xref: /openbmc/linux/tools/perf/builtin-stat.c (revision 8e74a48d)
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 (perf_cpu_map__nr(a->core.cpus) != perf_cpu_map__nr(b->core.cpus))
234 		return false;
235 
236 	for (int i = 0; i < perf_cpu_map__nr(a->core.cpus); i++) {
237 		if (perf_cpu_map__cpu(a->core.cpus, i).cpu !=
238 		    perf_cpu_map__cpu(b->core.cpus, i).cpu)
239 			return false;
240 	}
241 
242 	return true;
243 }
244 
245 static void evlist__check_cpu_maps(struct evlist *evlist)
246 {
247 	struct evsel *evsel, *pos, *leader;
248 	char buf[1024];
249 
250 	if (evlist__has_hybrid(evlist))
251 		evlist__warn_hybrid_group(evlist);
252 
253 	evlist__for_each_entry(evlist, evsel) {
254 		leader = evsel__leader(evsel);
255 
256 		/* Check that leader matches cpus with each member. */
257 		if (leader == evsel)
258 			continue;
259 		if (cpus_map_matched(leader, evsel))
260 			continue;
261 
262 		/* If there's mismatch disable the group and warn user. */
263 		WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
264 		evsel__group_desc(leader, buf, sizeof(buf));
265 		pr_warning("  %s\n", buf);
266 
267 		if (verbose) {
268 			cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
269 			pr_warning("     %s: %s\n", leader->name, buf);
270 			cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
271 			pr_warning("     %s: %s\n", evsel->name, buf);
272 		}
273 
274 		for_each_group_evsel(pos, leader) {
275 			evsel__set_leader(pos, pos);
276 			pos->core.nr_members = 0;
277 		}
278 		evsel->core.leader->nr_members = 0;
279 	}
280 }
281 
282 static inline void diff_timespec(struct timespec *r, struct timespec *a,
283 				 struct timespec *b)
284 {
285 	r->tv_sec = a->tv_sec - b->tv_sec;
286 	if (a->tv_nsec < b->tv_nsec) {
287 		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
288 		r->tv_sec--;
289 	} else {
290 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
291 	}
292 }
293 
294 static void perf_stat__reset_stats(void)
295 {
296 	int i;
297 
298 	evlist__reset_stats(evsel_list);
299 	perf_stat__reset_shadow_stats();
300 
301 	for (i = 0; i < stat_config.stats_num; i++)
302 		perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
303 }
304 
305 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
306 				     union perf_event *event,
307 				     struct perf_sample *sample __maybe_unused,
308 				     struct machine *machine __maybe_unused)
309 {
310 	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
311 		pr_err("failed to write perf data, error: %m\n");
312 		return -1;
313 	}
314 
315 	perf_stat.bytes_written += event->header.size;
316 	return 0;
317 }
318 
319 static int write_stat_round_event(u64 tm, u64 type)
320 {
321 	return perf_event__synthesize_stat_round(NULL, tm, type,
322 						 process_synthesized_event,
323 						 NULL);
324 }
325 
326 #define WRITE_STAT_ROUND_EVENT(time, interval) \
327 	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
328 
329 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
330 
331 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
332 				   struct perf_counts_values *count)
333 {
334 	struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
335 	struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
336 
337 	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
338 					   process_synthesized_event, NULL);
339 }
340 
341 static int read_single_counter(struct evsel *counter, int cpu_map_idx,
342 			       int thread, struct timespec *rs)
343 {
344 	if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
345 		u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
346 		struct perf_counts_values *count =
347 			perf_counts(counter->counts, cpu_map_idx, thread);
348 		count->ena = count->run = val;
349 		count->val = val;
350 		return 0;
351 	}
352 	return evsel__read_counter(counter, cpu_map_idx, thread);
353 }
354 
355 /*
356  * Read out the results of a single counter:
357  * do not aggregate counts across CPUs in system-wide mode
358  */
359 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx)
360 {
361 	int nthreads = perf_thread_map__nr(evsel_list->core.threads);
362 	int thread;
363 
364 	if (!counter->supported)
365 		return -ENOENT;
366 
367 	if (counter->core.system_wide)
368 		nthreads = 1;
369 
370 	for (thread = 0; thread < nthreads; thread++) {
371 		struct perf_counts_values *count;
372 
373 		count = perf_counts(counter->counts, cpu_map_idx, thread);
374 
375 		/*
376 		 * The leader's group read loads data into its group members
377 		 * (via evsel__read_counter()) and sets their count->loaded.
378 		 */
379 		if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
380 		    read_single_counter(counter, cpu_map_idx, thread, rs)) {
381 			counter->counts->scaled = -1;
382 			perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
383 			perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
384 			return -1;
385 		}
386 
387 		perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
388 
389 		if (STAT_RECORD) {
390 			if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
391 				pr_err("failed to write stat event\n");
392 				return -1;
393 			}
394 		}
395 
396 		if (verbose > 1) {
397 			fprintf(stat_config.output,
398 				"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
399 					evsel__name(counter),
400 					perf_cpu_map__cpu(evsel__cpus(counter),
401 							  cpu_map_idx).cpu,
402 					count->val, count->ena, count->run);
403 		}
404 	}
405 
406 	return 0;
407 }
408 
409 static int read_affinity_counters(struct timespec *rs)
410 {
411 	struct evlist_cpu_iterator evlist_cpu_itr;
412 	struct affinity saved_affinity, *affinity;
413 
414 	if (all_counters_use_bpf)
415 		return 0;
416 
417 	if (!target__has_cpu(&target) || target__has_per_thread(&target))
418 		affinity = NULL;
419 	else if (affinity__setup(&saved_affinity) < 0)
420 		return -1;
421 	else
422 		affinity = &saved_affinity;
423 
424 	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
425 		struct evsel *counter = evlist_cpu_itr.evsel;
426 
427 		if (evsel__is_bpf(counter))
428 			continue;
429 
430 		if (!counter->err) {
431 			counter->err = read_counter_cpu(counter, rs,
432 							evlist_cpu_itr.cpu_map_idx);
433 		}
434 	}
435 	if (affinity)
436 		affinity__cleanup(&saved_affinity);
437 
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 evlist_cpu_iterator evlist_cpu_itr;
792 	struct affinity saved_affinity, *affinity = NULL;
793 	int err;
794 	bool second_pass = false;
795 
796 	if (forks) {
797 		if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
798 			perror("failed to prepare workload");
799 			return -1;
800 		}
801 		child_pid = evsel_list->workload.pid;
802 	}
803 
804 	if (group)
805 		evlist__set_leader(evsel_list);
806 
807 	if (!cpu_map__is_dummy(evsel_list->core.cpus)) {
808 		if (affinity__setup(&saved_affinity) < 0)
809 			return -1;
810 		affinity = &saved_affinity;
811 	}
812 
813 	evlist__for_each_entry(evsel_list, counter) {
814 		if (bpf_counter__load(counter, &target))
815 			return -1;
816 		if (!evsel__is_bpf(counter))
817 			all_counters_use_bpf = false;
818 	}
819 
820 	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
821 		counter = evlist_cpu_itr.evsel;
822 
823 		/*
824 		 * bperf calls evsel__open_per_cpu() in bperf__load(), so
825 		 * no need to call it again here.
826 		 */
827 		if (target.use_bpf)
828 			break;
829 
830 		if (counter->reset_group || counter->errored)
831 			continue;
832 		if (evsel__is_bpf(counter))
833 			continue;
834 try_again:
835 		if (create_perf_stat_counter(counter, &stat_config, &target,
836 					     evlist_cpu_itr.cpu_map_idx) < 0) {
837 
838 			/*
839 			 * Weak group failed. We cannot just undo this here
840 			 * because earlier CPUs might be in group mode, and the kernel
841 			 * doesn't support mixing group and non group reads. Defer
842 			 * it to later.
843 			 * Don't close here because we're in the wrong affinity.
844 			 */
845 			if ((errno == EINVAL || errno == EBADF) &&
846 				evsel__leader(counter) != counter &&
847 				counter->weak_group) {
848 				evlist__reset_weak_group(evsel_list, counter, false);
849 				assert(counter->reset_group);
850 				second_pass = true;
851 				continue;
852 			}
853 
854 			switch (stat_handle_error(counter)) {
855 			case COUNTER_FATAL:
856 				return -1;
857 			case COUNTER_RETRY:
858 				goto try_again;
859 			case COUNTER_SKIP:
860 				continue;
861 			default:
862 				break;
863 			}
864 
865 		}
866 		counter->supported = true;
867 	}
868 
869 	if (second_pass) {
870 		/*
871 		 * Now redo all the weak group after closing them,
872 		 * and also close errored counters.
873 		 */
874 
875 		/* First close errored or weak retry */
876 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
877 			counter = evlist_cpu_itr.evsel;
878 
879 			if (!counter->reset_group && !counter->errored)
880 				continue;
881 
882 			perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
883 		}
884 		/* Now reopen weak */
885 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
886 			counter = evlist_cpu_itr.evsel;
887 
888 			if (!counter->reset_group && !counter->errored)
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 						     evlist_cpu_itr.cpu_map_idx) < 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 	affinity__cleanup(affinity);
912 
913 	evlist__for_each_entry(evsel_list, counter) {
914 		if (!counter->supported) {
915 			perf_evsel__free_fd(&counter->core);
916 			continue;
917 		}
918 
919 		l = strlen(counter->unit);
920 		if (l > stat_config.unit_width)
921 			stat_config.unit_width = l;
922 
923 		if (evsel__should_store_id(counter) &&
924 		    evsel__store_ids(counter, evsel_list))
925 			return -1;
926 	}
927 
928 	if (evlist__apply_filters(evsel_list, &counter)) {
929 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
930 			counter->filter, evsel__name(counter), errno,
931 			str_error_r(errno, msg, sizeof(msg)));
932 		return -1;
933 	}
934 
935 	if (STAT_RECORD) {
936 		int fd = perf_data__fd(&perf_stat.data);
937 
938 		if (is_pipe) {
939 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
940 		} else {
941 			err = perf_session__write_header(perf_stat.session, evsel_list,
942 							 fd, false);
943 		}
944 
945 		if (err < 0)
946 			return err;
947 
948 		err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
949 							 process_synthesized_event, is_pipe);
950 		if (err < 0)
951 			return err;
952 	}
953 
954 	/*
955 	 * Enable counters and exec the command:
956 	 */
957 	if (forks) {
958 		evlist__start_workload(evsel_list);
959 		err = enable_counters();
960 		if (err)
961 			return -1;
962 
963 		t0 = rdclock();
964 		clock_gettime(CLOCK_MONOTONIC, &ref_time);
965 
966 		if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
967 			status = dispatch_events(forks, timeout, interval, &times);
968 		if (child_pid != -1) {
969 			if (timeout)
970 				kill(child_pid, SIGTERM);
971 			wait4(child_pid, &status, 0, &stat_config.ru_data);
972 		}
973 
974 		if (workload_exec_errno) {
975 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
976 			pr_err("Workload failed: %s\n", emsg);
977 			return -1;
978 		}
979 
980 		if (WIFSIGNALED(status))
981 			psignal(WTERMSIG(status), argv[0]);
982 	} else {
983 		err = enable_counters();
984 		if (err)
985 			return -1;
986 
987 		t0 = rdclock();
988 		clock_gettime(CLOCK_MONOTONIC, &ref_time);
989 
990 		status = dispatch_events(forks, timeout, interval, &times);
991 	}
992 
993 	disable_counters();
994 
995 	t1 = rdclock();
996 
997 	if (stat_config.walltime_run_table)
998 		stat_config.walltime_run[run_idx] = t1 - t0;
999 
1000 	if (interval && stat_config.summary) {
1001 		stat_config.interval = 0;
1002 		stat_config.stop_read_counter = true;
1003 		init_stats(&walltime_nsecs_stats);
1004 		update_stats(&walltime_nsecs_stats, t1 - t0);
1005 
1006 		if (stat_config.aggr_mode == AGGR_GLOBAL)
1007 			evlist__save_aggr_prev_raw_counts(evsel_list);
1008 
1009 		evlist__copy_prev_raw_counts(evsel_list);
1010 		evlist__reset_prev_raw_counts(evsel_list);
1011 		runtime_stat_reset(&stat_config);
1012 		perf_stat__reset_shadow_per_stat(&rt_stat);
1013 	} else
1014 		update_stats(&walltime_nsecs_stats, t1 - t0);
1015 
1016 	/*
1017 	 * Closing a group leader splits the group, and as we only disable
1018 	 * group leaders, results in remaining events becoming enabled. To
1019 	 * avoid arbitrary skew, we must read all counters before closing any
1020 	 * group leaders.
1021 	 */
1022 	read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
1023 
1024 	/*
1025 	 * We need to keep evsel_list alive, because it's processed
1026 	 * later the evsel_list will be closed after.
1027 	 */
1028 	if (!STAT_RECORD)
1029 		evlist__close(evsel_list);
1030 
1031 	return WEXITSTATUS(status);
1032 }
1033 
1034 static int run_perf_stat(int argc, const char **argv, int run_idx)
1035 {
1036 	int ret;
1037 
1038 	if (pre_cmd) {
1039 		ret = system(pre_cmd);
1040 		if (ret)
1041 			return ret;
1042 	}
1043 
1044 	if (sync_run)
1045 		sync();
1046 
1047 	ret = __run_perf_stat(argc, argv, run_idx);
1048 	if (ret)
1049 		return ret;
1050 
1051 	if (post_cmd) {
1052 		ret = system(post_cmd);
1053 		if (ret)
1054 			return ret;
1055 	}
1056 
1057 	return ret;
1058 }
1059 
1060 static void print_counters(struct timespec *ts, int argc, const char **argv)
1061 {
1062 	/* Do not print anything if we record to the pipe. */
1063 	if (STAT_RECORD && perf_stat.data.is_pipe)
1064 		return;
1065 	if (stat_config.quiet)
1066 		return;
1067 
1068 	evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1069 }
1070 
1071 static volatile int signr = -1;
1072 
1073 static void skip_signal(int signo)
1074 {
1075 	if ((child_pid == -1) || stat_config.interval)
1076 		done = 1;
1077 
1078 	signr = signo;
1079 	/*
1080 	 * render child_pid harmless
1081 	 * won't send SIGTERM to a random
1082 	 * process in case of race condition
1083 	 * and fast PID recycling
1084 	 */
1085 	child_pid = -1;
1086 }
1087 
1088 static void sig_atexit(void)
1089 {
1090 	sigset_t set, oset;
1091 
1092 	/*
1093 	 * avoid race condition with SIGCHLD handler
1094 	 * in skip_signal() which is modifying child_pid
1095 	 * goal is to avoid send SIGTERM to a random
1096 	 * process
1097 	 */
1098 	sigemptyset(&set);
1099 	sigaddset(&set, SIGCHLD);
1100 	sigprocmask(SIG_BLOCK, &set, &oset);
1101 
1102 	if (child_pid != -1)
1103 		kill(child_pid, SIGTERM);
1104 
1105 	sigprocmask(SIG_SETMASK, &oset, NULL);
1106 
1107 	if (signr == -1)
1108 		return;
1109 
1110 	signal(signr, SIG_DFL);
1111 	kill(getpid(), signr);
1112 }
1113 
1114 void perf_stat__set_big_num(int set)
1115 {
1116 	stat_config.big_num = (set != 0);
1117 }
1118 
1119 void perf_stat__set_no_csv_summary(int set)
1120 {
1121 	stat_config.no_csv_summary = (set != 0);
1122 }
1123 
1124 static int stat__set_big_num(const struct option *opt __maybe_unused,
1125 			     const char *s __maybe_unused, int unset)
1126 {
1127 	big_num_opt = unset ? 0 : 1;
1128 	perf_stat__set_big_num(!unset);
1129 	return 0;
1130 }
1131 
1132 static int enable_metric_only(const struct option *opt __maybe_unused,
1133 			      const char *s __maybe_unused, int unset)
1134 {
1135 	force_metric_only = true;
1136 	stat_config.metric_only = !unset;
1137 	return 0;
1138 }
1139 
1140 static int parse_metric_groups(const struct option *opt,
1141 			       const char *str,
1142 			       int unset __maybe_unused)
1143 {
1144 	return metricgroup__parse_groups(opt, str,
1145 					 stat_config.metric_no_group,
1146 					 stat_config.metric_no_merge,
1147 					 &stat_config.metric_events);
1148 }
1149 
1150 static int parse_control_option(const struct option *opt,
1151 				const char *str,
1152 				int unset __maybe_unused)
1153 {
1154 	struct perf_stat_config *config = opt->value;
1155 
1156 	return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1157 }
1158 
1159 static int parse_stat_cgroups(const struct option *opt,
1160 			      const char *str, int unset)
1161 {
1162 	if (stat_config.cgroup_list) {
1163 		pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1164 		return -1;
1165 	}
1166 
1167 	return parse_cgroups(opt, str, unset);
1168 }
1169 
1170 static int parse_hybrid_type(const struct option *opt,
1171 			     const char *str,
1172 			     int unset __maybe_unused)
1173 {
1174 	struct evlist *evlist = *(struct evlist **)opt->value;
1175 
1176 	if (!list_empty(&evlist->core.entries)) {
1177 		fprintf(stderr, "Must define cputype before events/metrics\n");
1178 		return -1;
1179 	}
1180 
1181 	evlist->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu(str);
1182 	if (!evlist->hybrid_pmu_name) {
1183 		fprintf(stderr, "--cputype %s is not supported!\n", str);
1184 		return -1;
1185 	}
1186 
1187 	return 0;
1188 }
1189 
1190 static struct option stat_options[] = {
1191 	OPT_BOOLEAN('T', "transaction", &transaction_run,
1192 		    "hardware transaction statistics"),
1193 	OPT_CALLBACK('e', "event", &evsel_list, "event",
1194 		     "event selector. use 'perf list' to list available events",
1195 		     parse_events_option),
1196 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1197 		     "event filter", parse_filter),
1198 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1199 		    "child tasks do not inherit counters"),
1200 	OPT_STRING('p', "pid", &target.pid, "pid",
1201 		   "stat events on existing process id"),
1202 	OPT_STRING('t', "tid", &target.tid, "tid",
1203 		   "stat events on existing thread id"),
1204 #ifdef HAVE_BPF_SKEL
1205 	OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1206 		   "stat events on existing bpf program id"),
1207 	OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1208 		    "use bpf program to count events"),
1209 	OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1210 		   "path to perf_event_attr map"),
1211 #endif
1212 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1213 		    "system-wide collection from all CPUs"),
1214 	OPT_BOOLEAN('g', "group", &group,
1215 		    "put the counters into a counter group"),
1216 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
1217 		    "Use --no-scale to disable counter scaling for multiplexing"),
1218 	OPT_INCR('v', "verbose", &verbose,
1219 		    "be more verbose (show counter open errors, etc)"),
1220 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
1221 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1222 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1223 		    "display details about each run (only with -r option)"),
1224 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
1225 		    "null run - dont start any counters"),
1226 	OPT_INCR('d', "detailed", &detailed_run,
1227 		    "detailed run - start a lot of events"),
1228 	OPT_BOOLEAN('S', "sync", &sync_run,
1229 		    "call sync() before starting a run"),
1230 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1231 			   "print large numbers with thousands\' separators",
1232 			   stat__set_big_num),
1233 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1234 		    "list of cpus to monitor in system-wide"),
1235 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1236 		    "disable CPU count aggregation", AGGR_NONE),
1237 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1238 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1239 		   "print counts with custom separator"),
1240 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1241 		     "monitor event in cgroup name only", parse_stat_cgroups),
1242 	OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1243 		    "expand events for each cgroup"),
1244 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1245 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1246 	OPT_INTEGER(0, "log-fd", &output_fd,
1247 		    "log output to fd, instead of stderr"),
1248 	OPT_STRING(0, "pre", &pre_cmd, "command",
1249 			"command to run prior to the measured command"),
1250 	OPT_STRING(0, "post", &post_cmd, "command",
1251 			"command to run after to the measured command"),
1252 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1253 		    "print counts at regular interval in ms "
1254 		    "(overhead is possible for values <= 100ms)"),
1255 	OPT_INTEGER(0, "interval-count", &stat_config.times,
1256 		    "print counts for fixed number of times"),
1257 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1258 		    "clear screen in between new interval"),
1259 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1260 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1261 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1262 		     "aggregate counts per processor socket", AGGR_SOCKET),
1263 	OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1264 		     "aggregate counts per processor die", AGGR_DIE),
1265 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1266 		     "aggregate counts per physical processor core", AGGR_CORE),
1267 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1268 		     "aggregate counts per thread", AGGR_THREAD),
1269 	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1270 		     "aggregate counts per numa node", AGGR_NODE),
1271 	OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1272 		    "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1273 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1274 			"Only print computed metrics. No raw values", enable_metric_only),
1275 	OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1276 		       "don't group metric events, impacts multiplexing"),
1277 	OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1278 		       "don't try to share events between metrics in a group"),
1279 	OPT_BOOLEAN(0, "topdown", &topdown_run,
1280 			"measure top-down statistics"),
1281 	OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1282 			"Set the metrics level for the top-down statistics (0: max level)"),
1283 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1284 			"measure SMI cost"),
1285 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1286 		     "monitor specified metrics or metric groups (separated by ,)",
1287 		     parse_metric_groups),
1288 	OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1289 			 "Configure all used events to run in kernel space.",
1290 			 PARSE_OPT_EXCLUSIVE),
1291 	OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1292 			 "Configure all used events to run in user space.",
1293 			 PARSE_OPT_EXCLUSIVE),
1294 	OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1295 		    "Use with 'percore' event qualifier to show the event "
1296 		    "counts of one hardware thread by sum up total hardware "
1297 		    "threads of same physical core"),
1298 	OPT_BOOLEAN(0, "summary", &stat_config.summary,
1299 		       "print summary for interval mode"),
1300 	OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1301 		       "don't print 'summary' for CSV summary output"),
1302 	OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1303 			"don't print output (useful with record)"),
1304 	OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1305 		     "Only enable events on applying cpu with this type "
1306 		     "for hybrid platform (e.g. core or atom)",
1307 		     parse_hybrid_type),
1308 #ifdef HAVE_LIBPFM
1309 	OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1310 		"libpfm4 event selector. use 'perf list' to list available events",
1311 		parse_libpfm_events_option),
1312 #endif
1313 	OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1314 		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1315 		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1316 		     "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1317 		      parse_control_option),
1318 	OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1319 			    "measure I/O performance metrics provided by arch/platform",
1320 			    iostat_parse),
1321 	OPT_END()
1322 };
1323 
1324 static const char *const aggr_mode__string[] = {
1325 	[AGGR_CORE] = "core",
1326 	[AGGR_DIE] = "die",
1327 	[AGGR_GLOBAL] = "global",
1328 	[AGGR_NODE] = "node",
1329 	[AGGR_NONE] = "none",
1330 	[AGGR_SOCKET] = "socket",
1331 	[AGGR_THREAD] = "thread",
1332 	[AGGR_UNSET] = "unset",
1333 };
1334 
1335 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1336 						struct perf_cpu cpu)
1337 {
1338 	return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1339 }
1340 
1341 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1342 					     struct perf_cpu cpu)
1343 {
1344 	return aggr_cpu_id__die(cpu, /*data=*/NULL);
1345 }
1346 
1347 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1348 					      struct perf_cpu cpu)
1349 {
1350 	return aggr_cpu_id__core(cpu, /*data=*/NULL);
1351 }
1352 
1353 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1354 					      struct perf_cpu cpu)
1355 {
1356 	return aggr_cpu_id__node(cpu, /*data=*/NULL);
1357 }
1358 
1359 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1360 					      aggr_get_id_t get_id, struct perf_cpu cpu)
1361 {
1362 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1363 
1364 	if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1365 		config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1366 
1367 	id = config->cpus_aggr_map->map[cpu.cpu];
1368 	return id;
1369 }
1370 
1371 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1372 						       struct perf_cpu cpu)
1373 {
1374 	return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1375 }
1376 
1377 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1378 						    struct perf_cpu cpu)
1379 {
1380 	return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1381 }
1382 
1383 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1384 						     struct perf_cpu cpu)
1385 {
1386 	return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1387 }
1388 
1389 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1390 						     struct perf_cpu cpu)
1391 {
1392 	return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1393 }
1394 
1395 static bool term_percore_set(void)
1396 {
1397 	struct evsel *counter;
1398 
1399 	evlist__for_each_entry(evsel_list, counter) {
1400 		if (counter->percore)
1401 			return true;
1402 	}
1403 
1404 	return false;
1405 }
1406 
1407 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1408 {
1409 	switch (aggr_mode) {
1410 	case AGGR_SOCKET:
1411 		return aggr_cpu_id__socket;
1412 	case AGGR_DIE:
1413 		return aggr_cpu_id__die;
1414 	case AGGR_CORE:
1415 		return aggr_cpu_id__core;
1416 	case AGGR_NODE:
1417 		return aggr_cpu_id__node;
1418 	case AGGR_NONE:
1419 		if (term_percore_set())
1420 			return aggr_cpu_id__core;
1421 
1422 		return NULL;
1423 	case AGGR_GLOBAL:
1424 	case AGGR_THREAD:
1425 	case AGGR_UNSET:
1426 	default:
1427 		return NULL;
1428 	}
1429 }
1430 
1431 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1432 {
1433 	switch (aggr_mode) {
1434 	case AGGR_SOCKET:
1435 		return perf_stat__get_socket_cached;
1436 	case AGGR_DIE:
1437 		return perf_stat__get_die_cached;
1438 	case AGGR_CORE:
1439 		return perf_stat__get_core_cached;
1440 	case AGGR_NODE:
1441 		return perf_stat__get_node_cached;
1442 	case AGGR_NONE:
1443 		if (term_percore_set()) {
1444 			return perf_stat__get_core_cached;
1445 		}
1446 		return NULL;
1447 	case AGGR_GLOBAL:
1448 	case AGGR_THREAD:
1449 	case AGGR_UNSET:
1450 	default:
1451 		return NULL;
1452 	}
1453 }
1454 
1455 static int perf_stat_init_aggr_mode(void)
1456 {
1457 	int nr;
1458 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1459 
1460 	if (get_id) {
1461 		stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.cpus,
1462 							 get_id, /*data=*/NULL);
1463 		if (!stat_config.aggr_map) {
1464 			pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1465 			return -1;
1466 		}
1467 		stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1468 	}
1469 
1470 	/*
1471 	 * The evsel_list->cpus is the base we operate on,
1472 	 * taking the highest cpu number to be the size of
1473 	 * the aggregation translate cpumap.
1474 	 */
1475 	nr = perf_cpu_map__max(evsel_list->core.cpus).cpu;
1476 	stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1477 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1478 }
1479 
1480 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1481 {
1482 	if (map) {
1483 		WARN_ONCE(refcount_read(&map->refcnt) != 0,
1484 			  "cpu_aggr_map refcnt unbalanced\n");
1485 		free(map);
1486 	}
1487 }
1488 
1489 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1490 {
1491 	if (map && refcount_dec_and_test(&map->refcnt))
1492 		cpu_aggr_map__delete(map);
1493 }
1494 
1495 static void perf_stat__exit_aggr_mode(void)
1496 {
1497 	cpu_aggr_map__put(stat_config.aggr_map);
1498 	cpu_aggr_map__put(stat_config.cpus_aggr_map);
1499 	stat_config.aggr_map = NULL;
1500 	stat_config.cpus_aggr_map = NULL;
1501 }
1502 
1503 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1504 {
1505 	struct perf_env *env = data;
1506 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1507 
1508 	if (cpu.cpu != -1)
1509 		id.socket = env->cpu[cpu.cpu].socket_id;
1510 
1511 	return id;
1512 }
1513 
1514 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1515 {
1516 	struct perf_env *env = data;
1517 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1518 
1519 	if (cpu.cpu != -1) {
1520 		/*
1521 		 * die_id is relative to socket, so start
1522 		 * with the socket ID and then add die to
1523 		 * make a unique ID.
1524 		 */
1525 		id.socket = env->cpu[cpu.cpu].socket_id;
1526 		id.die = env->cpu[cpu.cpu].die_id;
1527 	}
1528 
1529 	return id;
1530 }
1531 
1532 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1533 {
1534 	struct perf_env *env = data;
1535 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1536 
1537 	if (cpu.cpu != -1) {
1538 		/*
1539 		 * core_id is relative to socket and die,
1540 		 * we need a global id. So we set
1541 		 * socket, die id and core id
1542 		 */
1543 		id.socket = env->cpu[cpu.cpu].socket_id;
1544 		id.die = env->cpu[cpu.cpu].die_id;
1545 		id.core = env->cpu[cpu.cpu].core_id;
1546 	}
1547 
1548 	return id;
1549 }
1550 
1551 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1552 {
1553 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1554 
1555 	id.node = perf_env__numa_node(data, cpu);
1556 	return id;
1557 }
1558 
1559 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1560 						     struct perf_cpu cpu)
1561 {
1562 	return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1563 }
1564 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1565 						  struct perf_cpu cpu)
1566 {
1567 	return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1568 }
1569 
1570 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1571 						   struct perf_cpu cpu)
1572 {
1573 	return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1574 }
1575 
1576 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1577 						   struct perf_cpu cpu)
1578 {
1579 	return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1580 }
1581 
1582 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1583 {
1584 	switch (aggr_mode) {
1585 	case AGGR_SOCKET:
1586 		return perf_env__get_socket_aggr_by_cpu;
1587 	case AGGR_DIE:
1588 		return perf_env__get_die_aggr_by_cpu;
1589 	case AGGR_CORE:
1590 		return perf_env__get_core_aggr_by_cpu;
1591 	case AGGR_NODE:
1592 		return perf_env__get_node_aggr_by_cpu;
1593 	case AGGR_NONE:
1594 	case AGGR_GLOBAL:
1595 	case AGGR_THREAD:
1596 	case AGGR_UNSET:
1597 	default:
1598 		return NULL;
1599 	}
1600 }
1601 
1602 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1603 {
1604 	switch (aggr_mode) {
1605 	case AGGR_SOCKET:
1606 		return perf_stat__get_socket_file;
1607 	case AGGR_DIE:
1608 		return perf_stat__get_die_file;
1609 	case AGGR_CORE:
1610 		return perf_stat__get_core_file;
1611 	case AGGR_NODE:
1612 		return perf_stat__get_node_file;
1613 	case AGGR_NONE:
1614 	case AGGR_GLOBAL:
1615 	case AGGR_THREAD:
1616 	case AGGR_UNSET:
1617 	default:
1618 		return NULL;
1619 	}
1620 }
1621 
1622 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1623 {
1624 	struct perf_env *env = &st->session->header.env;
1625 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1626 
1627 	if (!get_id)
1628 		return 0;
1629 
1630 	stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.cpus, get_id, env);
1631 	if (!stat_config.aggr_map) {
1632 		pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1633 		return -1;
1634 	}
1635 	stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1636 	return 0;
1637 }
1638 
1639 /*
1640  * Add default attributes, if there were no attributes specified or
1641  * if -d/--detailed, -d -d or -d -d -d is used:
1642  */
1643 static int add_default_attributes(void)
1644 {
1645 	int err;
1646 	struct perf_event_attr default_attrs0[] = {
1647 
1648   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1649   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1650   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1651   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1652 
1653   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1654 };
1655 	struct perf_event_attr frontend_attrs[] = {
1656   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1657 };
1658 	struct perf_event_attr backend_attrs[] = {
1659   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1660 };
1661 	struct perf_event_attr default_attrs1[] = {
1662   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1663   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1664   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1665 
1666 };
1667 	struct perf_event_attr default_sw_attrs[] = {
1668   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1669   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1670   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1671   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1672 };
1673 
1674 /*
1675  * Detailed stats (-d), covering the L1 and last level data caches:
1676  */
1677 	struct perf_event_attr detailed_attrs[] = {
1678 
1679   { .type = PERF_TYPE_HW_CACHE,
1680     .config =
1681 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1682 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1683 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1684 
1685   { .type = PERF_TYPE_HW_CACHE,
1686     .config =
1687 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1688 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1689 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1690 
1691   { .type = PERF_TYPE_HW_CACHE,
1692     .config =
1693 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1694 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1695 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1696 
1697   { .type = PERF_TYPE_HW_CACHE,
1698     .config =
1699 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1700 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1701 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1702 };
1703 
1704 /*
1705  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1706  */
1707 	struct perf_event_attr very_detailed_attrs[] = {
1708 
1709   { .type = PERF_TYPE_HW_CACHE,
1710     .config =
1711 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1712 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1713 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1714 
1715   { .type = PERF_TYPE_HW_CACHE,
1716     .config =
1717 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1718 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1719 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1720 
1721   { .type = PERF_TYPE_HW_CACHE,
1722     .config =
1723 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1724 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1725 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1726 
1727   { .type = PERF_TYPE_HW_CACHE,
1728     .config =
1729 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1730 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1731 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1732 
1733   { .type = PERF_TYPE_HW_CACHE,
1734     .config =
1735 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1736 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1737 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1738 
1739   { .type = PERF_TYPE_HW_CACHE,
1740     .config =
1741 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1742 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1743 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1744 
1745 };
1746 
1747 /*
1748  * Very, very detailed stats (-d -d -d), adding prefetch events:
1749  */
1750 	struct perf_event_attr very_very_detailed_attrs[] = {
1751 
1752   { .type = PERF_TYPE_HW_CACHE,
1753     .config =
1754 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1755 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1756 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1757 
1758   { .type = PERF_TYPE_HW_CACHE,
1759     .config =
1760 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1761 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1762 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1763 };
1764 	/* Set attrs if no event is selected and !null_run: */
1765 	if (stat_config.null_run)
1766 		return 0;
1767 
1768 	if (transaction_run) {
1769 		struct parse_events_error errinfo;
1770 		/* Handle -T as -M transaction. Once platform specific metrics
1771 		 * support has been added to the json files, all architectures
1772 		 * will use this approach. To determine transaction support
1773 		 * on an architecture test for such a metric name.
1774 		 */
1775 		if (metricgroup__has_metric("transaction")) {
1776 			struct option opt = { .value = &evsel_list };
1777 
1778 			return metricgroup__parse_groups(&opt, "transaction",
1779 							 stat_config.metric_no_group,
1780 							stat_config.metric_no_merge,
1781 							 &stat_config.metric_events);
1782 		}
1783 
1784 		parse_events_error__init(&errinfo);
1785 		if (pmu_have_event("cpu", "cycles-ct") &&
1786 		    pmu_have_event("cpu", "el-start"))
1787 			err = parse_events(evsel_list, transaction_attrs,
1788 					   &errinfo);
1789 		else
1790 			err = parse_events(evsel_list,
1791 					   transaction_limited_attrs,
1792 					   &errinfo);
1793 		if (err) {
1794 			fprintf(stderr, "Cannot set up transaction events\n");
1795 			parse_events_error__print(&errinfo, transaction_attrs);
1796 		}
1797 		parse_events_error__exit(&errinfo);
1798 		return err ? -1 : 0;
1799 	}
1800 
1801 	if (smi_cost) {
1802 		struct parse_events_error errinfo;
1803 		int smi;
1804 
1805 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1806 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1807 			return -1;
1808 		}
1809 
1810 		if (!smi) {
1811 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1812 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1813 				return -1;
1814 			}
1815 			smi_reset = true;
1816 		}
1817 
1818 		if (!pmu_have_event("msr", "aperf") ||
1819 		    !pmu_have_event("msr", "smi")) {
1820 			fprintf(stderr, "To measure SMI cost, it needs "
1821 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1822 			return -1;
1823 		}
1824 		if (!force_metric_only)
1825 			stat_config.metric_only = true;
1826 
1827 		parse_events_error__init(&errinfo);
1828 		err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1829 		if (err) {
1830 			parse_events_error__print(&errinfo, smi_cost_attrs);
1831 			fprintf(stderr, "Cannot set up SMI cost events\n");
1832 		}
1833 		parse_events_error__exit(&errinfo);
1834 		return err ? -1 : 0;
1835 	}
1836 
1837 	if (topdown_run) {
1838 		const char **metric_attrs = topdown_metric_attrs;
1839 		unsigned int max_level = 1;
1840 		char *str = NULL;
1841 		bool warn = false;
1842 
1843 		if (!force_metric_only)
1844 			stat_config.metric_only = true;
1845 
1846 		if (pmu_have_event("cpu", topdown_metric_L2_attrs[5])) {
1847 			metric_attrs = topdown_metric_L2_attrs;
1848 			max_level = 2;
1849 		}
1850 
1851 		if (stat_config.topdown_level > max_level) {
1852 			pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1853 			return -1;
1854 		} else if (!stat_config.topdown_level)
1855 			stat_config.topdown_level = max_level;
1856 
1857 		if (topdown_filter_events(metric_attrs, &str, 1) < 0) {
1858 			pr_err("Out of memory\n");
1859 			return -1;
1860 		}
1861 		if (metric_attrs[0] && str) {
1862 			if (!stat_config.interval && !stat_config.metric_only) {
1863 				fprintf(stat_config.output,
1864 					"Topdown accuracy may decrease when measuring long periods.\n"
1865 					"Please print the result regularly, e.g. -I1000\n");
1866 			}
1867 			goto setup_metrics;
1868 		}
1869 
1870 		zfree(&str);
1871 
1872 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1873 		    stat_config.aggr_mode != AGGR_CORE) {
1874 			pr_err("top down event configuration requires --per-core mode\n");
1875 			return -1;
1876 		}
1877 		stat_config.aggr_mode = AGGR_CORE;
1878 		if (nr_cgroups || !target__has_cpu(&target)) {
1879 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1880 			return -1;
1881 		}
1882 
1883 		if (topdown_filter_events(topdown_attrs, &str,
1884 				arch_topdown_check_group(&warn)) < 0) {
1885 			pr_err("Out of memory\n");
1886 			return -1;
1887 		}
1888 		if (topdown_attrs[0] && str) {
1889 			struct parse_events_error errinfo;
1890 			if (warn)
1891 				arch_topdown_group_warn();
1892 setup_metrics:
1893 			parse_events_error__init(&errinfo);
1894 			err = parse_events(evsel_list, str, &errinfo);
1895 			if (err) {
1896 				fprintf(stderr,
1897 					"Cannot set up top down events %s: %d\n",
1898 					str, err);
1899 				parse_events_error__print(&errinfo, str);
1900 				parse_events_error__exit(&errinfo);
1901 				free(str);
1902 				return -1;
1903 			}
1904 			parse_events_error__exit(&errinfo);
1905 		} else {
1906 			fprintf(stderr, "System does not support topdown\n");
1907 			return -1;
1908 		}
1909 		free(str);
1910 	}
1911 
1912 	if (!evsel_list->core.nr_entries) {
1913 		if (perf_pmu__has_hybrid()) {
1914 			struct parse_events_error errinfo;
1915 			const char *hybrid_str = "cycles,instructions,branches,branch-misses";
1916 
1917 			if (target__has_cpu(&target))
1918 				default_sw_attrs[0].config = PERF_COUNT_SW_CPU_CLOCK;
1919 
1920 			if (evlist__add_default_attrs(evsel_list,
1921 						      default_sw_attrs) < 0) {
1922 				return -1;
1923 			}
1924 
1925 			parse_events_error__init(&errinfo);
1926 			err = parse_events(evsel_list, hybrid_str, &errinfo);
1927 			if (err) {
1928 				fprintf(stderr,
1929 					"Cannot set up hybrid events %s: %d\n",
1930 					hybrid_str, err);
1931 				parse_events_error__print(&errinfo, hybrid_str);
1932 			}
1933 			parse_events_error__exit(&errinfo);
1934 			return err ? -1 : 0;
1935 		}
1936 
1937 		if (target__has_cpu(&target))
1938 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1939 
1940 		if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1941 			return -1;
1942 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1943 			if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1944 				return -1;
1945 		}
1946 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1947 			if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1948 				return -1;
1949 		}
1950 		if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1951 			return -1;
1952 
1953 		stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1954 		if (arch_evlist__add_default_attrs(evsel_list) < 0)
1955 			return -1;
1956 	}
1957 
1958 	/* Detailed events get appended to the event list: */
1959 
1960 	if (detailed_run <  1)
1961 		return 0;
1962 
1963 	/* Append detailed run extra attributes: */
1964 	if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1965 		return -1;
1966 
1967 	if (detailed_run < 2)
1968 		return 0;
1969 
1970 	/* Append very detailed run extra attributes: */
1971 	if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1972 		return -1;
1973 
1974 	if (detailed_run < 3)
1975 		return 0;
1976 
1977 	/* Append very, very detailed run extra attributes: */
1978 	return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1979 }
1980 
1981 static const char * const stat_record_usage[] = {
1982 	"perf stat record [<options>]",
1983 	NULL,
1984 };
1985 
1986 static void init_features(struct perf_session *session)
1987 {
1988 	int feat;
1989 
1990 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1991 		perf_header__set_feat(&session->header, feat);
1992 
1993 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1994 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1995 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1996 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1997 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1998 }
1999 
2000 static int __cmd_record(int argc, const char **argv)
2001 {
2002 	struct perf_session *session;
2003 	struct perf_data *data = &perf_stat.data;
2004 
2005 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
2006 			     PARSE_OPT_STOP_AT_NON_OPTION);
2007 
2008 	if (output_name)
2009 		data->path = output_name;
2010 
2011 	if (stat_config.run_count != 1 || forever) {
2012 		pr_err("Cannot use -r option with perf stat record.\n");
2013 		return -1;
2014 	}
2015 
2016 	session = perf_session__new(data, NULL);
2017 	if (IS_ERR(session)) {
2018 		pr_err("Perf session creation failed\n");
2019 		return PTR_ERR(session);
2020 	}
2021 
2022 	init_features(session);
2023 
2024 	session->evlist   = evsel_list;
2025 	perf_stat.session = session;
2026 	perf_stat.record  = true;
2027 	return argc;
2028 }
2029 
2030 static int process_stat_round_event(struct perf_session *session,
2031 				    union perf_event *event)
2032 {
2033 	struct perf_record_stat_round *stat_round = &event->stat_round;
2034 	struct evsel *counter;
2035 	struct timespec tsh, *ts = NULL;
2036 	const char **argv = session->header.env.cmdline_argv;
2037 	int argc = session->header.env.nr_cmdline;
2038 
2039 	evlist__for_each_entry(evsel_list, counter)
2040 		perf_stat_process_counter(&stat_config, counter);
2041 
2042 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2043 		update_stats(&walltime_nsecs_stats, stat_round->time);
2044 
2045 	if (stat_config.interval && stat_round->time) {
2046 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2047 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2048 		ts = &tsh;
2049 	}
2050 
2051 	print_counters(ts, argc, argv);
2052 	return 0;
2053 }
2054 
2055 static
2056 int process_stat_config_event(struct perf_session *session,
2057 			      union perf_event *event)
2058 {
2059 	struct perf_tool *tool = session->tool;
2060 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2061 
2062 	perf_event__read_stat_config(&stat_config, &event->stat_config);
2063 
2064 	if (perf_cpu_map__empty(st->cpus)) {
2065 		if (st->aggr_mode != AGGR_UNSET)
2066 			pr_warning("warning: processing task data, aggregation mode not set\n");
2067 		return 0;
2068 	}
2069 
2070 	if (st->aggr_mode != AGGR_UNSET)
2071 		stat_config.aggr_mode = st->aggr_mode;
2072 
2073 	if (perf_stat.data.is_pipe)
2074 		perf_stat_init_aggr_mode();
2075 	else
2076 		perf_stat_init_aggr_mode_file(st);
2077 
2078 	return 0;
2079 }
2080 
2081 static int set_maps(struct perf_stat *st)
2082 {
2083 	if (!st->cpus || !st->threads)
2084 		return 0;
2085 
2086 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2087 		return -EINVAL;
2088 
2089 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2090 
2091 	if (evlist__alloc_stats(evsel_list, true))
2092 		return -ENOMEM;
2093 
2094 	st->maps_allocated = true;
2095 	return 0;
2096 }
2097 
2098 static
2099 int process_thread_map_event(struct perf_session *session,
2100 			     union perf_event *event)
2101 {
2102 	struct perf_tool *tool = session->tool;
2103 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2104 
2105 	if (st->threads) {
2106 		pr_warning("Extra thread map event, ignoring.\n");
2107 		return 0;
2108 	}
2109 
2110 	st->threads = thread_map__new_event(&event->thread_map);
2111 	if (!st->threads)
2112 		return -ENOMEM;
2113 
2114 	return set_maps(st);
2115 }
2116 
2117 static
2118 int process_cpu_map_event(struct perf_session *session,
2119 			  union perf_event *event)
2120 {
2121 	struct perf_tool *tool = session->tool;
2122 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2123 	struct perf_cpu_map *cpus;
2124 
2125 	if (st->cpus) {
2126 		pr_warning("Extra cpu map event, ignoring.\n");
2127 		return 0;
2128 	}
2129 
2130 	cpus = cpu_map__new_data(&event->cpu_map.data);
2131 	if (!cpus)
2132 		return -ENOMEM;
2133 
2134 	st->cpus = cpus;
2135 	return set_maps(st);
2136 }
2137 
2138 static const char * const stat_report_usage[] = {
2139 	"perf stat report [<options>]",
2140 	NULL,
2141 };
2142 
2143 static struct perf_stat perf_stat = {
2144 	.tool = {
2145 		.attr		= perf_event__process_attr,
2146 		.event_update	= perf_event__process_event_update,
2147 		.thread_map	= process_thread_map_event,
2148 		.cpu_map	= process_cpu_map_event,
2149 		.stat_config	= process_stat_config_event,
2150 		.stat		= perf_event__process_stat_event,
2151 		.stat_round	= process_stat_round_event,
2152 	},
2153 	.aggr_mode = AGGR_UNSET,
2154 };
2155 
2156 static int __cmd_report(int argc, const char **argv)
2157 {
2158 	struct perf_session *session;
2159 	const struct option options[] = {
2160 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2161 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2162 		     "aggregate counts per processor socket", AGGR_SOCKET),
2163 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2164 		     "aggregate counts per processor die", AGGR_DIE),
2165 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2166 		     "aggregate counts per physical processor core", AGGR_CORE),
2167 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2168 		     "aggregate counts per numa node", AGGR_NODE),
2169 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2170 		     "disable CPU count aggregation", AGGR_NONE),
2171 	OPT_END()
2172 	};
2173 	struct stat st;
2174 	int ret;
2175 
2176 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2177 
2178 	if (!input_name || !strlen(input_name)) {
2179 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2180 			input_name = "-";
2181 		else
2182 			input_name = "perf.data";
2183 	}
2184 
2185 	perf_stat.data.path = input_name;
2186 	perf_stat.data.mode = PERF_DATA_MODE_READ;
2187 
2188 	session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2189 	if (IS_ERR(session))
2190 		return PTR_ERR(session);
2191 
2192 	perf_stat.session  = session;
2193 	stat_config.output = stderr;
2194 	evsel_list         = session->evlist;
2195 
2196 	ret = perf_session__process_events(session);
2197 	if (ret)
2198 		return ret;
2199 
2200 	perf_session__delete(session);
2201 	return 0;
2202 }
2203 
2204 static void setup_system_wide(int forks)
2205 {
2206 	/*
2207 	 * Make system wide (-a) the default target if
2208 	 * no target was specified and one of following
2209 	 * conditions is met:
2210 	 *
2211 	 *   - there's no workload specified
2212 	 *   - there is workload specified but all requested
2213 	 *     events are system wide events
2214 	 */
2215 	if (!target__none(&target))
2216 		return;
2217 
2218 	if (!forks)
2219 		target.system_wide = true;
2220 	else {
2221 		struct evsel *counter;
2222 
2223 		evlist__for_each_entry(evsel_list, counter) {
2224 			if (!counter->core.system_wide &&
2225 			    strcmp(counter->name, "duration_time")) {
2226 				return;
2227 			}
2228 		}
2229 
2230 		if (evsel_list->core.nr_entries)
2231 			target.system_wide = true;
2232 	}
2233 }
2234 
2235 int cmd_stat(int argc, const char **argv)
2236 {
2237 	const char * const stat_usage[] = {
2238 		"perf stat [<options>] [<command>]",
2239 		NULL
2240 	};
2241 	int status = -EINVAL, run_idx, err;
2242 	const char *mode;
2243 	FILE *output = stderr;
2244 	unsigned int interval, timeout;
2245 	const char * const stat_subcommands[] = { "record", "report" };
2246 	char errbuf[BUFSIZ];
2247 
2248 	setlocale(LC_ALL, "");
2249 
2250 	evsel_list = evlist__new();
2251 	if (evsel_list == NULL)
2252 		return -ENOMEM;
2253 
2254 	parse_events__shrink_config_terms();
2255 
2256 	/* String-parsing callback-based options would segfault when negated */
2257 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2258 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2259 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2260 
2261 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2262 					(const char **) stat_usage,
2263 					PARSE_OPT_STOP_AT_NON_OPTION);
2264 	perf_stat__collect_metric_expr(evsel_list);
2265 	perf_stat__init_shadow_stats();
2266 
2267 	if (stat_config.csv_sep) {
2268 		stat_config.csv_output = true;
2269 		if (!strcmp(stat_config.csv_sep, "\\t"))
2270 			stat_config.csv_sep = "\t";
2271 	} else
2272 		stat_config.csv_sep = DEFAULT_SEPARATOR;
2273 
2274 	if (argc && !strncmp(argv[0], "rec", 3)) {
2275 		argc = __cmd_record(argc, argv);
2276 		if (argc < 0)
2277 			return -1;
2278 	} else if (argc && !strncmp(argv[0], "rep", 3))
2279 		return __cmd_report(argc, argv);
2280 
2281 	interval = stat_config.interval;
2282 	timeout = stat_config.timeout;
2283 
2284 	/*
2285 	 * For record command the -o is already taken care of.
2286 	 */
2287 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2288 		output = NULL;
2289 
2290 	if (output_name && output_fd) {
2291 		fprintf(stderr, "cannot use both --output and --log-fd\n");
2292 		parse_options_usage(stat_usage, stat_options, "o", 1);
2293 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2294 		goto out;
2295 	}
2296 
2297 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2298 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2299 		goto out;
2300 	}
2301 
2302 	if (stat_config.metric_only && stat_config.run_count > 1) {
2303 		fprintf(stderr, "--metric-only is not supported with -r\n");
2304 		goto out;
2305 	}
2306 
2307 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2308 		fprintf(stderr, "--table is only supported with -r\n");
2309 		parse_options_usage(stat_usage, stat_options, "r", 1);
2310 		parse_options_usage(NULL, stat_options, "table", 0);
2311 		goto out;
2312 	}
2313 
2314 	if (output_fd < 0) {
2315 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2316 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2317 		goto out;
2318 	}
2319 
2320 	if (!output && !stat_config.quiet) {
2321 		struct timespec tm;
2322 		mode = append_file ? "a" : "w";
2323 
2324 		output = fopen(output_name, mode);
2325 		if (!output) {
2326 			perror("failed to create output file");
2327 			return -1;
2328 		}
2329 		clock_gettime(CLOCK_REALTIME, &tm);
2330 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2331 	} else if (output_fd > 0) {
2332 		mode = append_file ? "a" : "w";
2333 		output = fdopen(output_fd, mode);
2334 		if (!output) {
2335 			perror("Failed opening logfd");
2336 			return -errno;
2337 		}
2338 	}
2339 
2340 	stat_config.output = output;
2341 
2342 	/*
2343 	 * let the spreadsheet do the pretty-printing
2344 	 */
2345 	if (stat_config.csv_output) {
2346 		/* User explicitly passed -B? */
2347 		if (big_num_opt == 1) {
2348 			fprintf(stderr, "-B option not supported with -x\n");
2349 			parse_options_usage(stat_usage, stat_options, "B", 1);
2350 			parse_options_usage(NULL, stat_options, "x", 1);
2351 			goto out;
2352 		} else /* Nope, so disable big number formatting */
2353 			stat_config.big_num = false;
2354 	} else if (big_num_opt == 0) /* User passed --no-big-num */
2355 		stat_config.big_num = false;
2356 
2357 	err = target__validate(&target);
2358 	if (err) {
2359 		target__strerror(&target, err, errbuf, BUFSIZ);
2360 		pr_warning("%s\n", errbuf);
2361 	}
2362 
2363 	setup_system_wide(argc);
2364 
2365 	/*
2366 	 * Display user/system times only for single
2367 	 * run and when there's specified tracee.
2368 	 */
2369 	if ((stat_config.run_count == 1) && target__none(&target))
2370 		stat_config.ru_display = true;
2371 
2372 	if (stat_config.run_count < 0) {
2373 		pr_err("Run count must be a positive number\n");
2374 		parse_options_usage(stat_usage, stat_options, "r", 1);
2375 		goto out;
2376 	} else if (stat_config.run_count == 0) {
2377 		forever = true;
2378 		stat_config.run_count = 1;
2379 	}
2380 
2381 	if (stat_config.walltime_run_table) {
2382 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2383 		if (!stat_config.walltime_run) {
2384 			pr_err("failed to setup -r option");
2385 			goto out;
2386 		}
2387 	}
2388 
2389 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2390 		!target__has_task(&target)) {
2391 		if (!target.system_wide || target.cpu_list) {
2392 			fprintf(stderr, "The --per-thread option is only "
2393 				"available when monitoring via -p -t -a "
2394 				"options or only --per-thread.\n");
2395 			parse_options_usage(NULL, stat_options, "p", 1);
2396 			parse_options_usage(NULL, stat_options, "t", 1);
2397 			goto out;
2398 		}
2399 	}
2400 
2401 	/*
2402 	 * no_aggr, cgroup are for system-wide only
2403 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2404 	 */
2405 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2406 	      stat_config.aggr_mode != AGGR_THREAD) ||
2407 	     (nr_cgroups || stat_config.cgroup_list)) &&
2408 	    !target__has_cpu(&target)) {
2409 		fprintf(stderr, "both cgroup and no-aggregation "
2410 			"modes only available in system-wide mode\n");
2411 
2412 		parse_options_usage(stat_usage, stat_options, "G", 1);
2413 		parse_options_usage(NULL, stat_options, "A", 1);
2414 		parse_options_usage(NULL, stat_options, "a", 1);
2415 		parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2416 		goto out;
2417 	}
2418 
2419 	if (stat_config.iostat_run) {
2420 		status = iostat_prepare(evsel_list, &stat_config);
2421 		if (status)
2422 			goto out;
2423 		if (iostat_mode == IOSTAT_LIST) {
2424 			iostat_list(evsel_list, &stat_config);
2425 			goto out;
2426 		} else if (verbose)
2427 			iostat_list(evsel_list, &stat_config);
2428 		if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2429 			target.system_wide = true;
2430 	}
2431 
2432 	if (add_default_attributes())
2433 		goto out;
2434 
2435 	if (stat_config.cgroup_list) {
2436 		if (nr_cgroups > 0) {
2437 			pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2438 			parse_options_usage(stat_usage, stat_options, "G", 1);
2439 			parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2440 			goto out;
2441 		}
2442 
2443 		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2444 					  &stat_config.metric_events, true) < 0) {
2445 			parse_options_usage(stat_usage, stat_options,
2446 					    "for-each-cgroup", 0);
2447 			goto out;
2448 		}
2449 	}
2450 
2451 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2452 		target.per_thread = true;
2453 
2454 	if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2455 		pr_err("failed to use cpu list %s\n", target.cpu_list);
2456 		goto out;
2457 	}
2458 
2459 	target.hybrid = perf_pmu__has_hybrid();
2460 	if (evlist__create_maps(evsel_list, &target) < 0) {
2461 		if (target__has_task(&target)) {
2462 			pr_err("Problems finding threads of monitor\n");
2463 			parse_options_usage(stat_usage, stat_options, "p", 1);
2464 			parse_options_usage(NULL, stat_options, "t", 1);
2465 		} else if (target__has_cpu(&target)) {
2466 			perror("failed to parse CPUs map");
2467 			parse_options_usage(stat_usage, stat_options, "C", 1);
2468 			parse_options_usage(NULL, stat_options, "a", 1);
2469 		}
2470 		goto out;
2471 	}
2472 
2473 	evlist__check_cpu_maps(evsel_list);
2474 
2475 	/*
2476 	 * Initialize thread_map with comm names,
2477 	 * so we could print it out on output.
2478 	 */
2479 	if (stat_config.aggr_mode == AGGR_THREAD) {
2480 		thread_map__read_comms(evsel_list->core.threads);
2481 		if (target.system_wide) {
2482 			if (runtime_stat_new(&stat_config,
2483 				perf_thread_map__nr(evsel_list->core.threads))) {
2484 				goto out;
2485 			}
2486 		}
2487 	}
2488 
2489 	if (stat_config.aggr_mode == AGGR_NODE)
2490 		cpu__setup_cpunode_map();
2491 
2492 	if (stat_config.times && interval)
2493 		interval_count = true;
2494 	else if (stat_config.times && !interval) {
2495 		pr_err("interval-count option should be used together with "
2496 				"interval-print.\n");
2497 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2498 		parse_options_usage(stat_usage, stat_options, "I", 1);
2499 		goto out;
2500 	}
2501 
2502 	if (timeout && timeout < 100) {
2503 		if (timeout < 10) {
2504 			pr_err("timeout must be >= 10ms.\n");
2505 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2506 			goto out;
2507 		} else
2508 			pr_warning("timeout < 100ms. "
2509 				   "The overhead percentage could be high in some cases. "
2510 				   "Please proceed with caution.\n");
2511 	}
2512 	if (timeout && interval) {
2513 		pr_err("timeout option is not supported with interval-print.\n");
2514 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2515 		parse_options_usage(stat_usage, stat_options, "I", 1);
2516 		goto out;
2517 	}
2518 
2519 	if (evlist__alloc_stats(evsel_list, interval))
2520 		goto out;
2521 
2522 	if (perf_stat_init_aggr_mode())
2523 		goto out;
2524 
2525 	/*
2526 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2527 	 * while avoiding that older tools show confusing messages.
2528 	 *
2529 	 * However for pipe sessions we need to keep it zero,
2530 	 * because script's perf_evsel__check_attr is triggered
2531 	 * by attr->sample_type != 0, and we can't run it on
2532 	 * stat sessions.
2533 	 */
2534 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2535 
2536 	/*
2537 	 * We dont want to block the signals - that would cause
2538 	 * child tasks to inherit that and Ctrl-C would not work.
2539 	 * What we want is for Ctrl-C to work in the exec()-ed
2540 	 * task, but being ignored by perf stat itself:
2541 	 */
2542 	atexit(sig_atexit);
2543 	if (!forever)
2544 		signal(SIGINT,  skip_signal);
2545 	signal(SIGCHLD, skip_signal);
2546 	signal(SIGALRM, skip_signal);
2547 	signal(SIGABRT, skip_signal);
2548 
2549 	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2550 		goto out;
2551 
2552 	status = 0;
2553 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2554 		if (stat_config.run_count != 1 && verbose > 0)
2555 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2556 				run_idx + 1);
2557 
2558 		if (run_idx != 0)
2559 			evlist__reset_prev_raw_counts(evsel_list);
2560 
2561 		status = run_perf_stat(argc, argv, run_idx);
2562 		if (forever && status != -1 && !interval) {
2563 			print_counters(NULL, argc, argv);
2564 			perf_stat__reset_stats();
2565 		}
2566 	}
2567 
2568 	if (!forever && status != -1 && (!interval || stat_config.summary))
2569 		print_counters(NULL, argc, argv);
2570 
2571 	evlist__finalize_ctlfd(evsel_list);
2572 
2573 	if (STAT_RECORD) {
2574 		/*
2575 		 * We synthesize the kernel mmap record just so that older tools
2576 		 * don't emit warnings about not being able to resolve symbols
2577 		 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2578 		 * a saner message about no samples being in the perf.data file.
2579 		 *
2580 		 * This also serves to suppress a warning about f_header.data.size == 0
2581 		 * in header.c at the moment 'perf stat record' gets introduced, which
2582 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2583 		 * records, but the need to suppress the kptr_restrict messages in older
2584 		 * tools remain  -acme
2585 		 */
2586 		int fd = perf_data__fd(&perf_stat.data);
2587 
2588 		err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2589 							 process_synthesized_event,
2590 							 &perf_stat.session->machines.host);
2591 		if (err) {
2592 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2593 				   "older tools may produce warnings about this file\n.");
2594 		}
2595 
2596 		if (!interval) {
2597 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2598 				pr_err("failed to write stat round event\n");
2599 		}
2600 
2601 		if (!perf_stat.data.is_pipe) {
2602 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2603 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2604 		}
2605 
2606 		evlist__close(evsel_list);
2607 		perf_session__delete(perf_stat.session);
2608 	}
2609 
2610 	perf_stat__exit_aggr_mode();
2611 	evlist__free_stats(evsel_list);
2612 out:
2613 	if (stat_config.iostat_run)
2614 		iostat_release(evsel_list);
2615 
2616 	zfree(&stat_config.walltime_run);
2617 
2618 	if (smi_cost && smi_reset)
2619 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2620 
2621 	evlist__delete(evsel_list);
2622 
2623 	metricgroup__rblist_exit(&stat_config.metric_events);
2624 	runtime_stat_delete(&stat_config);
2625 	evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2626 
2627 	return status;
2628 }
2629