xref: /openbmc/linux/tools/perf/builtin-stat.c (revision 2dd6532e)
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 		if (bpf_counter__load(counter, &target))
830 			return -1;
831 		if (!evsel__is_bpf(counter))
832 			all_counters_use_bpf = false;
833 	}
834 
835 	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
836 		counter = evlist_cpu_itr.evsel;
837 
838 		/*
839 		 * bperf calls evsel__open_per_cpu() in bperf__load(), so
840 		 * no need to call it again here.
841 		 */
842 		if (target.use_bpf)
843 			break;
844 
845 		if (counter->reset_group || counter->errored)
846 			continue;
847 		if (evsel__is_bpf(counter))
848 			continue;
849 try_again:
850 		if (create_perf_stat_counter(counter, &stat_config, &target,
851 					     evlist_cpu_itr.cpu_map_idx) < 0) {
852 
853 			/*
854 			 * Weak group failed. We cannot just undo this here
855 			 * because earlier CPUs might be in group mode, and the kernel
856 			 * doesn't support mixing group and non group reads. Defer
857 			 * it to later.
858 			 * Don't close here because we're in the wrong affinity.
859 			 */
860 			if ((errno == EINVAL || errno == EBADF) &&
861 				evsel__leader(counter) != counter &&
862 				counter->weak_group) {
863 				evlist__reset_weak_group(evsel_list, counter, false);
864 				assert(counter->reset_group);
865 				second_pass = true;
866 				continue;
867 			}
868 
869 			switch (stat_handle_error(counter)) {
870 			case COUNTER_FATAL:
871 				return -1;
872 			case COUNTER_RETRY:
873 				goto try_again;
874 			case COUNTER_SKIP:
875 				continue;
876 			default:
877 				break;
878 			}
879 
880 		}
881 		counter->supported = true;
882 	}
883 
884 	if (second_pass) {
885 		/*
886 		 * Now redo all the weak group after closing them,
887 		 * and also close errored counters.
888 		 */
889 
890 		/* First close errored or weak retry */
891 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
892 			counter = evlist_cpu_itr.evsel;
893 
894 			if (!counter->reset_group && !counter->errored)
895 				continue;
896 
897 			perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
898 		}
899 		/* Now reopen weak */
900 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
901 			counter = evlist_cpu_itr.evsel;
902 
903 			if (!counter->reset_group && !counter->errored)
904 				continue;
905 			if (!counter->reset_group)
906 				continue;
907 try_again_reset:
908 			pr_debug2("reopening weak %s\n", evsel__name(counter));
909 			if (create_perf_stat_counter(counter, &stat_config, &target,
910 						     evlist_cpu_itr.cpu_map_idx) < 0) {
911 
912 				switch (stat_handle_error(counter)) {
913 				case COUNTER_FATAL:
914 					return -1;
915 				case COUNTER_RETRY:
916 					goto try_again_reset;
917 				case COUNTER_SKIP:
918 					continue;
919 				default:
920 					break;
921 				}
922 			}
923 			counter->supported = true;
924 		}
925 	}
926 	affinity__cleanup(affinity);
927 
928 	evlist__for_each_entry(evsel_list, counter) {
929 		if (!counter->supported) {
930 			perf_evsel__free_fd(&counter->core);
931 			continue;
932 		}
933 
934 		l = strlen(counter->unit);
935 		if (l > stat_config.unit_width)
936 			stat_config.unit_width = l;
937 
938 		if (evsel__should_store_id(counter) &&
939 		    evsel__store_ids(counter, evsel_list))
940 			return -1;
941 	}
942 
943 	if (evlist__apply_filters(evsel_list, &counter)) {
944 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
945 			counter->filter, evsel__name(counter), errno,
946 			str_error_r(errno, msg, sizeof(msg)));
947 		return -1;
948 	}
949 
950 	if (STAT_RECORD) {
951 		int fd = perf_data__fd(&perf_stat.data);
952 
953 		if (is_pipe) {
954 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
955 		} else {
956 			err = perf_session__write_header(perf_stat.session, evsel_list,
957 							 fd, false);
958 		}
959 
960 		if (err < 0)
961 			return err;
962 
963 		err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
964 							 process_synthesized_event, is_pipe);
965 		if (err < 0)
966 			return err;
967 	}
968 
969 	/*
970 	 * Enable counters and exec the command:
971 	 */
972 	if (forks) {
973 		err = enable_counters();
974 		if (err)
975 			return -1;
976 		evlist__start_workload(evsel_list);
977 
978 		t0 = rdclock();
979 		clock_gettime(CLOCK_MONOTONIC, &ref_time);
980 
981 		if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
982 			status = dispatch_events(forks, timeout, interval, &times);
983 		if (child_pid != -1) {
984 			if (timeout)
985 				kill(child_pid, SIGTERM);
986 			wait4(child_pid, &status, 0, &stat_config.ru_data);
987 		}
988 
989 		if (workload_exec_errno) {
990 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
991 			pr_err("Workload failed: %s\n", emsg);
992 			return -1;
993 		}
994 
995 		if (WIFSIGNALED(status))
996 			psignal(WTERMSIG(status), argv[0]);
997 	} else {
998 		err = enable_counters();
999 		if (err)
1000 			return -1;
1001 
1002 		t0 = rdclock();
1003 		clock_gettime(CLOCK_MONOTONIC, &ref_time);
1004 
1005 		status = dispatch_events(forks, timeout, interval, &times);
1006 	}
1007 
1008 	disable_counters();
1009 
1010 	t1 = rdclock();
1011 
1012 	if (stat_config.walltime_run_table)
1013 		stat_config.walltime_run[run_idx] = t1 - t0;
1014 
1015 	if (interval && stat_config.summary) {
1016 		stat_config.interval = 0;
1017 		stat_config.stop_read_counter = true;
1018 		init_stats(&walltime_nsecs_stats);
1019 		update_stats(&walltime_nsecs_stats, t1 - t0);
1020 
1021 		if (stat_config.aggr_mode == AGGR_GLOBAL)
1022 			evlist__save_aggr_prev_raw_counts(evsel_list);
1023 
1024 		evlist__copy_prev_raw_counts(evsel_list);
1025 		evlist__reset_prev_raw_counts(evsel_list);
1026 		runtime_stat_reset(&stat_config);
1027 		perf_stat__reset_shadow_per_stat(&rt_stat);
1028 	} else {
1029 		update_stats(&walltime_nsecs_stats, t1 - t0);
1030 		update_rusage_stats(&ru_stats, &stat_config.ru_data);
1031 	}
1032 
1033 	/*
1034 	 * Closing a group leader splits the group, and as we only disable
1035 	 * group leaders, results in remaining events becoming enabled. To
1036 	 * avoid arbitrary skew, we must read all counters before closing any
1037 	 * group leaders.
1038 	 */
1039 	read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
1040 
1041 	/*
1042 	 * We need to keep evsel_list alive, because it's processed
1043 	 * later the evsel_list will be closed after.
1044 	 */
1045 	if (!STAT_RECORD)
1046 		evlist__close(evsel_list);
1047 
1048 	return WEXITSTATUS(status);
1049 }
1050 
1051 static int run_perf_stat(int argc, const char **argv, int run_idx)
1052 {
1053 	int ret;
1054 
1055 	if (pre_cmd) {
1056 		ret = system(pre_cmd);
1057 		if (ret)
1058 			return ret;
1059 	}
1060 
1061 	if (sync_run)
1062 		sync();
1063 
1064 	ret = __run_perf_stat(argc, argv, run_idx);
1065 	if (ret)
1066 		return ret;
1067 
1068 	if (post_cmd) {
1069 		ret = system(post_cmd);
1070 		if (ret)
1071 			return ret;
1072 	}
1073 
1074 	return ret;
1075 }
1076 
1077 static void print_counters(struct timespec *ts, int argc, const char **argv)
1078 {
1079 	/* Do not print anything if we record to the pipe. */
1080 	if (STAT_RECORD && perf_stat.data.is_pipe)
1081 		return;
1082 	if (stat_config.quiet)
1083 		return;
1084 
1085 	evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1086 }
1087 
1088 static volatile int signr = -1;
1089 
1090 static void skip_signal(int signo)
1091 {
1092 	if ((child_pid == -1) || stat_config.interval)
1093 		done = 1;
1094 
1095 	signr = signo;
1096 	/*
1097 	 * render child_pid harmless
1098 	 * won't send SIGTERM to a random
1099 	 * process in case of race condition
1100 	 * and fast PID recycling
1101 	 */
1102 	child_pid = -1;
1103 }
1104 
1105 static void sig_atexit(void)
1106 {
1107 	sigset_t set, oset;
1108 
1109 	/*
1110 	 * avoid race condition with SIGCHLD handler
1111 	 * in skip_signal() which is modifying child_pid
1112 	 * goal is to avoid send SIGTERM to a random
1113 	 * process
1114 	 */
1115 	sigemptyset(&set);
1116 	sigaddset(&set, SIGCHLD);
1117 	sigprocmask(SIG_BLOCK, &set, &oset);
1118 
1119 	if (child_pid != -1)
1120 		kill(child_pid, SIGTERM);
1121 
1122 	sigprocmask(SIG_SETMASK, &oset, NULL);
1123 
1124 	if (signr == -1)
1125 		return;
1126 
1127 	signal(signr, SIG_DFL);
1128 	kill(getpid(), signr);
1129 }
1130 
1131 void perf_stat__set_big_num(int set)
1132 {
1133 	stat_config.big_num = (set != 0);
1134 }
1135 
1136 void perf_stat__set_no_csv_summary(int set)
1137 {
1138 	stat_config.no_csv_summary = (set != 0);
1139 }
1140 
1141 static int stat__set_big_num(const struct option *opt __maybe_unused,
1142 			     const char *s __maybe_unused, int unset)
1143 {
1144 	big_num_opt = unset ? 0 : 1;
1145 	perf_stat__set_big_num(!unset);
1146 	return 0;
1147 }
1148 
1149 static int enable_metric_only(const struct option *opt __maybe_unused,
1150 			      const char *s __maybe_unused, int unset)
1151 {
1152 	force_metric_only = true;
1153 	stat_config.metric_only = !unset;
1154 	return 0;
1155 }
1156 
1157 static int parse_metric_groups(const struct option *opt,
1158 			       const char *str,
1159 			       int unset __maybe_unused)
1160 {
1161 	return metricgroup__parse_groups(opt, str,
1162 					 stat_config.metric_no_group,
1163 					 stat_config.metric_no_merge,
1164 					 &stat_config.metric_events);
1165 }
1166 
1167 static int parse_control_option(const struct option *opt,
1168 				const char *str,
1169 				int unset __maybe_unused)
1170 {
1171 	struct perf_stat_config *config = opt->value;
1172 
1173 	return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1174 }
1175 
1176 static int parse_stat_cgroups(const struct option *opt,
1177 			      const char *str, int unset)
1178 {
1179 	if (stat_config.cgroup_list) {
1180 		pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1181 		return -1;
1182 	}
1183 
1184 	return parse_cgroups(opt, str, unset);
1185 }
1186 
1187 static int parse_hybrid_type(const struct option *opt,
1188 			     const char *str,
1189 			     int unset __maybe_unused)
1190 {
1191 	struct evlist *evlist = *(struct evlist **)opt->value;
1192 
1193 	if (!list_empty(&evlist->core.entries)) {
1194 		fprintf(stderr, "Must define cputype before events/metrics\n");
1195 		return -1;
1196 	}
1197 
1198 	evlist->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu(str);
1199 	if (!evlist->hybrid_pmu_name) {
1200 		fprintf(stderr, "--cputype %s is not supported!\n", str);
1201 		return -1;
1202 	}
1203 
1204 	return 0;
1205 }
1206 
1207 static struct option stat_options[] = {
1208 	OPT_BOOLEAN('T', "transaction", &transaction_run,
1209 		    "hardware transaction statistics"),
1210 	OPT_CALLBACK('e', "event", &evsel_list, "event",
1211 		     "event selector. use 'perf list' to list available events",
1212 		     parse_events_option),
1213 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1214 		     "event filter", parse_filter),
1215 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1216 		    "child tasks do not inherit counters"),
1217 	OPT_STRING('p', "pid", &target.pid, "pid",
1218 		   "stat events on existing process id"),
1219 	OPT_STRING('t', "tid", &target.tid, "tid",
1220 		   "stat events on existing thread id"),
1221 #ifdef HAVE_BPF_SKEL
1222 	OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1223 		   "stat events on existing bpf program id"),
1224 	OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1225 		    "use bpf program to count events"),
1226 	OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1227 		   "path to perf_event_attr map"),
1228 #endif
1229 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1230 		    "system-wide collection from all CPUs"),
1231 	OPT_BOOLEAN('g', "group", &group,
1232 		    "put the counters into a counter group"),
1233 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
1234 		    "Use --no-scale to disable counter scaling for multiplexing"),
1235 	OPT_INCR('v', "verbose", &verbose,
1236 		    "be more verbose (show counter open errors, etc)"),
1237 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
1238 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1239 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1240 		    "display details about each run (only with -r option)"),
1241 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
1242 		    "null run - dont start any counters"),
1243 	OPT_INCR('d', "detailed", &detailed_run,
1244 		    "detailed run - start a lot of events"),
1245 	OPT_BOOLEAN('S', "sync", &sync_run,
1246 		    "call sync() before starting a run"),
1247 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1248 			   "print large numbers with thousands\' separators",
1249 			   stat__set_big_num),
1250 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1251 		    "list of cpus to monitor in system-wide"),
1252 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1253 		    "disable CPU count aggregation", AGGR_NONE),
1254 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1255 	OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
1256 		    "Merge identical named hybrid events"),
1257 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1258 		   "print counts with custom separator"),
1259 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1260 		     "monitor event in cgroup name only", parse_stat_cgroups),
1261 	OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1262 		    "expand events for each cgroup"),
1263 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1264 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1265 	OPT_INTEGER(0, "log-fd", &output_fd,
1266 		    "log output to fd, instead of stderr"),
1267 	OPT_STRING(0, "pre", &pre_cmd, "command",
1268 			"command to run prior to the measured command"),
1269 	OPT_STRING(0, "post", &post_cmd, "command",
1270 			"command to run after to the measured command"),
1271 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1272 		    "print counts at regular interval in ms "
1273 		    "(overhead is possible for values <= 100ms)"),
1274 	OPT_INTEGER(0, "interval-count", &stat_config.times,
1275 		    "print counts for fixed number of times"),
1276 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1277 		    "clear screen in between new interval"),
1278 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1279 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1280 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1281 		     "aggregate counts per processor socket", AGGR_SOCKET),
1282 	OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1283 		     "aggregate counts per processor die", AGGR_DIE),
1284 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1285 		     "aggregate counts per physical processor core", AGGR_CORE),
1286 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1287 		     "aggregate counts per thread", AGGR_THREAD),
1288 	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1289 		     "aggregate counts per numa node", AGGR_NODE),
1290 	OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1291 		    "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1292 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1293 			"Only print computed metrics. No raw values", enable_metric_only),
1294 	OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1295 		       "don't group metric events, impacts multiplexing"),
1296 	OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1297 		       "don't try to share events between metrics in a group"),
1298 	OPT_BOOLEAN(0, "topdown", &topdown_run,
1299 			"measure top-down statistics"),
1300 	OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1301 			"Set the metrics level for the top-down statistics (0: max level)"),
1302 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1303 			"measure SMI cost"),
1304 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1305 		     "monitor specified metrics or metric groups (separated by ,)",
1306 		     parse_metric_groups),
1307 	OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1308 			 "Configure all used events to run in kernel space.",
1309 			 PARSE_OPT_EXCLUSIVE),
1310 	OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1311 			 "Configure all used events to run in user space.",
1312 			 PARSE_OPT_EXCLUSIVE),
1313 	OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1314 		    "Use with 'percore' event qualifier to show the event "
1315 		    "counts of one hardware thread by sum up total hardware "
1316 		    "threads of same physical core"),
1317 	OPT_BOOLEAN(0, "summary", &stat_config.summary,
1318 		       "print summary for interval mode"),
1319 	OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1320 		       "don't print 'summary' for CSV summary output"),
1321 	OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1322 			"don't print output (useful with record)"),
1323 	OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1324 		     "Only enable events on applying cpu with this type "
1325 		     "for hybrid platform (e.g. core or atom)",
1326 		     parse_hybrid_type),
1327 #ifdef HAVE_LIBPFM
1328 	OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1329 		"libpfm4 event selector. use 'perf list' to list available events",
1330 		parse_libpfm_events_option),
1331 #endif
1332 	OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1333 		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1334 		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1335 		     "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1336 		      parse_control_option),
1337 	OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1338 			    "measure I/O performance metrics provided by arch/platform",
1339 			    iostat_parse),
1340 	OPT_END()
1341 };
1342 
1343 static const char *const aggr_mode__string[] = {
1344 	[AGGR_CORE] = "core",
1345 	[AGGR_DIE] = "die",
1346 	[AGGR_GLOBAL] = "global",
1347 	[AGGR_NODE] = "node",
1348 	[AGGR_NONE] = "none",
1349 	[AGGR_SOCKET] = "socket",
1350 	[AGGR_THREAD] = "thread",
1351 	[AGGR_UNSET] = "unset",
1352 };
1353 
1354 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1355 						struct perf_cpu cpu)
1356 {
1357 	return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1358 }
1359 
1360 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1361 					     struct perf_cpu cpu)
1362 {
1363 	return aggr_cpu_id__die(cpu, /*data=*/NULL);
1364 }
1365 
1366 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1367 					      struct perf_cpu cpu)
1368 {
1369 	return aggr_cpu_id__core(cpu, /*data=*/NULL);
1370 }
1371 
1372 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1373 					      struct perf_cpu cpu)
1374 {
1375 	return aggr_cpu_id__node(cpu, /*data=*/NULL);
1376 }
1377 
1378 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1379 					      aggr_get_id_t get_id, struct perf_cpu cpu)
1380 {
1381 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1382 
1383 	if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1384 		config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1385 
1386 	id = config->cpus_aggr_map->map[cpu.cpu];
1387 	return id;
1388 }
1389 
1390 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1391 						       struct perf_cpu cpu)
1392 {
1393 	return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1394 }
1395 
1396 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1397 						    struct perf_cpu cpu)
1398 {
1399 	return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1400 }
1401 
1402 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1403 						     struct perf_cpu cpu)
1404 {
1405 	return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1406 }
1407 
1408 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1409 						     struct perf_cpu cpu)
1410 {
1411 	return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1412 }
1413 
1414 static bool term_percore_set(void)
1415 {
1416 	struct evsel *counter;
1417 
1418 	evlist__for_each_entry(evsel_list, counter) {
1419 		if (counter->percore)
1420 			return true;
1421 	}
1422 
1423 	return false;
1424 }
1425 
1426 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1427 {
1428 	switch (aggr_mode) {
1429 	case AGGR_SOCKET:
1430 		return aggr_cpu_id__socket;
1431 	case AGGR_DIE:
1432 		return aggr_cpu_id__die;
1433 	case AGGR_CORE:
1434 		return aggr_cpu_id__core;
1435 	case AGGR_NODE:
1436 		return aggr_cpu_id__node;
1437 	case AGGR_NONE:
1438 		if (term_percore_set())
1439 			return aggr_cpu_id__core;
1440 
1441 		return NULL;
1442 	case AGGR_GLOBAL:
1443 	case AGGR_THREAD:
1444 	case AGGR_UNSET:
1445 	default:
1446 		return NULL;
1447 	}
1448 }
1449 
1450 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1451 {
1452 	switch (aggr_mode) {
1453 	case AGGR_SOCKET:
1454 		return perf_stat__get_socket_cached;
1455 	case AGGR_DIE:
1456 		return perf_stat__get_die_cached;
1457 	case AGGR_CORE:
1458 		return perf_stat__get_core_cached;
1459 	case AGGR_NODE:
1460 		return perf_stat__get_node_cached;
1461 	case AGGR_NONE:
1462 		if (term_percore_set()) {
1463 			return perf_stat__get_core_cached;
1464 		}
1465 		return NULL;
1466 	case AGGR_GLOBAL:
1467 	case AGGR_THREAD:
1468 	case AGGR_UNSET:
1469 	default:
1470 		return NULL;
1471 	}
1472 }
1473 
1474 static int perf_stat_init_aggr_mode(void)
1475 {
1476 	int nr;
1477 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1478 
1479 	if (get_id) {
1480 		stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1481 							 get_id, /*data=*/NULL);
1482 		if (!stat_config.aggr_map) {
1483 			pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1484 			return -1;
1485 		}
1486 		stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1487 	}
1488 
1489 	/*
1490 	 * The evsel_list->cpus is the base we operate on,
1491 	 * taking the highest cpu number to be the size of
1492 	 * the aggregation translate cpumap.
1493 	 */
1494 	if (evsel_list->core.user_requested_cpus)
1495 		nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1496 	else
1497 		nr = 0;
1498 	stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1499 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1500 }
1501 
1502 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1503 {
1504 	if (map) {
1505 		WARN_ONCE(refcount_read(&map->refcnt) != 0,
1506 			  "cpu_aggr_map refcnt unbalanced\n");
1507 		free(map);
1508 	}
1509 }
1510 
1511 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1512 {
1513 	if (map && refcount_dec_and_test(&map->refcnt))
1514 		cpu_aggr_map__delete(map);
1515 }
1516 
1517 static void perf_stat__exit_aggr_mode(void)
1518 {
1519 	cpu_aggr_map__put(stat_config.aggr_map);
1520 	cpu_aggr_map__put(stat_config.cpus_aggr_map);
1521 	stat_config.aggr_map = NULL;
1522 	stat_config.cpus_aggr_map = NULL;
1523 }
1524 
1525 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1526 {
1527 	struct perf_env *env = data;
1528 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1529 
1530 	if (cpu.cpu != -1)
1531 		id.socket = env->cpu[cpu.cpu].socket_id;
1532 
1533 	return id;
1534 }
1535 
1536 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1537 {
1538 	struct perf_env *env = data;
1539 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1540 
1541 	if (cpu.cpu != -1) {
1542 		/*
1543 		 * die_id is relative to socket, so start
1544 		 * with the socket ID and then add die to
1545 		 * make a unique ID.
1546 		 */
1547 		id.socket = env->cpu[cpu.cpu].socket_id;
1548 		id.die = env->cpu[cpu.cpu].die_id;
1549 	}
1550 
1551 	return id;
1552 }
1553 
1554 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1555 {
1556 	struct perf_env *env = data;
1557 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1558 
1559 	if (cpu.cpu != -1) {
1560 		/*
1561 		 * core_id is relative to socket and die,
1562 		 * we need a global id. So we set
1563 		 * socket, die id and core id
1564 		 */
1565 		id.socket = env->cpu[cpu.cpu].socket_id;
1566 		id.die = env->cpu[cpu.cpu].die_id;
1567 		id.core = env->cpu[cpu.cpu].core_id;
1568 	}
1569 
1570 	return id;
1571 }
1572 
1573 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1574 {
1575 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1576 
1577 	id.node = perf_env__numa_node(data, cpu);
1578 	return id;
1579 }
1580 
1581 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1582 						     struct perf_cpu cpu)
1583 {
1584 	return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1585 }
1586 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1587 						  struct perf_cpu cpu)
1588 {
1589 	return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1590 }
1591 
1592 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1593 						   struct perf_cpu cpu)
1594 {
1595 	return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1596 }
1597 
1598 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1599 						   struct perf_cpu cpu)
1600 {
1601 	return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1602 }
1603 
1604 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1605 {
1606 	switch (aggr_mode) {
1607 	case AGGR_SOCKET:
1608 		return perf_env__get_socket_aggr_by_cpu;
1609 	case AGGR_DIE:
1610 		return perf_env__get_die_aggr_by_cpu;
1611 	case AGGR_CORE:
1612 		return perf_env__get_core_aggr_by_cpu;
1613 	case AGGR_NODE:
1614 		return perf_env__get_node_aggr_by_cpu;
1615 	case AGGR_NONE:
1616 	case AGGR_GLOBAL:
1617 	case AGGR_THREAD:
1618 	case AGGR_UNSET:
1619 	default:
1620 		return NULL;
1621 	}
1622 }
1623 
1624 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1625 {
1626 	switch (aggr_mode) {
1627 	case AGGR_SOCKET:
1628 		return perf_stat__get_socket_file;
1629 	case AGGR_DIE:
1630 		return perf_stat__get_die_file;
1631 	case AGGR_CORE:
1632 		return perf_stat__get_core_file;
1633 	case AGGR_NODE:
1634 		return perf_stat__get_node_file;
1635 	case AGGR_NONE:
1636 	case AGGR_GLOBAL:
1637 	case AGGR_THREAD:
1638 	case AGGR_UNSET:
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 	struct perf_event_attr default_sw_attrs[] = {
1690   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1691   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1692   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1693   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1694 };
1695 
1696 /*
1697  * Detailed stats (-d), covering the L1 and last level data caches:
1698  */
1699 	struct perf_event_attr detailed_attrs[] = {
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_ACCESS	<< 16)				},
1706 
1707   { .type = PERF_TYPE_HW_CACHE,
1708     .config =
1709 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1710 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1711 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 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_ACCESS	<< 16)				},
1718 
1719   { .type = PERF_TYPE_HW_CACHE,
1720     .config =
1721 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1722 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1723 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1724 };
1725 
1726 /*
1727  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1728  */
1729 	struct perf_event_attr very_detailed_attrs[] = {
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_ACCESS	<< 16)				},
1736 
1737   { .type = PERF_TYPE_HW_CACHE,
1738     .config =
1739 	 PERF_COUNT_HW_CACHE_L1I		<<  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_DTLB		<<  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_DTLB		<<  0  |
1752 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1753 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 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_ACCESS	<< 16)				},
1760 
1761   { .type = PERF_TYPE_HW_CACHE,
1762     .config =
1763 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1764 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1765 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1766 
1767 };
1768 
1769 /*
1770  * Very, very detailed stats (-d -d -d), adding prefetch events:
1771  */
1772 	struct perf_event_attr very_very_detailed_attrs[] = {
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_ACCESS	<< 16)				},
1779 
1780   { .type = PERF_TYPE_HW_CACHE,
1781     .config =
1782 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1783 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1784 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1785 };
1786 	/* Set attrs if no event is selected and !null_run: */
1787 	if (stat_config.null_run)
1788 		return 0;
1789 
1790 	if (transaction_run) {
1791 		struct parse_events_error errinfo;
1792 		/* Handle -T as -M transaction. Once platform specific metrics
1793 		 * support has been added to the json files, all architectures
1794 		 * will use this approach. To determine transaction support
1795 		 * on an architecture test for such a metric name.
1796 		 */
1797 		if (metricgroup__has_metric("transaction")) {
1798 			struct option opt = { .value = &evsel_list };
1799 
1800 			return metricgroup__parse_groups(&opt, "transaction",
1801 							 stat_config.metric_no_group,
1802 							stat_config.metric_no_merge,
1803 							 &stat_config.metric_events);
1804 		}
1805 
1806 		parse_events_error__init(&errinfo);
1807 		if (pmu_have_event("cpu", "cycles-ct") &&
1808 		    pmu_have_event("cpu", "el-start"))
1809 			err = parse_events(evsel_list, transaction_attrs,
1810 					   &errinfo);
1811 		else
1812 			err = parse_events(evsel_list,
1813 					   transaction_limited_attrs,
1814 					   &errinfo);
1815 		if (err) {
1816 			fprintf(stderr, "Cannot set up transaction events\n");
1817 			parse_events_error__print(&errinfo, transaction_attrs);
1818 		}
1819 		parse_events_error__exit(&errinfo);
1820 		return err ? -1 : 0;
1821 	}
1822 
1823 	if (smi_cost) {
1824 		struct parse_events_error errinfo;
1825 		int smi;
1826 
1827 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1828 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1829 			return -1;
1830 		}
1831 
1832 		if (!smi) {
1833 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1834 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1835 				return -1;
1836 			}
1837 			smi_reset = true;
1838 		}
1839 
1840 		if (!pmu_have_event("msr", "aperf") ||
1841 		    !pmu_have_event("msr", "smi")) {
1842 			fprintf(stderr, "To measure SMI cost, it needs "
1843 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1844 			return -1;
1845 		}
1846 		if (!force_metric_only)
1847 			stat_config.metric_only = true;
1848 
1849 		parse_events_error__init(&errinfo);
1850 		err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1851 		if (err) {
1852 			parse_events_error__print(&errinfo, smi_cost_attrs);
1853 			fprintf(stderr, "Cannot set up SMI cost events\n");
1854 		}
1855 		parse_events_error__exit(&errinfo);
1856 		return err ? -1 : 0;
1857 	}
1858 
1859 	if (topdown_run) {
1860 		const char **metric_attrs = topdown_metric_attrs;
1861 		unsigned int max_level = 1;
1862 		char *str = NULL;
1863 		bool warn = false;
1864 		const char *pmu_name = "cpu";
1865 
1866 		if (!force_metric_only)
1867 			stat_config.metric_only = true;
1868 
1869 		if (perf_pmu__has_hybrid()) {
1870 			if (!evsel_list->hybrid_pmu_name) {
1871 				pr_warning("WARNING: default to use cpu_core topdown events\n");
1872 				evsel_list->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu("core");
1873 			}
1874 
1875 			pmu_name = evsel_list->hybrid_pmu_name;
1876 			if (!pmu_name)
1877 				return -1;
1878 		}
1879 
1880 		if (pmu_have_event(pmu_name, topdown_metric_L2_attrs[5])) {
1881 			metric_attrs = topdown_metric_L2_attrs;
1882 			max_level = 2;
1883 		}
1884 
1885 		if (stat_config.topdown_level > max_level) {
1886 			pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1887 			return -1;
1888 		} else if (!stat_config.topdown_level)
1889 			stat_config.topdown_level = max_level;
1890 
1891 		if (topdown_filter_events(metric_attrs, &str, 1, pmu_name) < 0) {
1892 			pr_err("Out of memory\n");
1893 			return -1;
1894 		}
1895 
1896 		if (metric_attrs[0] && str) {
1897 			if (!stat_config.interval && !stat_config.metric_only) {
1898 				fprintf(stat_config.output,
1899 					"Topdown accuracy may decrease when measuring long periods.\n"
1900 					"Please print the result regularly, e.g. -I1000\n");
1901 			}
1902 			goto setup_metrics;
1903 		}
1904 
1905 		zfree(&str);
1906 
1907 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1908 		    stat_config.aggr_mode != AGGR_CORE) {
1909 			pr_err("top down event configuration requires --per-core mode\n");
1910 			return -1;
1911 		}
1912 		stat_config.aggr_mode = AGGR_CORE;
1913 		if (nr_cgroups || !target__has_cpu(&target)) {
1914 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1915 			return -1;
1916 		}
1917 
1918 		if (topdown_filter_events(topdown_attrs, &str,
1919 				arch_topdown_check_group(&warn),
1920 				pmu_name) < 0) {
1921 			pr_err("Out of memory\n");
1922 			return -1;
1923 		}
1924 
1925 		if (topdown_attrs[0] && str) {
1926 			struct parse_events_error errinfo;
1927 			if (warn)
1928 				arch_topdown_group_warn();
1929 setup_metrics:
1930 			parse_events_error__init(&errinfo);
1931 			err = parse_events(evsel_list, str, &errinfo);
1932 			if (err) {
1933 				fprintf(stderr,
1934 					"Cannot set up top down events %s: %d\n",
1935 					str, err);
1936 				parse_events_error__print(&errinfo, str);
1937 				parse_events_error__exit(&errinfo);
1938 				free(str);
1939 				return -1;
1940 			}
1941 			parse_events_error__exit(&errinfo);
1942 		} else {
1943 			fprintf(stderr, "System does not support topdown\n");
1944 			return -1;
1945 		}
1946 		free(str);
1947 	}
1948 
1949 	if (!evsel_list->core.nr_entries) {
1950 		if (perf_pmu__has_hybrid()) {
1951 			struct parse_events_error errinfo;
1952 			const char *hybrid_str = "cycles,instructions,branches,branch-misses";
1953 
1954 			if (target__has_cpu(&target))
1955 				default_sw_attrs[0].config = PERF_COUNT_SW_CPU_CLOCK;
1956 
1957 			if (evlist__add_default_attrs(evsel_list,
1958 						      default_sw_attrs) < 0) {
1959 				return -1;
1960 			}
1961 
1962 			parse_events_error__init(&errinfo);
1963 			err = parse_events(evsel_list, hybrid_str, &errinfo);
1964 			if (err) {
1965 				fprintf(stderr,
1966 					"Cannot set up hybrid events %s: %d\n",
1967 					hybrid_str, err);
1968 				parse_events_error__print(&errinfo, hybrid_str);
1969 			}
1970 			parse_events_error__exit(&errinfo);
1971 			return err ? -1 : 0;
1972 		}
1973 
1974 		if (target__has_cpu(&target))
1975 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1976 
1977 		if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1978 			return -1;
1979 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1980 			if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1981 				return -1;
1982 		}
1983 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1984 			if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1985 				return -1;
1986 		}
1987 		if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1988 			return -1;
1989 
1990 		stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1991 		if (arch_evlist__add_default_attrs(evsel_list) < 0)
1992 			return -1;
1993 	}
1994 
1995 	/* Detailed events get appended to the event list: */
1996 
1997 	if (detailed_run <  1)
1998 		return 0;
1999 
2000 	/* Append detailed run extra attributes: */
2001 	if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2002 		return -1;
2003 
2004 	if (detailed_run < 2)
2005 		return 0;
2006 
2007 	/* Append very detailed run extra attributes: */
2008 	if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2009 		return -1;
2010 
2011 	if (detailed_run < 3)
2012 		return 0;
2013 
2014 	/* Append very, very detailed run extra attributes: */
2015 	return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2016 }
2017 
2018 static const char * const stat_record_usage[] = {
2019 	"perf stat record [<options>]",
2020 	NULL,
2021 };
2022 
2023 static void init_features(struct perf_session *session)
2024 {
2025 	int feat;
2026 
2027 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2028 		perf_header__set_feat(&session->header, feat);
2029 
2030 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
2031 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2032 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2033 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2034 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2035 }
2036 
2037 static int __cmd_record(int argc, const char **argv)
2038 {
2039 	struct perf_session *session;
2040 	struct perf_data *data = &perf_stat.data;
2041 
2042 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
2043 			     PARSE_OPT_STOP_AT_NON_OPTION);
2044 
2045 	if (output_name)
2046 		data->path = output_name;
2047 
2048 	if (stat_config.run_count != 1 || forever) {
2049 		pr_err("Cannot use -r option with perf stat record.\n");
2050 		return -1;
2051 	}
2052 
2053 	session = perf_session__new(data, NULL);
2054 	if (IS_ERR(session)) {
2055 		pr_err("Perf session creation failed\n");
2056 		return PTR_ERR(session);
2057 	}
2058 
2059 	init_features(session);
2060 
2061 	session->evlist   = evsel_list;
2062 	perf_stat.session = session;
2063 	perf_stat.record  = true;
2064 	return argc;
2065 }
2066 
2067 static int process_stat_round_event(struct perf_session *session,
2068 				    union perf_event *event)
2069 {
2070 	struct perf_record_stat_round *stat_round = &event->stat_round;
2071 	struct evsel *counter;
2072 	struct timespec tsh, *ts = NULL;
2073 	const char **argv = session->header.env.cmdline_argv;
2074 	int argc = session->header.env.nr_cmdline;
2075 
2076 	evlist__for_each_entry(evsel_list, counter)
2077 		perf_stat_process_counter(&stat_config, counter);
2078 
2079 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2080 		update_stats(&walltime_nsecs_stats, stat_round->time);
2081 
2082 	if (stat_config.interval && stat_round->time) {
2083 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2084 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2085 		ts = &tsh;
2086 	}
2087 
2088 	print_counters(ts, argc, argv);
2089 	return 0;
2090 }
2091 
2092 static
2093 int process_stat_config_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 	perf_event__read_stat_config(&stat_config, &event->stat_config);
2100 
2101 	if (perf_cpu_map__empty(st->cpus)) {
2102 		if (st->aggr_mode != AGGR_UNSET)
2103 			pr_warning("warning: processing task data, aggregation mode not set\n");
2104 		return 0;
2105 	}
2106 
2107 	if (st->aggr_mode != AGGR_UNSET)
2108 		stat_config.aggr_mode = st->aggr_mode;
2109 
2110 	if (perf_stat.data.is_pipe)
2111 		perf_stat_init_aggr_mode();
2112 	else
2113 		perf_stat_init_aggr_mode_file(st);
2114 
2115 	return 0;
2116 }
2117 
2118 static int set_maps(struct perf_stat *st)
2119 {
2120 	if (!st->cpus || !st->threads)
2121 		return 0;
2122 
2123 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2124 		return -EINVAL;
2125 
2126 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2127 
2128 	if (evlist__alloc_stats(evsel_list, true))
2129 		return -ENOMEM;
2130 
2131 	st->maps_allocated = true;
2132 	return 0;
2133 }
2134 
2135 static
2136 int process_thread_map_event(struct perf_session *session,
2137 			     union perf_event *event)
2138 {
2139 	struct perf_tool *tool = session->tool;
2140 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2141 
2142 	if (st->threads) {
2143 		pr_warning("Extra thread map event, ignoring.\n");
2144 		return 0;
2145 	}
2146 
2147 	st->threads = thread_map__new_event(&event->thread_map);
2148 	if (!st->threads)
2149 		return -ENOMEM;
2150 
2151 	return set_maps(st);
2152 }
2153 
2154 static
2155 int process_cpu_map_event(struct perf_session *session,
2156 			  union perf_event *event)
2157 {
2158 	struct perf_tool *tool = session->tool;
2159 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2160 	struct perf_cpu_map *cpus;
2161 
2162 	if (st->cpus) {
2163 		pr_warning("Extra cpu map event, ignoring.\n");
2164 		return 0;
2165 	}
2166 
2167 	cpus = cpu_map__new_data(&event->cpu_map.data);
2168 	if (!cpus)
2169 		return -ENOMEM;
2170 
2171 	st->cpus = cpus;
2172 	return set_maps(st);
2173 }
2174 
2175 static const char * const stat_report_usage[] = {
2176 	"perf stat report [<options>]",
2177 	NULL,
2178 };
2179 
2180 static struct perf_stat perf_stat = {
2181 	.tool = {
2182 		.attr		= perf_event__process_attr,
2183 		.event_update	= perf_event__process_event_update,
2184 		.thread_map	= process_thread_map_event,
2185 		.cpu_map	= process_cpu_map_event,
2186 		.stat_config	= process_stat_config_event,
2187 		.stat		= perf_event__process_stat_event,
2188 		.stat_round	= process_stat_round_event,
2189 	},
2190 	.aggr_mode = AGGR_UNSET,
2191 };
2192 
2193 static int __cmd_report(int argc, const char **argv)
2194 {
2195 	struct perf_session *session;
2196 	const struct option options[] = {
2197 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2198 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2199 		     "aggregate counts per processor socket", AGGR_SOCKET),
2200 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2201 		     "aggregate counts per processor die", AGGR_DIE),
2202 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2203 		     "aggregate counts per physical processor core", AGGR_CORE),
2204 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2205 		     "aggregate counts per numa node", AGGR_NODE),
2206 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2207 		     "disable CPU count aggregation", AGGR_NONE),
2208 	OPT_END()
2209 	};
2210 	struct stat st;
2211 	int ret;
2212 
2213 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2214 
2215 	if (!input_name || !strlen(input_name)) {
2216 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2217 			input_name = "-";
2218 		else
2219 			input_name = "perf.data";
2220 	}
2221 
2222 	perf_stat.data.path = input_name;
2223 	perf_stat.data.mode = PERF_DATA_MODE_READ;
2224 
2225 	session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2226 	if (IS_ERR(session))
2227 		return PTR_ERR(session);
2228 
2229 	perf_stat.session  = session;
2230 	stat_config.output = stderr;
2231 	evsel_list         = session->evlist;
2232 
2233 	ret = perf_session__process_events(session);
2234 	if (ret)
2235 		return ret;
2236 
2237 	perf_session__delete(session);
2238 	return 0;
2239 }
2240 
2241 static void setup_system_wide(int forks)
2242 {
2243 	/*
2244 	 * Make system wide (-a) the default target if
2245 	 * no target was specified and one of following
2246 	 * conditions is met:
2247 	 *
2248 	 *   - there's no workload specified
2249 	 *   - there is workload specified but all requested
2250 	 *     events are system wide events
2251 	 */
2252 	if (!target__none(&target))
2253 		return;
2254 
2255 	if (!forks)
2256 		target.system_wide = true;
2257 	else {
2258 		struct evsel *counter;
2259 
2260 		evlist__for_each_entry(evsel_list, counter) {
2261 			if (!counter->core.requires_cpu &&
2262 			    strcmp(counter->name, "duration_time")) {
2263 				return;
2264 			}
2265 		}
2266 
2267 		if (evsel_list->core.nr_entries)
2268 			target.system_wide = true;
2269 	}
2270 }
2271 
2272 int cmd_stat(int argc, const char **argv)
2273 {
2274 	const char * const stat_usage[] = {
2275 		"perf stat [<options>] [<command>]",
2276 		NULL
2277 	};
2278 	int status = -EINVAL, run_idx, err;
2279 	const char *mode;
2280 	FILE *output = stderr;
2281 	unsigned int interval, timeout;
2282 	const char * const stat_subcommands[] = { "record", "report" };
2283 	char errbuf[BUFSIZ];
2284 
2285 	setlocale(LC_ALL, "");
2286 
2287 	evsel_list = evlist__new();
2288 	if (evsel_list == NULL)
2289 		return -ENOMEM;
2290 
2291 	parse_events__shrink_config_terms();
2292 
2293 	/* String-parsing callback-based options would segfault when negated */
2294 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2295 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2296 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2297 
2298 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2299 					(const char **) stat_usage,
2300 					PARSE_OPT_STOP_AT_NON_OPTION);
2301 	perf_stat__collect_metric_expr(evsel_list);
2302 	perf_stat__init_shadow_stats();
2303 
2304 	if (stat_config.csv_sep) {
2305 		stat_config.csv_output = true;
2306 		if (!strcmp(stat_config.csv_sep, "\\t"))
2307 			stat_config.csv_sep = "\t";
2308 	} else
2309 		stat_config.csv_sep = DEFAULT_SEPARATOR;
2310 
2311 	if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2312 		argc = __cmd_record(argc, argv);
2313 		if (argc < 0)
2314 			return -1;
2315 	} else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2316 		return __cmd_report(argc, argv);
2317 
2318 	interval = stat_config.interval;
2319 	timeout = stat_config.timeout;
2320 
2321 	/*
2322 	 * For record command the -o is already taken care of.
2323 	 */
2324 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2325 		output = NULL;
2326 
2327 	if (output_name && output_fd) {
2328 		fprintf(stderr, "cannot use both --output and --log-fd\n");
2329 		parse_options_usage(stat_usage, stat_options, "o", 1);
2330 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2331 		goto out;
2332 	}
2333 
2334 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2335 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2336 		goto out;
2337 	}
2338 
2339 	if (stat_config.metric_only && stat_config.run_count > 1) {
2340 		fprintf(stderr, "--metric-only is not supported with -r\n");
2341 		goto out;
2342 	}
2343 
2344 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2345 		fprintf(stderr, "--table is only supported with -r\n");
2346 		parse_options_usage(stat_usage, stat_options, "r", 1);
2347 		parse_options_usage(NULL, stat_options, "table", 0);
2348 		goto out;
2349 	}
2350 
2351 	if (output_fd < 0) {
2352 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2353 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2354 		goto out;
2355 	}
2356 
2357 	if (!output && !stat_config.quiet) {
2358 		struct timespec tm;
2359 		mode = append_file ? "a" : "w";
2360 
2361 		output = fopen(output_name, mode);
2362 		if (!output) {
2363 			perror("failed to create output file");
2364 			return -1;
2365 		}
2366 		clock_gettime(CLOCK_REALTIME, &tm);
2367 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2368 	} else if (output_fd > 0) {
2369 		mode = append_file ? "a" : "w";
2370 		output = fdopen(output_fd, mode);
2371 		if (!output) {
2372 			perror("Failed opening logfd");
2373 			return -errno;
2374 		}
2375 	}
2376 
2377 	stat_config.output = output;
2378 
2379 	/*
2380 	 * let the spreadsheet do the pretty-printing
2381 	 */
2382 	if (stat_config.csv_output) {
2383 		/* User explicitly passed -B? */
2384 		if (big_num_opt == 1) {
2385 			fprintf(stderr, "-B option not supported with -x\n");
2386 			parse_options_usage(stat_usage, stat_options, "B", 1);
2387 			parse_options_usage(NULL, stat_options, "x", 1);
2388 			goto out;
2389 		} else /* Nope, so disable big number formatting */
2390 			stat_config.big_num = false;
2391 	} else if (big_num_opt == 0) /* User passed --no-big-num */
2392 		stat_config.big_num = false;
2393 
2394 	err = target__validate(&target);
2395 	if (err) {
2396 		target__strerror(&target, err, errbuf, BUFSIZ);
2397 		pr_warning("%s\n", errbuf);
2398 	}
2399 
2400 	setup_system_wide(argc);
2401 
2402 	/*
2403 	 * Display user/system times only for single
2404 	 * run and when there's specified tracee.
2405 	 */
2406 	if ((stat_config.run_count == 1) && target__none(&target))
2407 		stat_config.ru_display = true;
2408 
2409 	if (stat_config.run_count < 0) {
2410 		pr_err("Run count must be a positive number\n");
2411 		parse_options_usage(stat_usage, stat_options, "r", 1);
2412 		goto out;
2413 	} else if (stat_config.run_count == 0) {
2414 		forever = true;
2415 		stat_config.run_count = 1;
2416 	}
2417 
2418 	if (stat_config.walltime_run_table) {
2419 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2420 		if (!stat_config.walltime_run) {
2421 			pr_err("failed to setup -r option");
2422 			goto out;
2423 		}
2424 	}
2425 
2426 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2427 		!target__has_task(&target)) {
2428 		if (!target.system_wide || target.cpu_list) {
2429 			fprintf(stderr, "The --per-thread option is only "
2430 				"available when monitoring via -p -t -a "
2431 				"options or only --per-thread.\n");
2432 			parse_options_usage(NULL, stat_options, "p", 1);
2433 			parse_options_usage(NULL, stat_options, "t", 1);
2434 			goto out;
2435 		}
2436 	}
2437 
2438 	/*
2439 	 * no_aggr, cgroup are for system-wide only
2440 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2441 	 */
2442 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2443 	      stat_config.aggr_mode != AGGR_THREAD) ||
2444 	     (nr_cgroups || stat_config.cgroup_list)) &&
2445 	    !target__has_cpu(&target)) {
2446 		fprintf(stderr, "both cgroup and no-aggregation "
2447 			"modes only available in system-wide mode\n");
2448 
2449 		parse_options_usage(stat_usage, stat_options, "G", 1);
2450 		parse_options_usage(NULL, stat_options, "A", 1);
2451 		parse_options_usage(NULL, stat_options, "a", 1);
2452 		parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2453 		goto out;
2454 	}
2455 
2456 	if (stat_config.iostat_run) {
2457 		status = iostat_prepare(evsel_list, &stat_config);
2458 		if (status)
2459 			goto out;
2460 		if (iostat_mode == IOSTAT_LIST) {
2461 			iostat_list(evsel_list, &stat_config);
2462 			goto out;
2463 		} else if (verbose)
2464 			iostat_list(evsel_list, &stat_config);
2465 		if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2466 			target.system_wide = true;
2467 	}
2468 
2469 	if (add_default_attributes())
2470 		goto out;
2471 
2472 	if (stat_config.cgroup_list) {
2473 		if (nr_cgroups > 0) {
2474 			pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2475 			parse_options_usage(stat_usage, stat_options, "G", 1);
2476 			parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2477 			goto out;
2478 		}
2479 
2480 		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2481 					  &stat_config.metric_events, true) < 0) {
2482 			parse_options_usage(stat_usage, stat_options,
2483 					    "for-each-cgroup", 0);
2484 			goto out;
2485 		}
2486 	}
2487 
2488 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2489 		target.per_thread = true;
2490 
2491 	if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2492 		pr_err("failed to use cpu list %s\n", target.cpu_list);
2493 		goto out;
2494 	}
2495 
2496 	target.hybrid = perf_pmu__has_hybrid();
2497 	if (evlist__create_maps(evsel_list, &target) < 0) {
2498 		if (target__has_task(&target)) {
2499 			pr_err("Problems finding threads of monitor\n");
2500 			parse_options_usage(stat_usage, stat_options, "p", 1);
2501 			parse_options_usage(NULL, stat_options, "t", 1);
2502 		} else if (target__has_cpu(&target)) {
2503 			perror("failed to parse CPUs map");
2504 			parse_options_usage(stat_usage, stat_options, "C", 1);
2505 			parse_options_usage(NULL, stat_options, "a", 1);
2506 		}
2507 		goto out;
2508 	}
2509 
2510 	evlist__check_cpu_maps(evsel_list);
2511 
2512 	/*
2513 	 * Initialize thread_map with comm names,
2514 	 * so we could print it out on output.
2515 	 */
2516 	if (stat_config.aggr_mode == AGGR_THREAD) {
2517 		thread_map__read_comms(evsel_list->core.threads);
2518 		if (target.system_wide) {
2519 			if (runtime_stat_new(&stat_config,
2520 				perf_thread_map__nr(evsel_list->core.threads))) {
2521 				goto out;
2522 			}
2523 		}
2524 	}
2525 
2526 	if (stat_config.aggr_mode == AGGR_NODE)
2527 		cpu__setup_cpunode_map();
2528 
2529 	if (stat_config.times && interval)
2530 		interval_count = true;
2531 	else if (stat_config.times && !interval) {
2532 		pr_err("interval-count option should be used together with "
2533 				"interval-print.\n");
2534 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2535 		parse_options_usage(stat_usage, stat_options, "I", 1);
2536 		goto out;
2537 	}
2538 
2539 	if (timeout && timeout < 100) {
2540 		if (timeout < 10) {
2541 			pr_err("timeout must be >= 10ms.\n");
2542 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2543 			goto out;
2544 		} else
2545 			pr_warning("timeout < 100ms. "
2546 				   "The overhead percentage could be high in some cases. "
2547 				   "Please proceed with caution.\n");
2548 	}
2549 	if (timeout && interval) {
2550 		pr_err("timeout option is not supported with interval-print.\n");
2551 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2552 		parse_options_usage(stat_usage, stat_options, "I", 1);
2553 		goto out;
2554 	}
2555 
2556 	if (evlist__alloc_stats(evsel_list, interval))
2557 		goto out;
2558 
2559 	if (perf_stat_init_aggr_mode())
2560 		goto out;
2561 
2562 	/*
2563 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2564 	 * while avoiding that older tools show confusing messages.
2565 	 *
2566 	 * However for pipe sessions we need to keep it zero,
2567 	 * because script's perf_evsel__check_attr is triggered
2568 	 * by attr->sample_type != 0, and we can't run it on
2569 	 * stat sessions.
2570 	 */
2571 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2572 
2573 	/*
2574 	 * We dont want to block the signals - that would cause
2575 	 * child tasks to inherit that and Ctrl-C would not work.
2576 	 * What we want is for Ctrl-C to work in the exec()-ed
2577 	 * task, but being ignored by perf stat itself:
2578 	 */
2579 	atexit(sig_atexit);
2580 	if (!forever)
2581 		signal(SIGINT,  skip_signal);
2582 	signal(SIGCHLD, skip_signal);
2583 	signal(SIGALRM, skip_signal);
2584 	signal(SIGABRT, skip_signal);
2585 
2586 	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2587 		goto out;
2588 
2589 	/* Enable ignoring missing threads when -p option is defined. */
2590 	evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2591 	status = 0;
2592 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2593 		if (stat_config.run_count != 1 && verbose > 0)
2594 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2595 				run_idx + 1);
2596 
2597 		if (run_idx != 0)
2598 			evlist__reset_prev_raw_counts(evsel_list);
2599 
2600 		status = run_perf_stat(argc, argv, run_idx);
2601 		if (forever && status != -1 && !interval) {
2602 			print_counters(NULL, argc, argv);
2603 			perf_stat__reset_stats();
2604 		}
2605 	}
2606 
2607 	if (!forever && status != -1 && (!interval || stat_config.summary))
2608 		print_counters(NULL, argc, argv);
2609 
2610 	evlist__finalize_ctlfd(evsel_list);
2611 
2612 	if (STAT_RECORD) {
2613 		/*
2614 		 * We synthesize the kernel mmap record just so that older tools
2615 		 * don't emit warnings about not being able to resolve symbols
2616 		 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2617 		 * a saner message about no samples being in the perf.data file.
2618 		 *
2619 		 * This also serves to suppress a warning about f_header.data.size == 0
2620 		 * in header.c at the moment 'perf stat record' gets introduced, which
2621 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2622 		 * records, but the need to suppress the kptr_restrict messages in older
2623 		 * tools remain  -acme
2624 		 */
2625 		int fd = perf_data__fd(&perf_stat.data);
2626 
2627 		err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2628 							 process_synthesized_event,
2629 							 &perf_stat.session->machines.host);
2630 		if (err) {
2631 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2632 				   "older tools may produce warnings about this file\n.");
2633 		}
2634 
2635 		if (!interval) {
2636 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2637 				pr_err("failed to write stat round event\n");
2638 		}
2639 
2640 		if (!perf_stat.data.is_pipe) {
2641 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2642 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2643 		}
2644 
2645 		evlist__close(evsel_list);
2646 		perf_session__delete(perf_stat.session);
2647 	}
2648 
2649 	perf_stat__exit_aggr_mode();
2650 	evlist__free_stats(evsel_list);
2651 out:
2652 	if (stat_config.iostat_run)
2653 		iostat_release(evsel_list);
2654 
2655 	zfree(&stat_config.walltime_run);
2656 
2657 	if (smi_cost && smi_reset)
2658 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2659 
2660 	evlist__delete(evsel_list);
2661 
2662 	metricgroup__rblist_exit(&stat_config.metric_events);
2663 	runtime_stat_delete(&stat_config);
2664 	evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2665 
2666 	return status;
2667 }
2668