xref: /openbmc/linux/tools/perf/builtin-stat.c (revision dfc66bef)
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 	.big_num		= true,
220 	.ctl_fd			= -1,
221 	.ctl_fd_ack		= -1,
222 	.iostat_run		= false,
223 };
224 
225 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
226 {
227 	if (!a->core.cpus && !b->core.cpus)
228 		return true;
229 
230 	if (!a->core.cpus || !b->core.cpus)
231 		return false;
232 
233 	if (a->core.cpus->nr != b->core.cpus->nr)
234 		return false;
235 
236 	for (int i = 0; i < a->core.cpus->nr; i++) {
237 		if (a->core.cpus->map[i] != b->core.cpus->map[i])
238 			return false;
239 	}
240 
241 	return true;
242 }
243 
244 static void evlist__check_cpu_maps(struct evlist *evlist)
245 {
246 	struct evsel *evsel, *pos, *leader;
247 	char buf[1024];
248 
249 	if (evlist__has_hybrid(evlist))
250 		evlist__warn_hybrid_group(evlist);
251 
252 	evlist__for_each_entry(evlist, evsel) {
253 		leader = evsel__leader(evsel);
254 
255 		/* Check that leader matches cpus with each member. */
256 		if (leader == evsel)
257 			continue;
258 		if (cpus_map_matched(leader, evsel))
259 			continue;
260 
261 		/* If there's mismatch disable the group and warn user. */
262 		WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
263 		evsel__group_desc(leader, buf, sizeof(buf));
264 		pr_warning("  %s\n", buf);
265 
266 		if (verbose) {
267 			cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
268 			pr_warning("     %s: %s\n", leader->name, buf);
269 			cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
270 			pr_warning("     %s: %s\n", evsel->name, buf);
271 		}
272 
273 		for_each_group_evsel(pos, leader) {
274 			evsel__set_leader(pos, pos);
275 			pos->core.nr_members = 0;
276 		}
277 		evsel->core.leader->nr_members = 0;
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, u32 cpu, u32 thread,
331 				   struct perf_counts_values *count)
332 {
333 	struct perf_sample_id *sid = SID(counter, cpu, thread);
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,
340 			       int thread, struct timespec *rs)
341 {
342 	if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
343 		u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
344 		struct perf_counts_values *count =
345 			perf_counts(counter->counts, cpu, thread);
346 		count->ena = count->run = val;
347 		count->val = val;
348 		return 0;
349 	}
350 	return evsel__read_counter(counter, cpu, thread);
351 }
352 
353 /*
354  * Read out the results of a single counter:
355  * do not aggregate counts across CPUs in system-wide mode
356  */
357 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu)
358 {
359 	int nthreads = perf_thread_map__nr(evsel_list->core.threads);
360 	int thread;
361 
362 	if (!counter->supported)
363 		return -ENOENT;
364 
365 	if (counter->core.system_wide)
366 		nthreads = 1;
367 
368 	for (thread = 0; thread < nthreads; thread++) {
369 		struct perf_counts_values *count;
370 
371 		count = perf_counts(counter->counts, cpu, thread);
372 
373 		/*
374 		 * The leader's group read loads data into its group members
375 		 * (via evsel__read_counter()) and sets their count->loaded.
376 		 */
377 		if (!perf_counts__is_loaded(counter->counts, cpu, thread) &&
378 		    read_single_counter(counter, cpu, thread, rs)) {
379 			counter->counts->scaled = -1;
380 			perf_counts(counter->counts, cpu, thread)->ena = 0;
381 			perf_counts(counter->counts, cpu, thread)->run = 0;
382 			return -1;
383 		}
384 
385 		perf_counts__set_loaded(counter->counts, cpu, thread, false);
386 
387 		if (STAT_RECORD) {
388 			if (evsel__write_stat_event(counter, cpu, thread, count)) {
389 				pr_err("failed to write stat event\n");
390 				return -1;
391 			}
392 		}
393 
394 		if (verbose > 1) {
395 			fprintf(stat_config.output,
396 				"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
397 					evsel__name(counter),
398 					cpu,
399 					count->val, count->ena, count->run);
400 		}
401 	}
402 
403 	return 0;
404 }
405 
406 static int read_affinity_counters(struct timespec *rs)
407 {
408 	struct evsel *counter;
409 	struct affinity affinity;
410 	int i, ncpus, cpu;
411 
412 	if (all_counters_use_bpf)
413 		return 0;
414 
415 	if (affinity__setup(&affinity) < 0)
416 		return -1;
417 
418 	ncpus = perf_cpu_map__nr(evsel_list->core.all_cpus);
419 	if (!target__has_cpu(&target) || target__has_per_thread(&target))
420 		ncpus = 1;
421 	evlist__for_each_cpu(evsel_list, i, cpu) {
422 		if (i >= ncpus)
423 			break;
424 		affinity__set(&affinity, cpu);
425 
426 		evlist__for_each_entry(evsel_list, counter) {
427 			if (evsel__cpu_iter_skip(counter, cpu))
428 				continue;
429 			if (evsel__is_bpf(counter))
430 				continue;
431 			if (!counter->err) {
432 				counter->err = read_counter_cpu(counter, rs,
433 								counter->cpu_iter - 1);
434 			}
435 		}
436 	}
437 	affinity__cleanup(&affinity);
438 	return 0;
439 }
440 
441 static int read_bpf_map_counters(void)
442 {
443 	struct evsel *counter;
444 	int err;
445 
446 	evlist__for_each_entry(evsel_list, counter) {
447 		if (!evsel__is_bpf(counter))
448 			continue;
449 
450 		err = bpf_counter__read(counter);
451 		if (err)
452 			return err;
453 	}
454 	return 0;
455 }
456 
457 static void read_counters(struct timespec *rs)
458 {
459 	struct evsel *counter;
460 
461 	if (!stat_config.stop_read_counter) {
462 		if (read_bpf_map_counters() ||
463 		    read_affinity_counters(rs))
464 			return;
465 	}
466 
467 	evlist__for_each_entry(evsel_list, counter) {
468 		if (counter->err)
469 			pr_debug("failed to read counter %s\n", counter->name);
470 		if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
471 			pr_warning("failed to process counter %s\n", counter->name);
472 		counter->err = 0;
473 	}
474 }
475 
476 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
477 {
478 	int i;
479 
480 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
481 	if (!config->stats)
482 		return -1;
483 
484 	config->stats_num = nthreads;
485 
486 	for (i = 0; i < nthreads; i++)
487 		runtime_stat__init(&config->stats[i]);
488 
489 	return 0;
490 }
491 
492 static void runtime_stat_delete(struct perf_stat_config *config)
493 {
494 	int i;
495 
496 	if (!config->stats)
497 		return;
498 
499 	for (i = 0; i < config->stats_num; i++)
500 		runtime_stat__exit(&config->stats[i]);
501 
502 	zfree(&config->stats);
503 }
504 
505 static void runtime_stat_reset(struct perf_stat_config *config)
506 {
507 	int i;
508 
509 	if (!config->stats)
510 		return;
511 
512 	for (i = 0; i < config->stats_num; i++)
513 		perf_stat__reset_shadow_per_stat(&config->stats[i]);
514 }
515 
516 static void process_interval(void)
517 {
518 	struct timespec ts, rs;
519 
520 	clock_gettime(CLOCK_MONOTONIC, &ts);
521 	diff_timespec(&rs, &ts, &ref_time);
522 
523 	perf_stat__reset_shadow_per_stat(&rt_stat);
524 	runtime_stat_reset(&stat_config);
525 	read_counters(&rs);
526 
527 	if (STAT_RECORD) {
528 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
529 			pr_err("failed to write stat round event\n");
530 	}
531 
532 	init_stats(&walltime_nsecs_stats);
533 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
534 	print_counters(&rs, 0, NULL);
535 }
536 
537 static bool handle_interval(unsigned int interval, int *times)
538 {
539 	if (interval) {
540 		process_interval();
541 		if (interval_count && !(--(*times)))
542 			return true;
543 	}
544 	return false;
545 }
546 
547 static int enable_counters(void)
548 {
549 	struct evsel *evsel;
550 	int err;
551 
552 	evlist__for_each_entry(evsel_list, evsel) {
553 		if (!evsel__is_bpf(evsel))
554 			continue;
555 
556 		err = bpf_counter__enable(evsel);
557 		if (err)
558 			return err;
559 	}
560 
561 	if (stat_config.initial_delay < 0) {
562 		pr_info(EVLIST_DISABLED_MSG);
563 		return 0;
564 	}
565 
566 	if (stat_config.initial_delay > 0) {
567 		pr_info(EVLIST_DISABLED_MSG);
568 		usleep(stat_config.initial_delay * USEC_PER_MSEC);
569 	}
570 
571 	/*
572 	 * We need to enable counters only if:
573 	 * - we don't have tracee (attaching to task or cpu)
574 	 * - we have initial delay configured
575 	 */
576 	if (!target__none(&target) || stat_config.initial_delay) {
577 		if (!all_counters_use_bpf)
578 			evlist__enable(evsel_list);
579 		if (stat_config.initial_delay > 0)
580 			pr_info(EVLIST_ENABLED_MSG);
581 	}
582 	return 0;
583 }
584 
585 static void disable_counters(void)
586 {
587 	struct evsel *counter;
588 
589 	/*
590 	 * If we don't have tracee (attaching to task or cpu), counters may
591 	 * still be running. To get accurate group ratios, we must stop groups
592 	 * from counting before reading their constituent counters.
593 	 */
594 	if (!target__none(&target)) {
595 		evlist__for_each_entry(evsel_list, counter)
596 			bpf_counter__disable(counter);
597 		if (!all_counters_use_bpf)
598 			evlist__disable(evsel_list);
599 	}
600 }
601 
602 static volatile int workload_exec_errno;
603 
604 /*
605  * evlist__prepare_workload will send a SIGUSR1
606  * if the fork fails, since we asked by setting its
607  * want_signal to true.
608  */
609 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
610 					void *ucontext __maybe_unused)
611 {
612 	workload_exec_errno = info->si_value.sival_int;
613 }
614 
615 static bool evsel__should_store_id(struct evsel *counter)
616 {
617 	return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
618 }
619 
620 static bool is_target_alive(struct target *_target,
621 			    struct perf_thread_map *threads)
622 {
623 	struct stat st;
624 	int i;
625 
626 	if (!target__has_task(_target))
627 		return true;
628 
629 	for (i = 0; i < threads->nr; i++) {
630 		char path[PATH_MAX];
631 
632 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
633 			  threads->map[i].pid);
634 
635 		if (!stat(path, &st))
636 			return true;
637 	}
638 
639 	return false;
640 }
641 
642 static void process_evlist(struct evlist *evlist, unsigned int interval)
643 {
644 	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
645 
646 	if (evlist__ctlfd_process(evlist, &cmd) > 0) {
647 		switch (cmd) {
648 		case EVLIST_CTL_CMD_ENABLE:
649 			if (interval)
650 				process_interval();
651 			break;
652 		case EVLIST_CTL_CMD_DISABLE:
653 			if (interval)
654 				process_interval();
655 			break;
656 		case EVLIST_CTL_CMD_SNAPSHOT:
657 		case EVLIST_CTL_CMD_ACK:
658 		case EVLIST_CTL_CMD_UNSUPPORTED:
659 		case EVLIST_CTL_CMD_EVLIST:
660 		case EVLIST_CTL_CMD_STOP:
661 		case EVLIST_CTL_CMD_PING:
662 		default:
663 			break;
664 		}
665 	}
666 }
667 
668 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
669 			int *time_to_sleep)
670 {
671 	int tts = *time_to_sleep;
672 	struct timespec time_diff;
673 
674 	diff_timespec(&time_diff, time_stop, time_start);
675 
676 	tts -= time_diff.tv_sec * MSEC_PER_SEC +
677 	       time_diff.tv_nsec / NSEC_PER_MSEC;
678 
679 	if (tts < 0)
680 		tts = 0;
681 
682 	*time_to_sleep = tts;
683 }
684 
685 static int dispatch_events(bool forks, int timeout, int interval, int *times)
686 {
687 	int child_exited = 0, status = 0;
688 	int time_to_sleep, sleep_time;
689 	struct timespec time_start, time_stop;
690 
691 	if (interval)
692 		sleep_time = interval;
693 	else if (timeout)
694 		sleep_time = timeout;
695 	else
696 		sleep_time = 1000;
697 
698 	time_to_sleep = sleep_time;
699 
700 	while (!done) {
701 		if (forks)
702 			child_exited = waitpid(child_pid, &status, WNOHANG);
703 		else
704 			child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
705 
706 		if (child_exited)
707 			break;
708 
709 		clock_gettime(CLOCK_MONOTONIC, &time_start);
710 		if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
711 			if (timeout || handle_interval(interval, times))
712 				break;
713 			time_to_sleep = sleep_time;
714 		} else { /* fd revent */
715 			process_evlist(evsel_list, interval);
716 			clock_gettime(CLOCK_MONOTONIC, &time_stop);
717 			compute_tts(&time_start, &time_stop, &time_to_sleep);
718 		}
719 	}
720 
721 	return status;
722 }
723 
724 enum counter_recovery {
725 	COUNTER_SKIP,
726 	COUNTER_RETRY,
727 	COUNTER_FATAL,
728 };
729 
730 static enum counter_recovery stat_handle_error(struct evsel *counter)
731 {
732 	char msg[BUFSIZ];
733 	/*
734 	 * PPC returns ENXIO for HW counters until 2.6.37
735 	 * (behavior changed with commit b0a873e).
736 	 */
737 	if (errno == EINVAL || errno == ENOSYS ||
738 	    errno == ENOENT || errno == EOPNOTSUPP ||
739 	    errno == ENXIO) {
740 		if (verbose > 0)
741 			ui__warning("%s event is not supported by the kernel.\n",
742 				    evsel__name(counter));
743 		counter->supported = false;
744 		/*
745 		 * errored is a sticky flag that means one of the counter's
746 		 * cpu event had a problem and needs to be reexamined.
747 		 */
748 		counter->errored = true;
749 
750 		if ((evsel__leader(counter) != counter) ||
751 		    !(counter->core.leader->nr_members > 1))
752 			return COUNTER_SKIP;
753 	} else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
754 		if (verbose > 0)
755 			ui__warning("%s\n", msg);
756 		return COUNTER_RETRY;
757 	} else if (target__has_per_thread(&target) &&
758 		   evsel_list->core.threads &&
759 		   evsel_list->core.threads->err_thread != -1) {
760 		/*
761 		 * For global --per-thread case, skip current
762 		 * error thread.
763 		 */
764 		if (!thread_map__remove(evsel_list->core.threads,
765 					evsel_list->core.threads->err_thread)) {
766 			evsel_list->core.threads->err_thread = -1;
767 			return COUNTER_RETRY;
768 		}
769 	}
770 
771 	evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
772 	ui__error("%s\n", msg);
773 
774 	if (child_pid != -1)
775 		kill(child_pid, SIGTERM);
776 	return COUNTER_FATAL;
777 }
778 
779 static int __run_perf_stat(int argc, const char **argv, int run_idx)
780 {
781 	int interval = stat_config.interval;
782 	int times = stat_config.times;
783 	int timeout = stat_config.timeout;
784 	char msg[BUFSIZ];
785 	unsigned long long t0, t1;
786 	struct evsel *counter;
787 	size_t l;
788 	int status = 0;
789 	const bool forks = (argc > 0);
790 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
791 	struct affinity affinity;
792 	int i, cpu, err;
793 	bool second_pass = false;
794 
795 	if (forks) {
796 		if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
797 			perror("failed to prepare workload");
798 			return -1;
799 		}
800 		child_pid = evsel_list->workload.pid;
801 	}
802 
803 	if (group)
804 		evlist__set_leader(evsel_list);
805 
806 	if (affinity__setup(&affinity) < 0)
807 		return -1;
808 
809 	evlist__for_each_entry(evsel_list, counter) {
810 		if (bpf_counter__load(counter, &target))
811 			return -1;
812 		if (!evsel__is_bpf(counter))
813 			all_counters_use_bpf = false;
814 	}
815 
816 	evlist__for_each_cpu (evsel_list, i, cpu) {
817 		/*
818 		 * bperf calls evsel__open_per_cpu() in bperf__load(), so
819 		 * no need to call it again here.
820 		 */
821 		if (target.use_bpf)
822 			break;
823 		affinity__set(&affinity, cpu);
824 
825 		evlist__for_each_entry(evsel_list, counter) {
826 			if (evsel__cpu_iter_skip(counter, cpu))
827 				continue;
828 			if (counter->reset_group || counter->errored)
829 				continue;
830 			if (evsel__is_bpf(counter))
831 				continue;
832 try_again:
833 			if (create_perf_stat_counter(counter, &stat_config, &target,
834 						     counter->cpu_iter - 1) < 0) {
835 
836 				/*
837 				 * Weak group failed. We cannot just undo this here
838 				 * because earlier CPUs might be in group mode, and the kernel
839 				 * doesn't support mixing group and non group reads. Defer
840 				 * it to later.
841 				 * Don't close here because we're in the wrong affinity.
842 				 */
843 				if ((errno == EINVAL || errno == EBADF) &&
844 				    evsel__leader(counter) != counter &&
845 				    counter->weak_group) {
846 					evlist__reset_weak_group(evsel_list, counter, false);
847 					assert(counter->reset_group);
848 					second_pass = true;
849 					continue;
850 				}
851 
852 				switch (stat_handle_error(counter)) {
853 				case COUNTER_FATAL:
854 					return -1;
855 				case COUNTER_RETRY:
856 					goto try_again;
857 				case COUNTER_SKIP:
858 					continue;
859 				default:
860 					break;
861 				}
862 
863 			}
864 			counter->supported = true;
865 		}
866 	}
867 
868 	if (second_pass) {
869 		/*
870 		 * Now redo all the weak group after closing them,
871 		 * and also close errored counters.
872 		 */
873 
874 		evlist__for_each_cpu(evsel_list, i, cpu) {
875 			affinity__set(&affinity, cpu);
876 			/* First close errored or weak retry */
877 			evlist__for_each_entry(evsel_list, counter) {
878 				if (!counter->reset_group && !counter->errored)
879 					continue;
880 				if (evsel__cpu_iter_skip_no_inc(counter, cpu))
881 					continue;
882 				perf_evsel__close_cpu(&counter->core, counter->cpu_iter);
883 			}
884 			/* Now reopen weak */
885 			evlist__for_each_entry(evsel_list, counter) {
886 				if (!counter->reset_group && !counter->errored)
887 					continue;
888 				if (evsel__cpu_iter_skip(counter, cpu))
889 					continue;
890 				if (!counter->reset_group)
891 					continue;
892 try_again_reset:
893 				pr_debug2("reopening weak %s\n", evsel__name(counter));
894 				if (create_perf_stat_counter(counter, &stat_config, &target,
895 							     counter->cpu_iter - 1) < 0) {
896 
897 					switch (stat_handle_error(counter)) {
898 					case COUNTER_FATAL:
899 						return -1;
900 					case COUNTER_RETRY:
901 						goto try_again_reset;
902 					case COUNTER_SKIP:
903 						continue;
904 					default:
905 						break;
906 					}
907 				}
908 				counter->supported = true;
909 			}
910 		}
911 	}
912 	affinity__cleanup(&affinity);
913 
914 	evlist__for_each_entry(evsel_list, counter) {
915 		if (!counter->supported) {
916 			perf_evsel__free_fd(&counter->core);
917 			continue;
918 		}
919 
920 		l = strlen(counter->unit);
921 		if (l > stat_config.unit_width)
922 			stat_config.unit_width = l;
923 
924 		if (evsel__should_store_id(counter) &&
925 		    evsel__store_ids(counter, evsel_list))
926 			return -1;
927 	}
928 
929 	if (evlist__apply_filters(evsel_list, &counter)) {
930 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
931 			counter->filter, evsel__name(counter), errno,
932 			str_error_r(errno, msg, sizeof(msg)));
933 		return -1;
934 	}
935 
936 	if (STAT_RECORD) {
937 		int fd = perf_data__fd(&perf_stat.data);
938 
939 		if (is_pipe) {
940 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
941 		} else {
942 			err = perf_session__write_header(perf_stat.session, evsel_list,
943 							 fd, false);
944 		}
945 
946 		if (err < 0)
947 			return err;
948 
949 		err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
950 							 process_synthesized_event, is_pipe);
951 		if (err < 0)
952 			return err;
953 	}
954 
955 	/*
956 	 * Enable counters and exec the command:
957 	 */
958 	if (forks) {
959 		evlist__start_workload(evsel_list);
960 		err = enable_counters();
961 		if (err)
962 			return -1;
963 
964 		t0 = rdclock();
965 		clock_gettime(CLOCK_MONOTONIC, &ref_time);
966 
967 		if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
968 			status = dispatch_events(forks, timeout, interval, &times);
969 		if (child_pid != -1) {
970 			if (timeout)
971 				kill(child_pid, SIGTERM);
972 			wait4(child_pid, &status, 0, &stat_config.ru_data);
973 		}
974 
975 		if (workload_exec_errno) {
976 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
977 			pr_err("Workload failed: %s\n", emsg);
978 			return -1;
979 		}
980 
981 		if (WIFSIGNALED(status))
982 			psignal(WTERMSIG(status), argv[0]);
983 	} else {
984 		err = enable_counters();
985 		if (err)
986 			return -1;
987 
988 		t0 = rdclock();
989 		clock_gettime(CLOCK_MONOTONIC, &ref_time);
990 
991 		status = dispatch_events(forks, timeout, interval, &times);
992 	}
993 
994 	disable_counters();
995 
996 	t1 = rdclock();
997 
998 	if (stat_config.walltime_run_table)
999 		stat_config.walltime_run[run_idx] = t1 - t0;
1000 
1001 	if (interval && stat_config.summary) {
1002 		stat_config.interval = 0;
1003 		stat_config.stop_read_counter = true;
1004 		init_stats(&walltime_nsecs_stats);
1005 		update_stats(&walltime_nsecs_stats, t1 - t0);
1006 
1007 		if (stat_config.aggr_mode == AGGR_GLOBAL)
1008 			evlist__save_aggr_prev_raw_counts(evsel_list);
1009 
1010 		evlist__copy_prev_raw_counts(evsel_list);
1011 		evlist__reset_prev_raw_counts(evsel_list);
1012 		runtime_stat_reset(&stat_config);
1013 		perf_stat__reset_shadow_per_stat(&rt_stat);
1014 	} else
1015 		update_stats(&walltime_nsecs_stats, t1 - t0);
1016 
1017 	/*
1018 	 * Closing a group leader splits the group, and as we only disable
1019 	 * group leaders, results in remaining events becoming enabled. To
1020 	 * avoid arbitrary skew, we must read all counters before closing any
1021 	 * group leaders.
1022 	 */
1023 	read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
1024 
1025 	/*
1026 	 * We need to keep evsel_list alive, because it's processed
1027 	 * later the evsel_list will be closed after.
1028 	 */
1029 	if (!STAT_RECORD)
1030 		evlist__close(evsel_list);
1031 
1032 	return WEXITSTATUS(status);
1033 }
1034 
1035 static int run_perf_stat(int argc, const char **argv, int run_idx)
1036 {
1037 	int ret;
1038 
1039 	if (pre_cmd) {
1040 		ret = system(pre_cmd);
1041 		if (ret)
1042 			return ret;
1043 	}
1044 
1045 	if (sync_run)
1046 		sync();
1047 
1048 	ret = __run_perf_stat(argc, argv, run_idx);
1049 	if (ret)
1050 		return ret;
1051 
1052 	if (post_cmd) {
1053 		ret = system(post_cmd);
1054 		if (ret)
1055 			return ret;
1056 	}
1057 
1058 	return ret;
1059 }
1060 
1061 static void print_counters(struct timespec *ts, int argc, const char **argv)
1062 {
1063 	/* Do not print anything if we record to the pipe. */
1064 	if (STAT_RECORD && perf_stat.data.is_pipe)
1065 		return;
1066 	if (stat_config.quiet)
1067 		return;
1068 
1069 	evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1070 }
1071 
1072 static volatile int signr = -1;
1073 
1074 static void skip_signal(int signo)
1075 {
1076 	if ((child_pid == -1) || stat_config.interval)
1077 		done = 1;
1078 
1079 	signr = signo;
1080 	/*
1081 	 * render child_pid harmless
1082 	 * won't send SIGTERM to a random
1083 	 * process in case of race condition
1084 	 * and fast PID recycling
1085 	 */
1086 	child_pid = -1;
1087 }
1088 
1089 static void sig_atexit(void)
1090 {
1091 	sigset_t set, oset;
1092 
1093 	/*
1094 	 * avoid race condition with SIGCHLD handler
1095 	 * in skip_signal() which is modifying child_pid
1096 	 * goal is to avoid send SIGTERM to a random
1097 	 * process
1098 	 */
1099 	sigemptyset(&set);
1100 	sigaddset(&set, SIGCHLD);
1101 	sigprocmask(SIG_BLOCK, &set, &oset);
1102 
1103 	if (child_pid != -1)
1104 		kill(child_pid, SIGTERM);
1105 
1106 	sigprocmask(SIG_SETMASK, &oset, NULL);
1107 
1108 	if (signr == -1)
1109 		return;
1110 
1111 	signal(signr, SIG_DFL);
1112 	kill(getpid(), signr);
1113 }
1114 
1115 void perf_stat__set_big_num(int set)
1116 {
1117 	stat_config.big_num = (set != 0);
1118 }
1119 
1120 void perf_stat__set_no_csv_summary(int set)
1121 {
1122 	stat_config.no_csv_summary = (set != 0);
1123 }
1124 
1125 static int stat__set_big_num(const struct option *opt __maybe_unused,
1126 			     const char *s __maybe_unused, int unset)
1127 {
1128 	big_num_opt = unset ? 0 : 1;
1129 	perf_stat__set_big_num(!unset);
1130 	return 0;
1131 }
1132 
1133 static int enable_metric_only(const struct option *opt __maybe_unused,
1134 			      const char *s __maybe_unused, int unset)
1135 {
1136 	force_metric_only = true;
1137 	stat_config.metric_only = !unset;
1138 	return 0;
1139 }
1140 
1141 static int parse_metric_groups(const struct option *opt,
1142 			       const char *str,
1143 			       int unset __maybe_unused)
1144 {
1145 	return metricgroup__parse_groups(opt, str,
1146 					 stat_config.metric_no_group,
1147 					 stat_config.metric_no_merge,
1148 					 &stat_config.metric_events);
1149 }
1150 
1151 static int parse_control_option(const struct option *opt,
1152 				const char *str,
1153 				int unset __maybe_unused)
1154 {
1155 	struct perf_stat_config *config = opt->value;
1156 
1157 	return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1158 }
1159 
1160 static int parse_stat_cgroups(const struct option *opt,
1161 			      const char *str, int unset)
1162 {
1163 	if (stat_config.cgroup_list) {
1164 		pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1165 		return -1;
1166 	}
1167 
1168 	return parse_cgroups(opt, str, unset);
1169 }
1170 
1171 static int parse_hybrid_type(const struct option *opt,
1172 			     const char *str,
1173 			     int unset __maybe_unused)
1174 {
1175 	struct evlist *evlist = *(struct evlist **)opt->value;
1176 
1177 	if (!list_empty(&evlist->core.entries)) {
1178 		fprintf(stderr, "Must define cputype before events/metrics\n");
1179 		return -1;
1180 	}
1181 
1182 	evlist->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu(str);
1183 	if (!evlist->hybrid_pmu_name) {
1184 		fprintf(stderr, "--cputype %s is not supported!\n", str);
1185 		return -1;
1186 	}
1187 
1188 	return 0;
1189 }
1190 
1191 static struct option stat_options[] = {
1192 	OPT_BOOLEAN('T', "transaction", &transaction_run,
1193 		    "hardware transaction statistics"),
1194 	OPT_CALLBACK('e', "event", &evsel_list, "event",
1195 		     "event selector. use 'perf list' to list available events",
1196 		     parse_events_option),
1197 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1198 		     "event filter", parse_filter),
1199 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1200 		    "child tasks do not inherit counters"),
1201 	OPT_STRING('p', "pid", &target.pid, "pid",
1202 		   "stat events on existing process id"),
1203 	OPT_STRING('t', "tid", &target.tid, "tid",
1204 		   "stat events on existing thread id"),
1205 #ifdef HAVE_BPF_SKEL
1206 	OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1207 		   "stat events on existing bpf program id"),
1208 	OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1209 		    "use bpf program to count events"),
1210 	OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1211 		   "path to perf_event_attr map"),
1212 #endif
1213 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1214 		    "system-wide collection from all CPUs"),
1215 	OPT_BOOLEAN('g', "group", &group,
1216 		    "put the counters into a counter group"),
1217 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
1218 		    "Use --no-scale to disable counter scaling for multiplexing"),
1219 	OPT_INCR('v', "verbose", &verbose,
1220 		    "be more verbose (show counter open errors, etc)"),
1221 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
1222 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1223 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1224 		    "display details about each run (only with -r option)"),
1225 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
1226 		    "null run - dont start any counters"),
1227 	OPT_INCR('d', "detailed", &detailed_run,
1228 		    "detailed run - start a lot of events"),
1229 	OPT_BOOLEAN('S', "sync", &sync_run,
1230 		    "call sync() before starting a run"),
1231 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1232 			   "print large numbers with thousands\' separators",
1233 			   stat__set_big_num),
1234 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1235 		    "list of cpus to monitor in system-wide"),
1236 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1237 		    "disable CPU count aggregation", AGGR_NONE),
1238 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1239 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1240 		   "print counts with custom separator"),
1241 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1242 		     "monitor event in cgroup name only", parse_stat_cgroups),
1243 	OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1244 		    "expand events for each cgroup"),
1245 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1246 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1247 	OPT_INTEGER(0, "log-fd", &output_fd,
1248 		    "log output to fd, instead of stderr"),
1249 	OPT_STRING(0, "pre", &pre_cmd, "command",
1250 			"command to run prior to the measured command"),
1251 	OPT_STRING(0, "post", &post_cmd, "command",
1252 			"command to run after to the measured command"),
1253 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1254 		    "print counts at regular interval in ms "
1255 		    "(overhead is possible for values <= 100ms)"),
1256 	OPT_INTEGER(0, "interval-count", &stat_config.times,
1257 		    "print counts for fixed number of times"),
1258 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1259 		    "clear screen in between new interval"),
1260 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1261 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1262 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1263 		     "aggregate counts per processor socket", AGGR_SOCKET),
1264 	OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1265 		     "aggregate counts per processor die", AGGR_DIE),
1266 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1267 		     "aggregate counts per physical processor core", AGGR_CORE),
1268 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1269 		     "aggregate counts per thread", AGGR_THREAD),
1270 	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1271 		     "aggregate counts per numa node", AGGR_NODE),
1272 	OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1273 		    "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1274 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1275 			"Only print computed metrics. No raw values", enable_metric_only),
1276 	OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1277 		       "don't group metric events, impacts multiplexing"),
1278 	OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1279 		       "don't try to share events between metrics in a group"),
1280 	OPT_BOOLEAN(0, "topdown", &topdown_run,
1281 			"measure top-down statistics"),
1282 	OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1283 			"Set the metrics level for the top-down statistics (0: max level)"),
1284 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1285 			"measure SMI cost"),
1286 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1287 		     "monitor specified metrics or metric groups (separated by ,)",
1288 		     parse_metric_groups),
1289 	OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1290 			 "Configure all used events to run in kernel space.",
1291 			 PARSE_OPT_EXCLUSIVE),
1292 	OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1293 			 "Configure all used events to run in user space.",
1294 			 PARSE_OPT_EXCLUSIVE),
1295 	OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1296 		    "Use with 'percore' event qualifier to show the event "
1297 		    "counts of one hardware thread by sum up total hardware "
1298 		    "threads of same physical core"),
1299 	OPT_BOOLEAN(0, "summary", &stat_config.summary,
1300 		       "print summary for interval mode"),
1301 	OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1302 		       "don't print 'summary' for CSV summary output"),
1303 	OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1304 			"don't print output (useful with record)"),
1305 	OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1306 		     "Only enable events on applying cpu with this type "
1307 		     "for hybrid platform (e.g. core or atom)",
1308 		     parse_hybrid_type),
1309 #ifdef HAVE_LIBPFM
1310 	OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1311 		"libpfm4 event selector. use 'perf list' to list available events",
1312 		parse_libpfm_events_option),
1313 #endif
1314 	OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1315 		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1316 		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1317 		     "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1318 		      parse_control_option),
1319 	OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1320 			    "measure I/O performance metrics provided by arch/platform",
1321 			    iostat_parse),
1322 	OPT_END()
1323 };
1324 
1325 static const char *const aggr_mode__string[] = {
1326 	[AGGR_CORE] = "core",
1327 	[AGGR_DIE] = "die",
1328 	[AGGR_GLOBAL] = "global",
1329 	[AGGR_NODE] = "node",
1330 	[AGGR_NONE] = "none",
1331 	[AGGR_SOCKET] = "socket",
1332 	[AGGR_THREAD] = "thread",
1333 	[AGGR_UNSET] = "unset",
1334 };
1335 
1336 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1337 						int cpu)
1338 {
1339 	return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1340 }
1341 
1342 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1343 					     int cpu)
1344 {
1345 	return aggr_cpu_id__die(cpu, /*data=*/NULL);
1346 }
1347 
1348 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1349 					      int cpu)
1350 {
1351 	return aggr_cpu_id__core(cpu, /*data=*/NULL);
1352 }
1353 
1354 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1355 					      int cpu)
1356 {
1357 	return aggr_cpu_id__node(cpu, /*data=*/NULL);
1358 }
1359 
1360 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1361 					      aggr_get_id_t get_id, int cpu)
1362 {
1363 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1364 
1365 	if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu]))
1366 		config->cpus_aggr_map->map[cpu] = get_id(config, cpu);
1367 
1368 	id = config->cpus_aggr_map->map[cpu];
1369 	return id;
1370 }
1371 
1372 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1373 						       int cpu)
1374 {
1375 	return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1376 }
1377 
1378 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1379 						    int cpu)
1380 {
1381 	return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1382 }
1383 
1384 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1385 						     int cpu)
1386 {
1387 	return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1388 }
1389 
1390 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1391 						     int cpu)
1392 {
1393 	return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1394 }
1395 
1396 static bool term_percore_set(void)
1397 {
1398 	struct evsel *counter;
1399 
1400 	evlist__for_each_entry(evsel_list, counter) {
1401 		if (counter->percore)
1402 			return true;
1403 	}
1404 
1405 	return false;
1406 }
1407 
1408 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1409 {
1410 	switch (aggr_mode) {
1411 	case AGGR_SOCKET:
1412 		return aggr_cpu_id__socket;
1413 	case AGGR_DIE:
1414 		return aggr_cpu_id__die;
1415 	case AGGR_CORE:
1416 		return aggr_cpu_id__core;
1417 	case AGGR_NODE:
1418 		return aggr_cpu_id__node;
1419 	case AGGR_NONE:
1420 		if (term_percore_set())
1421 			return aggr_cpu_id__core;
1422 
1423 		return NULL;
1424 	case AGGR_GLOBAL:
1425 	case AGGR_THREAD:
1426 	case AGGR_UNSET:
1427 	default:
1428 		return NULL;
1429 	}
1430 }
1431 
1432 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1433 {
1434 	switch (aggr_mode) {
1435 	case AGGR_SOCKET:
1436 		return perf_stat__get_socket_cached;
1437 	case AGGR_DIE:
1438 		return perf_stat__get_die_cached;
1439 	case AGGR_CORE:
1440 		return perf_stat__get_core_cached;
1441 	case AGGR_NODE:
1442 		return perf_stat__get_node_cached;
1443 	case AGGR_NONE:
1444 		if (term_percore_set()) {
1445 			return perf_stat__get_core_cached;
1446 		}
1447 		return NULL;
1448 	case AGGR_GLOBAL:
1449 	case AGGR_THREAD:
1450 	case AGGR_UNSET:
1451 	default:
1452 		return NULL;
1453 	}
1454 }
1455 
1456 static int perf_stat_init_aggr_mode(void)
1457 {
1458 	int nr;
1459 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1460 
1461 	if (get_id) {
1462 		stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.cpus,
1463 							 get_id, /*data=*/NULL);
1464 		if (!stat_config.aggr_map) {
1465 			pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1466 			return -1;
1467 		}
1468 		stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1469 	}
1470 
1471 	/*
1472 	 * The evsel_list->cpus is the base we operate on,
1473 	 * taking the highest cpu number to be the size of
1474 	 * the aggregation translate cpumap.
1475 	 */
1476 	nr = perf_cpu_map__max(evsel_list->core.cpus);
1477 	stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1478 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1479 }
1480 
1481 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1482 {
1483 	if (map) {
1484 		WARN_ONCE(refcount_read(&map->refcnt) != 0,
1485 			  "cpu_aggr_map refcnt unbalanced\n");
1486 		free(map);
1487 	}
1488 }
1489 
1490 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1491 {
1492 	if (map && refcount_dec_and_test(&map->refcnt))
1493 		cpu_aggr_map__delete(map);
1494 }
1495 
1496 static void perf_stat__exit_aggr_mode(void)
1497 {
1498 	cpu_aggr_map__put(stat_config.aggr_map);
1499 	cpu_aggr_map__put(stat_config.cpus_aggr_map);
1500 	stat_config.aggr_map = NULL;
1501 	stat_config.cpus_aggr_map = NULL;
1502 }
1503 
1504 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(int cpu, void *data)
1505 {
1506 	struct perf_env *env = data;
1507 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1508 
1509 	if (cpu != -1)
1510 		id.socket = env->cpu[cpu].socket_id;
1511 
1512 	return id;
1513 }
1514 
1515 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(int cpu, void *data)
1516 {
1517 	struct perf_env *env = data;
1518 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1519 
1520 	if (cpu != -1) {
1521 		/*
1522 		 * die_id is relative to socket, so start
1523 		 * with the socket ID and then add die to
1524 		 * make a unique ID.
1525 		 */
1526 		id.socket = env->cpu[cpu].socket_id;
1527 		id.die = env->cpu[cpu].die_id;
1528 	}
1529 
1530 	return id;
1531 }
1532 
1533 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(int cpu, void *data)
1534 {
1535 	struct perf_env *env = data;
1536 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1537 
1538 	if (cpu != -1) {
1539 		/*
1540 		 * core_id is relative to socket and die,
1541 		 * we need a global id. So we set
1542 		 * socket, die id and core id
1543 		 */
1544 		id.socket = env->cpu[cpu].socket_id;
1545 		id.die = env->cpu[cpu].die_id;
1546 		id.core = env->cpu[cpu].core_id;
1547 	}
1548 
1549 	return id;
1550 }
1551 
1552 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(int cpu, void *data)
1553 {
1554 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1555 
1556 	id.node = perf_env__numa_node(data, cpu);
1557 	return id;
1558 }
1559 
1560 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1561 						     int cpu)
1562 {
1563 	return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1564 }
1565 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1566 						  int cpu)
1567 {
1568 	return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1569 }
1570 
1571 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1572 						   int cpu)
1573 {
1574 	return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1575 }
1576 
1577 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1578 						   int cpu)
1579 {
1580 	return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1581 }
1582 
1583 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1584 {
1585 	switch (aggr_mode) {
1586 	case AGGR_SOCKET:
1587 		return perf_env__get_socket_aggr_by_cpu;
1588 	case AGGR_DIE:
1589 		return perf_env__get_die_aggr_by_cpu;
1590 	case AGGR_CORE:
1591 		return perf_env__get_core_aggr_by_cpu;
1592 	case AGGR_NODE:
1593 		return perf_env__get_node_aggr_by_cpu;
1594 	case AGGR_NONE:
1595 	case AGGR_GLOBAL:
1596 	case AGGR_THREAD:
1597 	case AGGR_UNSET:
1598 	default:
1599 		return NULL;
1600 	}
1601 }
1602 
1603 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1604 {
1605 	switch (aggr_mode) {
1606 	case AGGR_SOCKET:
1607 		return perf_stat__get_socket_file;
1608 	case AGGR_DIE:
1609 		return perf_stat__get_die_file;
1610 	case AGGR_CORE:
1611 		return perf_stat__get_core_file;
1612 	case AGGR_NODE:
1613 		return perf_stat__get_node_file;
1614 	case AGGR_NONE:
1615 	case AGGR_GLOBAL:
1616 	case AGGR_THREAD:
1617 	case AGGR_UNSET:
1618 	default:
1619 		return NULL;
1620 	}
1621 }
1622 
1623 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1624 {
1625 	struct perf_env *env = &st->session->header.env;
1626 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1627 
1628 	if (!get_id)
1629 		return 0;
1630 
1631 	stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.cpus, get_id, env);
1632 	if (!stat_config.aggr_map) {
1633 		pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1634 		return -1;
1635 	}
1636 	stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1637 	return 0;
1638 }
1639 
1640 /*
1641  * Add default attributes, if there were no attributes specified or
1642  * if -d/--detailed, -d -d or -d -d -d is used:
1643  */
1644 static int add_default_attributes(void)
1645 {
1646 	int err;
1647 	struct perf_event_attr default_attrs0[] = {
1648 
1649   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1650   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1651   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1652   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1653 
1654   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1655 };
1656 	struct perf_event_attr frontend_attrs[] = {
1657   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1658 };
1659 	struct perf_event_attr backend_attrs[] = {
1660   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1661 };
1662 	struct perf_event_attr default_attrs1[] = {
1663   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1664   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1665   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1666 
1667 };
1668 	struct perf_event_attr default_sw_attrs[] = {
1669   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1670   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1671   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1672   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1673 };
1674 
1675 /*
1676  * Detailed stats (-d), covering the L1 and last level data caches:
1677  */
1678 	struct perf_event_attr detailed_attrs[] = {
1679 
1680   { .type = PERF_TYPE_HW_CACHE,
1681     .config =
1682 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1683 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1684 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1685 
1686   { .type = PERF_TYPE_HW_CACHE,
1687     .config =
1688 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1689 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1690 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1691 
1692   { .type = PERF_TYPE_HW_CACHE,
1693     .config =
1694 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1695 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1696 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1697 
1698   { .type = PERF_TYPE_HW_CACHE,
1699     .config =
1700 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1701 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1702 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1703 };
1704 
1705 /*
1706  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1707  */
1708 	struct perf_event_attr very_detailed_attrs[] = {
1709 
1710   { .type = PERF_TYPE_HW_CACHE,
1711     .config =
1712 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1713 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1714 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1715 
1716   { .type = PERF_TYPE_HW_CACHE,
1717     .config =
1718 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1719 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1720 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1721 
1722   { .type = PERF_TYPE_HW_CACHE,
1723     .config =
1724 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1725 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1726 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1727 
1728   { .type = PERF_TYPE_HW_CACHE,
1729     .config =
1730 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1731 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1732 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1733 
1734   { .type = PERF_TYPE_HW_CACHE,
1735     .config =
1736 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1737 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1738 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1739 
1740   { .type = PERF_TYPE_HW_CACHE,
1741     .config =
1742 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1743 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1744 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1745 
1746 };
1747 
1748 /*
1749  * Very, very detailed stats (-d -d -d), adding prefetch events:
1750  */
1751 	struct perf_event_attr very_very_detailed_attrs[] = {
1752 
1753   { .type = PERF_TYPE_HW_CACHE,
1754     .config =
1755 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1756 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1757 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1758 
1759   { .type = PERF_TYPE_HW_CACHE,
1760     .config =
1761 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1762 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1763 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1764 };
1765 	/* Set attrs if no event is selected and !null_run: */
1766 	if (stat_config.null_run)
1767 		return 0;
1768 
1769 	if (transaction_run) {
1770 		struct parse_events_error errinfo;
1771 		/* Handle -T as -M transaction. Once platform specific metrics
1772 		 * support has been added to the json files, all architectures
1773 		 * will use this approach. To determine transaction support
1774 		 * on an architecture test for such a metric name.
1775 		 */
1776 		if (metricgroup__has_metric("transaction")) {
1777 			struct option opt = { .value = &evsel_list };
1778 
1779 			return metricgroup__parse_groups(&opt, "transaction",
1780 							 stat_config.metric_no_group,
1781 							stat_config.metric_no_merge,
1782 							 &stat_config.metric_events);
1783 		}
1784 
1785 		parse_events_error__init(&errinfo);
1786 		if (pmu_have_event("cpu", "cycles-ct") &&
1787 		    pmu_have_event("cpu", "el-start"))
1788 			err = parse_events(evsel_list, transaction_attrs,
1789 					   &errinfo);
1790 		else
1791 			err = parse_events(evsel_list,
1792 					   transaction_limited_attrs,
1793 					   &errinfo);
1794 		if (err) {
1795 			fprintf(stderr, "Cannot set up transaction events\n");
1796 			parse_events_error__print(&errinfo, transaction_attrs);
1797 		}
1798 		parse_events_error__exit(&errinfo);
1799 		return err ? -1 : 0;
1800 	}
1801 
1802 	if (smi_cost) {
1803 		struct parse_events_error errinfo;
1804 		int smi;
1805 
1806 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1807 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1808 			return -1;
1809 		}
1810 
1811 		if (!smi) {
1812 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1813 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1814 				return -1;
1815 			}
1816 			smi_reset = true;
1817 		}
1818 
1819 		if (!pmu_have_event("msr", "aperf") ||
1820 		    !pmu_have_event("msr", "smi")) {
1821 			fprintf(stderr, "To measure SMI cost, it needs "
1822 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1823 			return -1;
1824 		}
1825 		if (!force_metric_only)
1826 			stat_config.metric_only = true;
1827 
1828 		parse_events_error__init(&errinfo);
1829 		err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1830 		if (err) {
1831 			parse_events_error__print(&errinfo, smi_cost_attrs);
1832 			fprintf(stderr, "Cannot set up SMI cost events\n");
1833 		}
1834 		parse_events_error__exit(&errinfo);
1835 		return err ? -1 : 0;
1836 	}
1837 
1838 	if (topdown_run) {
1839 		const char **metric_attrs = topdown_metric_attrs;
1840 		unsigned int max_level = 1;
1841 		char *str = NULL;
1842 		bool warn = false;
1843 
1844 		if (!force_metric_only)
1845 			stat_config.metric_only = true;
1846 
1847 		if (pmu_have_event("cpu", topdown_metric_L2_attrs[5])) {
1848 			metric_attrs = topdown_metric_L2_attrs;
1849 			max_level = 2;
1850 		}
1851 
1852 		if (stat_config.topdown_level > max_level) {
1853 			pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1854 			return -1;
1855 		} else if (!stat_config.topdown_level)
1856 			stat_config.topdown_level = max_level;
1857 
1858 		if (topdown_filter_events(metric_attrs, &str, 1) < 0) {
1859 			pr_err("Out of memory\n");
1860 			return -1;
1861 		}
1862 		if (metric_attrs[0] && str) {
1863 			if (!stat_config.interval && !stat_config.metric_only) {
1864 				fprintf(stat_config.output,
1865 					"Topdown accuracy may decrease when measuring long periods.\n"
1866 					"Please print the result regularly, e.g. -I1000\n");
1867 			}
1868 			goto setup_metrics;
1869 		}
1870 
1871 		zfree(&str);
1872 
1873 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1874 		    stat_config.aggr_mode != AGGR_CORE) {
1875 			pr_err("top down event configuration requires --per-core mode\n");
1876 			return -1;
1877 		}
1878 		stat_config.aggr_mode = AGGR_CORE;
1879 		if (nr_cgroups || !target__has_cpu(&target)) {
1880 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1881 			return -1;
1882 		}
1883 
1884 		if (topdown_filter_events(topdown_attrs, &str,
1885 				arch_topdown_check_group(&warn)) < 0) {
1886 			pr_err("Out of memory\n");
1887 			return -1;
1888 		}
1889 		if (topdown_attrs[0] && str) {
1890 			struct parse_events_error errinfo;
1891 			if (warn)
1892 				arch_topdown_group_warn();
1893 setup_metrics:
1894 			parse_events_error__init(&errinfo);
1895 			err = parse_events(evsel_list, str, &errinfo);
1896 			if (err) {
1897 				fprintf(stderr,
1898 					"Cannot set up top down events %s: %d\n",
1899 					str, err);
1900 				parse_events_error__print(&errinfo, str);
1901 				parse_events_error__exit(&errinfo);
1902 				free(str);
1903 				return -1;
1904 			}
1905 			parse_events_error__exit(&errinfo);
1906 		} else {
1907 			fprintf(stderr, "System does not support topdown\n");
1908 			return -1;
1909 		}
1910 		free(str);
1911 	}
1912 
1913 	if (!evsel_list->core.nr_entries) {
1914 		if (perf_pmu__has_hybrid()) {
1915 			struct parse_events_error errinfo;
1916 			const char *hybrid_str = "cycles,instructions,branches,branch-misses";
1917 
1918 			if (target__has_cpu(&target))
1919 				default_sw_attrs[0].config = PERF_COUNT_SW_CPU_CLOCK;
1920 
1921 			if (evlist__add_default_attrs(evsel_list,
1922 						      default_sw_attrs) < 0) {
1923 				return -1;
1924 			}
1925 
1926 			parse_events_error__init(&errinfo);
1927 			err = parse_events(evsel_list, hybrid_str, &errinfo);
1928 			if (err) {
1929 				fprintf(stderr,
1930 					"Cannot set up hybrid events %s: %d\n",
1931 					hybrid_str, err);
1932 				parse_events_error__print(&errinfo, hybrid_str);
1933 			}
1934 			parse_events_error__exit(&errinfo);
1935 			return err ? -1 : 0;
1936 		}
1937 
1938 		if (target__has_cpu(&target))
1939 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1940 
1941 		if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1942 			return -1;
1943 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1944 			if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1945 				return -1;
1946 		}
1947 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1948 			if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1949 				return -1;
1950 		}
1951 		if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1952 			return -1;
1953 
1954 		stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1955 		if (arch_evlist__add_default_attrs(evsel_list) < 0)
1956 			return -1;
1957 	}
1958 
1959 	/* Detailed events get appended to the event list: */
1960 
1961 	if (detailed_run <  1)
1962 		return 0;
1963 
1964 	/* Append detailed run extra attributes: */
1965 	if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1966 		return -1;
1967 
1968 	if (detailed_run < 2)
1969 		return 0;
1970 
1971 	/* Append very detailed run extra attributes: */
1972 	if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1973 		return -1;
1974 
1975 	if (detailed_run < 3)
1976 		return 0;
1977 
1978 	/* Append very, very detailed run extra attributes: */
1979 	return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1980 }
1981 
1982 static const char * const stat_record_usage[] = {
1983 	"perf stat record [<options>]",
1984 	NULL,
1985 };
1986 
1987 static void init_features(struct perf_session *session)
1988 {
1989 	int feat;
1990 
1991 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1992 		perf_header__set_feat(&session->header, feat);
1993 
1994 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1995 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1996 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1997 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1998 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1999 }
2000 
2001 static int __cmd_record(int argc, const char **argv)
2002 {
2003 	struct perf_session *session;
2004 	struct perf_data *data = &perf_stat.data;
2005 
2006 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
2007 			     PARSE_OPT_STOP_AT_NON_OPTION);
2008 
2009 	if (output_name)
2010 		data->path = output_name;
2011 
2012 	if (stat_config.run_count != 1 || forever) {
2013 		pr_err("Cannot use -r option with perf stat record.\n");
2014 		return -1;
2015 	}
2016 
2017 	session = perf_session__new(data, NULL);
2018 	if (IS_ERR(session)) {
2019 		pr_err("Perf session creation failed\n");
2020 		return PTR_ERR(session);
2021 	}
2022 
2023 	init_features(session);
2024 
2025 	session->evlist   = evsel_list;
2026 	perf_stat.session = session;
2027 	perf_stat.record  = true;
2028 	return argc;
2029 }
2030 
2031 static int process_stat_round_event(struct perf_session *session,
2032 				    union perf_event *event)
2033 {
2034 	struct perf_record_stat_round *stat_round = &event->stat_round;
2035 	struct evsel *counter;
2036 	struct timespec tsh, *ts = NULL;
2037 	const char **argv = session->header.env.cmdline_argv;
2038 	int argc = session->header.env.nr_cmdline;
2039 
2040 	evlist__for_each_entry(evsel_list, counter)
2041 		perf_stat_process_counter(&stat_config, counter);
2042 
2043 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2044 		update_stats(&walltime_nsecs_stats, stat_round->time);
2045 
2046 	if (stat_config.interval && stat_round->time) {
2047 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2048 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2049 		ts = &tsh;
2050 	}
2051 
2052 	print_counters(ts, argc, argv);
2053 	return 0;
2054 }
2055 
2056 static
2057 int process_stat_config_event(struct perf_session *session,
2058 			      union perf_event *event)
2059 {
2060 	struct perf_tool *tool = session->tool;
2061 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2062 
2063 	perf_event__read_stat_config(&stat_config, &event->stat_config);
2064 
2065 	if (perf_cpu_map__empty(st->cpus)) {
2066 		if (st->aggr_mode != AGGR_UNSET)
2067 			pr_warning("warning: processing task data, aggregation mode not set\n");
2068 		return 0;
2069 	}
2070 
2071 	if (st->aggr_mode != AGGR_UNSET)
2072 		stat_config.aggr_mode = st->aggr_mode;
2073 
2074 	if (perf_stat.data.is_pipe)
2075 		perf_stat_init_aggr_mode();
2076 	else
2077 		perf_stat_init_aggr_mode_file(st);
2078 
2079 	return 0;
2080 }
2081 
2082 static int set_maps(struct perf_stat *st)
2083 {
2084 	if (!st->cpus || !st->threads)
2085 		return 0;
2086 
2087 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2088 		return -EINVAL;
2089 
2090 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2091 
2092 	if (evlist__alloc_stats(evsel_list, true))
2093 		return -ENOMEM;
2094 
2095 	st->maps_allocated = true;
2096 	return 0;
2097 }
2098 
2099 static
2100 int process_thread_map_event(struct perf_session *session,
2101 			     union perf_event *event)
2102 {
2103 	struct perf_tool *tool = session->tool;
2104 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2105 
2106 	if (st->threads) {
2107 		pr_warning("Extra thread map event, ignoring.\n");
2108 		return 0;
2109 	}
2110 
2111 	st->threads = thread_map__new_event(&event->thread_map);
2112 	if (!st->threads)
2113 		return -ENOMEM;
2114 
2115 	return set_maps(st);
2116 }
2117 
2118 static
2119 int process_cpu_map_event(struct perf_session *session,
2120 			  union perf_event *event)
2121 {
2122 	struct perf_tool *tool = session->tool;
2123 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2124 	struct perf_cpu_map *cpus;
2125 
2126 	if (st->cpus) {
2127 		pr_warning("Extra cpu map event, ignoring.\n");
2128 		return 0;
2129 	}
2130 
2131 	cpus = cpu_map__new_data(&event->cpu_map.data);
2132 	if (!cpus)
2133 		return -ENOMEM;
2134 
2135 	st->cpus = cpus;
2136 	return set_maps(st);
2137 }
2138 
2139 static const char * const stat_report_usage[] = {
2140 	"perf stat report [<options>]",
2141 	NULL,
2142 };
2143 
2144 static struct perf_stat perf_stat = {
2145 	.tool = {
2146 		.attr		= perf_event__process_attr,
2147 		.event_update	= perf_event__process_event_update,
2148 		.thread_map	= process_thread_map_event,
2149 		.cpu_map	= process_cpu_map_event,
2150 		.stat_config	= process_stat_config_event,
2151 		.stat		= perf_event__process_stat_event,
2152 		.stat_round	= process_stat_round_event,
2153 	},
2154 	.aggr_mode = AGGR_UNSET,
2155 };
2156 
2157 static int __cmd_report(int argc, const char **argv)
2158 {
2159 	struct perf_session *session;
2160 	const struct option options[] = {
2161 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2162 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2163 		     "aggregate counts per processor socket", AGGR_SOCKET),
2164 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2165 		     "aggregate counts per processor die", AGGR_DIE),
2166 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2167 		     "aggregate counts per physical processor core", AGGR_CORE),
2168 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2169 		     "aggregate counts per numa node", AGGR_NODE),
2170 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2171 		     "disable CPU count aggregation", AGGR_NONE),
2172 	OPT_END()
2173 	};
2174 	struct stat st;
2175 	int ret;
2176 
2177 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2178 
2179 	if (!input_name || !strlen(input_name)) {
2180 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2181 			input_name = "-";
2182 		else
2183 			input_name = "perf.data";
2184 	}
2185 
2186 	perf_stat.data.path = input_name;
2187 	perf_stat.data.mode = PERF_DATA_MODE_READ;
2188 
2189 	session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2190 	if (IS_ERR(session))
2191 		return PTR_ERR(session);
2192 
2193 	perf_stat.session  = session;
2194 	stat_config.output = stderr;
2195 	evsel_list         = session->evlist;
2196 
2197 	ret = perf_session__process_events(session);
2198 	if (ret)
2199 		return ret;
2200 
2201 	perf_session__delete(session);
2202 	return 0;
2203 }
2204 
2205 static void setup_system_wide(int forks)
2206 {
2207 	/*
2208 	 * Make system wide (-a) the default target if
2209 	 * no target was specified and one of following
2210 	 * conditions is met:
2211 	 *
2212 	 *   - there's no workload specified
2213 	 *   - there is workload specified but all requested
2214 	 *     events are system wide events
2215 	 */
2216 	if (!target__none(&target))
2217 		return;
2218 
2219 	if (!forks)
2220 		target.system_wide = true;
2221 	else {
2222 		struct evsel *counter;
2223 
2224 		evlist__for_each_entry(evsel_list, counter) {
2225 			if (!counter->core.system_wide &&
2226 			    strcmp(counter->name, "duration_time")) {
2227 				return;
2228 			}
2229 		}
2230 
2231 		if (evsel_list->core.nr_entries)
2232 			target.system_wide = true;
2233 	}
2234 }
2235 
2236 int cmd_stat(int argc, const char **argv)
2237 {
2238 	const char * const stat_usage[] = {
2239 		"perf stat [<options>] [<command>]",
2240 		NULL
2241 	};
2242 	int status = -EINVAL, run_idx, err;
2243 	const char *mode;
2244 	FILE *output = stderr;
2245 	unsigned int interval, timeout;
2246 	const char * const stat_subcommands[] = { "record", "report" };
2247 	char errbuf[BUFSIZ];
2248 
2249 	setlocale(LC_ALL, "");
2250 
2251 	evsel_list = evlist__new();
2252 	if (evsel_list == NULL)
2253 		return -ENOMEM;
2254 
2255 	parse_events__shrink_config_terms();
2256 
2257 	/* String-parsing callback-based options would segfault when negated */
2258 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2259 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2260 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2261 
2262 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2263 					(const char **) stat_usage,
2264 					PARSE_OPT_STOP_AT_NON_OPTION);
2265 	perf_stat__collect_metric_expr(evsel_list);
2266 	perf_stat__init_shadow_stats();
2267 
2268 	if (stat_config.csv_sep) {
2269 		stat_config.csv_output = true;
2270 		if (!strcmp(stat_config.csv_sep, "\\t"))
2271 			stat_config.csv_sep = "\t";
2272 	} else
2273 		stat_config.csv_sep = DEFAULT_SEPARATOR;
2274 
2275 	if (argc && !strncmp(argv[0], "rec", 3)) {
2276 		argc = __cmd_record(argc, argv);
2277 		if (argc < 0)
2278 			return -1;
2279 	} else if (argc && !strncmp(argv[0], "rep", 3))
2280 		return __cmd_report(argc, argv);
2281 
2282 	interval = stat_config.interval;
2283 	timeout = stat_config.timeout;
2284 
2285 	/*
2286 	 * For record command the -o is already taken care of.
2287 	 */
2288 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2289 		output = NULL;
2290 
2291 	if (output_name && output_fd) {
2292 		fprintf(stderr, "cannot use both --output and --log-fd\n");
2293 		parse_options_usage(stat_usage, stat_options, "o", 1);
2294 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2295 		goto out;
2296 	}
2297 
2298 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2299 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2300 		goto out;
2301 	}
2302 
2303 	if (stat_config.metric_only && stat_config.run_count > 1) {
2304 		fprintf(stderr, "--metric-only is not supported with -r\n");
2305 		goto out;
2306 	}
2307 
2308 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2309 		fprintf(stderr, "--table is only supported with -r\n");
2310 		parse_options_usage(stat_usage, stat_options, "r", 1);
2311 		parse_options_usage(NULL, stat_options, "table", 0);
2312 		goto out;
2313 	}
2314 
2315 	if (output_fd < 0) {
2316 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2317 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2318 		goto out;
2319 	}
2320 
2321 	if (!output && !stat_config.quiet) {
2322 		struct timespec tm;
2323 		mode = append_file ? "a" : "w";
2324 
2325 		output = fopen(output_name, mode);
2326 		if (!output) {
2327 			perror("failed to create output file");
2328 			return -1;
2329 		}
2330 		clock_gettime(CLOCK_REALTIME, &tm);
2331 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2332 	} else if (output_fd > 0) {
2333 		mode = append_file ? "a" : "w";
2334 		output = fdopen(output_fd, mode);
2335 		if (!output) {
2336 			perror("Failed opening logfd");
2337 			return -errno;
2338 		}
2339 	}
2340 
2341 	stat_config.output = output;
2342 
2343 	/*
2344 	 * let the spreadsheet do the pretty-printing
2345 	 */
2346 	if (stat_config.csv_output) {
2347 		/* User explicitly passed -B? */
2348 		if (big_num_opt == 1) {
2349 			fprintf(stderr, "-B option not supported with -x\n");
2350 			parse_options_usage(stat_usage, stat_options, "B", 1);
2351 			parse_options_usage(NULL, stat_options, "x", 1);
2352 			goto out;
2353 		} else /* Nope, so disable big number formatting */
2354 			stat_config.big_num = false;
2355 	} else if (big_num_opt == 0) /* User passed --no-big-num */
2356 		stat_config.big_num = false;
2357 
2358 	err = target__validate(&target);
2359 	if (err) {
2360 		target__strerror(&target, err, errbuf, BUFSIZ);
2361 		pr_warning("%s\n", errbuf);
2362 	}
2363 
2364 	setup_system_wide(argc);
2365 
2366 	/*
2367 	 * Display user/system times only for single
2368 	 * run and when there's specified tracee.
2369 	 */
2370 	if ((stat_config.run_count == 1) && target__none(&target))
2371 		stat_config.ru_display = true;
2372 
2373 	if (stat_config.run_count < 0) {
2374 		pr_err("Run count must be a positive number\n");
2375 		parse_options_usage(stat_usage, stat_options, "r", 1);
2376 		goto out;
2377 	} else if (stat_config.run_count == 0) {
2378 		forever = true;
2379 		stat_config.run_count = 1;
2380 	}
2381 
2382 	if (stat_config.walltime_run_table) {
2383 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2384 		if (!stat_config.walltime_run) {
2385 			pr_err("failed to setup -r option");
2386 			goto out;
2387 		}
2388 	}
2389 
2390 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2391 		!target__has_task(&target)) {
2392 		if (!target.system_wide || target.cpu_list) {
2393 			fprintf(stderr, "The --per-thread option is only "
2394 				"available when monitoring via -p -t -a "
2395 				"options or only --per-thread.\n");
2396 			parse_options_usage(NULL, stat_options, "p", 1);
2397 			parse_options_usage(NULL, stat_options, "t", 1);
2398 			goto out;
2399 		}
2400 	}
2401 
2402 	/*
2403 	 * no_aggr, cgroup are for system-wide only
2404 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2405 	 */
2406 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2407 	      stat_config.aggr_mode != AGGR_THREAD) ||
2408 	     (nr_cgroups || stat_config.cgroup_list)) &&
2409 	    !target__has_cpu(&target)) {
2410 		fprintf(stderr, "both cgroup and no-aggregation "
2411 			"modes only available in system-wide mode\n");
2412 
2413 		parse_options_usage(stat_usage, stat_options, "G", 1);
2414 		parse_options_usage(NULL, stat_options, "A", 1);
2415 		parse_options_usage(NULL, stat_options, "a", 1);
2416 		parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2417 		goto out;
2418 	}
2419 
2420 	if (stat_config.iostat_run) {
2421 		status = iostat_prepare(evsel_list, &stat_config);
2422 		if (status)
2423 			goto out;
2424 		if (iostat_mode == IOSTAT_LIST) {
2425 			iostat_list(evsel_list, &stat_config);
2426 			goto out;
2427 		} else if (verbose)
2428 			iostat_list(evsel_list, &stat_config);
2429 		if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2430 			target.system_wide = true;
2431 	}
2432 
2433 	if (add_default_attributes())
2434 		goto out;
2435 
2436 	if (stat_config.cgroup_list) {
2437 		if (nr_cgroups > 0) {
2438 			pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2439 			parse_options_usage(stat_usage, stat_options, "G", 1);
2440 			parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2441 			goto out;
2442 		}
2443 
2444 		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2445 					  &stat_config.metric_events, true) < 0) {
2446 			parse_options_usage(stat_usage, stat_options,
2447 					    "for-each-cgroup", 0);
2448 			goto out;
2449 		}
2450 	}
2451 
2452 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2453 		target.per_thread = true;
2454 
2455 	if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2456 		pr_err("failed to use cpu list %s\n", target.cpu_list);
2457 		goto out;
2458 	}
2459 
2460 	target.hybrid = perf_pmu__has_hybrid();
2461 	if (evlist__create_maps(evsel_list, &target) < 0) {
2462 		if (target__has_task(&target)) {
2463 			pr_err("Problems finding threads of monitor\n");
2464 			parse_options_usage(stat_usage, stat_options, "p", 1);
2465 			parse_options_usage(NULL, stat_options, "t", 1);
2466 		} else if (target__has_cpu(&target)) {
2467 			perror("failed to parse CPUs map");
2468 			parse_options_usage(stat_usage, stat_options, "C", 1);
2469 			parse_options_usage(NULL, stat_options, "a", 1);
2470 		}
2471 		goto out;
2472 	}
2473 
2474 	evlist__check_cpu_maps(evsel_list);
2475 
2476 	/*
2477 	 * Initialize thread_map with comm names,
2478 	 * so we could print it out on output.
2479 	 */
2480 	if (stat_config.aggr_mode == AGGR_THREAD) {
2481 		thread_map__read_comms(evsel_list->core.threads);
2482 		if (target.system_wide) {
2483 			if (runtime_stat_new(&stat_config,
2484 				perf_thread_map__nr(evsel_list->core.threads))) {
2485 				goto out;
2486 			}
2487 		}
2488 	}
2489 
2490 	if (stat_config.aggr_mode == AGGR_NODE)
2491 		cpu__setup_cpunode_map();
2492 
2493 	if (stat_config.times && interval)
2494 		interval_count = true;
2495 	else if (stat_config.times && !interval) {
2496 		pr_err("interval-count option should be used together with "
2497 				"interval-print.\n");
2498 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2499 		parse_options_usage(stat_usage, stat_options, "I", 1);
2500 		goto out;
2501 	}
2502 
2503 	if (timeout && timeout < 100) {
2504 		if (timeout < 10) {
2505 			pr_err("timeout must be >= 10ms.\n");
2506 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2507 			goto out;
2508 		} else
2509 			pr_warning("timeout < 100ms. "
2510 				   "The overhead percentage could be high in some cases. "
2511 				   "Please proceed with caution.\n");
2512 	}
2513 	if (timeout && interval) {
2514 		pr_err("timeout option is not supported with interval-print.\n");
2515 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2516 		parse_options_usage(stat_usage, stat_options, "I", 1);
2517 		goto out;
2518 	}
2519 
2520 	if (evlist__alloc_stats(evsel_list, interval))
2521 		goto out;
2522 
2523 	if (perf_stat_init_aggr_mode())
2524 		goto out;
2525 
2526 	/*
2527 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2528 	 * while avoiding that older tools show confusing messages.
2529 	 *
2530 	 * However for pipe sessions we need to keep it zero,
2531 	 * because script's perf_evsel__check_attr is triggered
2532 	 * by attr->sample_type != 0, and we can't run it on
2533 	 * stat sessions.
2534 	 */
2535 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2536 
2537 	/*
2538 	 * We dont want to block the signals - that would cause
2539 	 * child tasks to inherit that and Ctrl-C would not work.
2540 	 * What we want is for Ctrl-C to work in the exec()-ed
2541 	 * task, but being ignored by perf stat itself:
2542 	 */
2543 	atexit(sig_atexit);
2544 	if (!forever)
2545 		signal(SIGINT,  skip_signal);
2546 	signal(SIGCHLD, skip_signal);
2547 	signal(SIGALRM, skip_signal);
2548 	signal(SIGABRT, skip_signal);
2549 
2550 	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2551 		goto out;
2552 
2553 	status = 0;
2554 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2555 		if (stat_config.run_count != 1 && verbose > 0)
2556 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2557 				run_idx + 1);
2558 
2559 		if (run_idx != 0)
2560 			evlist__reset_prev_raw_counts(evsel_list);
2561 
2562 		status = run_perf_stat(argc, argv, run_idx);
2563 		if (forever && status != -1 && !interval) {
2564 			print_counters(NULL, argc, argv);
2565 			perf_stat__reset_stats();
2566 		}
2567 	}
2568 
2569 	if (!forever && status != -1 && (!interval || stat_config.summary))
2570 		print_counters(NULL, argc, argv);
2571 
2572 	evlist__finalize_ctlfd(evsel_list);
2573 
2574 	if (STAT_RECORD) {
2575 		/*
2576 		 * We synthesize the kernel mmap record just so that older tools
2577 		 * don't emit warnings about not being able to resolve symbols
2578 		 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2579 		 * a saner message about no samples being in the perf.data file.
2580 		 *
2581 		 * This also serves to suppress a warning about f_header.data.size == 0
2582 		 * in header.c at the moment 'perf stat record' gets introduced, which
2583 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2584 		 * records, but the need to suppress the kptr_restrict messages in older
2585 		 * tools remain  -acme
2586 		 */
2587 		int fd = perf_data__fd(&perf_stat.data);
2588 
2589 		err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2590 							 process_synthesized_event,
2591 							 &perf_stat.session->machines.host);
2592 		if (err) {
2593 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2594 				   "older tools may produce warnings about this file\n.");
2595 		}
2596 
2597 		if (!interval) {
2598 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2599 				pr_err("failed to write stat round event\n");
2600 		}
2601 
2602 		if (!perf_stat.data.is_pipe) {
2603 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2604 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2605 		}
2606 
2607 		evlist__close(evsel_list);
2608 		perf_session__delete(perf_stat.session);
2609 	}
2610 
2611 	perf_stat__exit_aggr_mode();
2612 	evlist__free_stats(evsel_list);
2613 out:
2614 	if (stat_config.iostat_run)
2615 		iostat_release(evsel_list);
2616 
2617 	zfree(&stat_config.walltime_run);
2618 
2619 	if (smi_cost && smi_reset)
2620 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2621 
2622 	evlist__delete(evsel_list);
2623 
2624 	metricgroup__rblist_exit(&stat_config.metric_events);
2625 	runtime_stat_delete(&stat_config);
2626 	evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2627 
2628 	return status;
2629 }
2630