xref: /openbmc/linux/tools/perf/builtin-stat.c (revision dff03381)
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 "util/topdown.h"
75 #include "asm/bug.h"
76 
77 #include <linux/time64.h>
78 #include <linux/zalloc.h>
79 #include <api/fs/fs.h>
80 #include <errno.h>
81 #include <signal.h>
82 #include <stdlib.h>
83 #include <sys/prctl.h>
84 #include <inttypes.h>
85 #include <locale.h>
86 #include <math.h>
87 #include <sys/types.h>
88 #include <sys/stat.h>
89 #include <sys/wait.h>
90 #include <unistd.h>
91 #include <sys/time.h>
92 #include <sys/resource.h>
93 #include <linux/err.h>
94 
95 #include <linux/ctype.h>
96 #include <perf/evlist.h>
97 
98 #define DEFAULT_SEPARATOR	" "
99 #define FREEZE_ON_SMI_PATH	"devices/cpu/freeze_on_smi"
100 
101 static void print_counters(struct timespec *ts, int argc, const char **argv);
102 
103 /* Default events used for perf stat -T */
104 static const char *transaction_attrs = {
105 	"task-clock,"
106 	"{"
107 	"instructions,"
108 	"cycles,"
109 	"cpu/cycles-t/,"
110 	"cpu/tx-start/,"
111 	"cpu/el-start/,"
112 	"cpu/cycles-ct/"
113 	"}"
114 };
115 
116 /* More limited version when the CPU does not have all events. */
117 static const char * transaction_limited_attrs = {
118 	"task-clock,"
119 	"{"
120 	"instructions,"
121 	"cycles,"
122 	"cpu/cycles-t/,"
123 	"cpu/tx-start/"
124 	"}"
125 };
126 
127 static const char * topdown_attrs[] = {
128 	"topdown-total-slots",
129 	"topdown-slots-retired",
130 	"topdown-recovery-bubbles",
131 	"topdown-fetch-bubbles",
132 	"topdown-slots-issued",
133 	NULL,
134 };
135 
136 static const char *topdown_metric_attrs[] = {
137 	"slots",
138 	"topdown-retiring",
139 	"topdown-bad-spec",
140 	"topdown-fe-bound",
141 	"topdown-be-bound",
142 	NULL,
143 };
144 
145 static const char *topdown_metric_L2_attrs[] = {
146 	"slots",
147 	"topdown-retiring",
148 	"topdown-bad-spec",
149 	"topdown-fe-bound",
150 	"topdown-be-bound",
151 	"topdown-heavy-ops",
152 	"topdown-br-mispredict",
153 	"topdown-fetch-lat",
154 	"topdown-mem-bound",
155 	NULL,
156 };
157 
158 #define TOPDOWN_MAX_LEVEL			2
159 
160 static const char *smi_cost_attrs = {
161 	"{"
162 	"msr/aperf/,"
163 	"msr/smi/,"
164 	"cycles"
165 	"}"
166 };
167 
168 static struct evlist	*evsel_list;
169 static bool all_counters_use_bpf = true;
170 
171 static struct target target = {
172 	.uid	= UINT_MAX,
173 };
174 
175 #define METRIC_ONLY_LEN 20
176 
177 static volatile pid_t		child_pid			= -1;
178 static int			detailed_run			=  0;
179 static bool			transaction_run;
180 static bool			topdown_run			= false;
181 static bool			smi_cost			= false;
182 static bool			smi_reset			= false;
183 static int			big_num_opt			=  -1;
184 static bool			group				= false;
185 static const char		*pre_cmd			= NULL;
186 static const char		*post_cmd			= NULL;
187 static bool			sync_run			= false;
188 static bool			forever				= false;
189 static bool			force_metric_only		= false;
190 static struct timespec		ref_time;
191 static bool			append_file;
192 static bool			interval_count;
193 static const char		*output_name;
194 static int			output_fd;
195 
196 struct perf_stat {
197 	bool			 record;
198 	struct perf_data	 data;
199 	struct perf_session	*session;
200 	u64			 bytes_written;
201 	struct perf_tool	 tool;
202 	bool			 maps_allocated;
203 	struct perf_cpu_map	*cpus;
204 	struct perf_thread_map *threads;
205 	enum aggr_mode		 aggr_mode;
206 };
207 
208 static struct perf_stat		perf_stat;
209 #define STAT_RECORD		perf_stat.record
210 
211 static volatile int done = 0;
212 
213 static struct perf_stat_config stat_config = {
214 	.aggr_mode		= AGGR_GLOBAL,
215 	.scale			= true,
216 	.unit_width		= 4, /* strlen("unit") */
217 	.run_count		= 1,
218 	.metric_only_len	= METRIC_ONLY_LEN,
219 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
220 	.ru_stats		= &ru_stats,
221 	.big_num		= true,
222 	.ctl_fd			= -1,
223 	.ctl_fd_ack		= -1,
224 	.iostat_run		= false,
225 };
226 
227 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
228 {
229 	if (!a->core.cpus && !b->core.cpus)
230 		return true;
231 
232 	if (!a->core.cpus || !b->core.cpus)
233 		return false;
234 
235 	if (perf_cpu_map__nr(a->core.cpus) != perf_cpu_map__nr(b->core.cpus))
236 		return false;
237 
238 	for (int i = 0; i < perf_cpu_map__nr(a->core.cpus); i++) {
239 		if (perf_cpu_map__cpu(a->core.cpus, i).cpu !=
240 		    perf_cpu_map__cpu(b->core.cpus, i).cpu)
241 			return false;
242 	}
243 
244 	return true;
245 }
246 
247 static void evlist__check_cpu_maps(struct evlist *evlist)
248 {
249 	struct evsel *evsel, *pos, *leader;
250 	char buf[1024];
251 
252 	if (evlist__has_hybrid(evlist))
253 		evlist__warn_hybrid_group(evlist);
254 
255 	evlist__for_each_entry(evlist, evsel) {
256 		leader = evsel__leader(evsel);
257 
258 		/* Check that leader matches cpus with each member. */
259 		if (leader == evsel)
260 			continue;
261 		if (cpus_map_matched(leader, evsel))
262 			continue;
263 
264 		/* If there's mismatch disable the group and warn user. */
265 		WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
266 		evsel__group_desc(leader, buf, sizeof(buf));
267 		pr_warning("  %s\n", buf);
268 
269 		if (verbose) {
270 			cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
271 			pr_warning("     %s: %s\n", leader->name, buf);
272 			cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
273 			pr_warning("     %s: %s\n", evsel->name, buf);
274 		}
275 
276 		for_each_group_evsel(pos, leader)
277 			evsel__remove_from_group(pos, leader);
278 	}
279 }
280 
281 static inline void diff_timespec(struct timespec *r, struct timespec *a,
282 				 struct timespec *b)
283 {
284 	r->tv_sec = a->tv_sec - b->tv_sec;
285 	if (a->tv_nsec < b->tv_nsec) {
286 		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
287 		r->tv_sec--;
288 	} else {
289 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
290 	}
291 }
292 
293 static void perf_stat__reset_stats(void)
294 {
295 	int i;
296 
297 	evlist__reset_stats(evsel_list);
298 	perf_stat__reset_shadow_stats();
299 
300 	for (i = 0; i < stat_config.stats_num; i++)
301 		perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
302 }
303 
304 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
305 				     union perf_event *event,
306 				     struct perf_sample *sample __maybe_unused,
307 				     struct machine *machine __maybe_unused)
308 {
309 	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
310 		pr_err("failed to write perf data, error: %m\n");
311 		return -1;
312 	}
313 
314 	perf_stat.bytes_written += event->header.size;
315 	return 0;
316 }
317 
318 static int write_stat_round_event(u64 tm, u64 type)
319 {
320 	return perf_event__synthesize_stat_round(NULL, tm, type,
321 						 process_synthesized_event,
322 						 NULL);
323 }
324 
325 #define WRITE_STAT_ROUND_EVENT(time, interval) \
326 	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
327 
328 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
329 
330 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
331 				   struct perf_counts_values *count)
332 {
333 	struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
334 	struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
335 
336 	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
337 					   process_synthesized_event, NULL);
338 }
339 
340 static int read_single_counter(struct evsel *counter, int cpu_map_idx,
341 			       int thread, struct timespec *rs)
342 {
343 	switch(counter->tool_event) {
344 		case PERF_TOOL_DURATION_TIME: {
345 			u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
346 			struct perf_counts_values *count =
347 				perf_counts(counter->counts, cpu_map_idx, thread);
348 			count->ena = count->run = val;
349 			count->val = val;
350 			return 0;
351 		}
352 		case PERF_TOOL_USER_TIME:
353 		case PERF_TOOL_SYSTEM_TIME: {
354 			u64 val;
355 			struct perf_counts_values *count =
356 				perf_counts(counter->counts, cpu_map_idx, thread);
357 			if (counter->tool_event == PERF_TOOL_USER_TIME)
358 				val = ru_stats.ru_utime_usec_stat.mean;
359 			else
360 				val = ru_stats.ru_stime_usec_stat.mean;
361 			count->ena = count->run = val;
362 			count->val = val;
363 			return 0;
364 		}
365 		default:
366 		case PERF_TOOL_NONE:
367 			return evsel__read_counter(counter, cpu_map_idx, thread);
368 		case PERF_TOOL_MAX:
369 			/* This should never be reached */
370 			return 0;
371 	}
372 }
373 
374 /*
375  * Read out the results of a single counter:
376  * do not aggregate counts across CPUs in system-wide mode
377  */
378 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx)
379 {
380 	int nthreads = perf_thread_map__nr(evsel_list->core.threads);
381 	int thread;
382 
383 	if (!counter->supported)
384 		return -ENOENT;
385 
386 	for (thread = 0; thread < nthreads; thread++) {
387 		struct perf_counts_values *count;
388 
389 		count = perf_counts(counter->counts, cpu_map_idx, thread);
390 
391 		/*
392 		 * The leader's group read loads data into its group members
393 		 * (via evsel__read_counter()) and sets their count->loaded.
394 		 */
395 		if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
396 		    read_single_counter(counter, cpu_map_idx, thread, rs)) {
397 			counter->counts->scaled = -1;
398 			perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
399 			perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
400 			return -1;
401 		}
402 
403 		perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
404 
405 		if (STAT_RECORD) {
406 			if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
407 				pr_err("failed to write stat event\n");
408 				return -1;
409 			}
410 		}
411 
412 		if (verbose > 1) {
413 			fprintf(stat_config.output,
414 				"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
415 					evsel__name(counter),
416 					perf_cpu_map__cpu(evsel__cpus(counter),
417 							  cpu_map_idx).cpu,
418 					count->val, count->ena, count->run);
419 		}
420 	}
421 
422 	return 0;
423 }
424 
425 static int read_affinity_counters(struct timespec *rs)
426 {
427 	struct evlist_cpu_iterator evlist_cpu_itr;
428 	struct affinity saved_affinity, *affinity;
429 
430 	if (all_counters_use_bpf)
431 		return 0;
432 
433 	if (!target__has_cpu(&target) || target__has_per_thread(&target))
434 		affinity = NULL;
435 	else if (affinity__setup(&saved_affinity) < 0)
436 		return -1;
437 	else
438 		affinity = &saved_affinity;
439 
440 	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
441 		struct evsel *counter = evlist_cpu_itr.evsel;
442 
443 		if (evsel__is_bpf(counter))
444 			continue;
445 
446 		if (!counter->err) {
447 			counter->err = read_counter_cpu(counter, rs,
448 							evlist_cpu_itr.cpu_map_idx);
449 		}
450 	}
451 	if (affinity)
452 		affinity__cleanup(&saved_affinity);
453 
454 	return 0;
455 }
456 
457 static int read_bpf_map_counters(void)
458 {
459 	struct evsel *counter;
460 	int err;
461 
462 	evlist__for_each_entry(evsel_list, counter) {
463 		if (!evsel__is_bpf(counter))
464 			continue;
465 
466 		err = bpf_counter__read(counter);
467 		if (err)
468 			return err;
469 	}
470 	return 0;
471 }
472 
473 static void read_counters(struct timespec *rs)
474 {
475 	struct evsel *counter;
476 
477 	if (!stat_config.stop_read_counter) {
478 		if (read_bpf_map_counters() ||
479 		    read_affinity_counters(rs))
480 			return;
481 	}
482 
483 	evlist__for_each_entry(evsel_list, counter) {
484 		if (counter->err)
485 			pr_debug("failed to read counter %s\n", counter->name);
486 		if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
487 			pr_warning("failed to process counter %s\n", counter->name);
488 		counter->err = 0;
489 	}
490 }
491 
492 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
493 {
494 	int i;
495 
496 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
497 	if (!config->stats)
498 		return -1;
499 
500 	config->stats_num = nthreads;
501 
502 	for (i = 0; i < nthreads; i++)
503 		runtime_stat__init(&config->stats[i]);
504 
505 	return 0;
506 }
507 
508 static void runtime_stat_delete(struct perf_stat_config *config)
509 {
510 	int i;
511 
512 	if (!config->stats)
513 		return;
514 
515 	for (i = 0; i < config->stats_num; i++)
516 		runtime_stat__exit(&config->stats[i]);
517 
518 	zfree(&config->stats);
519 }
520 
521 static void runtime_stat_reset(struct perf_stat_config *config)
522 {
523 	int i;
524 
525 	if (!config->stats)
526 		return;
527 
528 	for (i = 0; i < config->stats_num; i++)
529 		perf_stat__reset_shadow_per_stat(&config->stats[i]);
530 }
531 
532 static void process_interval(void)
533 {
534 	struct timespec ts, rs;
535 
536 	clock_gettime(CLOCK_MONOTONIC, &ts);
537 	diff_timespec(&rs, &ts, &ref_time);
538 
539 	perf_stat__reset_shadow_per_stat(&rt_stat);
540 	runtime_stat_reset(&stat_config);
541 	read_counters(&rs);
542 
543 	if (STAT_RECORD) {
544 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
545 			pr_err("failed to write stat round event\n");
546 	}
547 
548 	init_stats(&walltime_nsecs_stats);
549 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
550 	print_counters(&rs, 0, NULL);
551 }
552 
553 static bool handle_interval(unsigned int interval, int *times)
554 {
555 	if (interval) {
556 		process_interval();
557 		if (interval_count && !(--(*times)))
558 			return true;
559 	}
560 	return false;
561 }
562 
563 static int enable_counters(void)
564 {
565 	struct evsel *evsel;
566 	int err;
567 
568 	evlist__for_each_entry(evsel_list, evsel) {
569 		if (!evsel__is_bpf(evsel))
570 			continue;
571 
572 		err = bpf_counter__enable(evsel);
573 		if (err)
574 			return err;
575 	}
576 
577 	if (stat_config.initial_delay < 0) {
578 		pr_info(EVLIST_DISABLED_MSG);
579 		return 0;
580 	}
581 
582 	if (stat_config.initial_delay > 0) {
583 		pr_info(EVLIST_DISABLED_MSG);
584 		usleep(stat_config.initial_delay * USEC_PER_MSEC);
585 	}
586 
587 	/*
588 	 * We need to enable counters only if:
589 	 * - we don't have tracee (attaching to task or cpu)
590 	 * - we have initial delay configured
591 	 */
592 	if (!target__none(&target) || stat_config.initial_delay) {
593 		if (!all_counters_use_bpf)
594 			evlist__enable(evsel_list);
595 		if (stat_config.initial_delay > 0)
596 			pr_info(EVLIST_ENABLED_MSG);
597 	}
598 	return 0;
599 }
600 
601 static void disable_counters(void)
602 {
603 	struct evsel *counter;
604 
605 	/*
606 	 * If we don't have tracee (attaching to task or cpu), counters may
607 	 * still be running. To get accurate group ratios, we must stop groups
608 	 * from counting before reading their constituent counters.
609 	 */
610 	if (!target__none(&target)) {
611 		evlist__for_each_entry(evsel_list, counter)
612 			bpf_counter__disable(counter);
613 		if (!all_counters_use_bpf)
614 			evlist__disable(evsel_list);
615 	}
616 }
617 
618 static volatile int workload_exec_errno;
619 
620 /*
621  * evlist__prepare_workload will send a SIGUSR1
622  * if the fork fails, since we asked by setting its
623  * want_signal to true.
624  */
625 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
626 					void *ucontext __maybe_unused)
627 {
628 	workload_exec_errno = info->si_value.sival_int;
629 }
630 
631 static bool evsel__should_store_id(struct evsel *counter)
632 {
633 	return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
634 }
635 
636 static bool is_target_alive(struct target *_target,
637 			    struct perf_thread_map *threads)
638 {
639 	struct stat st;
640 	int i;
641 
642 	if (!target__has_task(_target))
643 		return true;
644 
645 	for (i = 0; i < threads->nr; i++) {
646 		char path[PATH_MAX];
647 
648 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
649 			  threads->map[i].pid);
650 
651 		if (!stat(path, &st))
652 			return true;
653 	}
654 
655 	return false;
656 }
657 
658 static void process_evlist(struct evlist *evlist, unsigned int interval)
659 {
660 	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
661 
662 	if (evlist__ctlfd_process(evlist, &cmd) > 0) {
663 		switch (cmd) {
664 		case EVLIST_CTL_CMD_ENABLE:
665 			if (interval)
666 				process_interval();
667 			break;
668 		case EVLIST_CTL_CMD_DISABLE:
669 			if (interval)
670 				process_interval();
671 			break;
672 		case EVLIST_CTL_CMD_SNAPSHOT:
673 		case EVLIST_CTL_CMD_ACK:
674 		case EVLIST_CTL_CMD_UNSUPPORTED:
675 		case EVLIST_CTL_CMD_EVLIST:
676 		case EVLIST_CTL_CMD_STOP:
677 		case EVLIST_CTL_CMD_PING:
678 		default:
679 			break;
680 		}
681 	}
682 }
683 
684 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
685 			int *time_to_sleep)
686 {
687 	int tts = *time_to_sleep;
688 	struct timespec time_diff;
689 
690 	diff_timespec(&time_diff, time_stop, time_start);
691 
692 	tts -= time_diff.tv_sec * MSEC_PER_SEC +
693 	       time_diff.tv_nsec / NSEC_PER_MSEC;
694 
695 	if (tts < 0)
696 		tts = 0;
697 
698 	*time_to_sleep = tts;
699 }
700 
701 static int dispatch_events(bool forks, int timeout, int interval, int *times)
702 {
703 	int child_exited = 0, status = 0;
704 	int time_to_sleep, sleep_time;
705 	struct timespec time_start, time_stop;
706 
707 	if (interval)
708 		sleep_time = interval;
709 	else if (timeout)
710 		sleep_time = timeout;
711 	else
712 		sleep_time = 1000;
713 
714 	time_to_sleep = sleep_time;
715 
716 	while (!done) {
717 		if (forks)
718 			child_exited = waitpid(child_pid, &status, WNOHANG);
719 		else
720 			child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
721 
722 		if (child_exited)
723 			break;
724 
725 		clock_gettime(CLOCK_MONOTONIC, &time_start);
726 		if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
727 			if (timeout || handle_interval(interval, times))
728 				break;
729 			time_to_sleep = sleep_time;
730 		} else { /* fd revent */
731 			process_evlist(evsel_list, interval);
732 			clock_gettime(CLOCK_MONOTONIC, &time_stop);
733 			compute_tts(&time_start, &time_stop, &time_to_sleep);
734 		}
735 	}
736 
737 	return status;
738 }
739 
740 enum counter_recovery {
741 	COUNTER_SKIP,
742 	COUNTER_RETRY,
743 	COUNTER_FATAL,
744 };
745 
746 static enum counter_recovery stat_handle_error(struct evsel *counter)
747 {
748 	char msg[BUFSIZ];
749 	/*
750 	 * PPC returns ENXIO for HW counters until 2.6.37
751 	 * (behavior changed with commit b0a873e).
752 	 */
753 	if (errno == EINVAL || errno == ENOSYS ||
754 	    errno == ENOENT || errno == EOPNOTSUPP ||
755 	    errno == ENXIO) {
756 		if (verbose > 0)
757 			ui__warning("%s event is not supported by the kernel.\n",
758 				    evsel__name(counter));
759 		counter->supported = false;
760 		/*
761 		 * errored is a sticky flag that means one of the counter's
762 		 * cpu event had a problem and needs to be reexamined.
763 		 */
764 		counter->errored = true;
765 
766 		if ((evsel__leader(counter) != counter) ||
767 		    !(counter->core.leader->nr_members > 1))
768 			return COUNTER_SKIP;
769 	} else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
770 		if (verbose > 0)
771 			ui__warning("%s\n", msg);
772 		return COUNTER_RETRY;
773 	} else if (target__has_per_thread(&target) &&
774 		   evsel_list->core.threads &&
775 		   evsel_list->core.threads->err_thread != -1) {
776 		/*
777 		 * For global --per-thread case, skip current
778 		 * error thread.
779 		 */
780 		if (!thread_map__remove(evsel_list->core.threads,
781 					evsel_list->core.threads->err_thread)) {
782 			evsel_list->core.threads->err_thread = -1;
783 			return COUNTER_RETRY;
784 		}
785 	}
786 
787 	evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
788 	ui__error("%s\n", msg);
789 
790 	if (child_pid != -1)
791 		kill(child_pid, SIGTERM);
792 	return COUNTER_FATAL;
793 }
794 
795 static int __run_perf_stat(int argc, const char **argv, int run_idx)
796 {
797 	int interval = stat_config.interval;
798 	int times = stat_config.times;
799 	int timeout = stat_config.timeout;
800 	char msg[BUFSIZ];
801 	unsigned long long t0, t1;
802 	struct evsel *counter;
803 	size_t l;
804 	int status = 0;
805 	const bool forks = (argc > 0);
806 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
807 	struct evlist_cpu_iterator evlist_cpu_itr;
808 	struct affinity saved_affinity, *affinity = NULL;
809 	int err;
810 	bool second_pass = false;
811 
812 	if (forks) {
813 		if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
814 			perror("failed to prepare workload");
815 			return -1;
816 		}
817 		child_pid = evsel_list->workload.pid;
818 	}
819 
820 	if (group)
821 		evlist__set_leader(evsel_list);
822 
823 	if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
824 		if (affinity__setup(&saved_affinity) < 0)
825 			return -1;
826 		affinity = &saved_affinity;
827 	}
828 
829 	evlist__for_each_entry(evsel_list, counter) {
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_CALLBACK('G', "cgroup", &evsel_list, "name",
1254 		     "monitor event in cgroup name only", parse_stat_cgroups),
1255 	OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1256 		    "expand events for each cgroup"),
1257 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1258 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1259 	OPT_INTEGER(0, "log-fd", &output_fd,
1260 		    "log output to fd, instead of stderr"),
1261 	OPT_STRING(0, "pre", &pre_cmd, "command",
1262 			"command to run prior to the measured command"),
1263 	OPT_STRING(0, "post", &post_cmd, "command",
1264 			"command to run after to the measured command"),
1265 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1266 		    "print counts at regular interval in ms "
1267 		    "(overhead is possible for values <= 100ms)"),
1268 	OPT_INTEGER(0, "interval-count", &stat_config.times,
1269 		    "print counts for fixed number of times"),
1270 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1271 		    "clear screen in between new interval"),
1272 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1273 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1274 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1275 		     "aggregate counts per processor socket", AGGR_SOCKET),
1276 	OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1277 		     "aggregate counts per processor die", AGGR_DIE),
1278 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1279 		     "aggregate counts per physical processor core", AGGR_CORE),
1280 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1281 		     "aggregate counts per thread", AGGR_THREAD),
1282 	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1283 		     "aggregate counts per numa node", AGGR_NODE),
1284 	OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1285 		    "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1286 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1287 			"Only print computed metrics. No raw values", enable_metric_only),
1288 	OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1289 		       "don't group metric events, impacts multiplexing"),
1290 	OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1291 		       "don't try to share events between metrics in a group"),
1292 	OPT_BOOLEAN(0, "topdown", &topdown_run,
1293 			"measure top-down statistics"),
1294 	OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1295 			"Set the metrics level for the top-down statistics (0: max level)"),
1296 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1297 			"measure SMI cost"),
1298 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1299 		     "monitor specified metrics or metric groups (separated by ,)",
1300 		     parse_metric_groups),
1301 	OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1302 			 "Configure all used events to run in kernel space.",
1303 			 PARSE_OPT_EXCLUSIVE),
1304 	OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1305 			 "Configure all used events to run in user space.",
1306 			 PARSE_OPT_EXCLUSIVE),
1307 	OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1308 		    "Use with 'percore' event qualifier to show the event "
1309 		    "counts of one hardware thread by sum up total hardware "
1310 		    "threads of same physical core"),
1311 	OPT_BOOLEAN(0, "summary", &stat_config.summary,
1312 		       "print summary for interval mode"),
1313 	OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1314 		       "don't print 'summary' for CSV summary output"),
1315 	OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1316 			"don't print output (useful with record)"),
1317 	OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1318 		     "Only enable events on applying cpu with this type "
1319 		     "for hybrid platform (e.g. core or atom)",
1320 		     parse_hybrid_type),
1321 #ifdef HAVE_LIBPFM
1322 	OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1323 		"libpfm4 event selector. use 'perf list' to list available events",
1324 		parse_libpfm_events_option),
1325 #endif
1326 	OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1327 		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1328 		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1329 		     "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1330 		      parse_control_option),
1331 	OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1332 			    "measure I/O performance metrics provided by arch/platform",
1333 			    iostat_parse),
1334 	OPT_END()
1335 };
1336 
1337 static const char *const aggr_mode__string[] = {
1338 	[AGGR_CORE] = "core",
1339 	[AGGR_DIE] = "die",
1340 	[AGGR_GLOBAL] = "global",
1341 	[AGGR_NODE] = "node",
1342 	[AGGR_NONE] = "none",
1343 	[AGGR_SOCKET] = "socket",
1344 	[AGGR_THREAD] = "thread",
1345 	[AGGR_UNSET] = "unset",
1346 };
1347 
1348 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1349 						struct perf_cpu cpu)
1350 {
1351 	return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1352 }
1353 
1354 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1355 					     struct perf_cpu cpu)
1356 {
1357 	return aggr_cpu_id__die(cpu, /*data=*/NULL);
1358 }
1359 
1360 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1361 					      struct perf_cpu cpu)
1362 {
1363 	return aggr_cpu_id__core(cpu, /*data=*/NULL);
1364 }
1365 
1366 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1367 					      struct perf_cpu cpu)
1368 {
1369 	return aggr_cpu_id__node(cpu, /*data=*/NULL);
1370 }
1371 
1372 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1373 					      aggr_get_id_t get_id, struct perf_cpu cpu)
1374 {
1375 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1376 
1377 	if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1378 		config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1379 
1380 	id = config->cpus_aggr_map->map[cpu.cpu];
1381 	return id;
1382 }
1383 
1384 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1385 						       struct perf_cpu cpu)
1386 {
1387 	return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1388 }
1389 
1390 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1391 						    struct perf_cpu cpu)
1392 {
1393 	return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1394 }
1395 
1396 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1397 						     struct perf_cpu cpu)
1398 {
1399 	return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1400 }
1401 
1402 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1403 						     struct perf_cpu cpu)
1404 {
1405 	return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1406 }
1407 
1408 static bool term_percore_set(void)
1409 {
1410 	struct evsel *counter;
1411 
1412 	evlist__for_each_entry(evsel_list, counter) {
1413 		if (counter->percore)
1414 			return true;
1415 	}
1416 
1417 	return false;
1418 }
1419 
1420 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1421 {
1422 	switch (aggr_mode) {
1423 	case AGGR_SOCKET:
1424 		return aggr_cpu_id__socket;
1425 	case AGGR_DIE:
1426 		return aggr_cpu_id__die;
1427 	case AGGR_CORE:
1428 		return aggr_cpu_id__core;
1429 	case AGGR_NODE:
1430 		return aggr_cpu_id__node;
1431 	case AGGR_NONE:
1432 		if (term_percore_set())
1433 			return aggr_cpu_id__core;
1434 
1435 		return NULL;
1436 	case AGGR_GLOBAL:
1437 	case AGGR_THREAD:
1438 	case AGGR_UNSET:
1439 	default:
1440 		return NULL;
1441 	}
1442 }
1443 
1444 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1445 {
1446 	switch (aggr_mode) {
1447 	case AGGR_SOCKET:
1448 		return perf_stat__get_socket_cached;
1449 	case AGGR_DIE:
1450 		return perf_stat__get_die_cached;
1451 	case AGGR_CORE:
1452 		return perf_stat__get_core_cached;
1453 	case AGGR_NODE:
1454 		return perf_stat__get_node_cached;
1455 	case AGGR_NONE:
1456 		if (term_percore_set()) {
1457 			return perf_stat__get_core_cached;
1458 		}
1459 		return NULL;
1460 	case AGGR_GLOBAL:
1461 	case AGGR_THREAD:
1462 	case AGGR_UNSET:
1463 	default:
1464 		return NULL;
1465 	}
1466 }
1467 
1468 static int perf_stat_init_aggr_mode(void)
1469 {
1470 	int nr;
1471 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1472 
1473 	if (get_id) {
1474 		stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1475 							 get_id, /*data=*/NULL);
1476 		if (!stat_config.aggr_map) {
1477 			pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1478 			return -1;
1479 		}
1480 		stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1481 	}
1482 
1483 	/*
1484 	 * The evsel_list->cpus is the base we operate on,
1485 	 * taking the highest cpu number to be the size of
1486 	 * the aggregation translate cpumap.
1487 	 */
1488 	if (evsel_list->core.user_requested_cpus)
1489 		nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1490 	else
1491 		nr = 0;
1492 	stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1493 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1494 }
1495 
1496 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1497 {
1498 	if (map) {
1499 		WARN_ONCE(refcount_read(&map->refcnt) != 0,
1500 			  "cpu_aggr_map refcnt unbalanced\n");
1501 		free(map);
1502 	}
1503 }
1504 
1505 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1506 {
1507 	if (map && refcount_dec_and_test(&map->refcnt))
1508 		cpu_aggr_map__delete(map);
1509 }
1510 
1511 static void perf_stat__exit_aggr_mode(void)
1512 {
1513 	cpu_aggr_map__put(stat_config.aggr_map);
1514 	cpu_aggr_map__put(stat_config.cpus_aggr_map);
1515 	stat_config.aggr_map = NULL;
1516 	stat_config.cpus_aggr_map = NULL;
1517 }
1518 
1519 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1520 {
1521 	struct perf_env *env = data;
1522 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1523 
1524 	if (cpu.cpu != -1)
1525 		id.socket = env->cpu[cpu.cpu].socket_id;
1526 
1527 	return id;
1528 }
1529 
1530 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1531 {
1532 	struct perf_env *env = data;
1533 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1534 
1535 	if (cpu.cpu != -1) {
1536 		/*
1537 		 * die_id is relative to socket, so start
1538 		 * with the socket ID and then add die to
1539 		 * make a unique ID.
1540 		 */
1541 		id.socket = env->cpu[cpu.cpu].socket_id;
1542 		id.die = env->cpu[cpu.cpu].die_id;
1543 	}
1544 
1545 	return id;
1546 }
1547 
1548 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1549 {
1550 	struct perf_env *env = data;
1551 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1552 
1553 	if (cpu.cpu != -1) {
1554 		/*
1555 		 * core_id is relative to socket and die,
1556 		 * we need a global id. So we set
1557 		 * socket, die id and core id
1558 		 */
1559 		id.socket = env->cpu[cpu.cpu].socket_id;
1560 		id.die = env->cpu[cpu.cpu].die_id;
1561 		id.core = env->cpu[cpu.cpu].core_id;
1562 	}
1563 
1564 	return id;
1565 }
1566 
1567 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1568 {
1569 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1570 
1571 	id.node = perf_env__numa_node(data, cpu);
1572 	return id;
1573 }
1574 
1575 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1576 						     struct perf_cpu cpu)
1577 {
1578 	return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1579 }
1580 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1581 						  struct perf_cpu cpu)
1582 {
1583 	return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1584 }
1585 
1586 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1587 						   struct perf_cpu cpu)
1588 {
1589 	return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1590 }
1591 
1592 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1593 						   struct perf_cpu cpu)
1594 {
1595 	return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1596 }
1597 
1598 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1599 {
1600 	switch (aggr_mode) {
1601 	case AGGR_SOCKET:
1602 		return perf_env__get_socket_aggr_by_cpu;
1603 	case AGGR_DIE:
1604 		return perf_env__get_die_aggr_by_cpu;
1605 	case AGGR_CORE:
1606 		return perf_env__get_core_aggr_by_cpu;
1607 	case AGGR_NODE:
1608 		return perf_env__get_node_aggr_by_cpu;
1609 	case AGGR_NONE:
1610 	case AGGR_GLOBAL:
1611 	case AGGR_THREAD:
1612 	case AGGR_UNSET:
1613 	default:
1614 		return NULL;
1615 	}
1616 }
1617 
1618 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1619 {
1620 	switch (aggr_mode) {
1621 	case AGGR_SOCKET:
1622 		return perf_stat__get_socket_file;
1623 	case AGGR_DIE:
1624 		return perf_stat__get_die_file;
1625 	case AGGR_CORE:
1626 		return perf_stat__get_core_file;
1627 	case AGGR_NODE:
1628 		return perf_stat__get_node_file;
1629 	case AGGR_NONE:
1630 	case AGGR_GLOBAL:
1631 	case AGGR_THREAD:
1632 	case AGGR_UNSET:
1633 	default:
1634 		return NULL;
1635 	}
1636 }
1637 
1638 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1639 {
1640 	struct perf_env *env = &st->session->header.env;
1641 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1642 
1643 	if (!get_id)
1644 		return 0;
1645 
1646 	stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, get_id, env);
1647 	if (!stat_config.aggr_map) {
1648 		pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1649 		return -1;
1650 	}
1651 	stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1652 	return 0;
1653 }
1654 
1655 /*
1656  * Add default attributes, if there were no attributes specified or
1657  * if -d/--detailed, -d -d or -d -d -d is used:
1658  */
1659 static int add_default_attributes(void)
1660 {
1661 	int err;
1662 	struct perf_event_attr default_attrs0[] = {
1663 
1664   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1665   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1666   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1667   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1668 
1669   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1670 };
1671 	struct perf_event_attr frontend_attrs[] = {
1672   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1673 };
1674 	struct perf_event_attr backend_attrs[] = {
1675   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1676 };
1677 	struct perf_event_attr default_attrs1[] = {
1678   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1679   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1680   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1681 
1682 };
1683 
1684 /*
1685  * Detailed stats (-d), covering the L1 and last level data caches:
1686  */
1687 	struct perf_event_attr detailed_attrs[] = {
1688 
1689   { .type = PERF_TYPE_HW_CACHE,
1690     .config =
1691 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1692 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1693 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
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_MISS	<< 16)				},
1700 
1701   { .type = PERF_TYPE_HW_CACHE,
1702     .config =
1703 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1704 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1705 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 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_MISS	<< 16)				},
1712 };
1713 
1714 /*
1715  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1716  */
1717 	struct perf_event_attr very_detailed_attrs[] = {
1718 
1719   { .type = PERF_TYPE_HW_CACHE,
1720     .config =
1721 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1722 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1723 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
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_MISS	<< 16)				},
1730 
1731   { .type = PERF_TYPE_HW_CACHE,
1732     .config =
1733 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1734 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1735 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 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_MISS	<< 16)				},
1742 
1743   { .type = PERF_TYPE_HW_CACHE,
1744     .config =
1745 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1746 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1747 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 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_MISS	<< 16)				},
1754 
1755 };
1756 
1757 /*
1758  * Very, very detailed stats (-d -d -d), adding prefetch events:
1759  */
1760 	struct perf_event_attr very_very_detailed_attrs[] = {
1761 
1762   { .type = PERF_TYPE_HW_CACHE,
1763     .config =
1764 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1765 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1766 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
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_MISS	<< 16)				},
1773 };
1774 
1775 	struct perf_event_attr default_null_attrs[] = {};
1776 
1777 	/* Set attrs if no event is selected and !null_run: */
1778 	if (stat_config.null_run)
1779 		return 0;
1780 
1781 	if (transaction_run) {
1782 		struct parse_events_error errinfo;
1783 		/* Handle -T as -M transaction. Once platform specific metrics
1784 		 * support has been added to the json files, all architectures
1785 		 * will use this approach. To determine transaction support
1786 		 * on an architecture test for such a metric name.
1787 		 */
1788 		if (metricgroup__has_metric("transaction")) {
1789 			struct option opt = { .value = &evsel_list };
1790 
1791 			return metricgroup__parse_groups(&opt, "transaction",
1792 							 stat_config.metric_no_group,
1793 							stat_config.metric_no_merge,
1794 							 &stat_config.metric_events);
1795 		}
1796 
1797 		parse_events_error__init(&errinfo);
1798 		if (pmu_have_event("cpu", "cycles-ct") &&
1799 		    pmu_have_event("cpu", "el-start"))
1800 			err = parse_events(evsel_list, transaction_attrs,
1801 					   &errinfo);
1802 		else
1803 			err = parse_events(evsel_list,
1804 					   transaction_limited_attrs,
1805 					   &errinfo);
1806 		if (err) {
1807 			fprintf(stderr, "Cannot set up transaction events\n");
1808 			parse_events_error__print(&errinfo, transaction_attrs);
1809 		}
1810 		parse_events_error__exit(&errinfo);
1811 		return err ? -1 : 0;
1812 	}
1813 
1814 	if (smi_cost) {
1815 		struct parse_events_error errinfo;
1816 		int smi;
1817 
1818 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1819 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1820 			return -1;
1821 		}
1822 
1823 		if (!smi) {
1824 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1825 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1826 				return -1;
1827 			}
1828 			smi_reset = true;
1829 		}
1830 
1831 		if (!pmu_have_event("msr", "aperf") ||
1832 		    !pmu_have_event("msr", "smi")) {
1833 			fprintf(stderr, "To measure SMI cost, it needs "
1834 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1835 			return -1;
1836 		}
1837 		if (!force_metric_only)
1838 			stat_config.metric_only = true;
1839 
1840 		parse_events_error__init(&errinfo);
1841 		err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1842 		if (err) {
1843 			parse_events_error__print(&errinfo, smi_cost_attrs);
1844 			fprintf(stderr, "Cannot set up SMI cost events\n");
1845 		}
1846 		parse_events_error__exit(&errinfo);
1847 		return err ? -1 : 0;
1848 	}
1849 
1850 	if (topdown_run) {
1851 		const char **metric_attrs = topdown_metric_attrs;
1852 		unsigned int max_level = 1;
1853 		char *str = NULL;
1854 		bool warn = false;
1855 		const char *pmu_name = arch_get_topdown_pmu_name(evsel_list, true);
1856 
1857 		if (!force_metric_only)
1858 			stat_config.metric_only = true;
1859 
1860 		if (pmu_have_event(pmu_name, topdown_metric_L2_attrs[5])) {
1861 			metric_attrs = topdown_metric_L2_attrs;
1862 			max_level = 2;
1863 		}
1864 
1865 		if (stat_config.topdown_level > max_level) {
1866 			pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1867 			return -1;
1868 		} else if (!stat_config.topdown_level)
1869 			stat_config.topdown_level = max_level;
1870 
1871 		if (topdown_filter_events(metric_attrs, &str, 1, pmu_name) < 0) {
1872 			pr_err("Out of memory\n");
1873 			return -1;
1874 		}
1875 
1876 		if (metric_attrs[0] && str) {
1877 			if (!stat_config.interval && !stat_config.metric_only) {
1878 				fprintf(stat_config.output,
1879 					"Topdown accuracy may decrease when measuring long periods.\n"
1880 					"Please print the result regularly, e.g. -I1000\n");
1881 			}
1882 			goto setup_metrics;
1883 		}
1884 
1885 		zfree(&str);
1886 
1887 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1888 		    stat_config.aggr_mode != AGGR_CORE) {
1889 			pr_err("top down event configuration requires --per-core mode\n");
1890 			return -1;
1891 		}
1892 		stat_config.aggr_mode = AGGR_CORE;
1893 		if (nr_cgroups || !target__has_cpu(&target)) {
1894 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1895 			return -1;
1896 		}
1897 
1898 		if (topdown_filter_events(topdown_attrs, &str,
1899 				arch_topdown_check_group(&warn),
1900 				pmu_name) < 0) {
1901 			pr_err("Out of memory\n");
1902 			return -1;
1903 		}
1904 
1905 		if (topdown_attrs[0] && str) {
1906 			struct parse_events_error errinfo;
1907 			if (warn)
1908 				arch_topdown_group_warn();
1909 setup_metrics:
1910 			parse_events_error__init(&errinfo);
1911 			err = parse_events(evsel_list, str, &errinfo);
1912 			if (err) {
1913 				fprintf(stderr,
1914 					"Cannot set up top down events %s: %d\n",
1915 					str, err);
1916 				parse_events_error__print(&errinfo, str);
1917 				parse_events_error__exit(&errinfo);
1918 				free(str);
1919 				return -1;
1920 			}
1921 			parse_events_error__exit(&errinfo);
1922 		} else {
1923 			fprintf(stderr, "System does not support topdown\n");
1924 			return -1;
1925 		}
1926 		free(str);
1927 	}
1928 
1929 	if (!evsel_list->core.nr_entries) {
1930 		if (target__has_cpu(&target))
1931 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1932 
1933 		if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1934 			return -1;
1935 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1936 			if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1937 				return -1;
1938 		}
1939 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1940 			if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1941 				return -1;
1942 		}
1943 		if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1944 			return -1;
1945 
1946 		stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1947 		/* Platform specific attrs */
1948 		if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0)
1949 			return -1;
1950 	}
1951 
1952 	/* Detailed events get appended to the event list: */
1953 
1954 	if (detailed_run <  1)
1955 		return 0;
1956 
1957 	/* Append detailed run extra attributes: */
1958 	if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1959 		return -1;
1960 
1961 	if (detailed_run < 2)
1962 		return 0;
1963 
1964 	/* Append very detailed run extra attributes: */
1965 	if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1966 		return -1;
1967 
1968 	if (detailed_run < 3)
1969 		return 0;
1970 
1971 	/* Append very, very detailed run extra attributes: */
1972 	return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1973 }
1974 
1975 static const char * const stat_record_usage[] = {
1976 	"perf stat record [<options>]",
1977 	NULL,
1978 };
1979 
1980 static void init_features(struct perf_session *session)
1981 {
1982 	int feat;
1983 
1984 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1985 		perf_header__set_feat(&session->header, feat);
1986 
1987 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1988 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1989 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1990 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1991 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1992 }
1993 
1994 static int __cmd_record(int argc, const char **argv)
1995 {
1996 	struct perf_session *session;
1997 	struct perf_data *data = &perf_stat.data;
1998 
1999 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
2000 			     PARSE_OPT_STOP_AT_NON_OPTION);
2001 
2002 	if (output_name)
2003 		data->path = output_name;
2004 
2005 	if (stat_config.run_count != 1 || forever) {
2006 		pr_err("Cannot use -r option with perf stat record.\n");
2007 		return -1;
2008 	}
2009 
2010 	session = perf_session__new(data, NULL);
2011 	if (IS_ERR(session)) {
2012 		pr_err("Perf session creation failed\n");
2013 		return PTR_ERR(session);
2014 	}
2015 
2016 	init_features(session);
2017 
2018 	session->evlist   = evsel_list;
2019 	perf_stat.session = session;
2020 	perf_stat.record  = true;
2021 	return argc;
2022 }
2023 
2024 static int process_stat_round_event(struct perf_session *session,
2025 				    union perf_event *event)
2026 {
2027 	struct perf_record_stat_round *stat_round = &event->stat_round;
2028 	struct evsel *counter;
2029 	struct timespec tsh, *ts = NULL;
2030 	const char **argv = session->header.env.cmdline_argv;
2031 	int argc = session->header.env.nr_cmdline;
2032 
2033 	evlist__for_each_entry(evsel_list, counter)
2034 		perf_stat_process_counter(&stat_config, counter);
2035 
2036 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2037 		update_stats(&walltime_nsecs_stats, stat_round->time);
2038 
2039 	if (stat_config.interval && stat_round->time) {
2040 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2041 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2042 		ts = &tsh;
2043 	}
2044 
2045 	print_counters(ts, argc, argv);
2046 	return 0;
2047 }
2048 
2049 static
2050 int process_stat_config_event(struct perf_session *session,
2051 			      union perf_event *event)
2052 {
2053 	struct perf_tool *tool = session->tool;
2054 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2055 
2056 	perf_event__read_stat_config(&stat_config, &event->stat_config);
2057 
2058 	if (perf_cpu_map__empty(st->cpus)) {
2059 		if (st->aggr_mode != AGGR_UNSET)
2060 			pr_warning("warning: processing task data, aggregation mode not set\n");
2061 		return 0;
2062 	}
2063 
2064 	if (st->aggr_mode != AGGR_UNSET)
2065 		stat_config.aggr_mode = st->aggr_mode;
2066 
2067 	if (perf_stat.data.is_pipe)
2068 		perf_stat_init_aggr_mode();
2069 	else
2070 		perf_stat_init_aggr_mode_file(st);
2071 
2072 	return 0;
2073 }
2074 
2075 static int set_maps(struct perf_stat *st)
2076 {
2077 	if (!st->cpus || !st->threads)
2078 		return 0;
2079 
2080 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2081 		return -EINVAL;
2082 
2083 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2084 
2085 	if (evlist__alloc_stats(evsel_list, true))
2086 		return -ENOMEM;
2087 
2088 	st->maps_allocated = true;
2089 	return 0;
2090 }
2091 
2092 static
2093 int process_thread_map_event(struct perf_session *session,
2094 			     union perf_event *event)
2095 {
2096 	struct perf_tool *tool = session->tool;
2097 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2098 
2099 	if (st->threads) {
2100 		pr_warning("Extra thread map event, ignoring.\n");
2101 		return 0;
2102 	}
2103 
2104 	st->threads = thread_map__new_event(&event->thread_map);
2105 	if (!st->threads)
2106 		return -ENOMEM;
2107 
2108 	return set_maps(st);
2109 }
2110 
2111 static
2112 int process_cpu_map_event(struct perf_session *session,
2113 			  union perf_event *event)
2114 {
2115 	struct perf_tool *tool = session->tool;
2116 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2117 	struct perf_cpu_map *cpus;
2118 
2119 	if (st->cpus) {
2120 		pr_warning("Extra cpu map event, ignoring.\n");
2121 		return 0;
2122 	}
2123 
2124 	cpus = cpu_map__new_data(&event->cpu_map.data);
2125 	if (!cpus)
2126 		return -ENOMEM;
2127 
2128 	st->cpus = cpus;
2129 	return set_maps(st);
2130 }
2131 
2132 static const char * const stat_report_usage[] = {
2133 	"perf stat report [<options>]",
2134 	NULL,
2135 };
2136 
2137 static struct perf_stat perf_stat = {
2138 	.tool = {
2139 		.attr		= perf_event__process_attr,
2140 		.event_update	= perf_event__process_event_update,
2141 		.thread_map	= process_thread_map_event,
2142 		.cpu_map	= process_cpu_map_event,
2143 		.stat_config	= process_stat_config_event,
2144 		.stat		= perf_event__process_stat_event,
2145 		.stat_round	= process_stat_round_event,
2146 	},
2147 	.aggr_mode = AGGR_UNSET,
2148 };
2149 
2150 static int __cmd_report(int argc, const char **argv)
2151 {
2152 	struct perf_session *session;
2153 	const struct option options[] = {
2154 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2155 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2156 		     "aggregate counts per processor socket", AGGR_SOCKET),
2157 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2158 		     "aggregate counts per processor die", AGGR_DIE),
2159 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2160 		     "aggregate counts per physical processor core", AGGR_CORE),
2161 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2162 		     "aggregate counts per numa node", AGGR_NODE),
2163 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2164 		     "disable CPU count aggregation", AGGR_NONE),
2165 	OPT_END()
2166 	};
2167 	struct stat st;
2168 	int ret;
2169 
2170 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2171 
2172 	if (!input_name || !strlen(input_name)) {
2173 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2174 			input_name = "-";
2175 		else
2176 			input_name = "perf.data";
2177 	}
2178 
2179 	perf_stat.data.path = input_name;
2180 	perf_stat.data.mode = PERF_DATA_MODE_READ;
2181 
2182 	session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2183 	if (IS_ERR(session))
2184 		return PTR_ERR(session);
2185 
2186 	perf_stat.session  = session;
2187 	stat_config.output = stderr;
2188 	evsel_list         = session->evlist;
2189 
2190 	ret = perf_session__process_events(session);
2191 	if (ret)
2192 		return ret;
2193 
2194 	perf_session__delete(session);
2195 	return 0;
2196 }
2197 
2198 static void setup_system_wide(int forks)
2199 {
2200 	/*
2201 	 * Make system wide (-a) the default target if
2202 	 * no target was specified and one of following
2203 	 * conditions is met:
2204 	 *
2205 	 *   - there's no workload specified
2206 	 *   - there is workload specified but all requested
2207 	 *     events are system wide events
2208 	 */
2209 	if (!target__none(&target))
2210 		return;
2211 
2212 	if (!forks)
2213 		target.system_wide = true;
2214 	else {
2215 		struct evsel *counter;
2216 
2217 		evlist__for_each_entry(evsel_list, counter) {
2218 			if (!counter->core.requires_cpu &&
2219 			    strcmp(counter->name, "duration_time")) {
2220 				return;
2221 			}
2222 		}
2223 
2224 		if (evsel_list->core.nr_entries)
2225 			target.system_wide = true;
2226 	}
2227 }
2228 
2229 int cmd_stat(int argc, const char **argv)
2230 {
2231 	const char * const stat_usage[] = {
2232 		"perf stat [<options>] [<command>]",
2233 		NULL
2234 	};
2235 	int status = -EINVAL, run_idx, err;
2236 	const char *mode;
2237 	FILE *output = stderr;
2238 	unsigned int interval, timeout;
2239 	const char * const stat_subcommands[] = { "record", "report" };
2240 	char errbuf[BUFSIZ];
2241 
2242 	setlocale(LC_ALL, "");
2243 
2244 	evsel_list = evlist__new();
2245 	if (evsel_list == NULL)
2246 		return -ENOMEM;
2247 
2248 	parse_events__shrink_config_terms();
2249 
2250 	/* String-parsing callback-based options would segfault when negated */
2251 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2252 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2253 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2254 
2255 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2256 					(const char **) stat_usage,
2257 					PARSE_OPT_STOP_AT_NON_OPTION);
2258 	perf_stat__collect_metric_expr(evsel_list);
2259 	perf_stat__init_shadow_stats();
2260 
2261 	if (stat_config.csv_sep) {
2262 		stat_config.csv_output = true;
2263 		if (!strcmp(stat_config.csv_sep, "\\t"))
2264 			stat_config.csv_sep = "\t";
2265 	} else
2266 		stat_config.csv_sep = DEFAULT_SEPARATOR;
2267 
2268 	if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2269 		argc = __cmd_record(argc, argv);
2270 		if (argc < 0)
2271 			return -1;
2272 	} else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2273 		return __cmd_report(argc, argv);
2274 
2275 	interval = stat_config.interval;
2276 	timeout = stat_config.timeout;
2277 
2278 	/*
2279 	 * For record command the -o is already taken care of.
2280 	 */
2281 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2282 		output = NULL;
2283 
2284 	if (output_name && output_fd) {
2285 		fprintf(stderr, "cannot use both --output and --log-fd\n");
2286 		parse_options_usage(stat_usage, stat_options, "o", 1);
2287 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2288 		goto out;
2289 	}
2290 
2291 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2292 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2293 		goto out;
2294 	}
2295 
2296 	if (stat_config.metric_only && stat_config.run_count > 1) {
2297 		fprintf(stderr, "--metric-only is not supported with -r\n");
2298 		goto out;
2299 	}
2300 
2301 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2302 		fprintf(stderr, "--table is only supported with -r\n");
2303 		parse_options_usage(stat_usage, stat_options, "r", 1);
2304 		parse_options_usage(NULL, stat_options, "table", 0);
2305 		goto out;
2306 	}
2307 
2308 	if (output_fd < 0) {
2309 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2310 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2311 		goto out;
2312 	}
2313 
2314 	if (!output && !stat_config.quiet) {
2315 		struct timespec tm;
2316 		mode = append_file ? "a" : "w";
2317 
2318 		output = fopen(output_name, mode);
2319 		if (!output) {
2320 			perror("failed to create output file");
2321 			return -1;
2322 		}
2323 		clock_gettime(CLOCK_REALTIME, &tm);
2324 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2325 	} else if (output_fd > 0) {
2326 		mode = append_file ? "a" : "w";
2327 		output = fdopen(output_fd, mode);
2328 		if (!output) {
2329 			perror("Failed opening logfd");
2330 			return -errno;
2331 		}
2332 	}
2333 
2334 	stat_config.output = output;
2335 
2336 	/*
2337 	 * let the spreadsheet do the pretty-printing
2338 	 */
2339 	if (stat_config.csv_output) {
2340 		/* User explicitly passed -B? */
2341 		if (big_num_opt == 1) {
2342 			fprintf(stderr, "-B option not supported with -x\n");
2343 			parse_options_usage(stat_usage, stat_options, "B", 1);
2344 			parse_options_usage(NULL, stat_options, "x", 1);
2345 			goto out;
2346 		} else /* Nope, so disable big number formatting */
2347 			stat_config.big_num = false;
2348 	} else if (big_num_opt == 0) /* User passed --no-big-num */
2349 		stat_config.big_num = false;
2350 
2351 	err = target__validate(&target);
2352 	if (err) {
2353 		target__strerror(&target, err, errbuf, BUFSIZ);
2354 		pr_warning("%s\n", errbuf);
2355 	}
2356 
2357 	setup_system_wide(argc);
2358 
2359 	/*
2360 	 * Display user/system times only for single
2361 	 * run and when there's specified tracee.
2362 	 */
2363 	if ((stat_config.run_count == 1) && target__none(&target))
2364 		stat_config.ru_display = true;
2365 
2366 	if (stat_config.run_count < 0) {
2367 		pr_err("Run count must be a positive number\n");
2368 		parse_options_usage(stat_usage, stat_options, "r", 1);
2369 		goto out;
2370 	} else if (stat_config.run_count == 0) {
2371 		forever = true;
2372 		stat_config.run_count = 1;
2373 	}
2374 
2375 	if (stat_config.walltime_run_table) {
2376 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2377 		if (!stat_config.walltime_run) {
2378 			pr_err("failed to setup -r option");
2379 			goto out;
2380 		}
2381 	}
2382 
2383 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2384 		!target__has_task(&target)) {
2385 		if (!target.system_wide || target.cpu_list) {
2386 			fprintf(stderr, "The --per-thread option is only "
2387 				"available when monitoring via -p -t -a "
2388 				"options or only --per-thread.\n");
2389 			parse_options_usage(NULL, stat_options, "p", 1);
2390 			parse_options_usage(NULL, stat_options, "t", 1);
2391 			goto out;
2392 		}
2393 	}
2394 
2395 	/*
2396 	 * no_aggr, cgroup are for system-wide only
2397 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2398 	 */
2399 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2400 	      stat_config.aggr_mode != AGGR_THREAD) ||
2401 	     (nr_cgroups || stat_config.cgroup_list)) &&
2402 	    !target__has_cpu(&target)) {
2403 		fprintf(stderr, "both cgroup and no-aggregation "
2404 			"modes only available in system-wide mode\n");
2405 
2406 		parse_options_usage(stat_usage, stat_options, "G", 1);
2407 		parse_options_usage(NULL, stat_options, "A", 1);
2408 		parse_options_usage(NULL, stat_options, "a", 1);
2409 		parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2410 		goto out;
2411 	}
2412 
2413 	if (stat_config.iostat_run) {
2414 		status = iostat_prepare(evsel_list, &stat_config);
2415 		if (status)
2416 			goto out;
2417 		if (iostat_mode == IOSTAT_LIST) {
2418 			iostat_list(evsel_list, &stat_config);
2419 			goto out;
2420 		} else if (verbose)
2421 			iostat_list(evsel_list, &stat_config);
2422 		if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2423 			target.system_wide = true;
2424 	}
2425 
2426 	if (add_default_attributes())
2427 		goto out;
2428 
2429 	if (stat_config.cgroup_list) {
2430 		if (nr_cgroups > 0) {
2431 			pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2432 			parse_options_usage(stat_usage, stat_options, "G", 1);
2433 			parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2434 			goto out;
2435 		}
2436 
2437 		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2438 					  &stat_config.metric_events, true) < 0) {
2439 			parse_options_usage(stat_usage, stat_options,
2440 					    "for-each-cgroup", 0);
2441 			goto out;
2442 		}
2443 	}
2444 
2445 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2446 		target.per_thread = true;
2447 
2448 	if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2449 		pr_err("failed to use cpu list %s\n", target.cpu_list);
2450 		goto out;
2451 	}
2452 
2453 	target.hybrid = perf_pmu__has_hybrid();
2454 	if (evlist__create_maps(evsel_list, &target) < 0) {
2455 		if (target__has_task(&target)) {
2456 			pr_err("Problems finding threads of monitor\n");
2457 			parse_options_usage(stat_usage, stat_options, "p", 1);
2458 			parse_options_usage(NULL, stat_options, "t", 1);
2459 		} else if (target__has_cpu(&target)) {
2460 			perror("failed to parse CPUs map");
2461 			parse_options_usage(stat_usage, stat_options, "C", 1);
2462 			parse_options_usage(NULL, stat_options, "a", 1);
2463 		}
2464 		goto out;
2465 	}
2466 
2467 	evlist__check_cpu_maps(evsel_list);
2468 
2469 	/*
2470 	 * Initialize thread_map with comm names,
2471 	 * so we could print it out on output.
2472 	 */
2473 	if (stat_config.aggr_mode == AGGR_THREAD) {
2474 		thread_map__read_comms(evsel_list->core.threads);
2475 		if (target.system_wide) {
2476 			if (runtime_stat_new(&stat_config,
2477 				perf_thread_map__nr(evsel_list->core.threads))) {
2478 				goto out;
2479 			}
2480 		}
2481 	}
2482 
2483 	if (stat_config.aggr_mode == AGGR_NODE)
2484 		cpu__setup_cpunode_map();
2485 
2486 	if (stat_config.times && interval)
2487 		interval_count = true;
2488 	else if (stat_config.times && !interval) {
2489 		pr_err("interval-count option should be used together with "
2490 				"interval-print.\n");
2491 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2492 		parse_options_usage(stat_usage, stat_options, "I", 1);
2493 		goto out;
2494 	}
2495 
2496 	if (timeout && timeout < 100) {
2497 		if (timeout < 10) {
2498 			pr_err("timeout must be >= 10ms.\n");
2499 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2500 			goto out;
2501 		} else
2502 			pr_warning("timeout < 100ms. "
2503 				   "The overhead percentage could be high in some cases. "
2504 				   "Please proceed with caution.\n");
2505 	}
2506 	if (timeout && interval) {
2507 		pr_err("timeout option is not supported with interval-print.\n");
2508 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2509 		parse_options_usage(stat_usage, stat_options, "I", 1);
2510 		goto out;
2511 	}
2512 
2513 	if (evlist__alloc_stats(evsel_list, interval))
2514 		goto out;
2515 
2516 	if (perf_stat_init_aggr_mode())
2517 		goto out;
2518 
2519 	/*
2520 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2521 	 * while avoiding that older tools show confusing messages.
2522 	 *
2523 	 * However for pipe sessions we need to keep it zero,
2524 	 * because script's perf_evsel__check_attr is triggered
2525 	 * by attr->sample_type != 0, and we can't run it on
2526 	 * stat sessions.
2527 	 */
2528 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2529 
2530 	/*
2531 	 * We dont want to block the signals - that would cause
2532 	 * child tasks to inherit that and Ctrl-C would not work.
2533 	 * What we want is for Ctrl-C to work in the exec()-ed
2534 	 * task, but being ignored by perf stat itself:
2535 	 */
2536 	atexit(sig_atexit);
2537 	if (!forever)
2538 		signal(SIGINT,  skip_signal);
2539 	signal(SIGCHLD, skip_signal);
2540 	signal(SIGALRM, skip_signal);
2541 	signal(SIGABRT, skip_signal);
2542 
2543 	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2544 		goto out;
2545 
2546 	/* Enable ignoring missing threads when -p option is defined. */
2547 	evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2548 	status = 0;
2549 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2550 		if (stat_config.run_count != 1 && verbose > 0)
2551 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2552 				run_idx + 1);
2553 
2554 		if (run_idx != 0)
2555 			evlist__reset_prev_raw_counts(evsel_list);
2556 
2557 		status = run_perf_stat(argc, argv, run_idx);
2558 		if (forever && status != -1 && !interval) {
2559 			print_counters(NULL, argc, argv);
2560 			perf_stat__reset_stats();
2561 		}
2562 	}
2563 
2564 	if (!forever && status != -1 && (!interval || stat_config.summary))
2565 		print_counters(NULL, argc, argv);
2566 
2567 	evlist__finalize_ctlfd(evsel_list);
2568 
2569 	if (STAT_RECORD) {
2570 		/*
2571 		 * We synthesize the kernel mmap record just so that older tools
2572 		 * don't emit warnings about not being able to resolve symbols
2573 		 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2574 		 * a saner message about no samples being in the perf.data file.
2575 		 *
2576 		 * This also serves to suppress a warning about f_header.data.size == 0
2577 		 * in header.c at the moment 'perf stat record' gets introduced, which
2578 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2579 		 * records, but the need to suppress the kptr_restrict messages in older
2580 		 * tools remain  -acme
2581 		 */
2582 		int fd = perf_data__fd(&perf_stat.data);
2583 
2584 		err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2585 							 process_synthesized_event,
2586 							 &perf_stat.session->machines.host);
2587 		if (err) {
2588 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2589 				   "older tools may produce warnings about this file\n.");
2590 		}
2591 
2592 		if (!interval) {
2593 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2594 				pr_err("failed to write stat round event\n");
2595 		}
2596 
2597 		if (!perf_stat.data.is_pipe) {
2598 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2599 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2600 		}
2601 
2602 		evlist__close(evsel_list);
2603 		perf_session__delete(perf_stat.session);
2604 	}
2605 
2606 	perf_stat__exit_aggr_mode();
2607 	evlist__free_stats(evsel_list);
2608 out:
2609 	if (stat_config.iostat_run)
2610 		iostat_release(evsel_list);
2611 
2612 	zfree(&stat_config.walltime_run);
2613 
2614 	if (smi_cost && smi_reset)
2615 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2616 
2617 	evlist__delete(evsel_list);
2618 
2619 	metricgroup__rblist_exit(&stat_config.metric_events);
2620 	runtime_stat_delete(&stat_config);
2621 	evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2622 
2623 	return status;
2624 }
2625