal(int signo)
1085 {
1086 	if ((child_pid == -1) || stat_config.interval)
1087 		done = 1;
1088 
1089 	signr = signo;
1090 	/*
1091 	 * render child_pid harmless
1092 	 * won't send SIGTERM to a random
1093 	 * process in case of race condition
1094 	 * and fast PID recycling
1095 	 */
1096 	child_pid = -1;
1097 }
1098 
1099 static void sig_atexit(void)
1100 {
1101 	sigset_t set, oset;
1102 
1103 	/*
1104 	 * avoid race condition with SIGCHLD handler
1105 	 * in skip_signal() which is modifying child_pid
1106 	 * goal is to avoid send SIGTERM to a random
1107 	 * process
1108 	 */
1109 	sigemptyset(&set);
1110 	sigaddset(&set, SIGCHLD);
1111 	sigprocmask(SIG_BLOCK, &set, &oset);
1112 
1113 	if (child_pid != -1)
1114 		kill(child_pid, SIGTERM);
1115 
1116 	sigprocmask(SIG_SETMASK, &oset, NULL);
1117 
1118 	if (signr == -1)
1119 		return;
1120 
1121 	signal(signr, SIG_DFL);
1122 	kill(getpid(), signr);
1123 }
1124 
1125 void perf_stat__set_big_num(int set)
1126 {
1127 	stat_config.big_num = (set != 0);
1128 }
1129 
1130 void perf_stat__set_no_csv_summary(int set)
1131 {
1132 	stat_config.no_csv_summary = (set != 0);
1133 }
1134 
1135 static int stat__set_big_num(const struct option *opt __maybe_unused,
1136 			     const char *s __maybe_unused, int unset)
1137 {
1138 	big_num_opt = unset ? 0 : 1;
1139 	perf_stat__set_big_num(!unset);
1140 	return 0;
1141 }
1142 
1143 static int enable_metric_only(const struct option *opt __maybe_unused,
1144 			      const char *s __maybe_unused, int unset)
1145 {
1146 	force_metric_only = true;
1147 	stat_config.metric_only = !unset;
1148 	return 0;
1149 }
1150 
1151 static int parse_metric_groups(const struct option *opt,
1152 			       const char *str,
1153 			       int unset __maybe_unused)
1154 {
1155 	return metricgroup__parse_groups(opt, str,
1156 					 stat_config.metric_no_group,
1157 					 stat_config.metric_no_merge,
1158 					 &stat_config.metric_events);
1159 }
1160 
1161 static int parse_control_option(const struct option *opt,
1162 				const char *str,
1163 				int unset __maybe_unused)
1164 {
1165 	struct perf_stat_config *config = opt->value;
1166 
1167 	return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1168 }
1169 
1170 static int parse_stat_cgroups(const struct option *opt,
1171 			      const char *str, int unset)
1172 {
1173 	if (stat_config.cgroup_list) {
1174 		pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1175 		return -1;
1176 	}
1177 
1178 	return parse_cgroups(opt, str, unset);
1179 }
1180 
1181 static int parse_hybrid_type(const struct option *opt,
1182 			     const char *str,
1183 			     int unset __maybe_unused)
1184 {
1185 	struct evlist *evlist = *(struct evlist **)opt->value;
1186 
1187 	if (!list_empty(&evlist->core.entries)) {
1188 		fprintf(stderr, "Must define cputype before events/metrics\n");
1189 		return -1;
1190 	}
1191 
1192 	evlist->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu(str);
1193 	if (!evlist->hybrid_pmu_name) {
1194 		fprintf(stderr, "--cputype %s is not supported!\n", str);
1195 		return -1;
1196 	}
1197 
1198 	return 0;
1199 }
1200 
1201 static struct option stat_options[] = {
1202 	OPT_BOOLEAN('T', "transaction", &transaction_run,
1203 		    "hardware transaction statistics"),
1204 	OPT_CALLBACK('e', "event", &evsel_list, "event",
1205 		     "event selector. use 'perf list' to list available events",
1206 		     parse_events_option),
1207 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1208 		     "event filter", parse_filter),
1209 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1210 		    "child tasks do not inherit counters"),
1211 	OPT_STRING('p', "pid", &target.pid, "pid",
1212 		   "stat events on existing process id"),
1213 	OPT_STRING('t', "tid", &target.tid, "tid",
1214 		   "stat events on existing thread id"),
1215 #ifdef HAVE_BPF_SKEL
1216 	OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1217 		   "stat events on existing bpf program id"),
1218 	OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1219 		    "use bpf program to count events"),
1220 	OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1221 		   "path to perf_event_attr map"),
1222 #endif
1223 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1224 		    "system-wide collection from all CPUs"),
1225 	OPT_BOOLEAN('g', "group", &group,
1226 		    "put the counters into a counter group"),
1227 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
1228 		    "Use --no-scale to disable counter scaling for multiplexing"),
1229 	OPT_INCR('v', "verbose", &verbose,
1230 		    "be more verbose (show counter open errors, etc)"),
1231 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
1232 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1233 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1234 		    "display details about each run (only with -r option)"),
1235 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
1236 		    "null run - dont start any counters"),
1237 	OPT_INCR('d', "detailed", &detailed_run,
1238 		    "detailed run - start a lot of events"),
1239 	OPT_BOOLEAN('S', "sync", &sync_run,
1240 		    "call sync() before starting a run"),
1241 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1242 			   "print large numbers with thousands\' separators",
1243 			   stat__set_big_num),
1244 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1245 		    "list of cpus to monitor in system-wide"),
1246 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1247 		    "disable CPU count aggregation", AGGR_NONE),
1248 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1249 	OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
1250 		    "Merge identical named hybrid events"),
1251 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1252 		   "print counts with custom separator"),
1253 	OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
1254 		   "print counts in JSON format"),
1255 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1256 		     "monitor event in cgroup name only", parse_stat_cgroups),
1257 	OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1258 		    "expand events for each cgroup"),
1259 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1260 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1261 	OPT_INTEGER(0, "log-fd", &output_fd,
1262 		    "log output to fd, instead of stderr"),
1263 	OPT_STRING(0, "pre", &pre_cmd, "command",
1264 			"command to run prior to the measured command"),
1265 	OPT_STRING(0, "post", &post_cmd, "command",
1266 			"command to run after to the measured command"),
1267 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1268 		    "print counts at regular interval in ms "
1269 		    "(overhead is possible for values <= 100ms)"),
1270 	OPT_INTEGER(0, "interval-count", &stat_config.times,
1271 		    "print counts for fixed number of times"),
1272 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1273 		    "clear screen in between new interval"),
1274 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1275 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1276 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1277 		     "aggregate counts per processor socket", AGGR_SOCKET),
1278 	OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1279 		     "aggregate counts per processor die", AGGR_DIE),
1280 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1281 		     "aggregate counts per physical processor core", AGGR_CORE),
1282 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1283 		     "aggregate counts per thread", AGGR_THREAD),
1284 	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1285 		     "aggregate counts per numa node", AGGR_NODE),
1286 	OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1287 		    "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1288 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1289 			"Only print computed metrics. No raw values", enable_metric_only),
1290 	OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1291 		       "don't group metric events, impacts multiplexing"),
1292 	OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1293 		       "don't try to share events between metrics in a group"),
1294 	OPT_BOOLEAN(0, "topdown", &topdown_run,
1295 			"measure top-down statistics"),
1296 	OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1297 			"Set the metrics level for the top-down statistics (0: max level)"),
1298 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1299 			"measure SMI cost"),
1300 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1301 		     "monitor specified metrics or metric groups (separated by ,)",
1302 		     parse_metric_groups),
1303 	OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1304 			 "Configure all used events to run in kernel space.",
1305 			 PARSE_OPT_EXCLUSIVE),
1306 	OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1307 			 "Configure all used events to run in user space.",
1308 			 PARSE_OPT_EXCLUSIVE),
1309 	OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1310 		    "Use with 'percore' event qualifier to show the event "
1311 		    "counts of one hardware thread by sum up total hardware "
1312 		    "threads of same physical core"),
1313 	OPT_BOOLEAN(0, "summary", &stat_config.summary,
1314 		       "print summary for interval mode"),
1315 	OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1316 		       "don't print 'summary' for CSV summary output"),
1317 	OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1318 			"don't print output (useful with record)"),
1319 	OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1320 		     "Only enable events on applying cpu with this type "
1321 		     "for hybrid platform (e.g. core or atom)",
1322 		     parse_hybrid_type),
1323 #ifdef HAVE_LIBPFM
1324 	OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1325 		"libpfm4 event selector. use 'perf list' to list available events",
1326 		parse_libpfm_events_option),
1327 #endif
1328 	OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1329 		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1330 		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1331 		     "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1332 		      parse_control_option),
1333 	OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1334 			    "measure I/O performance metrics provided by arch/platform",
1335 			    iostat_parse),
1336 	OPT_END()
1337 };
1338 
1339 static const char *const aggr_mode__string[] = {
1340 	[AGGR_CORE] = "core",
1341 	[AGGR_DIE] = "die",
1342 	[AGGR_GLOBAL] = "global",
1343 	[AGGR_NODE] = "node",
1344 	[AGGR_NONE] = "none",
1345 	[AGGR_SOCKET] = "socket",
1346 	[AGGR_THREAD] = "thread",
1347 	[AGGR_UNSET] = "unset",
1348 };
1349 
1350 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1351 						struct perf_cpu cpu)
1352 {
1353 	return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1354 }
1355 
1356 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1357 					     struct perf_cpu cpu)
1358 {
1359 	return aggr_cpu_id__die(cpu, /*data=*/NULL);
1360 }
1361 
1362 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1363 					      struct perf_cpu cpu)
1364 {
1365 	return aggr_cpu_id__core(cpu, /*data=*/NULL);
1366 }
1367 
1368 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1369 					      struct perf_cpu cpu)
1370 {
1371 	return aggr_cpu_id__node(cpu, /*data=*/NULL);
1372 }
1373 
1374 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1375 					      aggr_get_id_t get_id, struct perf_cpu cpu)
1376 {
1377 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1378 
1379 	if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1380 		config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1381 
1382 	id = config->cpus_aggr_map->map[cpu.cpu];
1383 	return id;
1384 }
1385 
1386 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1387 						       struct perf_cpu cpu)
1388 {
1389 	return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1390 }
1391 
1392 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1393 						    struct perf_cpu cpu)
1394 {
1395 	return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1396 }
1397 
1398 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1399 						     struct perf_cpu cpu)
1400 {
1401 	return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1402 }
1403 
1404 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1405 						     struct perf_cpu cpu)
1406 {
1407 	return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1408 }
1409 
1410 static bool term_percore_set(void)
1411 {
1412 	struct evsel *counter;
1413 
1414 	evlist__for_each_entry(evsel_list, counter) {
1415 		if (counter->percore)
1416 			return true;
1417 	}
1418 
1419 	return false;
1420 }
1421 
1422 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1423 {
1424 	switch (aggr_mode) {
1425 	case AGGR_SOCKET:
1426 		return aggr_cpu_id__socket;
1427 	case AGGR_DIE:
1428 		return aggr_cpu_id__die;
1429 	case AGGR_CORE:
1430 		return aggr_cpu_id__core;
1431 	case AGGR_NODE:
1432 		return aggr_cpu_id__node;
1433 	case AGGR_NONE:
1434 		if (term_percore_set())
1435 			return aggr_cpu_id__core;
1436 
1437 		return NULL;
1438 	case AGGR_GLOBAL:
1439 	case AGGR_THREAD:
1440 	case AGGR_UNSET:
1441 	case AGGR_MAX:
1442 	default:
1443 		return NULL;
1444 	}
1445 }
1446 
1447 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1448 {
1449 	switch (aggr_mode) {
1450 	case AGGR_SOCKET:
1451 		return perf_stat__get_socket_cached;
1452 	case AGGR_DIE:
1453 		return perf_stat__get_die_cached;
1454 	case AGGR_CORE:
1455 		return perf_stat__get_core_cached;
1456 	case AGGR_NODE:
1457 		return perf_stat__get_node_cached;
1458 	case AGGR_NONE:
1459 		if (term_percore_set()) {
1460 			return perf_stat__get_core_cached;
1461 		}
1462 		return NULL;
1463 	case AGGR_GLOBAL:
1464 	case AGGR_THREAD:
1465 	case AGGR_UNSET:
1466 	case AGGR_MAX:
1467 	default:
1468 		return NULL;
1469 	}
1470 }
1471 
1472 static int perf_stat_init_aggr_mode(void)
1473 {
1474 	int nr;
1475 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1476 
1477 	if (get_id) {
1478 		stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1479 							 get_id, /*data=*/NULL);
1480 		if (!stat_config.aggr_map) {
1481 			pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1482 			return -1;
1483 		}
1484 		stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1485 	}
1486 
1487 	/*
1488 	 * The evsel_list->cpus is the base we operate on,
1489 	 * taking the highest cpu number to be the size of
1490 	 * the aggregation translate cpumap.
1491 	 */
1492 	if (evsel_list->core.user_requested_cpus)
1493 		nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1494 	else
1495 		nr = 0;
1496 	stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1497 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1498 }
1499 
1500 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1501 {
1502 	if (map) {
1503 		WARN_ONCE(refcount_read(&map->refcnt) != 0,
1504 			  "cpu_aggr_map refcnt unbalanced\n");
1505 		free(map);
1506 	}
1507 }
1508 
1509 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1510 {
1511 	if (map && refcount_dec_and_test(&map->refcnt))
1512 		cpu_aggr_map__delete(map);
1513 }
1514 
1515 static void perf_stat__exit_aggr_mode(void)
1516 {
1517 	cpu_aggr_map__put(stat_config.aggr_map);
1518 	cpu_aggr_map__put(stat_config.cpus_aggr_map);
1519 	stat_config.aggr_map = NULL;
1520 	stat_config.cpus_aggr_map = NULL;
1521 }
1522 
1523 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1524 {
1525 	struct perf_env *env = data;
1526 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1527 
1528 	if (cpu.cpu != -1)
1529 		id.socket = env->cpu[cpu.cpu].socket_id;
1530 
1531 	return id;
1532 }
1533 
1534 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1535 {
1536 	struct perf_env *env = data;
1537 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1538 
1539 	if (cpu.cpu != -1) {
1540 		/*
1541 		 * die_id is relative to socket, so start
1542 		 * with the socket ID and then add die to
1543 		 * make a unique ID.
1544 		 */
1545 		id.socket = env->cpu[cpu.cpu].socket_id;
1546 		id.die = env->cpu[cpu.cpu].die_id;
1547 	}
1548 
1549 	return id;
1550 }
1551 
1552 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1553 {
1554 	struct perf_env *env = data;
1555 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1556 
1557 	if (cpu.cpu != -1) {
1558 		/*
1559 		 * core_id is relative to socket and die,
1560 		 * we need a global id. So we set
1561 		 * socket, die id and core id
1562 		 */
1563 		id.socket = env->cpu[cpu.cpu].socket_id;
1564 		id.die = env->cpu[cpu.cpu].die_id;
1565 		id.core = env->cpu[cpu.cpu].core_id;
1566 	}
1567 
1568 	return id;
1569 }
1570 
1571 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1572 {
1573 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1574 
1575 	id.node = perf_env__numa_node(data, cpu);
1576 	return id;
1577 }
1578 
1579 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1580 						     struct perf_cpu cpu)
1581 {
1582 	return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1583 }
1584 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1585 						  struct perf_cpu cpu)
1586 {
1587 	return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1588 }
1589 
1590 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1591 						   struct perf_cpu cpu)
1592 {
1593 	return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1594 }
1595 
1596 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1597 						   struct perf_cpu cpu)
1598 {
1599 	return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1600 }
1601 
1602 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1603 {
1604 	switch (aggr_mode) {
1605 	case AGGR_SOCKET:
1606 		return perf_env__get_socket_aggr_by_cpu;
1607 	case AGGR_DIE:
1608 		return perf_env__get_die_aggr_by_cpu;
1609 	case AGGR_CORE:
1610 		return perf_env__get_core_aggr_by_cpu;
1611 	case AGGR_NODE:
1612 		return perf_env__get_node_aggr_by_cpu;
1613 	case AGGR_NONE:
1614 	case AGGR_GLOBAL:
1615 	case AGGR_THREAD:
1616 	case AGGR_UNSET:
1617 	case AGGR_MAX:
1618 	default:
1619 		return NULL;
1620 	}
1621 }
1622 
1623 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1624 {
1625 	switch (aggr_mode) {
1626 	case AGGR_SOCKET:
1627 		return perf_stat__get_socket_file;
1628 	case AGGR_DIE:
1629 		return perf_stat__get_die_file;
1630 	case AGGR_CORE:
1631 		return perf_stat__get_core_file;
1632 	case AGGR_NODE:
1633 		return perf_stat__get_node_file;
1634 	case AGGR_NONE:
1635 	case AGGR_GLOBAL:
1636 	case AGGR_THREAD:
1637 	case AGGR_UNSET:
1638 	case AGGR_MAX:
1639 	default:
1640 		return NULL;
1641 	}
1642 }
1643 
1644 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1645 {
1646 	struct perf_env *env = &st->session->header.env;
1647 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1648 
1649 	if (!get_id)
1650 		return 0;
1651 
1652 	stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, get_id, env);
1653 	if (!stat_config.aggr_map) {
1654 		pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1655 		return -1;
1656 	}
1657 	stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1658 	return 0;
1659 }
1660 
1661 /*
1662  * Add default attributes, if there were no attributes specified or
1663  * if -d/--detailed, -d -d or -d -d -d is used:
1664  */
1665 static int add_default_attributes(void)
1666 {
1667 	int err;
1668 	struct perf_event_attr default_attrs0[] = {
1669 
1670   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1671   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1672   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1673   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1674 
1675   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1676 };
1677 	struct perf_event_attr frontend_attrs[] = {
1678   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1679 };
1680 	struct perf_event_attr backend_attrs[] = {
1681   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1682 };
1683 	struct perf_event_attr default_attrs1[] = {
1684   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1685   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1686   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1687 
1688 };
1689 
1690 /*
1691  * Detailed stats (-d), covering the L1 and last level data caches:
1692  */
1693 	struct perf_event_attr detailed_attrs[] = {
1694 
1695   { .type = PERF_TYPE_HW_CACHE,
1696     .config =
1697 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1698 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1699 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1700 
1701   { .type = PERF_TYPE_HW_CACHE,
1702     .config =
1703 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1704 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1705 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1706 
1707   { .type = PERF_TYPE_HW_CACHE,
1708     .config =
1709 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1710 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1711 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1712 
1713   { .type = PERF_TYPE_HW_CACHE,
1714     .config =
1715 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1716 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1717 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1718 };
1719 
1720 /*
1721  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1722  */
1723 	struct perf_event_attr very_detailed_attrs[] = {
1724 
1725   { .type = PERF_TYPE_HW_CACHE,
1726     .config =
1727 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1728 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1729 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1730 
1731   { .type = PERF_TYPE_HW_CACHE,
1732     .config =
1733 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1734 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1735 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1736 
1737   { .type = PERF_TYPE_HW_CACHE,
1738     .config =
1739 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1740 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1741 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1742 
1743   { .type = PERF_TYPE_HW_CACHE,
1744     .config =
1745 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1746 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1747 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1748 
1749   { .type = PERF_TYPE_HW_CACHE,
1750     .config =
1751 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1752 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1753 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1754 
1755   { .type = PERF_TYPE_HW_CACHE,
1756     .config =
1757 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1758 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1759 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1760 
1761 };
1762 
1763 /*
1764  * Very, very detailed stats (-d -d -d), adding prefetch events:
1765  */
1766 	struct perf_event_attr very_very_detailed_attrs[] = {
1767 
1768   { .type = PERF_TYPE_HW_CACHE,
1769     .config =
1770 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1771 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1772 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1773 
1774   { .type = PERF_TYPE_HW_CACHE,
1775     .config =
1776 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1777 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1778 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1779 };
1780 
1781 	struct perf_event_attr default_null_attrs[] = {};
1782 
1783 	/* Set attrs if no event is selected and !null_run: */
1784 	if (stat_config.null_run)
1785 		return 0;
1786 
1787 	if (transaction_run) {
1788 		struct parse_events_error errinfo;
1789 		/* Handle -T as -M transaction. Once platform specific metrics
1790 		 * support has been added to the json files, all architectures
1791 		 * will use this approach. To determine transaction support
1792 		 * on an architecture test for such a metric name.
1793 		 */
1794 		if (metricgroup__has_metric("transaction")) {
1795 			struct option opt = { .value = &evsel_list };
1796 
1797 			return metricgroup__parse_groups(&opt, "transaction",
1798 							 stat_config.metric_no_group,
1799 							stat_config.metric_no_merge,
1800 							 &stat_config.metric_events);
1801 		}
1802 
1803 		parse_events_error__init(&errinfo);
1804 		if (pmu_have_event("cpu", "cycles-ct") &&
1805 		    pmu_have_event("cpu", "el-start"))
1806 			err = parse_events(evsel_list, transaction_attrs,
1807 					   &errinfo);
1808 		else
1809 			err = parse_events(evsel_list,
1810 					   transaction_limited_attrs,
1811 					   &errinfo);
1812 		if (err) {
1813 			fprintf(stderr, "Cannot set up transaction events\n");
1814 			parse_events_error__print(&errinfo, transaction_attrs);
1815 		}
1816 		parse_events_error__exit(&errinfo);
1817 		return err ? -1 : 0;
1818 	}
1819 
1820 	if (smi_cost) {
1821 		struct parse_events_error errinfo;
1822 		int smi;
1823 
1824 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1825 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1826 			return -1;
1827 		}
1828 
1829 		if (!smi) {
1830 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1831 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1832 				return -1;
1833 			}
1834 			smi_reset = true;
1835 		}
1836 
1837 		if (!pmu_have_event("msr", "aperf") ||
1838 		    !pmu_have_event("msr", "smi")) {
1839 			fprintf(stderr, "To measure SMI cost, it needs "
1840 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1841 			return -1;
1842 		}
1843 		if (!force_metric_only)
1844 			stat_config.metric_only = true;
1845 
1846 		parse_events_error__init(&errinfo);
1847 		err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1848 		if (err) {
1849 			parse_events_error__print(&errinfo, smi_cost_attrs);
1850 			fprintf(stderr, "Cannot set up SMI cost events\n");
1851 		}
1852 		parse_events_error__exit(&errinfo);
1853 		return err ? -1 : 0;
1854 	}
1855 
1856 	if (topdown_run) {
1857 		const char **metric_attrs = topdown_metric_attrs;
1858 		unsigned int max_level = 1;
1859 		char *str = NULL;
1860 		bool warn = false;
1861 		const char *pmu_name = arch_get_topdown_pmu_name(evsel_list, true);
1862 
1863 		if (!force_metric_only)
1864 			stat_config.metric_only = true;
1865 
1866 		if (pmu_have_event(pmu_name, topdown_metric_L2_attrs[5])) {
1867 			metric_attrs = topdown_metric_L2_attrs;
1868 			max_level = 2;
1869 		}
1870 
1871 		if (stat_config.topdown_level > max_level) {
1872 			pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1873 			return -1;
1874 		} else if (!stat_config.topdown_level)
1875 			stat_config.topdown_level = max_level;
1876 
1877 		if (topdown_filter_events(metric_attrs, &str, 1, pmu_name) < 0) {
1878 			pr_err("Out of memory\n");
1879 			return -1;
1880 		}
1881 
1882 		if (metric_attrs[0] && str) {
1883 			if (!stat_config.interval && !stat_config.metric_only) {
1884 				fprintf(stat_config.output,
1885 					"Topdown accuracy may decrease when measuring long periods.\n"
1886 					"Please print the result regularly, e.g. -I1000\n");
1887 			}
1888 			goto setup_metrics;
1889 		}
1890 
1891 		zfree(&str);
1892 
1893 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1894 		    stat_config.aggr_mode != AGGR_CORE) {
1895 			pr_err("top down event configuration requires --per-core mode\n");
1896 			return -1;
1897 		}
1898 		stat_config.aggr_mode = AGGR_CORE;
1899 		if (nr_cgroups || !target__has_cpu(&target)) {
1900 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1901 			return -1;
1902 		}
1903 
1904 		if (topdown_filter_events(topdown_attrs, &str,
1905 				arch_topdown_check_group(&warn),
1906 				pmu_name) < 0) {
1907 			pr_err("Out of memory\n");
1908 			return -1;
1909 		}
1910 
1911 		if (topdown_attrs[0] && str) {
1912 			struct parse_events_error errinfo;
1913 			if (warn)
1914 				arch_topdown_group_warn();
1915 setup_metrics:
1916 			parse_events_error__init(&errinfo);
1917 			err = parse_events(evsel_list, str, &errinfo);
1918 			if (err) {
1919 				fprintf(stderr,
1920 					"Cannot set up top down events %s: %d\n",
1921 					str, err);
1922 				parse_events_error__print(&errinfo, str);
1923 				parse_events_error__exit(&errinfo);
1924 				free(str);
1925 				return -1;
1926 			}
1927 			parse_events_error__exit(&errinfo);
1928 		} else {
1929 			fprintf(stderr, "System does not support topdown\n");
1930 			return -1;
1931 		}
1932 		free(str);
1933 	}
1934 
1935 	if (!evsel_list->core.nr_entries) {
1936 		if (target__has_cpu(&target))
1937 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1938 
1939 		if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1940 			return -1;
1941 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1942 			if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1943 				return -1;
1944 		}
1945 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1946 			if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1947 				return -1;
1948 		}
1949 		if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1950 			return -1;
1951 
1952 		stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1953 		/* Platform specific attrs */
1954 		if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0)
1955 			return -1;
1956 	}
1957 
1958 	/* Detailed events get appended to the event list: */
1959 
1960 	if (detailed_run <  1)
1961 		return 0;
1962 
1963 	/* Append detailed run extra attributes: */
1964 	if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1965 		return -1;
1966 
1967 	if (detailed_run < 2)
1968 		return 0;
1969 
1970 	/* Append very detailed run extra attributes: */
1971 	if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1972 		return -1;
1973 
1974 	if (detailed_run < 3)
1975 		return 0;
1976 
1977 	/* Append very, very detailed run extra attributes: */
1978 	return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1979 }
1980 
1981 static const char * const stat_record_usage[] = {
1982 	"perf stat record [<options>]",
1983 	NULL,
1984 };
1985 
1986 static void init_features(struct perf_session *session)
1987 {
1988 	int feat;
1989 
1990 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1991 		perf_header__set_feat(&session->header, feat);
1992 
1993 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1994 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1995 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1996 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1997 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1998 }
1999 
2000 static int __cmd_record(int argc, const char **argv)
2001 {
2002 	struct perf_session *session;
2003 	struct perf_data *data = &perf_stat.data;
2004 
2005 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
2006 			     PARSE_OPT_STOP_AT_NON_OPTION);
2007 
2008 	if (output_name)
2009 		data->path = output_name;
2010 
2011 	if (stat_config.run_count != 1 || forever) {
2012 		pr_err("Cannot use -r option with perf stat record.\n");
2013 		return -1;
2014 	}
2015 
2016 	session = perf_session__new(data, NULL);
2017 	if (IS_ERR(session)) {
2018 		pr_err("Perf session creation failed\n");
2019 		return PTR_ERR(session);
2020 	}
2021 
2022 	init_features(session);
2023 
2024 	session->evlist   = evsel_list;
2025 	perf_stat.session = session;
2026 	perf_stat.record  = true;
2027 	return argc;
2028 }
2029 
2030 static int process_stat_round_event(struct perf_session *session,
2031 				    union perf_event *event)
2032 {
2033 	struct perf_record_stat_round *stat_round = &event->stat_round;
2034 	struct evsel *counter;
2035 	struct timespec tsh, *ts = NULL;
2036 	const char **argv = session->header.env.cmdline_argv;
2037 	int argc = session->header.env.nr_cmdline;
2038 
2039 	evlist__for_each_entry(evsel_list, counter)
2040 		perf_stat_process_counter(&stat_config, counter);
2041 
2042 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2043 		update_stats(&walltime_nsecs_stats, stat_round->time);
2044 
2045 	if (stat_config.interval && stat_round->time) {
2046 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2047 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2048 		ts = &tsh;
2049 	}
2050 
2051 	print_counters(ts, argc, argv);
2052 	return 0;
2053 }
2054 
2055 static
2056 int process_stat_config_event(struct perf_session *session,
2057 			      union perf_event *event)
2058 {
2059 	struct perf_tool *tool = session->tool;
2060 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2061 
2062 	perf_event__read_stat_config(&stat_config, &event->stat_config);
2063 
2064 	if (perf_cpu_map__empty(st->cpus)) {
2065 		if (st->aggr_mode != AGGR_UNSET)
2066 			pr_warning("warning: processing task data, aggregation mode not set\n");
2067 		return 0;
2068 	}
2069 
2070 	if (st->aggr_mode != AGGR_UNSET)
2071 		stat_config.aggr_mode = st->aggr_mode;
2072 
2073 	if (perf_stat.data.is_pipe)
2074 		perf_stat_init_aggr_mode();
2075 	else
2076 		perf_stat_init_aggr_mode_file(st);
2077 
2078 	return 0;
2079 }
2080 
2081 static int set_maps(struct perf_stat *st)
2082 {
2083 	if (!st->cpus || !st->threads)
2084 		return 0;
2085 
2086 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2087 		return -EINVAL;
2088 
2089 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2090 
2091 	if (evlist__alloc_stats(evsel_list, true))
2092 		return -ENOMEM;
2093 
2094 	st->maps_allocated = true;
2095 	return 0;
2096 }
2097 
2098 static
2099 int process_thread_map_event(struct perf_session *session,
2100 			     union perf_event *event)
2101 {
2102 	struct perf_tool *tool = session->tool;
2103 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2104 
2105 	if (st->threads) {
2106 		pr_warning("Extra thread map event, ignoring.\n");
2107 		return 0;
2108 	}
2109 
2110 	st->threads = thread_map__new_event(&event->thread_map);
2111 	if (!st->threads)
2112 		return -ENOMEM;
2113 
2114 	return set_maps(st);
2115 }
2116 
2117 static
2118 int process_cpu_map_event(struct perf_session *session,
2119 			  union perf_event *event)
2120 {
2121 	struct perf_tool *tool = session->tool;
2122 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2123 	struct perf_cpu_map *cpus;
2124 
2125 	if (st->cpus) {
2126 		pr_warning("Extra cpu map event, ignoring.\n");
2127 		return 0;
2128 	}
2129 
2130 	cpus = cpu_map__new_data(&event->cpu_map.data);
2131 	if (!cpus)
2132 		return -ENOMEM;
2133 
2134 	st->cpus = cpus;
2135 	return set_maps(st);
2136 }
2137 
2138 static const char * const stat_report_usage[] = {
2139 	"perf stat report [<options>]",
2140 	NULL,
2141 };
2142 
2143 static struct perf_stat perf_stat = {
2144 	.tool = {
2145 		.attr		= perf_event__process_attr,
2146 		.event_update	= perf_event__process_event_update,
2147 		.thread_map	= process_thread_map_event,
2148 		.cpu_map	= process_cpu_map_event,
2149 		.stat_config	= process_stat_config_event,
2150 		.stat		= perf_event__process_stat_event,
2151 		.stat_round	= process_stat_round_event,
2152 	},
2153 	.aggr_mode = AGGR_UNSET,
2154 };
2155 
2156 static int __cmd_report(int argc, const char **argv)
2157 {
2158 	struct perf_session *session;
2159 	const struct option options[] = {
2160 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2161 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2162 		     "aggregate counts per processor socket", AGGR_SOCKET),
2163 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2164 		     "aggregate counts per processor die", AGGR_DIE),
2165 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2166 		     "aggregate counts per physical processor core", AGGR_CORE),
2167 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2168 		     "aggregate counts per numa node", AGGR_NODE),
2169 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2170 		     "disable CPU count aggregation", AGGR_NONE),
2171 	OPT_END()
2172 	};
2173 	struct stat st;
2174 	int ret;
2175 
2176 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2177 
2178 	if (!input_name || !strlen(input_name)) {
2179 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2180 			input_name = "-";
2181 		else
2182 			input_name = "perf.data";
2183 	}
2184 
2185 	perf_stat.data.path = input_name;
2186 	perf_stat.data.mode = PERF_DATA_MODE_READ;
2187 
2188 	session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2189 	if (IS_ERR(session))
2190 		return PTR_ERR(session);
2191 
2192 	perf_stat.session  = session;
2193 	stat_config.output = stderr;
2194 	evsel_list         = session->evlist;
2195 
2196 	ret = perf_session__process_events(session);
2197 	if (ret)
2198 		return ret;
2199 
2200 	perf_session__delete(session);
2201 	return 0;
2202 }
2203 
2204 static void setup_system_wide(int forks)
2205 {
2206 	/*
2207 	 * Make system wide (-a) the default target if
2208 	 * no target was specified and one of following
2209 	 * conditions is met:
2210 	 *
2211 	 *   - there's no workload specified
2212 	 *   - there is workload specified but all requested
2213 	 *     events are system wide events
2214 	 */
2215 	if (!target__none(&target))
2216 		return;
2217 
2218 	if (!forks)
2219 		target.system_wide = true;
2220 	else {
2221 		struct evsel *counter;
2222 
2223 		evlist__for_each_entry(evsel_list, counter) {
2224 			if (!counter->core.requires_cpu &&
2225 			    strcmp(counter->name, "duration_time")) {
2226 				return;
2227 			}
2228 		}
2229 
2230 		if (evsel_list->core.nr_entries)
2231 			target.system_wide = true;
2232 	}
2233 }
2234 
2235 int cmd_stat(int argc, const char **argv)
2236 {
2237 	const char * const stat_usage[] = {
2238 		"perf stat [<options>] [<command>]",
2239 		NULL
2240 	};
2241 	int status = -EINVAL, run_idx, err;
2242 	const char *mode;
2243 	FILE *output = stderr;
2244 	unsigned int interval, timeout;
2245 	const char * const stat_subcommands[] = { "record", "report" };
2246 	char errbuf[BUFSIZ];
2247 
2248 	setlocale(LC_ALL, "");
2249 
2250 	evsel_list = evlist__new();
2251 	if (evsel_list == NULL)
2252 		return -ENOMEM;
2253 
2254 	parse_events__shrink_config_terms();
2255 
2256 	/* String-parsing callback-based options would segfault when negated */
2257 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2258 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2259 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2260 
2261 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2262 					(const char **) stat_usage,
2263 					PARSE_OPT_STOP_AT_NON_OPTION);
2264 	perf_stat__collect_metric_expr(evsel_list);
2265 	perf_stat__init_shadow_stats();
2266 
2267 	if (stat_config.csv_sep) {
2268 		stat_config.csv_output = true;
2269 		if (!strcmp(stat_config.csv_sep, "\\t"))
2270 			stat_config.csv_sep = "\t";
2271 	} else
2272 		stat_config.csv_sep = DEFAULT_SEPARATOR;
2273 
2274 	if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2275 		argc = __cmd_record(argc, argv);
2276 		if (argc < 0)
2277 			return -1;
2278 	} else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2279 		return __cmd_report(argc, argv);
2280 
2281 	interval = stat_config.interval;
2282 	timeout = stat_config.timeout;
2283 
2284 	/*
2285 	 * For record command the -o is already taken care of.
2286 	 */
2287 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2288 		output = NULL;
2289 
2290 	if (output_name && output_fd) {
2291 		fprintf(stderr, "cannot use both --output and --log-fd\n");
2292 		parse_options_usage(stat_usage, stat_options, "o", 1);
2293 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2294 		goto out;
2295 	}
2296 
2297 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2298 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2299 		goto out;
2300 	}
2301 
2302 	if (stat_config.metric_only && stat_config.run_count > 1) {
2303 		fprintf(stderr, "--metric-only is not supported with -r\n");
2304 		goto out;
2305 	}
2306 
2307 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2308 		fprintf(stderr, "--table is only supported with -r\n");
2309 		parse_options_usage(stat_usage, stat_options, "r", 1);
2310 		parse_options_usage(NULL, stat_options, "table", 0);
2311 		goto out;
2312 	}
2313 
2314 	if (output_fd < 0) {
2315 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2316 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2317 		goto out;
2318 	}
2319 
2320 	if (!output && !stat_config.quiet) {
2321 		struct timespec tm;
2322 		mode = append_file ? "a" : "w";
2323 
2324 		output = fopen(output_name, mode);
2325 		if (!output) {
2326 			perror("failed to create output file");
2327 			return -1;
2328 		}
2329 		clock_gettime(CLOCK_REALTIME, &tm);
2330 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2331 	} else if (output_fd > 0) {
2332 		mode = append_file ? "a" : "w";
2333 		output = fdopen(output_fd, mode);
2334 		if (!output) {
2335 			perror("Failed opening logfd");
2336 			return -errno;
2337 		}
2338 	}
2339 
2340 	stat_config.output = output;
2341 
2342 	/*
2343 	 * let the spreadsheet do the pretty-printing
2344 	 */
2345 	if (stat_config.csv_output) {
2346 		/* User explicitly passed -B? */
2347 		if (big_num_opt == 1) {
2348 			fprintf(stderr, "-B option not supported with -x\n");
2349 			parse_options_usage(stat_usage, stat_options, "B", 1);
2350 			parse_options_usage(NULL, stat_options, "x", 1);
2351 			goto out;
2352 		} else /* Nope, so disable big number formatting */
2353 			stat_config.big_num = false;
2354 	} else if (big_num_opt == 0) /* User passed --no-big-num */
2355 		stat_config.big_num = false;
2356 
2357 	err = target__validate(&target);
2358 	if (err) {
2359 		target__strerror(&target, err, errbuf, BUFSIZ);
2360 		pr_warning("%s\n", errbuf);
2361 	}
2362 
2363 	setup_system_wide(argc);
2364 
2365 	/*
2366 	 * Display user/system times only for single
2367 	 * run and when there's specified tracee.
2368 	 */
2369 	if ((stat_config.run_count == 1) && target__none(&target))
2370 		stat_config.ru_display = true;
2371 
2372 	if (stat_config.run_count < 0) {
2373 		pr_err("Run count must be a positive number\n");
2374 		parse_options_usage(stat_usage, stat_options, "r", 1);
2375 		goto out;
2376 	} else if (stat_config.run_count == 0) {
2377 		forever = true;
2378 		stat_config.run_count = 1;
2379 	}
2380 
2381 	if (stat_config.walltime_run_table) {
2382 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2383 		if (!stat_config.walltime_run) {
2384 			pr_err("failed to setup -r option");
2385 			goto out;
2386 		}
2387 	}
2388 
2389 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2390 		!target__has_task(&target)) {
2391 		if (!target.system_wide || target.cpu_list) {
2392 			fprintf(stderr, "The --per-thread option is only "
2393 				"available when monitoring via -p -t -a "
2394 				"options or only --per-thread.\n");
2395 			parse_options_usage(NULL, stat_options, "p", 1);
2396 			parse_options_usage(NULL, stat_options, "t", 1);
2397 			goto out;
2398 		}
2399 	}
2400 
2401 	/*
2402 	 * no_aggr, cgroup are for system-wide only
2403 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2404 	 */
2405 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2406 	      stat_config.aggr_mode != AGGR_THREAD) ||
2407 	     (nr_cgroups || stat_config.cgroup_list)) &&
2408 	    !target__has_cpu(&target)) {
2409 		fprintf(stderr, "both cgroup and no-aggregation "
2410 			"modes only available in system-wide mode\n");
2411 
2412 		parse_options_usage(stat_usage, stat_options, "G", 1);
2413 		parse_options_usage(NULL, stat_options, "A", 1);
2414 		parse_options_usage(NULL, stat_options, "a", 1);
2415 		parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2416 		goto out;
2417 	}
2418 
2419 	if (stat_config.iostat_run) {
2420 		status = iostat_prepare(evsel_list, &stat_config);
2421 		if (status)
2422 			goto out;
2423 		if (iostat_mode == IOSTAT_LIST) {
2424 			iostat_list(evsel_list, &stat_config);
2425 			goto out;
2426 		} else if (verbose)
2427 			iostat_list(evsel_list, &stat_config);
2428 		if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2429 			target.system_wide = true;
2430 	}
2431 
2432 	if (add_default_attributes())
2433 		goto out;
2434 
2435 	if (stat_config.cgroup_list) {
2436 		if (nr_cgroups > 0) {
2437 			pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2438 			parse_options_usage(stat_usage, stat_options, "G", 1);
2439 			parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2440 			goto out;
2441 		}
2442 
2443 		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2444 					  &stat_config.metric_events, true) < 0) {
2445 			parse_options_usage(stat_usage, stat_options,
2446 					    "for-each-cgroup", 0);
2447 			goto out;
2448 		}
2449 	}
2450 
2451 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2452 		target.per_thread = true;
2453 
2454 	if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2455 		pr_err("failed to use cpu list %s\n", target.cpu_list);
2456 		goto out;
2457 	}
2458 
2459 	target.hybrid = perf_pmu__has_hybrid();
2460 	if (evlist__create_maps(evsel_list, &target) < 0) {
2461 		if (target__has_task(&target)) {
2462 			pr_err("Problems finding threads of monitor\n");
2463 			parse_options_usage(stat_usage, stat_options, "p", 1);
2464 			parse_options_usage(NULL, stat_options, "t", 1);
2465 		} else if (target__has_cpu(&target)) {
2466 			perror("failed to parse CPUs map");
2467 			parse_options_usage(stat_usage, stat_options, "C", 1);
2468 			parse_options_usage(NULL, stat_options, "a", 1);
2469 		}
2470 		goto out;
2471 	}
2472 
2473 	evlist__check_cpu_maps(evsel_list);
2474 
2475 	/*
2476 	 * Initialize thread_map with comm names,
2477 	 * so we could print it out on output.
2478 	 */
2479 	if (stat_config.aggr_mode == AGGR_THREAD) {
2480 		thread_map__read_comms(evsel_list->core.threads);
2481 		if (target.system_wide) {
2482 			if (runtime_stat_new(&stat_config,
2483 				perf_thread_map__nr(evsel_list->core.threads))) {
2484 				goto out;
2485 			}
2486 		}
2487 	}
2488 
2489 	if (stat_config.aggr_mode == AGGR_NODE)
2490 		cpu__setup_cpunode_map();
2491 
2492 	if (stat_config.times && interval)
2493 		interval_count = true;
2494 	else if (stat_config.times && !interval) {
2495 		pr_err("interval-count option should be used together with "
2496 				"interval-print.\n");
2497 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2498 		parse_options_usage(stat_usage, stat_options, "I", 1);
2499 		goto out;
2500 	}
2501 
2502 	if (timeout && timeout < 100) {
2503 		if (timeout < 10) {
2504 			pr_err("timeout must be >= 10ms.\n");
2505 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2506 			goto out;
2507 		} else
2508 			pr_warning("timeout < 100ms. "
2509 				   "The overhead percentage could be high in some cases. "
2510 				   "Please proceed with caution.\n");
2511 	}
2512 	if (timeout && interval) {
2513 		pr_err("timeout option is not supported with interval-print.\n");
2514 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2515 		parse_options_usage(stat_usage, stat_options, "I", 1);
2516 		goto out;
2517 	}
2518 
2519 	if (evlist__alloc_stats(evsel_list, interval))
2520 		goto out;
2521 
2522 	if (perf_stat_init_aggr_mode())
2523 		goto out;
2524 
2525 	/*
2526 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2527 	 * while avoiding that older tools show confusing messages.
2528 	 *
2529 	 * However for pipe sessions we need to keep it zero,
2530 	 * because script's perf_evsel__check_attr is triggered
2531 	 * by attr->sample_type != 0, and we can't run it on
2532 	 * stat sessions.
2533 	 */
2534 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2535 
2536 	/*
2537 	 * We dont want to block the signals - that would cause
2538 	 * child tasks to inherit that and Ctrl-C would not work.
2539 	 * What we want is for Ctrl-C to work in the exec()-ed
2540 	 * task, but being ignored by perf stat itself:
2541 	 */
2542 	atexit(sig_atexit);
2543 	if (!forever)
2544 		signal(SIGINT,  skip_signal);
2545 	signal(SIGCHLD, skip_signal);
2546 	signal(SIGALRM, skip_signal);
2547 	signal(SIGABRT, skip_signal);
2548 
2549 	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2550 		goto out;
2551 
2552 	/* Enable ignoring missing threads when -p option is defined. */
2553 	evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2554 	status = 0;
2555 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2556 		if (stat_config.run_count != 1 && verbose > 0)
2557 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2558 				run_idx + 1);
2559 
2560 		if (run_idx != 0)
2561 			evlist__reset_prev_raw_counts(evsel_list);
2562 
2563 		status = run_perf_stat(argc, argv, run_idx);
2564 		if (forever && status != -1 && !interval) {
2565 			print_counters(NULL, argc, argv);
2566 			perf_stat__reset_stats();
2567 		}
2568 	}
2569 
2570 	if (!forever && status != -1 && (!interval || stat_config.summary))
2571 		print_counters(NULL, argc, argv);
2572 
2573 	evlist__finalize_ctlfd(evsel_list);
2574 
2575 	if (STAT_RECORD) {
2576 		/*
2577 		 * We synthesize the kernel mmap record just so that older tools
2578 		 * don't emit warnings about not being able to resolve symbols
2579 		 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2580 		 * a saner message about no samples being in the perf.data file.
2581 		 *
2582 		 * This also serves to suppress a warning about f_header.data.size == 0
2583 		 * in header.c at the moment 'perf stat record' gets introduced, which
2584 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2585 		 * records, but the need to suppress the kptr_restrict messages in older
2586 		 * tools remain  -acme
2587 		 */
2588 		int fd = perf_data__fd(&perf_stat.data);
2589 
2590 		err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2591 							 process_synthesized_event,
2592 							 &perf_stat.session->machines.host);
2593 		if (err) {
2594 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2595 				   "older tools may produce warnings about this file\n.");
2596 		}
2597 
2598 		if (!interval) {
2599 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2600 				pr_err("failed to write stat round event\n");
2601 		}
2602 
2603 		if (!perf_stat.data.is_pipe) {
2604 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2605 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2606 		}
2607 
2608 		evlist__close(evsel_list);
2609 		perf_session__delete(perf_stat.session);
2610 	}
2611 
2612 	perf_stat__exit_aggr_mode();
2613 	evlist__free_stats(evsel_list);
2614 out:
2615 	if (stat_config.iostat_run)
2616 		iostat_release(evsel_list);
2617 
2618 	zfree(&stat_config.walltime_run);
2619 
2620 	if (smi_cost && smi_reset)
2621 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2622 
2623 	evlist__delete(evsel_list);
2624 
2625 	metricgroup__rblist_exit(&stat_config.metric_events);
2626 	runtime_stat_delete(&stat_config);
2627 	evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2628 
2629 	return status;
2630 }
2631