xref: /openbmc/linux/tools/perf/builtin-stat.c (revision fbb6b31a)
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 	if (counter->core.system_wide)
386 		nthreads = 1;
387 
388 	for (thread = 0; thread < nthreads; thread++) {
389 		struct perf_counts_values *count;
390 
391 		count = perf_counts(counter->counts, cpu_map_idx, thread);
392 
393 		/*
394 		 * The leader's group read loads data into its group members
395 		 * (via evsel__read_counter()) and sets their count->loaded.
396 		 */
397 		if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
398 		    read_single_counter(counter, cpu_map_idx, thread, rs)) {
399 			counter->counts->scaled = -1;
400 			perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
401 			perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
402 			return -1;
403 		}
404 
405 		perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
406 
407 		if (STAT_RECORD) {
408 			if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
409 				pr_err("failed to write stat event\n");
410 				return -1;
411 			}
412 		}
413 
414 		if (verbose > 1) {
415 			fprintf(stat_config.output,
416 				"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
417 					evsel__name(counter),
418 					perf_cpu_map__cpu(evsel__cpus(counter),
419 							  cpu_map_idx).cpu,
420 					count->val, count->ena, count->run);
421 		}
422 	}
423 
424 	return 0;
425 }
426 
427 static int read_affinity_counters(struct timespec *rs)
428 {
429 	struct evlist_cpu_iterator evlist_cpu_itr;
430 	struct affinity saved_affinity, *affinity;
431 
432 	if (all_counters_use_bpf)
433 		return 0;
434 
435 	if (!target__has_cpu(&target) || target__has_per_thread(&target))
436 		affinity = NULL;
437 	else if (affinity__setup(&saved_affinity) < 0)
438 		return -1;
439 	else
440 		affinity = &saved_affinity;
441 
442 	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
443 		struct evsel *counter = evlist_cpu_itr.evsel;
444 
445 		if (evsel__is_bpf(counter))
446 			continue;
447 
448 		if (!counter->err) {
449 			counter->err = read_counter_cpu(counter, rs,
450 							evlist_cpu_itr.cpu_map_idx);
451 		}
452 	}
453 	if (affinity)
454 		affinity__cleanup(&saved_affinity);
455 
456 	return 0;
457 }
458 
459 static int read_bpf_map_counters(void)
460 {
461 	struct evsel *counter;
462 	int err;
463 
464 	evlist__for_each_entry(evsel_list, counter) {
465 		if (!evsel__is_bpf(counter))
466 			continue;
467 
468 		err = bpf_counter__read(counter);
469 		if (err)
470 			return err;
471 	}
472 	return 0;
473 }
474 
475 static void read_counters(struct timespec *rs)
476 {
477 	struct evsel *counter;
478 
479 	if (!stat_config.stop_read_counter) {
480 		if (read_bpf_map_counters() ||
481 		    read_affinity_counters(rs))
482 			return;
483 	}
484 
485 	evlist__for_each_entry(evsel_list, counter) {
486 		if (counter->err)
487 			pr_debug("failed to read counter %s\n", counter->name);
488 		if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
489 			pr_warning("failed to process counter %s\n", counter->name);
490 		counter->err = 0;
491 	}
492 }
493 
494 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
495 {
496 	int i;
497 
498 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
499 	if (!config->stats)
500 		return -1;
501 
502 	config->stats_num = nthreads;
503 
504 	for (i = 0; i < nthreads; i++)
505 		runtime_stat__init(&config->stats[i]);
506 
507 	return 0;
508 }
509 
510 static void runtime_stat_delete(struct perf_stat_config *config)
511 {
512 	int i;
513 
514 	if (!config->stats)
515 		return;
516 
517 	for (i = 0; i < config->stats_num; i++)
518 		runtime_stat__exit(&config->stats[i]);
519 
520 	zfree(&config->stats);
521 }
522 
523 static void runtime_stat_reset(struct perf_stat_config *config)
524 {
525 	int i;
526 
527 	if (!config->stats)
528 		return;
529 
530 	for (i = 0; i < config->stats_num; i++)
531 		perf_stat__reset_shadow_per_stat(&config->stats[i]);
532 }
533 
534 static void process_interval(void)
535 {
536 	struct timespec ts, rs;
537 
538 	clock_gettime(CLOCK_MONOTONIC, &ts);
539 	diff_timespec(&rs, &ts, &ref_time);
540 
541 	perf_stat__reset_shadow_per_stat(&rt_stat);
542 	runtime_stat_reset(&stat_config);
543 	read_counters(&rs);
544 
545 	if (STAT_RECORD) {
546 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
547 			pr_err("failed to write stat round event\n");
548 	}
549 
550 	init_stats(&walltime_nsecs_stats);
551 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
552 	print_counters(&rs, 0, NULL);
553 }
554 
555 static bool handle_interval(unsigned int interval, int *times)
556 {
557 	if (interval) {
558 		process_interval();
559 		if (interval_count && !(--(*times)))
560 			return true;
561 	}
562 	return false;
563 }
564 
565 static int enable_counters(void)
566 {
567 	struct evsel *evsel;
568 	int err;
569 
570 	evlist__for_each_entry(evsel_list, evsel) {
571 		if (!evsel__is_bpf(evsel))
572 			continue;
573 
574 		err = bpf_counter__enable(evsel);
575 		if (err)
576 			return err;
577 	}
578 
579 	if (stat_config.initial_delay < 0) {
580 		pr_info(EVLIST_DISABLED_MSG);
581 		return 0;
582 	}
583 
584 	if (stat_config.initial_delay > 0) {
585 		pr_info(EVLIST_DISABLED_MSG);
586 		usleep(stat_config.initial_delay * USEC_PER_MSEC);
587 	}
588 
589 	/*
590 	 * We need to enable counters only if:
591 	 * - we don't have tracee (attaching to task or cpu)
592 	 * - we have initial delay configured
593 	 */
594 	if (!target__none(&target) || stat_config.initial_delay) {
595 		if (!all_counters_use_bpf)
596 			evlist__enable(evsel_list);
597 		if (stat_config.initial_delay > 0)
598 			pr_info(EVLIST_ENABLED_MSG);
599 	}
600 	return 0;
601 }
602 
603 static void disable_counters(void)
604 {
605 	struct evsel *counter;
606 
607 	/*
608 	 * If we don't have tracee (attaching to task or cpu), counters may
609 	 * still be running. To get accurate group ratios, we must stop groups
610 	 * from counting before reading their constituent counters.
611 	 */
612 	if (!target__none(&target)) {
613 		evlist__for_each_entry(evsel_list, counter)
614 			bpf_counter__disable(counter);
615 		if (!all_counters_use_bpf)
616 			evlist__disable(evsel_list);
617 	}
618 }
619 
620 static volatile int workload_exec_errno;
621 
622 /*
623  * evlist__prepare_workload will send a SIGUSR1
624  * if the fork fails, since we asked by setting its
625  * want_signal to true.
626  */
627 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
628 					void *ucontext __maybe_unused)
629 {
630 	workload_exec_errno = info->si_value.sival_int;
631 }
632 
633 static bool evsel__should_store_id(struct evsel *counter)
634 {
635 	return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
636 }
637 
638 static bool is_target_alive(struct target *_target,
639 			    struct perf_thread_map *threads)
640 {
641 	struct stat st;
642 	int i;
643 
644 	if (!target__has_task(_target))
645 		return true;
646 
647 	for (i = 0; i < threads->nr; i++) {
648 		char path[PATH_MAX];
649 
650 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
651 			  threads->map[i].pid);
652 
653 		if (!stat(path, &st))
654 			return true;
655 	}
656 
657 	return false;
658 }
659 
660 static void process_evlist(struct evlist *evlist, unsigned int interval)
661 {
662 	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
663 
664 	if (evlist__ctlfd_process(evlist, &cmd) > 0) {
665 		switch (cmd) {
666 		case EVLIST_CTL_CMD_ENABLE:
667 			if (interval)
668 				process_interval();
669 			break;
670 		case EVLIST_CTL_CMD_DISABLE:
671 			if (interval)
672 				process_interval();
673 			break;
674 		case EVLIST_CTL_CMD_SNAPSHOT:
675 		case EVLIST_CTL_CMD_ACK:
676 		case EVLIST_CTL_CMD_UNSUPPORTED:
677 		case EVLIST_CTL_CMD_EVLIST:
678 		case EVLIST_CTL_CMD_STOP:
679 		case EVLIST_CTL_CMD_PING:
680 		default:
681 			break;
682 		}
683 	}
684 }
685 
686 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
687 			int *time_to_sleep)
688 {
689 	int tts = *time_to_sleep;
690 	struct timespec time_diff;
691 
692 	diff_timespec(&time_diff, time_stop, time_start);
693 
694 	tts -= time_diff.tv_sec * MSEC_PER_SEC +
695 	       time_diff.tv_nsec / NSEC_PER_MSEC;
696 
697 	if (tts < 0)
698 		tts = 0;
699 
700 	*time_to_sleep = tts;
701 }
702 
703 static int dispatch_events(bool forks, int timeout, int interval, int *times)
704 {
705 	int child_exited = 0, status = 0;
706 	int time_to_sleep, sleep_time;
707 	struct timespec time_start, time_stop;
708 
709 	if (interval)
710 		sleep_time = interval;
711 	else if (timeout)
712 		sleep_time = timeout;
713 	else
714 		sleep_time = 1000;
715 
716 	time_to_sleep = sleep_time;
717 
718 	while (!done) {
719 		if (forks)
720 			child_exited = waitpid(child_pid, &status, WNOHANG);
721 		else
722 			child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
723 
724 		if (child_exited)
725 			break;
726 
727 		clock_gettime(CLOCK_MONOTONIC, &time_start);
728 		if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
729 			if (timeout || handle_interval(interval, times))
730 				break;
731 			time_to_sleep = sleep_time;
732 		} else { /* fd revent */
733 			process_evlist(evsel_list, interval);
734 			clock_gettime(CLOCK_MONOTONIC, &time_stop);
735 			compute_tts(&time_start, &time_stop, &time_to_sleep);
736 		}
737 	}
738 
739 	return status;
740 }
741 
742 enum counter_recovery {
743 	COUNTER_SKIP,
744 	COUNTER_RETRY,
745 	COUNTER_FATAL,
746 };
747 
748 static enum counter_recovery stat_handle_error(struct evsel *counter)
749 {
750 	char msg[BUFSIZ];
751 	/*
752 	 * PPC returns ENXIO for HW counters until 2.6.37
753 	 * (behavior changed with commit b0a873e).
754 	 */
755 	if (errno == EINVAL || errno == ENOSYS ||
756 	    errno == ENOENT || errno == EOPNOTSUPP ||
757 	    errno == ENXIO) {
758 		if (verbose > 0)
759 			ui__warning("%s event is not supported by the kernel.\n",
760 				    evsel__name(counter));
761 		counter->supported = false;
762 		/*
763 		 * errored is a sticky flag that means one of the counter's
764 		 * cpu event had a problem and needs to be reexamined.
765 		 */
766 		counter->errored = true;
767 
768 		if ((evsel__leader(counter) != counter) ||
769 		    !(counter->core.leader->nr_members > 1))
770 			return COUNTER_SKIP;
771 	} else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
772 		if (verbose > 0)
773 			ui__warning("%s\n", msg);
774 		return COUNTER_RETRY;
775 	} else if (target__has_per_thread(&target) &&
776 		   evsel_list->core.threads &&
777 		   evsel_list->core.threads->err_thread != -1) {
778 		/*
779 		 * For global --per-thread case, skip current
780 		 * error thread.
781 		 */
782 		if (!thread_map__remove(evsel_list->core.threads,
783 					evsel_list->core.threads->err_thread)) {
784 			evsel_list->core.threads->err_thread = -1;
785 			return COUNTER_RETRY;
786 		}
787 	}
788 
789 	evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
790 	ui__error("%s\n", msg);
791 
792 	if (child_pid != -1)
793 		kill(child_pid, SIGTERM);
794 	return COUNTER_FATAL;
795 }
796 
797 static int __run_perf_stat(int argc, const char **argv, int run_idx)
798 {
799 	int interval = stat_config.interval;
800 	int times = stat_config.times;
801 	int timeout = stat_config.timeout;
802 	char msg[BUFSIZ];
803 	unsigned long long t0, t1;
804 	struct evsel *counter;
805 	size_t l;
806 	int status = 0;
807 	const bool forks = (argc > 0);
808 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
809 	struct evlist_cpu_iterator evlist_cpu_itr;
810 	struct affinity saved_affinity, *affinity = NULL;
811 	int err;
812 	bool second_pass = false;
813 
814 	if (forks) {
815 		if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
816 			perror("failed to prepare workload");
817 			return -1;
818 		}
819 		child_pid = evsel_list->workload.pid;
820 	}
821 
822 	if (group)
823 		evlist__set_leader(evsel_list);
824 
825 	if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
826 		if (affinity__setup(&saved_affinity) < 0)
827 			return -1;
828 		affinity = &saved_affinity;
829 	}
830 
831 	evlist__for_each_entry(evsel_list, counter) {
832 		if (bpf_counter__load(counter, &target))
833 			return -1;
834 		if (!evsel__is_bpf(counter))
835 			all_counters_use_bpf = false;
836 	}
837 
838 	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
839 		counter = evlist_cpu_itr.evsel;
840 
841 		/*
842 		 * bperf calls evsel__open_per_cpu() in bperf__load(), so
843 		 * no need to call it again here.
844 		 */
845 		if (target.use_bpf)
846 			break;
847 
848 		if (counter->reset_group || counter->errored)
849 			continue;
850 		if (evsel__is_bpf(counter))
851 			continue;
852 try_again:
853 		if (create_perf_stat_counter(counter, &stat_config, &target,
854 					     evlist_cpu_itr.cpu_map_idx) < 0) {
855 
856 			/*
857 			 * Weak group failed. We cannot just undo this here
858 			 * because earlier CPUs might be in group mode, and the kernel
859 			 * doesn't support mixing group and non group reads. Defer
860 			 * it to later.
861 			 * Don't close here because we're in the wrong affinity.
862 			 */
863 			if ((errno == EINVAL || errno == EBADF) &&
864 				evsel__leader(counter) != counter &&
865 				counter->weak_group) {
866 				evlist__reset_weak_group(evsel_list, counter, false);
867 				assert(counter->reset_group);
868 				second_pass = true;
869 				continue;
870 			}
871 
872 			switch (stat_handle_error(counter)) {
873 			case COUNTER_FATAL:
874 				return -1;
875 			case COUNTER_RETRY:
876 				goto try_again;
877 			case COUNTER_SKIP:
878 				continue;
879 			default:
880 				break;
881 			}
882 
883 		}
884 		counter->supported = true;
885 	}
886 
887 	if (second_pass) {
888 		/*
889 		 * Now redo all the weak group after closing them,
890 		 * and also close errored counters.
891 		 */
892 
893 		/* First close errored or weak retry */
894 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
895 			counter = evlist_cpu_itr.evsel;
896 
897 			if (!counter->reset_group && !counter->errored)
898 				continue;
899 
900 			perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
901 		}
902 		/* Now reopen weak */
903 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
904 			counter = evlist_cpu_itr.evsel;
905 
906 			if (!counter->reset_group && !counter->errored)
907 				continue;
908 			if (!counter->reset_group)
909 				continue;
910 try_again_reset:
911 			pr_debug2("reopening weak %s\n", evsel__name(counter));
912 			if (create_perf_stat_counter(counter, &stat_config, &target,
913 						     evlist_cpu_itr.cpu_map_idx) < 0) {
914 
915 				switch (stat_handle_error(counter)) {
916 				case COUNTER_FATAL:
917 					return -1;
918 				case COUNTER_RETRY:
919 					goto try_again_reset;
920 				case COUNTER_SKIP:
921 					continue;
922 				default:
923 					break;
924 				}
925 			}
926 			counter->supported = true;
927 		}
928 	}
929 	affinity__cleanup(affinity);
930 
931 	evlist__for_each_entry(evsel_list, counter) {
932 		if (!counter->supported) {
933 			perf_evsel__free_fd(&counter->core);
934 			continue;
935 		}
936 
937 		l = strlen(counter->unit);
938 		if (l > stat_config.unit_width)
939 			stat_config.unit_width = l;
940 
941 		if (evsel__should_store_id(counter) &&
942 		    evsel__store_ids(counter, evsel_list))
943 			return -1;
944 	}
945 
946 	if (evlist__apply_filters(evsel_list, &counter)) {
947 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
948 			counter->filter, evsel__name(counter), errno,
949 			str_error_r(errno, msg, sizeof(msg)));
950 		return -1;
951 	}
952 
953 	if (STAT_RECORD) {
954 		int fd = perf_data__fd(&perf_stat.data);
955 
956 		if (is_pipe) {
957 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
958 		} else {
959 			err = perf_session__write_header(perf_stat.session, evsel_list,
960 							 fd, false);
961 		}
962 
963 		if (err < 0)
964 			return err;
965 
966 		err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
967 							 process_synthesized_event, is_pipe);
968 		if (err < 0)
969 			return err;
970 	}
971 
972 	/*
973 	 * Enable counters and exec the command:
974 	 */
975 	if (forks) {
976 		err = enable_counters();
977 		if (err)
978 			return -1;
979 		evlist__start_workload(evsel_list);
980 
981 		t0 = rdclock();
982 		clock_gettime(CLOCK_MONOTONIC, &ref_time);
983 
984 		if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
985 			status = dispatch_events(forks, timeout, interval, &times);
986 		if (child_pid != -1) {
987 			if (timeout)
988 				kill(child_pid, SIGTERM);
989 			wait4(child_pid, &status, 0, &stat_config.ru_data);
990 		}
991 
992 		if (workload_exec_errno) {
993 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
994 			pr_err("Workload failed: %s\n", emsg);
995 			return -1;
996 		}
997 
998 		if (WIFSIGNALED(status))
999 			psignal(WTERMSIG(status), argv[0]);
1000 	} else {
1001 		err = enable_counters();
1002 		if (err)
1003 			return -1;
1004 
1005 		t0 = rdclock();
1006 		clock_gettime(CLOCK_MONOTONIC, &ref_time);
1007 
1008 		status = dispatch_events(forks, timeout, interval, &times);
1009 	}
1010 
1011 	disable_counters();
1012 
1013 	t1 = rdclock();
1014 
1015 	if (stat_config.walltime_run_table)
1016 		stat_config.walltime_run[run_idx] = t1 - t0;
1017 
1018 	if (interval && stat_config.summary) {
1019 		stat_config.interval = 0;
1020 		stat_config.stop_read_counter = true;
1021 		init_stats(&walltime_nsecs_stats);
1022 		update_stats(&walltime_nsecs_stats, t1 - t0);
1023 
1024 		if (stat_config.aggr_mode == AGGR_GLOBAL)
1025 			evlist__save_aggr_prev_raw_counts(evsel_list);
1026 
1027 		evlist__copy_prev_raw_counts(evsel_list);
1028 		evlist__reset_prev_raw_counts(evsel_list);
1029 		runtime_stat_reset(&stat_config);
1030 		perf_stat__reset_shadow_per_stat(&rt_stat);
1031 	} else {
1032 		update_stats(&walltime_nsecs_stats, t1 - t0);
1033 		update_rusage_stats(&ru_stats, &stat_config.ru_data);
1034 	}
1035 
1036 	/*
1037 	 * Closing a group leader splits the group, and as we only disable
1038 	 * group leaders, results in remaining events becoming enabled. To
1039 	 * avoid arbitrary skew, we must read all counters before closing any
1040 	 * group leaders.
1041 	 */
1042 	read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
1043 
1044 	/*
1045 	 * We need to keep evsel_list alive, because it's processed
1046 	 * later the evsel_list will be closed after.
1047 	 */
1048 	if (!STAT_RECORD)
1049 		evlist__close(evsel_list);
1050 
1051 	return WEXITSTATUS(status);
1052 }
1053 
1054 static int run_perf_stat(int argc, const char **argv, int run_idx)
1055 {
1056 	int ret;
1057 
1058 	if (pre_cmd) {
1059 		ret = system(pre_cmd);
1060 		if (ret)
1061 			return ret;
1062 	}
1063 
1064 	if (sync_run)
1065 		sync();
1066 
1067 	ret = __run_perf_stat(argc, argv, run_idx);
1068 	if (ret)
1069 		return ret;
1070 
1071 	if (post_cmd) {
1072 		ret = system(post_cmd);
1073 		if (ret)
1074 			return ret;
1075 	}
1076 
1077 	return ret;
1078 }
1079 
1080 static void print_counters(struct timespec *ts, int argc, const char **argv)
1081 {
1082 	/* Do not print anything if we record to the pipe. */
1083 	if (STAT_RECORD && perf_stat.data.is_pipe)
1084 		return;
1085 	if (stat_config.quiet)
1086 		return;
1087 
1088 	evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1089 }
1090 
1091 static volatile int signr = -1;
1092 
1093 static void skip_signal(int signo)
1094 {
1095 	if ((child_pid == -1) || stat_config.interval)
1096 		done = 1;
1097 
1098 	signr = signo;
1099 	/*
1100 	 * render child_pid harmless
1101 	 * won't send SIGTERM to a random
1102 	 * process in case of race condition
1103 	 * and fast PID recycling
1104 	 */
1105 	child_pid = -1;
1106 }
1107 
1108 static void sig_atexit(void)
1109 {
1110 	sigset_t set, oset;
1111 
1112 	/*
1113 	 * avoid race condition with SIGCHLD handler
1114 	 * in skip_signal() which is modifying child_pid
1115 	 * goal is to avoid send SIGTERM to a random
1116 	 * process
1117 	 */
1118 	sigemptyset(&set);
1119 	sigaddset(&set, SIGCHLD);
1120 	sigprocmask(SIG_BLOCK, &set, &oset);
1121 
1122 	if (child_pid != -1)
1123 		kill(child_pid, SIGTERM);
1124 
1125 	sigprocmask(SIG_SETMASK, &oset, NULL);
1126 
1127 	if (signr == -1)
1128 		return;
1129 
1130 	signal(signr, SIG_DFL);
1131 	kill(getpid(), signr);
1132 }
1133 
1134 void perf_stat__set_big_num(int set)
1135 {
1136 	stat_config.big_num = (set != 0);
1137 }
1138 
1139 void perf_stat__set_no_csv_summary(int set)
1140 {
1141 	stat_config.no_csv_summary = (set != 0);
1142 }
1143 
1144 static int stat__set_big_num(const struct option *opt __maybe_unused,
1145 			     const char *s __maybe_unused, int unset)
1146 {
1147 	big_num_opt = unset ? 0 : 1;
1148 	perf_stat__set_big_num(!unset);
1149 	return 0;
1150 }
1151 
1152 static int enable_metric_only(const struct option *opt __maybe_unused,
1153 			      const char *s __maybe_unused, int unset)
1154 {
1155 	force_metric_only = true;
1156 	stat_config.metric_only = !unset;
1157 	return 0;
1158 }
1159 
1160 static int parse_metric_groups(const struct option *opt,
1161 			       const char *str,
1162 			       int unset __maybe_unused)
1163 {
1164 	return metricgroup__parse_groups(opt, str,
1165 					 stat_config.metric_no_group,
1166 					 stat_config.metric_no_merge,
1167 					 &stat_config.metric_events);
1168 }
1169 
1170 static int parse_control_option(const struct option *opt,
1171 				const char *str,
1172 				int unset __maybe_unused)
1173 {
1174 	struct perf_stat_config *config = opt->value;
1175 
1176 	return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1177 }
1178 
1179 static int parse_stat_cgroups(const struct option *opt,
1180 			      const char *str, int unset)
1181 {
1182 	if (stat_config.cgroup_list) {
1183 		pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1184 		return -1;
1185 	}
1186 
1187 	return parse_cgroups(opt, str, unset);
1188 }
1189 
1190 static int parse_hybrid_type(const struct option *opt,
1191 			     const char *str,
1192 			     int unset __maybe_unused)
1193 {
1194 	struct evlist *evlist = *(struct evlist **)opt->value;
1195 
1196 	if (!list_empty(&evlist->core.entries)) {
1197 		fprintf(stderr, "Must define cputype before events/metrics\n");
1198 		return -1;
1199 	}
1200 
1201 	evlist->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu(str);
1202 	if (!evlist->hybrid_pmu_name) {
1203 		fprintf(stderr, "--cputype %s is not supported!\n", str);
1204 		return -1;
1205 	}
1206 
1207 	return 0;
1208 }
1209 
1210 static struct option stat_options[] = {
1211 	OPT_BOOLEAN('T', "transaction", &transaction_run,
1212 		    "hardware transaction statistics"),
1213 	OPT_CALLBACK('e', "event", &evsel_list, "event",
1214 		     "event selector. use 'perf list' to list available events",
1215 		     parse_events_option),
1216 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1217 		     "event filter", parse_filter),
1218 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1219 		    "child tasks do not inherit counters"),
1220 	OPT_STRING('p', "pid", &target.pid, "pid",
1221 		   "stat events on existing process id"),
1222 	OPT_STRING('t', "tid", &target.tid, "tid",
1223 		   "stat events on existing thread id"),
1224 #ifdef HAVE_BPF_SKEL
1225 	OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1226 		   "stat events on existing bpf program id"),
1227 	OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1228 		    "use bpf program to count events"),
1229 	OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1230 		   "path to perf_event_attr map"),
1231 #endif
1232 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1233 		    "system-wide collection from all CPUs"),
1234 	OPT_BOOLEAN('g', "group", &group,
1235 		    "put the counters into a counter group"),
1236 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
1237 		    "Use --no-scale to disable counter scaling for multiplexing"),
1238 	OPT_INCR('v', "verbose", &verbose,
1239 		    "be more verbose (show counter open errors, etc)"),
1240 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
1241 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1242 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1243 		    "display details about each run (only with -r option)"),
1244 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
1245 		    "null run - dont start any counters"),
1246 	OPT_INCR('d', "detailed", &detailed_run,
1247 		    "detailed run - start a lot of events"),
1248 	OPT_BOOLEAN('S', "sync", &sync_run,
1249 		    "call sync() before starting a run"),
1250 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1251 			   "print large numbers with thousands\' separators",
1252 			   stat__set_big_num),
1253 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1254 		    "list of cpus to monitor in system-wide"),
1255 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1256 		    "disable CPU count aggregation", AGGR_NONE),
1257 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1258 	OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
1259 		    "Merge identical named hybrid events"),
1260 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1261 		   "print counts with custom separator"),
1262 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1263 		     "monitor event in cgroup name only", parse_stat_cgroups),
1264 	OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1265 		    "expand events for each cgroup"),
1266 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1267 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1268 	OPT_INTEGER(0, "log-fd", &output_fd,
1269 		    "log output to fd, instead of stderr"),
1270 	OPT_STRING(0, "pre", &pre_cmd, "command",
1271 			"command to run prior to the measured command"),
1272 	OPT_STRING(0, "post", &post_cmd, "command",
1273 			"command to run after to the measured command"),
1274 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1275 		    "print counts at regular interval in ms "
1276 		    "(overhead is possible for values <= 100ms)"),
1277 	OPT_INTEGER(0, "interval-count", &stat_config.times,
1278 		    "print counts for fixed number of times"),
1279 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1280 		    "clear screen in between new interval"),
1281 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1282 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1283 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1284 		     "aggregate counts per processor socket", AGGR_SOCKET),
1285 	OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1286 		     "aggregate counts per processor die", AGGR_DIE),
1287 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1288 		     "aggregate counts per physical processor core", AGGR_CORE),
1289 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1290 		     "aggregate counts per thread", AGGR_THREAD),
1291 	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1292 		     "aggregate counts per numa node", AGGR_NODE),
1293 	OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1294 		    "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1295 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1296 			"Only print computed metrics. No raw values", enable_metric_only),
1297 	OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1298 		       "don't group metric events, impacts multiplexing"),
1299 	OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1300 		       "don't try to share events between metrics in a group"),
1301 	OPT_BOOLEAN(0, "topdown", &topdown_run,
1302 			"measure top-down statistics"),
1303 	OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1304 			"Set the metrics level for the top-down statistics (0: max level)"),
1305 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1306 			"measure SMI cost"),
1307 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1308 		     "monitor specified metrics or metric groups (separated by ,)",
1309 		     parse_metric_groups),
1310 	OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1311 			 "Configure all used events to run in kernel space.",
1312 			 PARSE_OPT_EXCLUSIVE),
1313 	OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1314 			 "Configure all used events to run in user space.",
1315 			 PARSE_OPT_EXCLUSIVE),
1316 	OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1317 		    "Use with 'percore' event qualifier to show the event "
1318 		    "counts of one hardware thread by sum up total hardware "
1319 		    "threads of same physical core"),
1320 	OPT_BOOLEAN(0, "summary", &stat_config.summary,
1321 		       "print summary for interval mode"),
1322 	OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1323 		       "don't print 'summary' for CSV summary output"),
1324 	OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1325 			"don't print output (useful with record)"),
1326 	OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1327 		     "Only enable events on applying cpu with this type "
1328 		     "for hybrid platform (e.g. core or atom)",
1329 		     parse_hybrid_type),
1330 #ifdef HAVE_LIBPFM
1331 	OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1332 		"libpfm4 event selector. use 'perf list' to list available events",
1333 		parse_libpfm_events_option),
1334 #endif
1335 	OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1336 		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1337 		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1338 		     "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1339 		      parse_control_option),
1340 	OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1341 			    "measure I/O performance metrics provided by arch/platform",
1342 			    iostat_parse),
1343 	OPT_END()
1344 };
1345 
1346 static const char *const aggr_mode__string[] = {
1347 	[AGGR_CORE] = "core",
1348 	[AGGR_DIE] = "die",
1349 	[AGGR_GLOBAL] = "global",
1350 	[AGGR_NODE] = "node",
1351 	[AGGR_NONE] = "none",
1352 	[AGGR_SOCKET] = "socket",
1353 	[AGGR_THREAD] = "thread",
1354 	[AGGR_UNSET] = "unset",
1355 };
1356 
1357 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1358 						struct perf_cpu cpu)
1359 {
1360 	return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1361 }
1362 
1363 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1364 					     struct perf_cpu cpu)
1365 {
1366 	return aggr_cpu_id__die(cpu, /*data=*/NULL);
1367 }
1368 
1369 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1370 					      struct perf_cpu cpu)
1371 {
1372 	return aggr_cpu_id__core(cpu, /*data=*/NULL);
1373 }
1374 
1375 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1376 					      struct perf_cpu cpu)
1377 {
1378 	return aggr_cpu_id__node(cpu, /*data=*/NULL);
1379 }
1380 
1381 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1382 					      aggr_get_id_t get_id, struct perf_cpu cpu)
1383 {
1384 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1385 
1386 	if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1387 		config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1388 
1389 	id = config->cpus_aggr_map->map[cpu.cpu];
1390 	return id;
1391 }
1392 
1393 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1394 						       struct perf_cpu cpu)
1395 {
1396 	return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1397 }
1398 
1399 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1400 						    struct perf_cpu cpu)
1401 {
1402 	return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1403 }
1404 
1405 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1406 						     struct perf_cpu cpu)
1407 {
1408 	return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1409 }
1410 
1411 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1412 						     struct perf_cpu cpu)
1413 {
1414 	return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1415 }
1416 
1417 static bool term_percore_set(void)
1418 {
1419 	struct evsel *counter;
1420 
1421 	evlist__for_each_entry(evsel_list, counter) {
1422 		if (counter->percore)
1423 			return true;
1424 	}
1425 
1426 	return false;
1427 }
1428 
1429 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1430 {
1431 	switch (aggr_mode) {
1432 	case AGGR_SOCKET:
1433 		return aggr_cpu_id__socket;
1434 	case AGGR_DIE:
1435 		return aggr_cpu_id__die;
1436 	case AGGR_CORE:
1437 		return aggr_cpu_id__core;
1438 	case AGGR_NODE:
1439 		return aggr_cpu_id__node;
1440 	case AGGR_NONE:
1441 		if (term_percore_set())
1442 			return aggr_cpu_id__core;
1443 
1444 		return NULL;
1445 	case AGGR_GLOBAL:
1446 	case AGGR_THREAD:
1447 	case AGGR_UNSET:
1448 	default:
1449 		return NULL;
1450 	}
1451 }
1452 
1453 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1454 {
1455 	switch (aggr_mode) {
1456 	case AGGR_SOCKET:
1457 		return perf_stat__get_socket_cached;
1458 	case AGGR_DIE:
1459 		return perf_stat__get_die_cached;
1460 	case AGGR_CORE:
1461 		return perf_stat__get_core_cached;
1462 	case AGGR_NODE:
1463 		return perf_stat__get_node_cached;
1464 	case AGGR_NONE:
1465 		if (term_percore_set()) {
1466 			return perf_stat__get_core_cached;
1467 		}
1468 		return NULL;
1469 	case AGGR_GLOBAL:
1470 	case AGGR_THREAD:
1471 	case AGGR_UNSET:
1472 	default:
1473 		return NULL;
1474 	}
1475 }
1476 
1477 static int perf_stat_init_aggr_mode(void)
1478 {
1479 	int nr;
1480 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1481 
1482 	if (get_id) {
1483 		stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1484 							 get_id, /*data=*/NULL);
1485 		if (!stat_config.aggr_map) {
1486 			pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1487 			return -1;
1488 		}
1489 		stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1490 	}
1491 
1492 	/*
1493 	 * The evsel_list->cpus is the base we operate on,
1494 	 * taking the highest cpu number to be the size of
1495 	 * the aggregation translate cpumap.
1496 	 */
1497 	if (evsel_list->core.user_requested_cpus)
1498 		nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1499 	else
1500 		nr = 0;
1501 	stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1502 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1503 }
1504 
1505 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1506 {
1507 	if (map) {
1508 		WARN_ONCE(refcount_read(&map->refcnt) != 0,
1509 			  "cpu_aggr_map refcnt unbalanced\n");
1510 		free(map);
1511 	}
1512 }
1513 
1514 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1515 {
1516 	if (map && refcount_dec_and_test(&map->refcnt))
1517 		cpu_aggr_map__delete(map);
1518 }
1519 
1520 static void perf_stat__exit_aggr_mode(void)
1521 {
1522 	cpu_aggr_map__put(stat_config.aggr_map);
1523 	cpu_aggr_map__put(stat_config.cpus_aggr_map);
1524 	stat_config.aggr_map = NULL;
1525 	stat_config.cpus_aggr_map = NULL;
1526 }
1527 
1528 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1529 {
1530 	struct perf_env *env = data;
1531 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1532 
1533 	if (cpu.cpu != -1)
1534 		id.socket = env->cpu[cpu.cpu].socket_id;
1535 
1536 	return id;
1537 }
1538 
1539 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1540 {
1541 	struct perf_env *env = data;
1542 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1543 
1544 	if (cpu.cpu != -1) {
1545 		/*
1546 		 * die_id is relative to socket, so start
1547 		 * with the socket ID and then add die to
1548 		 * make a unique ID.
1549 		 */
1550 		id.socket = env->cpu[cpu.cpu].socket_id;
1551 		id.die = env->cpu[cpu.cpu].die_id;
1552 	}
1553 
1554 	return id;
1555 }
1556 
1557 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1558 {
1559 	struct perf_env *env = data;
1560 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1561 
1562 	if (cpu.cpu != -1) {
1563 		/*
1564 		 * core_id is relative to socket and die,
1565 		 * we need a global id. So we set
1566 		 * socket, die id and core id
1567 		 */
1568 		id.socket = env->cpu[cpu.cpu].socket_id;
1569 		id.die = env->cpu[cpu.cpu].die_id;
1570 		id.core = env->cpu[cpu.cpu].core_id;
1571 	}
1572 
1573 	return id;
1574 }
1575 
1576 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1577 {
1578 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1579 
1580 	id.node = perf_env__numa_node(data, cpu);
1581 	return id;
1582 }
1583 
1584 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1585 						     struct perf_cpu cpu)
1586 {
1587 	return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1588 }
1589 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1590 						  struct perf_cpu cpu)
1591 {
1592 	return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1593 }
1594 
1595 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1596 						   struct perf_cpu cpu)
1597 {
1598 	return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1599 }
1600 
1601 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1602 						   struct perf_cpu cpu)
1603 {
1604 	return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1605 }
1606 
1607 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1608 {
1609 	switch (aggr_mode) {
1610 	case AGGR_SOCKET:
1611 		return perf_env__get_socket_aggr_by_cpu;
1612 	case AGGR_DIE:
1613 		return perf_env__get_die_aggr_by_cpu;
1614 	case AGGR_CORE:
1615 		return perf_env__get_core_aggr_by_cpu;
1616 	case AGGR_NODE:
1617 		return perf_env__get_node_aggr_by_cpu;
1618 	case AGGR_NONE:
1619 	case AGGR_GLOBAL:
1620 	case AGGR_THREAD:
1621 	case AGGR_UNSET:
1622 	default:
1623 		return NULL;
1624 	}
1625 }
1626 
1627 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1628 {
1629 	switch (aggr_mode) {
1630 	case AGGR_SOCKET:
1631 		return perf_stat__get_socket_file;
1632 	case AGGR_DIE:
1633 		return perf_stat__get_die_file;
1634 	case AGGR_CORE:
1635 		return perf_stat__get_core_file;
1636 	case AGGR_NODE:
1637 		return perf_stat__get_node_file;
1638 	case AGGR_NONE:
1639 	case AGGR_GLOBAL:
1640 	case AGGR_THREAD:
1641 	case AGGR_UNSET:
1642 	default:
1643 		return NULL;
1644 	}
1645 }
1646 
1647 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1648 {
1649 	struct perf_env *env = &st->session->header.env;
1650 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1651 
1652 	if (!get_id)
1653 		return 0;
1654 
1655 	stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, get_id, env);
1656 	if (!stat_config.aggr_map) {
1657 		pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1658 		return -1;
1659 	}
1660 	stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1661 	return 0;
1662 }
1663 
1664 /*
1665  * Add default attributes, if there were no attributes specified or
1666  * if -d/--detailed, -d -d or -d -d -d is used:
1667  */
1668 static int add_default_attributes(void)
1669 {
1670 	int err;
1671 	struct perf_event_attr default_attrs0[] = {
1672 
1673   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1674   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1675   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1676   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1677 
1678   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1679 };
1680 	struct perf_event_attr frontend_attrs[] = {
1681   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1682 };
1683 	struct perf_event_attr backend_attrs[] = {
1684   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1685 };
1686 	struct perf_event_attr default_attrs1[] = {
1687   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1688   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1689   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1690 
1691 };
1692 	struct perf_event_attr default_sw_attrs[] = {
1693   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1694   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1695   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1696   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1697 };
1698 
1699 /*
1700  * Detailed stats (-d), covering the L1 and last level data caches:
1701  */
1702 	struct perf_event_attr detailed_attrs[] = {
1703 
1704   { .type = PERF_TYPE_HW_CACHE,
1705     .config =
1706 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1707 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1708 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1709 
1710   { .type = PERF_TYPE_HW_CACHE,
1711     .config =
1712 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1713 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1714 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1715 
1716   { .type = PERF_TYPE_HW_CACHE,
1717     .config =
1718 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1719 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1720 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1721 
1722   { .type = PERF_TYPE_HW_CACHE,
1723     .config =
1724 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1725 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1726 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1727 };
1728 
1729 /*
1730  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1731  */
1732 	struct perf_event_attr very_detailed_attrs[] = {
1733 
1734   { .type = PERF_TYPE_HW_CACHE,
1735     .config =
1736 	 PERF_COUNT_HW_CACHE_L1I		<<  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_L1I		<<  0  |
1743 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1744 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1745 
1746   { .type = PERF_TYPE_HW_CACHE,
1747     .config =
1748 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1749 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1750 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1751 
1752   { .type = PERF_TYPE_HW_CACHE,
1753     .config =
1754 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1755 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1756 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1757 
1758   { .type = PERF_TYPE_HW_CACHE,
1759     .config =
1760 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1761 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1762 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1763 
1764   { .type = PERF_TYPE_HW_CACHE,
1765     .config =
1766 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1767 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1768 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1769 
1770 };
1771 
1772 /*
1773  * Very, very detailed stats (-d -d -d), adding prefetch events:
1774  */
1775 	struct perf_event_attr very_very_detailed_attrs[] = {
1776 
1777   { .type = PERF_TYPE_HW_CACHE,
1778     .config =
1779 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1780 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1781 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1782 
1783   { .type = PERF_TYPE_HW_CACHE,
1784     .config =
1785 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1786 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1787 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1788 };
1789 	/* Set attrs if no event is selected and !null_run: */
1790 	if (stat_config.null_run)
1791 		return 0;
1792 
1793 	if (transaction_run) {
1794 		struct parse_events_error errinfo;
1795 		/* Handle -T as -M transaction. Once platform specific metrics
1796 		 * support has been added to the json files, all architectures
1797 		 * will use this approach. To determine transaction support
1798 		 * on an architecture test for such a metric name.
1799 		 */
1800 		if (metricgroup__has_metric("transaction")) {
1801 			struct option opt = { .value = &evsel_list };
1802 
1803 			return metricgroup__parse_groups(&opt, "transaction",
1804 							 stat_config.metric_no_group,
1805 							stat_config.metric_no_merge,
1806 							 &stat_config.metric_events);
1807 		}
1808 
1809 		parse_events_error__init(&errinfo);
1810 		if (pmu_have_event("cpu", "cycles-ct") &&
1811 		    pmu_have_event("cpu", "el-start"))
1812 			err = parse_events(evsel_list, transaction_attrs,
1813 					   &errinfo);
1814 		else
1815 			err = parse_events(evsel_list,
1816 					   transaction_limited_attrs,
1817 					   &errinfo);
1818 		if (err) {
1819 			fprintf(stderr, "Cannot set up transaction events\n");
1820 			parse_events_error__print(&errinfo, transaction_attrs);
1821 		}
1822 		parse_events_error__exit(&errinfo);
1823 		return err ? -1 : 0;
1824 	}
1825 
1826 	if (smi_cost) {
1827 		struct parse_events_error errinfo;
1828 		int smi;
1829 
1830 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1831 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1832 			return -1;
1833 		}
1834 
1835 		if (!smi) {
1836 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1837 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1838 				return -1;
1839 			}
1840 			smi_reset = true;
1841 		}
1842 
1843 		if (!pmu_have_event("msr", "aperf") ||
1844 		    !pmu_have_event("msr", "smi")) {
1845 			fprintf(stderr, "To measure SMI cost, it needs "
1846 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1847 			return -1;
1848 		}
1849 		if (!force_metric_only)
1850 			stat_config.metric_only = true;
1851 
1852 		parse_events_error__init(&errinfo);
1853 		err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1854 		if (err) {
1855 			parse_events_error__print(&errinfo, smi_cost_attrs);
1856 			fprintf(stderr, "Cannot set up SMI cost events\n");
1857 		}
1858 		parse_events_error__exit(&errinfo);
1859 		return err ? -1 : 0;
1860 	}
1861 
1862 	if (topdown_run) {
1863 		const char **metric_attrs = topdown_metric_attrs;
1864 		unsigned int max_level = 1;
1865 		char *str = NULL;
1866 		bool warn = false;
1867 		const char *pmu_name = "cpu";
1868 
1869 		if (!force_metric_only)
1870 			stat_config.metric_only = true;
1871 
1872 		if (perf_pmu__has_hybrid()) {
1873 			if (!evsel_list->hybrid_pmu_name) {
1874 				pr_warning("WARNING: default to use cpu_core topdown events\n");
1875 				evsel_list->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu("core");
1876 			}
1877 
1878 			pmu_name = evsel_list->hybrid_pmu_name;
1879 			if (!pmu_name)
1880 				return -1;
1881 		}
1882 
1883 		if (pmu_have_event(pmu_name, topdown_metric_L2_attrs[5])) {
1884 			metric_attrs = topdown_metric_L2_attrs;
1885 			max_level = 2;
1886 		}
1887 
1888 		if (stat_config.topdown_level > max_level) {
1889 			pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1890 			return -1;
1891 		} else if (!stat_config.topdown_level)
1892 			stat_config.topdown_level = max_level;
1893 
1894 		if (topdown_filter_events(metric_attrs, &str, 1, pmu_name) < 0) {
1895 			pr_err("Out of memory\n");
1896 			return -1;
1897 		}
1898 
1899 		if (metric_attrs[0] && str) {
1900 			if (!stat_config.interval && !stat_config.metric_only) {
1901 				fprintf(stat_config.output,
1902 					"Topdown accuracy may decrease when measuring long periods.\n"
1903 					"Please print the result regularly, e.g. -I1000\n");
1904 			}
1905 			goto setup_metrics;
1906 		}
1907 
1908 		zfree(&str);
1909 
1910 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1911 		    stat_config.aggr_mode != AGGR_CORE) {
1912 			pr_err("top down event configuration requires --per-core mode\n");
1913 			return -1;
1914 		}
1915 		stat_config.aggr_mode = AGGR_CORE;
1916 		if (nr_cgroups || !target__has_cpu(&target)) {
1917 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1918 			return -1;
1919 		}
1920 
1921 		if (topdown_filter_events(topdown_attrs, &str,
1922 				arch_topdown_check_group(&warn),
1923 				pmu_name) < 0) {
1924 			pr_err("Out of memory\n");
1925 			return -1;
1926 		}
1927 
1928 		if (topdown_attrs[0] && str) {
1929 			struct parse_events_error errinfo;
1930 			if (warn)
1931 				arch_topdown_group_warn();
1932 setup_metrics:
1933 			parse_events_error__init(&errinfo);
1934 			err = parse_events(evsel_list, str, &errinfo);
1935 			if (err) {
1936 				fprintf(stderr,
1937 					"Cannot set up top down events %s: %d\n",
1938 					str, err);
1939 				parse_events_error__print(&errinfo, str);
1940 				parse_events_error__exit(&errinfo);
1941 				free(str);
1942 				return -1;
1943 			}
1944 			parse_events_error__exit(&errinfo);
1945 		} else {
1946 			fprintf(stderr, "System does not support topdown\n");
1947 			return -1;
1948 		}
1949 		free(str);
1950 	}
1951 
1952 	if (!evsel_list->core.nr_entries) {
1953 		if (perf_pmu__has_hybrid()) {
1954 			struct parse_events_error errinfo;
1955 			const char *hybrid_str = "cycles,instructions,branches,branch-misses";
1956 
1957 			if (target__has_cpu(&target))
1958 				default_sw_attrs[0].config = PERF_COUNT_SW_CPU_CLOCK;
1959 
1960 			if (evlist__add_default_attrs(evsel_list,
1961 						      default_sw_attrs) < 0) {
1962 				return -1;
1963 			}
1964 
1965 			parse_events_error__init(&errinfo);
1966 			err = parse_events(evsel_list, hybrid_str, &errinfo);
1967 			if (err) {
1968 				fprintf(stderr,
1969 					"Cannot set up hybrid events %s: %d\n",
1970 					hybrid_str, err);
1971 				parse_events_error__print(&errinfo, hybrid_str);
1972 			}
1973 			parse_events_error__exit(&errinfo);
1974 			return err ? -1 : 0;
1975 		}
1976 
1977 		if (target__has_cpu(&target))
1978 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1979 
1980 		if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1981 			return -1;
1982 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1983 			if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1984 				return -1;
1985 		}
1986 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1987 			if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1988 				return -1;
1989 		}
1990 		if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1991 			return -1;
1992 
1993 		stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1994 		if (arch_evlist__add_default_attrs(evsel_list) < 0)
1995 			return -1;
1996 	}
1997 
1998 	/* Detailed events get appended to the event list: */
1999 
2000 	if (detailed_run <  1)
2001 		return 0;
2002 
2003 	/* Append detailed run extra attributes: */
2004 	if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2005 		return -1;
2006 
2007 	if (detailed_run < 2)
2008 		return 0;
2009 
2010 	/* Append very detailed run extra attributes: */
2011 	if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2012 		return -1;
2013 
2014 	if (detailed_run < 3)
2015 		return 0;
2016 
2017 	/* Append very, very detailed run extra attributes: */
2018 	return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2019 }
2020 
2021 static const char * const stat_record_usage[] = {
2022 	"perf stat record [<options>]",
2023 	NULL,
2024 };
2025 
2026 static void init_features(struct perf_session *session)
2027 {
2028 	int feat;
2029 
2030 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2031 		perf_header__set_feat(&session->header, feat);
2032 
2033 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
2034 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2035 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2036 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2037 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2038 }
2039 
2040 static int __cmd_record(int argc, const char **argv)
2041 {
2042 	struct perf_session *session;
2043 	struct perf_data *data = &perf_stat.data;
2044 
2045 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
2046 			     PARSE_OPT_STOP_AT_NON_OPTION);
2047 
2048 	if (output_name)
2049 		data->path = output_name;
2050 
2051 	if (stat_config.run_count != 1 || forever) {
2052 		pr_err("Cannot use -r option with perf stat record.\n");
2053 		return -1;
2054 	}
2055 
2056 	session = perf_session__new(data, NULL);
2057 	if (IS_ERR(session)) {
2058 		pr_err("Perf session creation failed\n");
2059 		return PTR_ERR(session);
2060 	}
2061 
2062 	init_features(session);
2063 
2064 	session->evlist   = evsel_list;
2065 	perf_stat.session = session;
2066 	perf_stat.record  = true;
2067 	return argc;
2068 }
2069 
2070 static int process_stat_round_event(struct perf_session *session,
2071 				    union perf_event *event)
2072 {
2073 	struct perf_record_stat_round *stat_round = &event->stat_round;
2074 	struct evsel *counter;
2075 	struct timespec tsh, *ts = NULL;
2076 	const char **argv = session->header.env.cmdline_argv;
2077 	int argc = session->header.env.nr_cmdline;
2078 
2079 	evlist__for_each_entry(evsel_list, counter)
2080 		perf_stat_process_counter(&stat_config, counter);
2081 
2082 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2083 		update_stats(&walltime_nsecs_stats, stat_round->time);
2084 
2085 	if (stat_config.interval && stat_round->time) {
2086 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2087 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2088 		ts = &tsh;
2089 	}
2090 
2091 	print_counters(ts, argc, argv);
2092 	return 0;
2093 }
2094 
2095 static
2096 int process_stat_config_event(struct perf_session *session,
2097 			      union perf_event *event)
2098 {
2099 	struct perf_tool *tool = session->tool;
2100 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2101 
2102 	perf_event__read_stat_config(&stat_config, &event->stat_config);
2103 
2104 	if (perf_cpu_map__empty(st->cpus)) {
2105 		if (st->aggr_mode != AGGR_UNSET)
2106 			pr_warning("warning: processing task data, aggregation mode not set\n");
2107 		return 0;
2108 	}
2109 
2110 	if (st->aggr_mode != AGGR_UNSET)
2111 		stat_config.aggr_mode = st->aggr_mode;
2112 
2113 	if (perf_stat.data.is_pipe)
2114 		perf_stat_init_aggr_mode();
2115 	else
2116 		perf_stat_init_aggr_mode_file(st);
2117 
2118 	return 0;
2119 }
2120 
2121 static int set_maps(struct perf_stat *st)
2122 {
2123 	if (!st->cpus || !st->threads)
2124 		return 0;
2125 
2126 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2127 		return -EINVAL;
2128 
2129 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2130 
2131 	if (evlist__alloc_stats(evsel_list, true))
2132 		return -ENOMEM;
2133 
2134 	st->maps_allocated = true;
2135 	return 0;
2136 }
2137 
2138 static
2139 int process_thread_map_event(struct perf_session *session,
2140 			     union perf_event *event)
2141 {
2142 	struct perf_tool *tool = session->tool;
2143 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2144 
2145 	if (st->threads) {
2146 		pr_warning("Extra thread map event, ignoring.\n");
2147 		return 0;
2148 	}
2149 
2150 	st->threads = thread_map__new_event(&event->thread_map);
2151 	if (!st->threads)
2152 		return -ENOMEM;
2153 
2154 	return set_maps(st);
2155 }
2156 
2157 static
2158 int process_cpu_map_event(struct perf_session *session,
2159 			  union perf_event *event)
2160 {
2161 	struct perf_tool *tool = session->tool;
2162 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2163 	struct perf_cpu_map *cpus;
2164 
2165 	if (st->cpus) {
2166 		pr_warning("Extra cpu map event, ignoring.\n");
2167 		return 0;
2168 	}
2169 
2170 	cpus = cpu_map__new_data(&event->cpu_map.data);
2171 	if (!cpus)
2172 		return -ENOMEM;
2173 
2174 	st->cpus = cpus;
2175 	return set_maps(st);
2176 }
2177 
2178 static const char * const stat_report_usage[] = {
2179 	"perf stat report [<options>]",
2180 	NULL,
2181 };
2182 
2183 static struct perf_stat perf_stat = {
2184 	.tool = {
2185 		.attr		= perf_event__process_attr,
2186 		.event_update	= perf_event__process_event_update,
2187 		.thread_map	= process_thread_map_event,
2188 		.cpu_map	= process_cpu_map_event,
2189 		.stat_config	= process_stat_config_event,
2190 		.stat		= perf_event__process_stat_event,
2191 		.stat_round	= process_stat_round_event,
2192 	},
2193 	.aggr_mode = AGGR_UNSET,
2194 };
2195 
2196 static int __cmd_report(int argc, const char **argv)
2197 {
2198 	struct perf_session *session;
2199 	const struct option options[] = {
2200 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2201 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2202 		     "aggregate counts per processor socket", AGGR_SOCKET),
2203 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2204 		     "aggregate counts per processor die", AGGR_DIE),
2205 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2206 		     "aggregate counts per physical processor core", AGGR_CORE),
2207 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2208 		     "aggregate counts per numa node", AGGR_NODE),
2209 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2210 		     "disable CPU count aggregation", AGGR_NONE),
2211 	OPT_END()
2212 	};
2213 	struct stat st;
2214 	int ret;
2215 
2216 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2217 
2218 	if (!input_name || !strlen(input_name)) {
2219 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2220 			input_name = "-";
2221 		else
2222 			input_name = "perf.data";
2223 	}
2224 
2225 	perf_stat.data.path = input_name;
2226 	perf_stat.data.mode = PERF_DATA_MODE_READ;
2227 
2228 	session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2229 	if (IS_ERR(session))
2230 		return PTR_ERR(session);
2231 
2232 	perf_stat.session  = session;
2233 	stat_config.output = stderr;
2234 	evsel_list         = session->evlist;
2235 
2236 	ret = perf_session__process_events(session);
2237 	if (ret)
2238 		return ret;
2239 
2240 	perf_session__delete(session);
2241 	return 0;
2242 }
2243 
2244 static void setup_system_wide(int forks)
2245 {
2246 	/*
2247 	 * Make system wide (-a) the default target if
2248 	 * no target was specified and one of following
2249 	 * conditions is met:
2250 	 *
2251 	 *   - there's no workload specified
2252 	 *   - there is workload specified but all requested
2253 	 *     events are system wide events
2254 	 */
2255 	if (!target__none(&target))
2256 		return;
2257 
2258 	if (!forks)
2259 		target.system_wide = true;
2260 	else {
2261 		struct evsel *counter;
2262 
2263 		evlist__for_each_entry(evsel_list, counter) {
2264 			if (!counter->core.system_wide &&
2265 			    strcmp(counter->name, "duration_time")) {
2266 				return;
2267 			}
2268 		}
2269 
2270 		if (evsel_list->core.nr_entries)
2271 			target.system_wide = true;
2272 	}
2273 }
2274 
2275 int cmd_stat(int argc, const char **argv)
2276 {
2277 	const char * const stat_usage[] = {
2278 		"perf stat [<options>] [<command>]",
2279 		NULL
2280 	};
2281 	int status = -EINVAL, run_idx, err;
2282 	const char *mode;
2283 	FILE *output = stderr;
2284 	unsigned int interval, timeout;
2285 	const char * const stat_subcommands[] = { "record", "report" };
2286 	char errbuf[BUFSIZ];
2287 
2288 	setlocale(LC_ALL, "");
2289 
2290 	evsel_list = evlist__new();
2291 	if (evsel_list == NULL)
2292 		return -ENOMEM;
2293 
2294 	parse_events__shrink_config_terms();
2295 
2296 	/* String-parsing callback-based options would segfault when negated */
2297 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2298 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2299 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2300 
2301 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2302 					(const char **) stat_usage,
2303 					PARSE_OPT_STOP_AT_NON_OPTION);
2304 	perf_stat__collect_metric_expr(evsel_list);
2305 	perf_stat__init_shadow_stats();
2306 
2307 	if (stat_config.csv_sep) {
2308 		stat_config.csv_output = true;
2309 		if (!strcmp(stat_config.csv_sep, "\\t"))
2310 			stat_config.csv_sep = "\t";
2311 	} else
2312 		stat_config.csv_sep = DEFAULT_SEPARATOR;
2313 
2314 	if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2315 		argc = __cmd_record(argc, argv);
2316 		if (argc < 0)
2317 			return -1;
2318 	} else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2319 		return __cmd_report(argc, argv);
2320 
2321 	interval = stat_config.interval;
2322 	timeout = stat_config.timeout;
2323 
2324 	/*
2325 	 * For record command the -o is already taken care of.
2326 	 */
2327 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2328 		output = NULL;
2329 
2330 	if (output_name && output_fd) {
2331 		fprintf(stderr, "cannot use both --output and --log-fd\n");
2332 		parse_options_usage(stat_usage, stat_options, "o", 1);
2333 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2334 		goto out;
2335 	}
2336 
2337 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2338 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2339 		goto out;
2340 	}
2341 
2342 	if (stat_config.metric_only && stat_config.run_count > 1) {
2343 		fprintf(stderr, "--metric-only is not supported with -r\n");
2344 		goto out;
2345 	}
2346 
2347 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2348 		fprintf(stderr, "--table is only supported with -r\n");
2349 		parse_options_usage(stat_usage, stat_options, "r", 1);
2350 		parse_options_usage(NULL, stat_options, "table", 0);
2351 		goto out;
2352 	}
2353 
2354 	if (output_fd < 0) {
2355 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2356 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2357 		goto out;
2358 	}
2359 
2360 	if (!output && !stat_config.quiet) {
2361 		struct timespec tm;
2362 		mode = append_file ? "a" : "w";
2363 
2364 		output = fopen(output_name, mode);
2365 		if (!output) {
2366 			perror("failed to create output file");
2367 			return -1;
2368 		}
2369 		clock_gettime(CLOCK_REALTIME, &tm);
2370 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2371 	} else if (output_fd > 0) {
2372 		mode = append_file ? "a" : "w";
2373 		output = fdopen(output_fd, mode);
2374 		if (!output) {
2375 			perror("Failed opening logfd");
2376 			return -errno;
2377 		}
2378 	}
2379 
2380 	stat_config.output = output;
2381 
2382 	/*
2383 	 * let the spreadsheet do the pretty-printing
2384 	 */
2385 	if (stat_config.csv_output) {
2386 		/* User explicitly passed -B? */
2387 		if (big_num_opt == 1) {
2388 			fprintf(stderr, "-B option not supported with -x\n");
2389 			parse_options_usage(stat_usage, stat_options, "B", 1);
2390 			parse_options_usage(NULL, stat_options, "x", 1);
2391 			goto out;
2392 		} else /* Nope, so disable big number formatting */
2393 			stat_config.big_num = false;
2394 	} else if (big_num_opt == 0) /* User passed --no-big-num */
2395 		stat_config.big_num = false;
2396 
2397 	err = target__validate(&target);
2398 	if (err) {
2399 		target__strerror(&target, err, errbuf, BUFSIZ);
2400 		pr_warning("%s\n", errbuf);
2401 	}
2402 
2403 	setup_system_wide(argc);
2404 
2405 	/*
2406 	 * Display user/system times only for single
2407 	 * run and when there's specified tracee.
2408 	 */
2409 	if ((stat_config.run_count == 1) && target__none(&target))
2410 		stat_config.ru_display = true;
2411 
2412 	if (stat_config.run_count < 0) {
2413 		pr_err("Run count must be a positive number\n");
2414 		parse_options_usage(stat_usage, stat_options, "r", 1);
2415 		goto out;
2416 	} else if (stat_config.run_count == 0) {
2417 		forever = true;
2418 		stat_config.run_count = 1;
2419 	}
2420 
2421 	if (stat_config.walltime_run_table) {
2422 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2423 		if (!stat_config.walltime_run) {
2424 			pr_err("failed to setup -r option");
2425 			goto out;
2426 		}
2427 	}
2428 
2429 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2430 		!target__has_task(&target)) {
2431 		if (!target.system_wide || target.cpu_list) {
2432 			fprintf(stderr, "The --per-thread option is only "
2433 				"available when monitoring via -p -t -a "
2434 				"options or only --per-thread.\n");
2435 			parse_options_usage(NULL, stat_options, "p", 1);
2436 			parse_options_usage(NULL, stat_options, "t", 1);
2437 			goto out;
2438 		}
2439 	}
2440 
2441 	/*
2442 	 * no_aggr, cgroup are for system-wide only
2443 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2444 	 */
2445 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2446 	      stat_config.aggr_mode != AGGR_THREAD) ||
2447 	     (nr_cgroups || stat_config.cgroup_list)) &&
2448 	    !target__has_cpu(&target)) {
2449 		fprintf(stderr, "both cgroup and no-aggregation "
2450 			"modes only available in system-wide mode\n");
2451 
2452 		parse_options_usage(stat_usage, stat_options, "G", 1);
2453 		parse_options_usage(NULL, stat_options, "A", 1);
2454 		parse_options_usage(NULL, stat_options, "a", 1);
2455 		parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2456 		goto out;
2457 	}
2458 
2459 	if (stat_config.iostat_run) {
2460 		status = iostat_prepare(evsel_list, &stat_config);
2461 		if (status)
2462 			goto out;
2463 		if (iostat_mode == IOSTAT_LIST) {
2464 			iostat_list(evsel_list, &stat_config);
2465 			goto out;
2466 		} else if (verbose)
2467 			iostat_list(evsel_list, &stat_config);
2468 		if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2469 			target.system_wide = true;
2470 	}
2471 
2472 	if (add_default_attributes())
2473 		goto out;
2474 
2475 	if (stat_config.cgroup_list) {
2476 		if (nr_cgroups > 0) {
2477 			pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2478 			parse_options_usage(stat_usage, stat_options, "G", 1);
2479 			parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2480 			goto out;
2481 		}
2482 
2483 		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2484 					  &stat_config.metric_events, true) < 0) {
2485 			parse_options_usage(stat_usage, stat_options,
2486 					    "for-each-cgroup", 0);
2487 			goto out;
2488 		}
2489 	}
2490 
2491 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2492 		target.per_thread = true;
2493 
2494 	if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2495 		pr_err("failed to use cpu list %s\n", target.cpu_list);
2496 		goto out;
2497 	}
2498 
2499 	target.hybrid = perf_pmu__has_hybrid();
2500 	if (evlist__create_maps(evsel_list, &target) < 0) {
2501 		if (target__has_task(&target)) {
2502 			pr_err("Problems finding threads of monitor\n");
2503 			parse_options_usage(stat_usage, stat_options, "p", 1);
2504 			parse_options_usage(NULL, stat_options, "t", 1);
2505 		} else if (target__has_cpu(&target)) {
2506 			perror("failed to parse CPUs map");
2507 			parse_options_usage(stat_usage, stat_options, "C", 1);
2508 			parse_options_usage(NULL, stat_options, "a", 1);
2509 		}
2510 		goto out;
2511 	}
2512 
2513 	evlist__check_cpu_maps(evsel_list);
2514 
2515 	/*
2516 	 * Initialize thread_map with comm names,
2517 	 * so we could print it out on output.
2518 	 */
2519 	if (stat_config.aggr_mode == AGGR_THREAD) {
2520 		thread_map__read_comms(evsel_list->core.threads);
2521 		if (target.system_wide) {
2522 			if (runtime_stat_new(&stat_config,
2523 				perf_thread_map__nr(evsel_list->core.threads))) {
2524 				goto out;
2525 			}
2526 		}
2527 	}
2528 
2529 	if (stat_config.aggr_mode == AGGR_NODE)
2530 		cpu__setup_cpunode_map();
2531 
2532 	if (stat_config.times && interval)
2533 		interval_count = true;
2534 	else if (stat_config.times && !interval) {
2535 		pr_err("interval-count option should be used together with "
2536 				"interval-print.\n");
2537 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2538 		parse_options_usage(stat_usage, stat_options, "I", 1);
2539 		goto out;
2540 	}
2541 
2542 	if (timeout && timeout < 100) {
2543 		if (timeout < 10) {
2544 			pr_err("timeout must be >= 10ms.\n");
2545 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2546 			goto out;
2547 		} else
2548 			pr_warning("timeout < 100ms. "
2549 				   "The overhead percentage could be high in some cases. "
2550 				   "Please proceed with caution.\n");
2551 	}
2552 	if (timeout && interval) {
2553 		pr_err("timeout option is not supported with interval-print.\n");
2554 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2555 		parse_options_usage(stat_usage, stat_options, "I", 1);
2556 		goto out;
2557 	}
2558 
2559 	if (evlist__alloc_stats(evsel_list, interval))
2560 		goto out;
2561 
2562 	if (perf_stat_init_aggr_mode())
2563 		goto out;
2564 
2565 	/*
2566 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2567 	 * while avoiding that older tools show confusing messages.
2568 	 *
2569 	 * However for pipe sessions we need to keep it zero,
2570 	 * because script's perf_evsel__check_attr is triggered
2571 	 * by attr->sample_type != 0, and we can't run it on
2572 	 * stat sessions.
2573 	 */
2574 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2575 
2576 	/*
2577 	 * We dont want to block the signals - that would cause
2578 	 * child tasks to inherit that and Ctrl-C would not work.
2579 	 * What we want is for Ctrl-C to work in the exec()-ed
2580 	 * task, but being ignored by perf stat itself:
2581 	 */
2582 	atexit(sig_atexit);
2583 	if (!forever)
2584 		signal(SIGINT,  skip_signal);
2585 	signal(SIGCHLD, skip_signal);
2586 	signal(SIGALRM, skip_signal);
2587 	signal(SIGABRT, skip_signal);
2588 
2589 	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2590 		goto out;
2591 
2592 	status = 0;
2593 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2594 		if (stat_config.run_count != 1 && verbose > 0)
2595 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2596 				run_idx + 1);
2597 
2598 		if (run_idx != 0)
2599 			evlist__reset_prev_raw_counts(evsel_list);
2600 
2601 		status = run_perf_stat(argc, argv, run_idx);
2602 		if (forever && status != -1 && !interval) {
2603 			print_counters(NULL, argc, argv);
2604 			perf_stat__reset_stats();
2605 		}
2606 	}
2607 
2608 	if (!forever && status != -1 && (!interval || stat_config.summary))
2609 		print_counters(NULL, argc, argv);
2610 
2611 	evlist__finalize_ctlfd(evsel_list);
2612 
2613 	if (STAT_RECORD) {
2614 		/*
2615 		 * We synthesize the kernel mmap record just so that older tools
2616 		 * don't emit warnings about not being able to resolve symbols
2617 		 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2618 		 * a saner message about no samples being in the perf.data file.
2619 		 *
2620 		 * This also serves to suppress a warning about f_header.data.size == 0
2621 		 * in header.c at the moment 'perf stat record' gets introduced, which
2622 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2623 		 * records, but the need to suppress the kptr_restrict messages in older
2624 		 * tools remain  -acme
2625 		 */
2626 		int fd = perf_data__fd(&perf_stat.data);
2627 
2628 		err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2629 							 process_synthesized_event,
2630 							 &perf_stat.session->machines.host);
2631 		if (err) {
2632 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2633 				   "older tools may produce warnings about this file\n.");
2634 		}
2635 
2636 		if (!interval) {
2637 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2638 				pr_err("failed to write stat round event\n");
2639 		}
2640 
2641 		if (!perf_stat.data.is_pipe) {
2642 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2643 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2644 		}
2645 
2646 		evlist__close(evsel_list);
2647 		perf_session__delete(perf_stat.session);
2648 	}
2649 
2650 	perf_stat__exit_aggr_mode();
2651 	evlist__free_stats(evsel_list);
2652 out:
2653 	if (stat_config.iostat_run)
2654 		iostat_release(evsel_list);
2655 
2656 	zfree(&stat_config.walltime_run);
2657 
2658 	if (smi_cost && smi_reset)
2659 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2660 
2661 	evlist__delete(evsel_list);
2662 
2663 	metricgroup__rblist_exit(&stat_config.metric_events);
2664 	runtime_stat_delete(&stat_config);
2665 	evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2666 
2667 	return status;
2668 }
2669