xref: /openbmc/linux/tools/perf/builtin-stat.c (revision 3d40aed8)
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 "util/cgroup.h"
45 #include <subcmd/parse-options.h>
46 #include "util/parse-events.h"
47 #include "util/pmus.h"
48 #include "util/pmu.h"
49 #include "util/event.h"
50 #include "util/evlist.h"
51 #include "util/evsel.h"
52 #include "util/debug.h"
53 #include "util/color.h"
54 #include "util/stat.h"
55 #include "util/header.h"
56 #include "util/cpumap.h"
57 #include "util/thread_map.h"
58 #include "util/counts.h"
59 #include "util/topdown.h"
60 #include "util/session.h"
61 #include "util/tool.h"
62 #include "util/string2.h"
63 #include "util/metricgroup.h"
64 #include "util/synthetic-events.h"
65 #include "util/target.h"
66 #include "util/time-utils.h"
67 #include "util/top.h"
68 #include "util/affinity.h"
69 #include "util/pfm.h"
70 #include "util/bpf_counter.h"
71 #include "util/iostat.h"
72 #include "util/util.h"
73 #include "asm/bug.h"
74 
75 #include <linux/time64.h>
76 #include <linux/zalloc.h>
77 #include <api/fs/fs.h>
78 #include <errno.h>
79 #include <signal.h>
80 #include <stdlib.h>
81 #include <sys/prctl.h>
82 #include <inttypes.h>
83 #include <locale.h>
84 #include <math.h>
85 #include <sys/types.h>
86 #include <sys/stat.h>
87 #include <sys/wait.h>
88 #include <unistd.h>
89 #include <sys/time.h>
90 #include <sys/resource.h>
91 #include <linux/err.h>
92 
93 #include <linux/ctype.h>
94 #include <perf/evlist.h>
95 #include <internal/threadmap.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 static struct evlist	*evsel_list;
103 static struct parse_events_option_args parse_events_option_args = {
104 	.evlistp = &evsel_list,
105 };
106 
107 static bool all_counters_use_bpf = true;
108 
109 static struct target target = {
110 	.uid	= UINT_MAX,
111 };
112 
113 #define METRIC_ONLY_LEN 20
114 
115 static volatile sig_atomic_t	child_pid			= -1;
116 static int			detailed_run			=  0;
117 static bool			transaction_run;
118 static bool			topdown_run			= false;
119 static bool			smi_cost			= false;
120 static bool			smi_reset			= false;
121 static int			big_num_opt			=  -1;
122 static const char		*pre_cmd			= NULL;
123 static const char		*post_cmd			= NULL;
124 static bool			sync_run			= false;
125 static bool			forever				= false;
126 static bool			force_metric_only		= false;
127 static struct timespec		ref_time;
128 static bool			append_file;
129 static bool			interval_count;
130 static const char		*output_name;
131 static int			output_fd;
132 static char			*metrics;
133 
134 struct perf_stat {
135 	bool			 record;
136 	struct perf_data	 data;
137 	struct perf_session	*session;
138 	u64			 bytes_written;
139 	struct perf_tool	 tool;
140 	bool			 maps_allocated;
141 	struct perf_cpu_map	*cpus;
142 	struct perf_thread_map *threads;
143 	enum aggr_mode		 aggr_mode;
144 	u32			 aggr_level;
145 };
146 
147 static struct perf_stat		perf_stat;
148 #define STAT_RECORD		perf_stat.record
149 
150 static volatile sig_atomic_t done = 0;
151 
152 static struct perf_stat_config stat_config = {
153 	.aggr_mode		= AGGR_GLOBAL,
154 	.aggr_level		= MAX_CACHE_LVL + 1,
155 	.scale			= true,
156 	.unit_width		= 4, /* strlen("unit") */
157 	.run_count		= 1,
158 	.metric_only_len	= METRIC_ONLY_LEN,
159 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
160 	.ru_stats		= &ru_stats,
161 	.big_num		= true,
162 	.ctl_fd			= -1,
163 	.ctl_fd_ack		= -1,
164 	.iostat_run		= false,
165 };
166 
167 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
168 {
169 	if (!a->core.cpus && !b->core.cpus)
170 		return true;
171 
172 	if (!a->core.cpus || !b->core.cpus)
173 		return false;
174 
175 	if (perf_cpu_map__nr(a->core.cpus) != perf_cpu_map__nr(b->core.cpus))
176 		return false;
177 
178 	for (int i = 0; i < perf_cpu_map__nr(a->core.cpus); i++) {
179 		if (perf_cpu_map__cpu(a->core.cpus, i).cpu !=
180 		    perf_cpu_map__cpu(b->core.cpus, i).cpu)
181 			return false;
182 	}
183 
184 	return true;
185 }
186 
187 static void evlist__check_cpu_maps(struct evlist *evlist)
188 {
189 	struct evsel *evsel, *warned_leader = NULL;
190 
191 	evlist__for_each_entry(evlist, evsel) {
192 		struct evsel *leader = evsel__leader(evsel);
193 
194 		/* Check that leader matches cpus with each member. */
195 		if (leader == evsel)
196 			continue;
197 		if (cpus_map_matched(leader, evsel))
198 			continue;
199 
200 		/* If there's mismatch disable the group and warn user. */
201 		if (warned_leader != leader) {
202 			char buf[200];
203 
204 			pr_warning("WARNING: grouped events cpus do not match.\n"
205 				"Events with CPUs not matching the leader will "
206 				"be removed from the group.\n");
207 			evsel__group_desc(leader, buf, sizeof(buf));
208 			pr_warning("  %s\n", buf);
209 			warned_leader = leader;
210 		}
211 		if (verbose > 0) {
212 			char buf[200];
213 
214 			cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
215 			pr_warning("     %s: %s\n", leader->name, buf);
216 			cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
217 			pr_warning("     %s: %s\n", evsel->name, buf);
218 		}
219 
220 		evsel__remove_from_group(evsel, leader);
221 	}
222 }
223 
224 static inline void diff_timespec(struct timespec *r, struct timespec *a,
225 				 struct timespec *b)
226 {
227 	r->tv_sec = a->tv_sec - b->tv_sec;
228 	if (a->tv_nsec < b->tv_nsec) {
229 		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
230 		r->tv_sec--;
231 	} else {
232 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
233 	}
234 }
235 
236 static void perf_stat__reset_stats(void)
237 {
238 	evlist__reset_stats(evsel_list);
239 	perf_stat__reset_shadow_stats();
240 }
241 
242 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
243 				     union perf_event *event,
244 				     struct perf_sample *sample __maybe_unused,
245 				     struct machine *machine __maybe_unused)
246 {
247 	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
248 		pr_err("failed to write perf data, error: %m\n");
249 		return -1;
250 	}
251 
252 	perf_stat.bytes_written += event->header.size;
253 	return 0;
254 }
255 
256 static int write_stat_round_event(u64 tm, u64 type)
257 {
258 	return perf_event__synthesize_stat_round(NULL, tm, type,
259 						 process_synthesized_event,
260 						 NULL);
261 }
262 
263 #define WRITE_STAT_ROUND_EVENT(time, interval) \
264 	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
265 
266 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
267 
268 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
269 				   struct perf_counts_values *count)
270 {
271 	struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
272 	struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
273 
274 	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
275 					   process_synthesized_event, NULL);
276 }
277 
278 static int read_single_counter(struct evsel *counter, int cpu_map_idx,
279 			       int thread, struct timespec *rs)
280 {
281 	switch(counter->tool_event) {
282 		case PERF_TOOL_DURATION_TIME: {
283 			u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
284 			struct perf_counts_values *count =
285 				perf_counts(counter->counts, cpu_map_idx, thread);
286 			count->ena = count->run = val;
287 			count->val = val;
288 			return 0;
289 		}
290 		case PERF_TOOL_USER_TIME:
291 		case PERF_TOOL_SYSTEM_TIME: {
292 			u64 val;
293 			struct perf_counts_values *count =
294 				perf_counts(counter->counts, cpu_map_idx, thread);
295 			if (counter->tool_event == PERF_TOOL_USER_TIME)
296 				val = ru_stats.ru_utime_usec_stat.mean;
297 			else
298 				val = ru_stats.ru_stime_usec_stat.mean;
299 			count->ena = count->run = val;
300 			count->val = val;
301 			return 0;
302 		}
303 		default:
304 		case PERF_TOOL_NONE:
305 			return evsel__read_counter(counter, cpu_map_idx, thread);
306 		case PERF_TOOL_MAX:
307 			/* This should never be reached */
308 			return 0;
309 	}
310 }
311 
312 /*
313  * Read out the results of a single counter:
314  * do not aggregate counts across CPUs in system-wide mode
315  */
316 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx)
317 {
318 	int nthreads = perf_thread_map__nr(evsel_list->core.threads);
319 	int thread;
320 
321 	if (!counter->supported)
322 		return -ENOENT;
323 
324 	for (thread = 0; thread < nthreads; thread++) {
325 		struct perf_counts_values *count;
326 
327 		count = perf_counts(counter->counts, cpu_map_idx, thread);
328 
329 		/*
330 		 * The leader's group read loads data into its group members
331 		 * (via evsel__read_counter()) and sets their count->loaded.
332 		 */
333 		if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
334 		    read_single_counter(counter, cpu_map_idx, thread, rs)) {
335 			counter->counts->scaled = -1;
336 			perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
337 			perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
338 			return -1;
339 		}
340 
341 		perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
342 
343 		if (STAT_RECORD) {
344 			if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
345 				pr_err("failed to write stat event\n");
346 				return -1;
347 			}
348 		}
349 
350 		if (verbose > 1) {
351 			fprintf(stat_config.output,
352 				"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
353 					evsel__name(counter),
354 					perf_cpu_map__cpu(evsel__cpus(counter),
355 							  cpu_map_idx).cpu,
356 					count->val, count->ena, count->run);
357 		}
358 	}
359 
360 	return 0;
361 }
362 
363 static int read_affinity_counters(struct timespec *rs)
364 {
365 	struct evlist_cpu_iterator evlist_cpu_itr;
366 	struct affinity saved_affinity, *affinity;
367 
368 	if (all_counters_use_bpf)
369 		return 0;
370 
371 	if (!target__has_cpu(&target) || target__has_per_thread(&target))
372 		affinity = NULL;
373 	else if (affinity__setup(&saved_affinity) < 0)
374 		return -1;
375 	else
376 		affinity = &saved_affinity;
377 
378 	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
379 		struct evsel *counter = evlist_cpu_itr.evsel;
380 
381 		if (evsel__is_bpf(counter))
382 			continue;
383 
384 		if (!counter->err) {
385 			counter->err = read_counter_cpu(counter, rs,
386 							evlist_cpu_itr.cpu_map_idx);
387 		}
388 	}
389 	if (affinity)
390 		affinity__cleanup(&saved_affinity);
391 
392 	return 0;
393 }
394 
395 static int read_bpf_map_counters(void)
396 {
397 	struct evsel *counter;
398 	int err;
399 
400 	evlist__for_each_entry(evsel_list, counter) {
401 		if (!evsel__is_bpf(counter))
402 			continue;
403 
404 		err = bpf_counter__read(counter);
405 		if (err)
406 			return err;
407 	}
408 	return 0;
409 }
410 
411 static int read_counters(struct timespec *rs)
412 {
413 	if (!stat_config.stop_read_counter) {
414 		if (read_bpf_map_counters() ||
415 		    read_affinity_counters(rs))
416 			return -1;
417 	}
418 	return 0;
419 }
420 
421 static void process_counters(void)
422 {
423 	struct evsel *counter;
424 
425 	evlist__for_each_entry(evsel_list, counter) {
426 		if (counter->err)
427 			pr_debug("failed to read counter %s\n", counter->name);
428 		if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
429 			pr_warning("failed to process counter %s\n", counter->name);
430 		counter->err = 0;
431 	}
432 
433 	perf_stat_merge_counters(&stat_config, evsel_list);
434 	perf_stat_process_percore(&stat_config, evsel_list);
435 }
436 
437 static void process_interval(void)
438 {
439 	struct timespec ts, rs;
440 
441 	clock_gettime(CLOCK_MONOTONIC, &ts);
442 	diff_timespec(&rs, &ts, &ref_time);
443 
444 	evlist__reset_aggr_stats(evsel_list);
445 
446 	if (read_counters(&rs) == 0)
447 		process_counters();
448 
449 	if (STAT_RECORD) {
450 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
451 			pr_err("failed to write stat round event\n");
452 	}
453 
454 	init_stats(&walltime_nsecs_stats);
455 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
456 	print_counters(&rs, 0, NULL);
457 }
458 
459 static bool handle_interval(unsigned int interval, int *times)
460 {
461 	if (interval) {
462 		process_interval();
463 		if (interval_count && !(--(*times)))
464 			return true;
465 	}
466 	return false;
467 }
468 
469 static int enable_counters(void)
470 {
471 	struct evsel *evsel;
472 	int err;
473 
474 	evlist__for_each_entry(evsel_list, evsel) {
475 		if (!evsel__is_bpf(evsel))
476 			continue;
477 
478 		err = bpf_counter__enable(evsel);
479 		if (err)
480 			return err;
481 	}
482 
483 	if (!target__enable_on_exec(&target)) {
484 		if (!all_counters_use_bpf)
485 			evlist__enable(evsel_list);
486 	}
487 	return 0;
488 }
489 
490 static void disable_counters(void)
491 {
492 	struct evsel *counter;
493 
494 	/*
495 	 * If we don't have tracee (attaching to task or cpu), counters may
496 	 * still be running. To get accurate group ratios, we must stop groups
497 	 * from counting before reading their constituent counters.
498 	 */
499 	if (!target__none(&target)) {
500 		evlist__for_each_entry(evsel_list, counter)
501 			bpf_counter__disable(counter);
502 		if (!all_counters_use_bpf)
503 			evlist__disable(evsel_list);
504 	}
505 }
506 
507 static volatile sig_atomic_t workload_exec_errno;
508 
509 /*
510  * evlist__prepare_workload will send a SIGUSR1
511  * if the fork fails, since we asked by setting its
512  * want_signal to true.
513  */
514 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
515 					void *ucontext __maybe_unused)
516 {
517 	workload_exec_errno = info->si_value.sival_int;
518 }
519 
520 static bool evsel__should_store_id(struct evsel *counter)
521 {
522 	return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
523 }
524 
525 static bool is_target_alive(struct target *_target,
526 			    struct perf_thread_map *threads)
527 {
528 	struct stat st;
529 	int i;
530 
531 	if (!target__has_task(_target))
532 		return true;
533 
534 	for (i = 0; i < threads->nr; i++) {
535 		char path[PATH_MAX];
536 
537 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
538 			  threads->map[i].pid);
539 
540 		if (!stat(path, &st))
541 			return true;
542 	}
543 
544 	return false;
545 }
546 
547 static void process_evlist(struct evlist *evlist, unsigned int interval)
548 {
549 	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
550 
551 	if (evlist__ctlfd_process(evlist, &cmd) > 0) {
552 		switch (cmd) {
553 		case EVLIST_CTL_CMD_ENABLE:
554 			fallthrough;
555 		case EVLIST_CTL_CMD_DISABLE:
556 			if (interval)
557 				process_interval();
558 			break;
559 		case EVLIST_CTL_CMD_SNAPSHOT:
560 		case EVLIST_CTL_CMD_ACK:
561 		case EVLIST_CTL_CMD_UNSUPPORTED:
562 		case EVLIST_CTL_CMD_EVLIST:
563 		case EVLIST_CTL_CMD_STOP:
564 		case EVLIST_CTL_CMD_PING:
565 		default:
566 			break;
567 		}
568 	}
569 }
570 
571 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
572 			int *time_to_sleep)
573 {
574 	int tts = *time_to_sleep;
575 	struct timespec time_diff;
576 
577 	diff_timespec(&time_diff, time_stop, time_start);
578 
579 	tts -= time_diff.tv_sec * MSEC_PER_SEC +
580 	       time_diff.tv_nsec / NSEC_PER_MSEC;
581 
582 	if (tts < 0)
583 		tts = 0;
584 
585 	*time_to_sleep = tts;
586 }
587 
588 static int dispatch_events(bool forks, int timeout, int interval, int *times)
589 {
590 	int child_exited = 0, status = 0;
591 	int time_to_sleep, sleep_time;
592 	struct timespec time_start, time_stop;
593 
594 	if (interval)
595 		sleep_time = interval;
596 	else if (timeout)
597 		sleep_time = timeout;
598 	else
599 		sleep_time = 1000;
600 
601 	time_to_sleep = sleep_time;
602 
603 	while (!done) {
604 		if (forks)
605 			child_exited = waitpid(child_pid, &status, WNOHANG);
606 		else
607 			child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
608 
609 		if (child_exited)
610 			break;
611 
612 		clock_gettime(CLOCK_MONOTONIC, &time_start);
613 		if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
614 			if (timeout || handle_interval(interval, times))
615 				break;
616 			time_to_sleep = sleep_time;
617 		} else { /* fd revent */
618 			process_evlist(evsel_list, interval);
619 			clock_gettime(CLOCK_MONOTONIC, &time_stop);
620 			compute_tts(&time_start, &time_stop, &time_to_sleep);
621 		}
622 	}
623 
624 	return status;
625 }
626 
627 enum counter_recovery {
628 	COUNTER_SKIP,
629 	COUNTER_RETRY,
630 	COUNTER_FATAL,
631 };
632 
633 static enum counter_recovery stat_handle_error(struct evsel *counter)
634 {
635 	char msg[BUFSIZ];
636 	/*
637 	 * PPC returns ENXIO for HW counters until 2.6.37
638 	 * (behavior changed with commit b0a873e).
639 	 */
640 	if (errno == EINVAL || errno == ENOSYS ||
641 	    errno == ENOENT || errno == EOPNOTSUPP ||
642 	    errno == ENXIO) {
643 		if (verbose > 0)
644 			ui__warning("%s event is not supported by the kernel.\n",
645 				    evsel__name(counter));
646 		counter->supported = false;
647 		/*
648 		 * errored is a sticky flag that means one of the counter's
649 		 * cpu event had a problem and needs to be reexamined.
650 		 */
651 		counter->errored = true;
652 
653 		if ((evsel__leader(counter) != counter) ||
654 		    !(counter->core.leader->nr_members > 1))
655 			return COUNTER_SKIP;
656 	} else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
657 		if (verbose > 0)
658 			ui__warning("%s\n", msg);
659 		return COUNTER_RETRY;
660 	} else if (target__has_per_thread(&target) &&
661 		   evsel_list->core.threads &&
662 		   evsel_list->core.threads->err_thread != -1) {
663 		/*
664 		 * For global --per-thread case, skip current
665 		 * error thread.
666 		 */
667 		if (!thread_map__remove(evsel_list->core.threads,
668 					evsel_list->core.threads->err_thread)) {
669 			evsel_list->core.threads->err_thread = -1;
670 			return COUNTER_RETRY;
671 		}
672 	} else if (counter->skippable) {
673 		if (verbose > 0)
674 			ui__warning("skipping event %s that kernel failed to open .\n",
675 				    evsel__name(counter));
676 		counter->supported = false;
677 		counter->errored = true;
678 		return COUNTER_SKIP;
679 	}
680 
681 	evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
682 	ui__error("%s\n", msg);
683 
684 	if (child_pid != -1)
685 		kill(child_pid, SIGTERM);
686 	return COUNTER_FATAL;
687 }
688 
689 static int __run_perf_stat(int argc, const char **argv, int run_idx)
690 {
691 	int interval = stat_config.interval;
692 	int times = stat_config.times;
693 	int timeout = stat_config.timeout;
694 	char msg[BUFSIZ];
695 	unsigned long long t0, t1;
696 	struct evsel *counter;
697 	size_t l;
698 	int status = 0;
699 	const bool forks = (argc > 0);
700 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
701 	struct evlist_cpu_iterator evlist_cpu_itr;
702 	struct affinity saved_affinity, *affinity = NULL;
703 	int err;
704 	bool second_pass = false;
705 
706 	if (forks) {
707 		if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
708 			perror("failed to prepare workload");
709 			return -1;
710 		}
711 		child_pid = evsel_list->workload.pid;
712 	}
713 
714 	if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
715 		if (affinity__setup(&saved_affinity) < 0)
716 			return -1;
717 		affinity = &saved_affinity;
718 	}
719 
720 	evlist__for_each_entry(evsel_list, counter) {
721 		counter->reset_group = false;
722 		if (bpf_counter__load(counter, &target))
723 			return -1;
724 		if (!(evsel__is_bperf(counter)))
725 			all_counters_use_bpf = false;
726 	}
727 
728 	evlist__reset_aggr_stats(evsel_list);
729 
730 	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
731 		counter = evlist_cpu_itr.evsel;
732 
733 		/*
734 		 * bperf calls evsel__open_per_cpu() in bperf__load(), so
735 		 * no need to call it again here.
736 		 */
737 		if (target.use_bpf)
738 			break;
739 
740 		if (counter->reset_group || counter->errored)
741 			continue;
742 		if (evsel__is_bperf(counter))
743 			continue;
744 try_again:
745 		if (create_perf_stat_counter(counter, &stat_config, &target,
746 					     evlist_cpu_itr.cpu_map_idx) < 0) {
747 
748 			/*
749 			 * Weak group failed. We cannot just undo this here
750 			 * because earlier CPUs might be in group mode, and the kernel
751 			 * doesn't support mixing group and non group reads. Defer
752 			 * it to later.
753 			 * Don't close here because we're in the wrong affinity.
754 			 */
755 			if ((errno == EINVAL || errno == EBADF) &&
756 				evsel__leader(counter) != counter &&
757 				counter->weak_group) {
758 				evlist__reset_weak_group(evsel_list, counter, false);
759 				assert(counter->reset_group);
760 				second_pass = true;
761 				continue;
762 			}
763 
764 			switch (stat_handle_error(counter)) {
765 			case COUNTER_FATAL:
766 				return -1;
767 			case COUNTER_RETRY:
768 				goto try_again;
769 			case COUNTER_SKIP:
770 				continue;
771 			default:
772 				break;
773 			}
774 
775 		}
776 		counter->supported = true;
777 	}
778 
779 	if (second_pass) {
780 		/*
781 		 * Now redo all the weak group after closing them,
782 		 * and also close errored counters.
783 		 */
784 
785 		/* First close errored or weak retry */
786 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
787 			counter = evlist_cpu_itr.evsel;
788 
789 			if (!counter->reset_group && !counter->errored)
790 				continue;
791 
792 			perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
793 		}
794 		/* Now reopen weak */
795 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
796 			counter = evlist_cpu_itr.evsel;
797 
798 			if (!counter->reset_group)
799 				continue;
800 try_again_reset:
801 			pr_debug2("reopening weak %s\n", evsel__name(counter));
802 			if (create_perf_stat_counter(counter, &stat_config, &target,
803 						     evlist_cpu_itr.cpu_map_idx) < 0) {
804 
805 				switch (stat_handle_error(counter)) {
806 				case COUNTER_FATAL:
807 					return -1;
808 				case COUNTER_RETRY:
809 					goto try_again_reset;
810 				case COUNTER_SKIP:
811 					continue;
812 				default:
813 					break;
814 				}
815 			}
816 			counter->supported = true;
817 		}
818 	}
819 	affinity__cleanup(affinity);
820 
821 	evlist__for_each_entry(evsel_list, counter) {
822 		if (!counter->supported) {
823 			perf_evsel__free_fd(&counter->core);
824 			continue;
825 		}
826 
827 		l = strlen(counter->unit);
828 		if (l > stat_config.unit_width)
829 			stat_config.unit_width = l;
830 
831 		if (evsel__should_store_id(counter) &&
832 		    evsel__store_ids(counter, evsel_list))
833 			return -1;
834 	}
835 
836 	if (evlist__apply_filters(evsel_list, &counter)) {
837 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
838 			counter->filter, evsel__name(counter), errno,
839 			str_error_r(errno, msg, sizeof(msg)));
840 		return -1;
841 	}
842 
843 	if (STAT_RECORD) {
844 		int fd = perf_data__fd(&perf_stat.data);
845 
846 		if (is_pipe) {
847 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
848 		} else {
849 			err = perf_session__write_header(perf_stat.session, evsel_list,
850 							 fd, false);
851 		}
852 
853 		if (err < 0)
854 			return err;
855 
856 		err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
857 							 process_synthesized_event, is_pipe);
858 		if (err < 0)
859 			return err;
860 	}
861 
862 	if (target.initial_delay) {
863 		pr_info(EVLIST_DISABLED_MSG);
864 	} else {
865 		err = enable_counters();
866 		if (err)
867 			return -1;
868 	}
869 
870 	/* Exec the command, if any */
871 	if (forks)
872 		evlist__start_workload(evsel_list);
873 
874 	if (target.initial_delay > 0) {
875 		usleep(target.initial_delay * USEC_PER_MSEC);
876 		err = enable_counters();
877 		if (err)
878 			return -1;
879 
880 		pr_info(EVLIST_ENABLED_MSG);
881 	}
882 
883 	t0 = rdclock();
884 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
885 
886 	if (forks) {
887 		if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
888 			status = dispatch_events(forks, timeout, interval, &times);
889 		if (child_pid != -1) {
890 			if (timeout)
891 				kill(child_pid, SIGTERM);
892 			wait4(child_pid, &status, 0, &stat_config.ru_data);
893 		}
894 
895 		if (workload_exec_errno) {
896 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
897 			pr_err("Workload failed: %s\n", emsg);
898 			return -1;
899 		}
900 
901 		if (WIFSIGNALED(status))
902 			psignal(WTERMSIG(status), argv[0]);
903 	} else {
904 		status = dispatch_events(forks, timeout, interval, &times);
905 	}
906 
907 	disable_counters();
908 
909 	t1 = rdclock();
910 
911 	if (stat_config.walltime_run_table)
912 		stat_config.walltime_run[run_idx] = t1 - t0;
913 
914 	if (interval && stat_config.summary) {
915 		stat_config.interval = 0;
916 		stat_config.stop_read_counter = true;
917 		init_stats(&walltime_nsecs_stats);
918 		update_stats(&walltime_nsecs_stats, t1 - t0);
919 
920 		evlist__copy_prev_raw_counts(evsel_list);
921 		evlist__reset_prev_raw_counts(evsel_list);
922 		evlist__reset_aggr_stats(evsel_list);
923 	} else {
924 		update_stats(&walltime_nsecs_stats, t1 - t0);
925 		update_rusage_stats(&ru_stats, &stat_config.ru_data);
926 	}
927 
928 	/*
929 	 * Closing a group leader splits the group, and as we only disable
930 	 * group leaders, results in remaining events becoming enabled. To
931 	 * avoid arbitrary skew, we must read all counters before closing any
932 	 * group leaders.
933 	 */
934 	if (read_counters(&(struct timespec) { .tv_nsec = t1-t0 }) == 0)
935 		process_counters();
936 
937 	/*
938 	 * We need to keep evsel_list alive, because it's processed
939 	 * later the evsel_list will be closed after.
940 	 */
941 	if (!STAT_RECORD)
942 		evlist__close(evsel_list);
943 
944 	return WEXITSTATUS(status);
945 }
946 
947 static int run_perf_stat(int argc, const char **argv, int run_idx)
948 {
949 	int ret;
950 
951 	if (pre_cmd) {
952 		ret = system(pre_cmd);
953 		if (ret)
954 			return ret;
955 	}
956 
957 	if (sync_run)
958 		sync();
959 
960 	ret = __run_perf_stat(argc, argv, run_idx);
961 	if (ret)
962 		return ret;
963 
964 	if (post_cmd) {
965 		ret = system(post_cmd);
966 		if (ret)
967 			return ret;
968 	}
969 
970 	return ret;
971 }
972 
973 static void print_counters(struct timespec *ts, int argc, const char **argv)
974 {
975 	/* Do not print anything if we record to the pipe. */
976 	if (STAT_RECORD && perf_stat.data.is_pipe)
977 		return;
978 	if (quiet)
979 		return;
980 
981 	evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
982 }
983 
984 static volatile sig_atomic_t signr = -1;
985 
986 static void skip_signal(int signo)
987 {
988 	if ((child_pid == -1) || stat_config.interval)
989 		done = 1;
990 
991 	signr = signo;
992 	/*
993 	 * render child_pid harmless
994 	 * won't send SIGTERM to a random
995 	 * process in case of race condition
996 	 * and fast PID recycling
997 	 */
998 	child_pid = -1;
999 }
1000 
1001 static void sig_atexit(void)
1002 {
1003 	sigset_t set, oset;
1004 
1005 	/*
1006 	 * avoid race condition with SIGCHLD handler
1007 	 * in skip_signal() which is modifying child_pid
1008 	 * goal is to avoid send SIGTERM to a random
1009 	 * process
1010 	 */
1011 	sigemptyset(&set);
1012 	sigaddset(&set, SIGCHLD);
1013 	sigprocmask(SIG_BLOCK, &set, &oset);
1014 
1015 	if (child_pid != -1)
1016 		kill(child_pid, SIGTERM);
1017 
1018 	sigprocmask(SIG_SETMASK, &oset, NULL);
1019 
1020 	if (signr == -1)
1021 		return;
1022 
1023 	signal(signr, SIG_DFL);
1024 	kill(getpid(), signr);
1025 }
1026 
1027 void perf_stat__set_big_num(int set)
1028 {
1029 	stat_config.big_num = (set != 0);
1030 }
1031 
1032 void perf_stat__set_no_csv_summary(int set)
1033 {
1034 	stat_config.no_csv_summary = (set != 0);
1035 }
1036 
1037 static int stat__set_big_num(const struct option *opt __maybe_unused,
1038 			     const char *s __maybe_unused, int unset)
1039 {
1040 	big_num_opt = unset ? 0 : 1;
1041 	perf_stat__set_big_num(!unset);
1042 	return 0;
1043 }
1044 
1045 static int enable_metric_only(const struct option *opt __maybe_unused,
1046 			      const char *s __maybe_unused, int unset)
1047 {
1048 	force_metric_only = true;
1049 	stat_config.metric_only = !unset;
1050 	return 0;
1051 }
1052 
1053 static int append_metric_groups(const struct option *opt __maybe_unused,
1054 			       const char *str,
1055 			       int unset __maybe_unused)
1056 {
1057 	if (metrics) {
1058 		char *tmp;
1059 
1060 		if (asprintf(&tmp, "%s,%s", metrics, str) < 0)
1061 			return -ENOMEM;
1062 		free(metrics);
1063 		metrics = tmp;
1064 	} else {
1065 		metrics = strdup(str);
1066 		if (!metrics)
1067 			return -ENOMEM;
1068 	}
1069 	return 0;
1070 }
1071 
1072 static int parse_control_option(const struct option *opt,
1073 				const char *str,
1074 				int unset __maybe_unused)
1075 {
1076 	struct perf_stat_config *config = opt->value;
1077 
1078 	return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1079 }
1080 
1081 static int parse_stat_cgroups(const struct option *opt,
1082 			      const char *str, int unset)
1083 {
1084 	if (stat_config.cgroup_list) {
1085 		pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1086 		return -1;
1087 	}
1088 
1089 	return parse_cgroups(opt, str, unset);
1090 }
1091 
1092 static int parse_cputype(const struct option *opt,
1093 			     const char *str,
1094 			     int unset __maybe_unused)
1095 {
1096 	const struct perf_pmu *pmu;
1097 	struct evlist *evlist = *(struct evlist **)opt->value;
1098 
1099 	if (!list_empty(&evlist->core.entries)) {
1100 		fprintf(stderr, "Must define cputype before events/metrics\n");
1101 		return -1;
1102 	}
1103 
1104 	pmu = perf_pmus__pmu_for_pmu_filter(str);
1105 	if (!pmu) {
1106 		fprintf(stderr, "--cputype %s is not supported!\n", str);
1107 		return -1;
1108 	}
1109 	parse_events_option_args.pmu_filter = pmu->name;
1110 
1111 	return 0;
1112 }
1113 
1114 static int parse_cache_level(const struct option *opt,
1115 			     const char *str,
1116 			     int unset __maybe_unused)
1117 {
1118 	int level;
1119 	u32 *aggr_mode = (u32 *)opt->value;
1120 	u32 *aggr_level = (u32 *)opt->data;
1121 
1122 	/*
1123 	 * If no string is specified, aggregate based on the topology of
1124 	 * Last Level Cache (LLC). Since the LLC level can change from
1125 	 * architecture to architecture, set level greater than
1126 	 * MAX_CACHE_LVL which will be interpreted as LLC.
1127 	 */
1128 	if (str == NULL) {
1129 		level = MAX_CACHE_LVL + 1;
1130 		goto out;
1131 	}
1132 
1133 	/*
1134 	 * The format to specify cache level is LX or lX where X is the
1135 	 * cache level.
1136 	 */
1137 	if (strlen(str) != 2 || (str[0] != 'l' && str[0] != 'L')) {
1138 		pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1139 		       MAX_CACHE_LVL,
1140 		       MAX_CACHE_LVL);
1141 		return -EINVAL;
1142 	}
1143 
1144 	level = atoi(&str[1]);
1145 	if (level < 1) {
1146 		pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1147 		       MAX_CACHE_LVL,
1148 		       MAX_CACHE_LVL);
1149 		return -EINVAL;
1150 	}
1151 
1152 	if (level > MAX_CACHE_LVL) {
1153 		pr_err("perf only supports max cache level of %d.\n"
1154 		       "Consider increasing MAX_CACHE_LVL\n", MAX_CACHE_LVL);
1155 		return -EINVAL;
1156 	}
1157 out:
1158 	*aggr_mode = AGGR_CACHE;
1159 	*aggr_level = level;
1160 	return 0;
1161 }
1162 
1163 static struct option stat_options[] = {
1164 	OPT_BOOLEAN('T', "transaction", &transaction_run,
1165 		    "hardware transaction statistics"),
1166 	OPT_CALLBACK('e', "event", &parse_events_option_args, "event",
1167 		     "event selector. use 'perf list' to list available events",
1168 		     parse_events_option),
1169 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1170 		     "event filter", parse_filter),
1171 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1172 		    "child tasks do not inherit counters"),
1173 	OPT_STRING('p', "pid", &target.pid, "pid",
1174 		   "stat events on existing process id"),
1175 	OPT_STRING('t', "tid", &target.tid, "tid",
1176 		   "stat events on existing thread id"),
1177 #ifdef HAVE_BPF_SKEL
1178 	OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1179 		   "stat events on existing bpf program id"),
1180 	OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1181 		    "use bpf program to count events"),
1182 	OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1183 		   "path to perf_event_attr map"),
1184 #endif
1185 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1186 		    "system-wide collection from all CPUs"),
1187 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
1188 		    "Use --no-scale to disable counter scaling for multiplexing"),
1189 	OPT_INCR('v', "verbose", &verbose,
1190 		    "be more verbose (show counter open errors, etc)"),
1191 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
1192 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1193 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1194 		    "display details about each run (only with -r option)"),
1195 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
1196 		    "null run - dont start any counters"),
1197 	OPT_INCR('d', "detailed", &detailed_run,
1198 		    "detailed run - start a lot of events"),
1199 	OPT_BOOLEAN('S', "sync", &sync_run,
1200 		    "call sync() before starting a run"),
1201 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1202 			   "print large numbers with thousands\' separators",
1203 			   stat__set_big_num),
1204 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1205 		    "list of cpus to monitor in system-wide"),
1206 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1207 		    "disable CPU count aggregation", AGGR_NONE),
1208 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1209 	OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
1210 		    "Merge identical named hybrid events"),
1211 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1212 		   "print counts with custom separator"),
1213 	OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
1214 		   "print counts in JSON format"),
1215 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1216 		     "monitor event in cgroup name only", parse_stat_cgroups),
1217 	OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1218 		    "expand events for each cgroup"),
1219 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1220 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1221 	OPT_INTEGER(0, "log-fd", &output_fd,
1222 		    "log output to fd, instead of stderr"),
1223 	OPT_STRING(0, "pre", &pre_cmd, "command",
1224 			"command to run prior to the measured command"),
1225 	OPT_STRING(0, "post", &post_cmd, "command",
1226 			"command to run after to the measured command"),
1227 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1228 		    "print counts at regular interval in ms "
1229 		    "(overhead is possible for values <= 100ms)"),
1230 	OPT_INTEGER(0, "interval-count", &stat_config.times,
1231 		    "print counts for fixed number of times"),
1232 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1233 		    "clear screen in between new interval"),
1234 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1235 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1236 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1237 		     "aggregate counts per processor socket", AGGR_SOCKET),
1238 	OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1239 		     "aggregate counts per processor die", AGGR_DIE),
1240 	OPT_CALLBACK_OPTARG(0, "per-cache", &stat_config.aggr_mode, &stat_config.aggr_level,
1241 			    "cache level", "aggregate count at this cache level (Default: LLC)",
1242 			    parse_cache_level),
1243 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1244 		     "aggregate counts per physical processor core", AGGR_CORE),
1245 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1246 		     "aggregate counts per thread", AGGR_THREAD),
1247 	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1248 		     "aggregate counts per numa node", AGGR_NODE),
1249 	OPT_INTEGER('D', "delay", &target.initial_delay,
1250 		    "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1251 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1252 			"Only print computed metrics. No raw values", enable_metric_only),
1253 	OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1254 		       "don't group metric events, impacts multiplexing"),
1255 	OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1256 		       "don't try to share events between metrics in a group"),
1257 	OPT_BOOLEAN(0, "metric-no-threshold", &stat_config.metric_no_threshold,
1258 		       "don't try to share events between metrics in a group  "),
1259 	OPT_BOOLEAN(0, "topdown", &topdown_run,
1260 			"measure top-down statistics"),
1261 	OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1262 			"Set the metrics level for the top-down statistics (0: max level)"),
1263 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1264 			"measure SMI cost"),
1265 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1266 		     "monitor specified metrics or metric groups (separated by ,)",
1267 		     append_metric_groups),
1268 	OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1269 			 "Configure all used events to run in kernel space.",
1270 			 PARSE_OPT_EXCLUSIVE),
1271 	OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1272 			 "Configure all used events to run in user space.",
1273 			 PARSE_OPT_EXCLUSIVE),
1274 	OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1275 		    "Use with 'percore' event qualifier to show the event "
1276 		    "counts of one hardware thread by sum up total hardware "
1277 		    "threads of same physical core"),
1278 	OPT_BOOLEAN(0, "summary", &stat_config.summary,
1279 		       "print summary for interval mode"),
1280 	OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1281 		       "don't print 'summary' for CSV summary output"),
1282 	OPT_BOOLEAN(0, "quiet", &quiet,
1283 			"don't print any output, messages or warnings (useful with record)"),
1284 	OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1285 		     "Only enable events on applying cpu with this type "
1286 		     "for hybrid platform (e.g. core or atom)",
1287 		     parse_cputype),
1288 #ifdef HAVE_LIBPFM
1289 	OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1290 		"libpfm4 event selector. use 'perf list' to list available events",
1291 		parse_libpfm_events_option),
1292 #endif
1293 	OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1294 		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1295 		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1296 		     "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1297 		      parse_control_option),
1298 	OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1299 			    "measure I/O performance metrics provided by arch/platform",
1300 			    iostat_parse),
1301 	OPT_END()
1302 };
1303 
1304 /**
1305  * Calculate the cache instance ID from the map in
1306  * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1307  * Cache instance ID is the first CPU reported in the shared_cpu_list file.
1308  */
1309 static int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map)
1310 {
1311 	int id;
1312 	struct perf_cpu_map *cpu_map = perf_cpu_map__new(map);
1313 
1314 	/*
1315 	 * If the map contains no CPU, consider the current CPU to
1316 	 * be the first online CPU in the cache domain else use the
1317 	 * first online CPU of the cache domain as the ID.
1318 	 */
1319 	if (perf_cpu_map__empty(cpu_map))
1320 		id = cpu.cpu;
1321 	else
1322 		id = perf_cpu_map__cpu(cpu_map, 0).cpu;
1323 
1324 	/* Free the perf_cpu_map used to find the cache ID */
1325 	perf_cpu_map__put(cpu_map);
1326 
1327 	return id;
1328 }
1329 
1330 /**
1331  * cpu__get_cache_id - Returns 0 if successful in populating the
1332  * cache level and cache id. Cache level is read from
1333  * /sys/devices/system/cpu/cpuX/cache/indexY/level where as cache instance ID
1334  * is the first CPU reported by
1335  * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1336  */
1337 static int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache)
1338 {
1339 	int ret = 0;
1340 	u32 cache_level = stat_config.aggr_level;
1341 	struct cpu_cache_level caches[MAX_CACHE_LVL];
1342 	u32 i = 0, caches_cnt = 0;
1343 
1344 	cache->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1345 	cache->cache = -1;
1346 
1347 	ret = build_caches_for_cpu(cpu.cpu, caches, &caches_cnt);
1348 	if (ret) {
1349 		/*
1350 		 * If caches_cnt is not 0, cpu_cache_level data
1351 		 * was allocated when building the topology.
1352 		 * Free the allocated data before returning.
1353 		 */
1354 		if (caches_cnt)
1355 			goto free_caches;
1356 
1357 		return ret;
1358 	}
1359 
1360 	if (!caches_cnt)
1361 		return -1;
1362 
1363 	/*
1364 	 * Save the data for the highest level if no
1365 	 * level was specified by the user.
1366 	 */
1367 	if (cache_level > MAX_CACHE_LVL) {
1368 		int max_level_index = 0;
1369 
1370 		for (i = 1; i < caches_cnt; ++i) {
1371 			if (caches[i].level > caches[max_level_index].level)
1372 				max_level_index = i;
1373 		}
1374 
1375 		cache->cache_lvl = caches[max_level_index].level;
1376 		cache->cache = cpu__get_cache_id_from_map(cpu, caches[max_level_index].map);
1377 
1378 		/* Reset i to 0 to free entire caches[] */
1379 		i = 0;
1380 		goto free_caches;
1381 	}
1382 
1383 	for (i = 0; i < caches_cnt; ++i) {
1384 		if (caches[i].level == cache_level) {
1385 			cache->cache_lvl = cache_level;
1386 			cache->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1387 		}
1388 
1389 		cpu_cache_level__free(&caches[i]);
1390 	}
1391 
1392 free_caches:
1393 	/*
1394 	 * Free all the allocated cpu_cache_level data.
1395 	 */
1396 	while (i < caches_cnt)
1397 		cpu_cache_level__free(&caches[i++]);
1398 
1399 	return ret;
1400 }
1401 
1402 /**
1403  * aggr_cpu_id__cache - Create an aggr_cpu_id with cache instache ID, cache
1404  * level, die and socket populated with the cache instache ID, cache level,
1405  * die and socket for cpu. The function signature is compatible with
1406  * aggr_cpu_id_get_t.
1407  */
1408 static struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data)
1409 {
1410 	int ret;
1411 	struct aggr_cpu_id id;
1412 	struct perf_cache cache;
1413 
1414 	id = aggr_cpu_id__die(cpu, data);
1415 	if (aggr_cpu_id__is_empty(&id))
1416 		return id;
1417 
1418 	ret = cpu__get_cache_details(cpu, &cache);
1419 	if (ret)
1420 		return id;
1421 
1422 	id.cache_lvl = cache.cache_lvl;
1423 	id.cache = cache.cache;
1424 	return id;
1425 }
1426 
1427 static const char *const aggr_mode__string[] = {
1428 	[AGGR_CORE] = "core",
1429 	[AGGR_CACHE] = "cache",
1430 	[AGGR_DIE] = "die",
1431 	[AGGR_GLOBAL] = "global",
1432 	[AGGR_NODE] = "node",
1433 	[AGGR_NONE] = "none",
1434 	[AGGR_SOCKET] = "socket",
1435 	[AGGR_THREAD] = "thread",
1436 	[AGGR_UNSET] = "unset",
1437 };
1438 
1439 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1440 						struct perf_cpu cpu)
1441 {
1442 	return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1443 }
1444 
1445 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1446 					     struct perf_cpu cpu)
1447 {
1448 	return aggr_cpu_id__die(cpu, /*data=*/NULL);
1449 }
1450 
1451 static struct aggr_cpu_id perf_stat__get_cache_id(struct perf_stat_config *config __maybe_unused,
1452 						  struct perf_cpu cpu)
1453 {
1454 	return aggr_cpu_id__cache(cpu, /*data=*/NULL);
1455 }
1456 
1457 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1458 					      struct perf_cpu cpu)
1459 {
1460 	return aggr_cpu_id__core(cpu, /*data=*/NULL);
1461 }
1462 
1463 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1464 					      struct perf_cpu cpu)
1465 {
1466 	return aggr_cpu_id__node(cpu, /*data=*/NULL);
1467 }
1468 
1469 static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused,
1470 						struct perf_cpu cpu)
1471 {
1472 	return aggr_cpu_id__global(cpu, /*data=*/NULL);
1473 }
1474 
1475 static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused,
1476 					     struct perf_cpu cpu)
1477 {
1478 	return aggr_cpu_id__cpu(cpu, /*data=*/NULL);
1479 }
1480 
1481 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1482 					      aggr_get_id_t get_id, struct perf_cpu cpu)
1483 {
1484 	struct aggr_cpu_id id;
1485 
1486 	/* per-process mode - should use global aggr mode */
1487 	if (cpu.cpu == -1)
1488 		return get_id(config, cpu);
1489 
1490 	if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1491 		config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1492 
1493 	id = config->cpus_aggr_map->map[cpu.cpu];
1494 	return id;
1495 }
1496 
1497 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1498 						       struct perf_cpu cpu)
1499 {
1500 	return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1501 }
1502 
1503 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1504 						    struct perf_cpu cpu)
1505 {
1506 	return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1507 }
1508 
1509 static struct aggr_cpu_id perf_stat__get_cache_id_cached(struct perf_stat_config *config,
1510 							 struct perf_cpu cpu)
1511 {
1512 	return perf_stat__get_aggr(config, perf_stat__get_cache_id, cpu);
1513 }
1514 
1515 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1516 						     struct perf_cpu cpu)
1517 {
1518 	return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1519 }
1520 
1521 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1522 						     struct perf_cpu cpu)
1523 {
1524 	return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1525 }
1526 
1527 static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config,
1528 						       struct perf_cpu cpu)
1529 {
1530 	return perf_stat__get_aggr(config, perf_stat__get_global, cpu);
1531 }
1532 
1533 static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config,
1534 						    struct perf_cpu cpu)
1535 {
1536 	return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu);
1537 }
1538 
1539 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1540 {
1541 	switch (aggr_mode) {
1542 	case AGGR_SOCKET:
1543 		return aggr_cpu_id__socket;
1544 	case AGGR_DIE:
1545 		return aggr_cpu_id__die;
1546 	case AGGR_CACHE:
1547 		return aggr_cpu_id__cache;
1548 	case AGGR_CORE:
1549 		return aggr_cpu_id__core;
1550 	case AGGR_NODE:
1551 		return aggr_cpu_id__node;
1552 	case AGGR_NONE:
1553 		return aggr_cpu_id__cpu;
1554 	case AGGR_GLOBAL:
1555 		return aggr_cpu_id__global;
1556 	case AGGR_THREAD:
1557 	case AGGR_UNSET:
1558 	case AGGR_MAX:
1559 	default:
1560 		return NULL;
1561 	}
1562 }
1563 
1564 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1565 {
1566 	switch (aggr_mode) {
1567 	case AGGR_SOCKET:
1568 		return perf_stat__get_socket_cached;
1569 	case AGGR_DIE:
1570 		return perf_stat__get_die_cached;
1571 	case AGGR_CACHE:
1572 		return perf_stat__get_cache_id_cached;
1573 	case AGGR_CORE:
1574 		return perf_stat__get_core_cached;
1575 	case AGGR_NODE:
1576 		return perf_stat__get_node_cached;
1577 	case AGGR_NONE:
1578 		return perf_stat__get_cpu_cached;
1579 	case AGGR_GLOBAL:
1580 		return perf_stat__get_global_cached;
1581 	case AGGR_THREAD:
1582 	case AGGR_UNSET:
1583 	case AGGR_MAX:
1584 	default:
1585 		return NULL;
1586 	}
1587 }
1588 
1589 static int perf_stat_init_aggr_mode(void)
1590 {
1591 	int nr;
1592 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1593 
1594 	if (get_id) {
1595 		bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1596 		stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1597 							 get_id, /*data=*/NULL, needs_sort);
1598 		if (!stat_config.aggr_map) {
1599 			pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]);
1600 			return -1;
1601 		}
1602 		stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1603 	}
1604 
1605 	if (stat_config.aggr_mode == AGGR_THREAD) {
1606 		nr = perf_thread_map__nr(evsel_list->core.threads);
1607 		stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1608 		if (stat_config.aggr_map == NULL)
1609 			return -ENOMEM;
1610 
1611 		for (int s = 0; s < nr; s++) {
1612 			struct aggr_cpu_id id = aggr_cpu_id__empty();
1613 
1614 			id.thread_idx = s;
1615 			stat_config.aggr_map->map[s] = id;
1616 		}
1617 		return 0;
1618 	}
1619 
1620 	/*
1621 	 * The evsel_list->cpus is the base we operate on,
1622 	 * taking the highest cpu number to be the size of
1623 	 * the aggregation translate cpumap.
1624 	 */
1625 	if (evsel_list->core.user_requested_cpus)
1626 		nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1627 	else
1628 		nr = 0;
1629 	stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1630 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1631 }
1632 
1633 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1634 {
1635 	if (map) {
1636 		WARN_ONCE(refcount_read(&map->refcnt) != 0,
1637 			  "cpu_aggr_map refcnt unbalanced\n");
1638 		free(map);
1639 	}
1640 }
1641 
1642 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1643 {
1644 	if (map && refcount_dec_and_test(&map->refcnt))
1645 		cpu_aggr_map__delete(map);
1646 }
1647 
1648 static void perf_stat__exit_aggr_mode(void)
1649 {
1650 	cpu_aggr_map__put(stat_config.aggr_map);
1651 	cpu_aggr_map__put(stat_config.cpus_aggr_map);
1652 	stat_config.aggr_map = NULL;
1653 	stat_config.cpus_aggr_map = NULL;
1654 }
1655 
1656 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1657 {
1658 	struct perf_env *env = data;
1659 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1660 
1661 	if (cpu.cpu != -1)
1662 		id.socket = env->cpu[cpu.cpu].socket_id;
1663 
1664 	return id;
1665 }
1666 
1667 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1668 {
1669 	struct perf_env *env = data;
1670 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1671 
1672 	if (cpu.cpu != -1) {
1673 		/*
1674 		 * die_id is relative to socket, so start
1675 		 * with the socket ID and then add die to
1676 		 * make a unique ID.
1677 		 */
1678 		id.socket = env->cpu[cpu.cpu].socket_id;
1679 		id.die = env->cpu[cpu.cpu].die_id;
1680 	}
1681 
1682 	return id;
1683 }
1684 
1685 static void perf_env__get_cache_id_for_cpu(struct perf_cpu cpu, struct perf_env *env,
1686 					   u32 cache_level, struct aggr_cpu_id *id)
1687 {
1688 	int i;
1689 	int caches_cnt = env->caches_cnt;
1690 	struct cpu_cache_level *caches = env->caches;
1691 
1692 	id->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1693 	id->cache = -1;
1694 
1695 	if (!caches_cnt)
1696 		return;
1697 
1698 	for (i = caches_cnt - 1; i > -1; --i) {
1699 		struct perf_cpu_map *cpu_map;
1700 		int map_contains_cpu;
1701 
1702 		/*
1703 		 * If user has not specified a level, find the fist level with
1704 		 * the cpu in the map. Since building the map is expensive, do
1705 		 * this only if levels match.
1706 		 */
1707 		if (cache_level <= MAX_CACHE_LVL && caches[i].level != cache_level)
1708 			continue;
1709 
1710 		cpu_map = perf_cpu_map__new(caches[i].map);
1711 		map_contains_cpu = perf_cpu_map__idx(cpu_map, cpu);
1712 		perf_cpu_map__put(cpu_map);
1713 
1714 		if (map_contains_cpu != -1) {
1715 			id->cache_lvl = caches[i].level;
1716 			id->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1717 			return;
1718 		}
1719 	}
1720 }
1721 
1722 static struct aggr_cpu_id perf_env__get_cache_aggr_by_cpu(struct perf_cpu cpu,
1723 							  void *data)
1724 {
1725 	struct perf_env *env = data;
1726 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1727 
1728 	if (cpu.cpu != -1) {
1729 		u32 cache_level = (perf_stat.aggr_level) ?: stat_config.aggr_level;
1730 
1731 		id.socket = env->cpu[cpu.cpu].socket_id;
1732 		id.die = env->cpu[cpu.cpu].die_id;
1733 		perf_env__get_cache_id_for_cpu(cpu, env, cache_level, &id);
1734 	}
1735 
1736 	return id;
1737 }
1738 
1739 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1740 {
1741 	struct perf_env *env = data;
1742 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1743 
1744 	if (cpu.cpu != -1) {
1745 		/*
1746 		 * core_id is relative to socket and die,
1747 		 * we need a global id. So we set
1748 		 * socket, die id and core id
1749 		 */
1750 		id.socket = env->cpu[cpu.cpu].socket_id;
1751 		id.die = env->cpu[cpu.cpu].die_id;
1752 		id.core = env->cpu[cpu.cpu].core_id;
1753 	}
1754 
1755 	return id;
1756 }
1757 
1758 static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data)
1759 {
1760 	struct perf_env *env = data;
1761 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1762 
1763 	if (cpu.cpu != -1) {
1764 		/*
1765 		 * core_id is relative to socket and die,
1766 		 * we need a global id. So we set
1767 		 * socket, die id and core id
1768 		 */
1769 		id.socket = env->cpu[cpu.cpu].socket_id;
1770 		id.die = env->cpu[cpu.cpu].die_id;
1771 		id.core = env->cpu[cpu.cpu].core_id;
1772 		id.cpu = cpu;
1773 	}
1774 
1775 	return id;
1776 }
1777 
1778 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1779 {
1780 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1781 
1782 	id.node = perf_env__numa_node(data, cpu);
1783 	return id;
1784 }
1785 
1786 static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused,
1787 							   void *data __maybe_unused)
1788 {
1789 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1790 
1791 	/* it always aggregates to the cpu 0 */
1792 	id.cpu = (struct perf_cpu){ .cpu = 0 };
1793 	return id;
1794 }
1795 
1796 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1797 						     struct perf_cpu cpu)
1798 {
1799 	return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1800 }
1801 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1802 						  struct perf_cpu cpu)
1803 {
1804 	return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1805 }
1806 
1807 static struct aggr_cpu_id perf_stat__get_cache_file(struct perf_stat_config *config __maybe_unused,
1808 						    struct perf_cpu cpu)
1809 {
1810 	return perf_env__get_cache_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1811 }
1812 
1813 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1814 						   struct perf_cpu cpu)
1815 {
1816 	return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1817 }
1818 
1819 static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused,
1820 						  struct perf_cpu cpu)
1821 {
1822 	return perf_env__get_cpu_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1823 }
1824 
1825 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1826 						   struct perf_cpu cpu)
1827 {
1828 	return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1829 }
1830 
1831 static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused,
1832 						     struct perf_cpu cpu)
1833 {
1834 	return perf_env__get_global_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1835 }
1836 
1837 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1838 {
1839 	switch (aggr_mode) {
1840 	case AGGR_SOCKET:
1841 		return perf_env__get_socket_aggr_by_cpu;
1842 	case AGGR_DIE:
1843 		return perf_env__get_die_aggr_by_cpu;
1844 	case AGGR_CACHE:
1845 		return perf_env__get_cache_aggr_by_cpu;
1846 	case AGGR_CORE:
1847 		return perf_env__get_core_aggr_by_cpu;
1848 	case AGGR_NODE:
1849 		return perf_env__get_node_aggr_by_cpu;
1850 	case AGGR_GLOBAL:
1851 		return perf_env__get_global_aggr_by_cpu;
1852 	case AGGR_NONE:
1853 		return perf_env__get_cpu_aggr_by_cpu;
1854 	case AGGR_THREAD:
1855 	case AGGR_UNSET:
1856 	case AGGR_MAX:
1857 	default:
1858 		return NULL;
1859 	}
1860 }
1861 
1862 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1863 {
1864 	switch (aggr_mode) {
1865 	case AGGR_SOCKET:
1866 		return perf_stat__get_socket_file;
1867 	case AGGR_DIE:
1868 		return perf_stat__get_die_file;
1869 	case AGGR_CACHE:
1870 		return perf_stat__get_cache_file;
1871 	case AGGR_CORE:
1872 		return perf_stat__get_core_file;
1873 	case AGGR_NODE:
1874 		return perf_stat__get_node_file;
1875 	case AGGR_GLOBAL:
1876 		return perf_stat__get_global_file;
1877 	case AGGR_NONE:
1878 		return perf_stat__get_cpu_file;
1879 	case AGGR_THREAD:
1880 	case AGGR_UNSET:
1881 	case AGGR_MAX:
1882 	default:
1883 		return NULL;
1884 	}
1885 }
1886 
1887 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1888 {
1889 	struct perf_env *env = &st->session->header.env;
1890 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1891 	bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1892 
1893 	if (stat_config.aggr_mode == AGGR_THREAD) {
1894 		int nr = perf_thread_map__nr(evsel_list->core.threads);
1895 
1896 		stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1897 		if (stat_config.aggr_map == NULL)
1898 			return -ENOMEM;
1899 
1900 		for (int s = 0; s < nr; s++) {
1901 			struct aggr_cpu_id id = aggr_cpu_id__empty();
1902 
1903 			id.thread_idx = s;
1904 			stat_config.aggr_map->map[s] = id;
1905 		}
1906 		return 0;
1907 	}
1908 
1909 	if (!get_id)
1910 		return 0;
1911 
1912 	stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1913 						 get_id, env, needs_sort);
1914 	if (!stat_config.aggr_map) {
1915 		pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]);
1916 		return -1;
1917 	}
1918 	stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1919 	return 0;
1920 }
1921 
1922 /*
1923  * Add default attributes, if there were no attributes specified or
1924  * if -d/--detailed, -d -d or -d -d -d is used:
1925  */
1926 static int add_default_attributes(void)
1927 {
1928 	struct perf_event_attr default_attrs0[] = {
1929 
1930   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1931   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1932   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1933   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1934 
1935   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1936 };
1937 	struct perf_event_attr frontend_attrs[] = {
1938   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1939 };
1940 	struct perf_event_attr backend_attrs[] = {
1941   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1942 };
1943 	struct perf_event_attr default_attrs1[] = {
1944   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1945   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1946   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1947 
1948 };
1949 
1950 /*
1951  * Detailed stats (-d), covering the L1 and last level data caches:
1952  */
1953 	struct perf_event_attr detailed_attrs[] = {
1954 
1955   { .type = PERF_TYPE_HW_CACHE,
1956     .config =
1957 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1958 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1959 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1960 
1961   { .type = PERF_TYPE_HW_CACHE,
1962     .config =
1963 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1964 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1965 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1966 
1967   { .type = PERF_TYPE_HW_CACHE,
1968     .config =
1969 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1970 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1971 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1972 
1973   { .type = PERF_TYPE_HW_CACHE,
1974     .config =
1975 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1976 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1977 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1978 };
1979 
1980 /*
1981  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1982  */
1983 	struct perf_event_attr very_detailed_attrs[] = {
1984 
1985   { .type = PERF_TYPE_HW_CACHE,
1986     .config =
1987 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1988 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1989 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1990 
1991   { .type = PERF_TYPE_HW_CACHE,
1992     .config =
1993 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1994 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1995 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1996 
1997   { .type = PERF_TYPE_HW_CACHE,
1998     .config =
1999 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
2000 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
2001 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
2002 
2003   { .type = PERF_TYPE_HW_CACHE,
2004     .config =
2005 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
2006 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
2007 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
2008 
2009   { .type = PERF_TYPE_HW_CACHE,
2010     .config =
2011 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
2012 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
2013 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
2014 
2015   { .type = PERF_TYPE_HW_CACHE,
2016     .config =
2017 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
2018 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
2019 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
2020 
2021 };
2022 
2023 /*
2024  * Very, very detailed stats (-d -d -d), adding prefetch events:
2025  */
2026 	struct perf_event_attr very_very_detailed_attrs[] = {
2027 
2028   { .type = PERF_TYPE_HW_CACHE,
2029     .config =
2030 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
2031 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
2032 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
2033 
2034   { .type = PERF_TYPE_HW_CACHE,
2035     .config =
2036 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
2037 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
2038 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
2039 };
2040 
2041 	struct perf_event_attr default_null_attrs[] = {};
2042 	const char *pmu = parse_events_option_args.pmu_filter ?: "all";
2043 
2044 	/* Set attrs if no event is selected and !null_run: */
2045 	if (stat_config.null_run)
2046 		return 0;
2047 
2048 	if (transaction_run) {
2049 		/* Handle -T as -M transaction. Once platform specific metrics
2050 		 * support has been added to the json files, all architectures
2051 		 * will use this approach. To determine transaction support
2052 		 * on an architecture test for such a metric name.
2053 		 */
2054 		if (!metricgroup__has_metric(pmu, "transaction")) {
2055 			pr_err("Missing transaction metrics\n");
2056 			return -1;
2057 		}
2058 		return metricgroup__parse_groups(evsel_list, pmu, "transaction",
2059 						stat_config.metric_no_group,
2060 						stat_config.metric_no_merge,
2061 						stat_config.metric_no_threshold,
2062 						stat_config.user_requested_cpu_list,
2063 						stat_config.system_wide,
2064 						&stat_config.metric_events);
2065 	}
2066 
2067 	if (smi_cost) {
2068 		int smi;
2069 
2070 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2071 			pr_err("freeze_on_smi is not supported.\n");
2072 			return -1;
2073 		}
2074 
2075 		if (!smi) {
2076 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2077 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
2078 				return -1;
2079 			}
2080 			smi_reset = true;
2081 		}
2082 
2083 		if (!metricgroup__has_metric(pmu, "smi")) {
2084 			pr_err("Missing smi metrics\n");
2085 			return -1;
2086 		}
2087 
2088 		if (!force_metric_only)
2089 			stat_config.metric_only = true;
2090 
2091 		return metricgroup__parse_groups(evsel_list, pmu, "smi",
2092 						stat_config.metric_no_group,
2093 						stat_config.metric_no_merge,
2094 						stat_config.metric_no_threshold,
2095 						stat_config.user_requested_cpu_list,
2096 						stat_config.system_wide,
2097 						&stat_config.metric_events);
2098 	}
2099 
2100 	if (topdown_run) {
2101 		unsigned int max_level = metricgroups__topdown_max_level();
2102 		char str[] = "TopdownL1";
2103 
2104 		if (!force_metric_only)
2105 			stat_config.metric_only = true;
2106 
2107 		if (!max_level) {
2108 			pr_err("Topdown requested but the topdown metric groups aren't present.\n"
2109 				"(See perf list the metric groups have names like TopdownL1)\n");
2110 			return -1;
2111 		}
2112 		if (stat_config.topdown_level > max_level) {
2113 			pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
2114 			return -1;
2115 		} else if (!stat_config.topdown_level)
2116 			stat_config.topdown_level = 1;
2117 
2118 		if (!stat_config.interval && !stat_config.metric_only) {
2119 			fprintf(stat_config.output,
2120 				"Topdown accuracy may decrease when measuring long periods.\n"
2121 				"Please print the result regularly, e.g. -I1000\n");
2122 		}
2123 		str[8] = stat_config.topdown_level + '0';
2124 		if (metricgroup__parse_groups(evsel_list,
2125 						pmu, str,
2126 						/*metric_no_group=*/false,
2127 						/*metric_no_merge=*/false,
2128 						/*metric_no_threshold=*/true,
2129 						stat_config.user_requested_cpu_list,
2130 						stat_config.system_wide,
2131 						&stat_config.metric_events) < 0)
2132 			return -1;
2133 	}
2134 
2135 	if (!stat_config.topdown_level)
2136 		stat_config.topdown_level = 1;
2137 
2138 	if (!evsel_list->core.nr_entries) {
2139 		/* No events so add defaults. */
2140 		if (target__has_cpu(&target))
2141 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2142 
2143 		if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2144 			return -1;
2145 		if (perf_pmus__have_event("cpu", "stalled-cycles-frontend")) {
2146 			if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
2147 				return -1;
2148 		}
2149 		if (perf_pmus__have_event("cpu", "stalled-cycles-backend")) {
2150 			if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
2151 				return -1;
2152 		}
2153 		if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2154 			return -1;
2155 		/*
2156 		 * Add TopdownL1 metrics if they exist. To minimize
2157 		 * multiplexing, don't request threshold computation.
2158 		 */
2159 		if (metricgroup__has_metric(pmu, "Default")) {
2160 			struct evlist *metric_evlist = evlist__new();
2161 			struct evsel *metric_evsel;
2162 
2163 			if (!metric_evlist)
2164 				return -1;
2165 
2166 			if (metricgroup__parse_groups(metric_evlist, pmu, "Default",
2167 							/*metric_no_group=*/false,
2168 							/*metric_no_merge=*/false,
2169 							/*metric_no_threshold=*/true,
2170 							stat_config.user_requested_cpu_list,
2171 							stat_config.system_wide,
2172 							&stat_config.metric_events) < 0)
2173 				return -1;
2174 
2175 			evlist__for_each_entry(metric_evlist, metric_evsel) {
2176 				metric_evsel->skippable = true;
2177 				metric_evsel->default_metricgroup = true;
2178 			}
2179 			evlist__splice_list_tail(evsel_list, &metric_evlist->core.entries);
2180 			evlist__delete(metric_evlist);
2181 		}
2182 
2183 		/* Platform specific attrs */
2184 		if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0)
2185 			return -1;
2186 	}
2187 
2188 	/* Detailed events get appended to the event list: */
2189 
2190 	if (detailed_run <  1)
2191 		return 0;
2192 
2193 	/* Append detailed run extra attributes: */
2194 	if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2195 		return -1;
2196 
2197 	if (detailed_run < 2)
2198 		return 0;
2199 
2200 	/* Append very detailed run extra attributes: */
2201 	if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2202 		return -1;
2203 
2204 	if (detailed_run < 3)
2205 		return 0;
2206 
2207 	/* Append very, very detailed run extra attributes: */
2208 	return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2209 }
2210 
2211 static const char * const stat_record_usage[] = {
2212 	"perf stat record [<options>]",
2213 	NULL,
2214 };
2215 
2216 static void init_features(struct perf_session *session)
2217 {
2218 	int feat;
2219 
2220 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2221 		perf_header__set_feat(&session->header, feat);
2222 
2223 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
2224 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2225 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2226 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2227 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2228 }
2229 
2230 static int __cmd_record(int argc, const char **argv)
2231 {
2232 	struct perf_session *session;
2233 	struct perf_data *data = &perf_stat.data;
2234 
2235 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
2236 			     PARSE_OPT_STOP_AT_NON_OPTION);
2237 
2238 	if (output_name)
2239 		data->path = output_name;
2240 
2241 	if (stat_config.run_count != 1 || forever) {
2242 		pr_err("Cannot use -r option with perf stat record.\n");
2243 		return -1;
2244 	}
2245 
2246 	session = perf_session__new(data, NULL);
2247 	if (IS_ERR(session)) {
2248 		pr_err("Perf session creation failed\n");
2249 		return PTR_ERR(session);
2250 	}
2251 
2252 	init_features(session);
2253 
2254 	session->evlist   = evsel_list;
2255 	perf_stat.session = session;
2256 	perf_stat.record  = true;
2257 	return argc;
2258 }
2259 
2260 static int process_stat_round_event(struct perf_session *session,
2261 				    union perf_event *event)
2262 {
2263 	struct perf_record_stat_round *stat_round = &event->stat_round;
2264 	struct timespec tsh, *ts = NULL;
2265 	const char **argv = session->header.env.cmdline_argv;
2266 	int argc = session->header.env.nr_cmdline;
2267 
2268 	process_counters();
2269 
2270 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2271 		update_stats(&walltime_nsecs_stats, stat_round->time);
2272 
2273 	if (stat_config.interval && stat_round->time) {
2274 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2275 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2276 		ts = &tsh;
2277 	}
2278 
2279 	print_counters(ts, argc, argv);
2280 	return 0;
2281 }
2282 
2283 static
2284 int process_stat_config_event(struct perf_session *session,
2285 			      union perf_event *event)
2286 {
2287 	struct perf_tool *tool = session->tool;
2288 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2289 
2290 	perf_event__read_stat_config(&stat_config, &event->stat_config);
2291 
2292 	if (perf_cpu_map__empty(st->cpus)) {
2293 		if (st->aggr_mode != AGGR_UNSET)
2294 			pr_warning("warning: processing task data, aggregation mode not set\n");
2295 	} else if (st->aggr_mode != AGGR_UNSET) {
2296 		stat_config.aggr_mode = st->aggr_mode;
2297 	}
2298 
2299 	if (perf_stat.data.is_pipe)
2300 		perf_stat_init_aggr_mode();
2301 	else
2302 		perf_stat_init_aggr_mode_file(st);
2303 
2304 	if (stat_config.aggr_map) {
2305 		int nr_aggr = stat_config.aggr_map->nr;
2306 
2307 		if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) {
2308 			pr_err("cannot allocate aggr counts\n");
2309 			return -1;
2310 		}
2311 	}
2312 	return 0;
2313 }
2314 
2315 static int set_maps(struct perf_stat *st)
2316 {
2317 	if (!st->cpus || !st->threads)
2318 		return 0;
2319 
2320 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2321 		return -EINVAL;
2322 
2323 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2324 
2325 	if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true))
2326 		return -ENOMEM;
2327 
2328 	st->maps_allocated = true;
2329 	return 0;
2330 }
2331 
2332 static
2333 int process_thread_map_event(struct perf_session *session,
2334 			     union perf_event *event)
2335 {
2336 	struct perf_tool *tool = session->tool;
2337 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2338 
2339 	if (st->threads) {
2340 		pr_warning("Extra thread map event, ignoring.\n");
2341 		return 0;
2342 	}
2343 
2344 	st->threads = thread_map__new_event(&event->thread_map);
2345 	if (!st->threads)
2346 		return -ENOMEM;
2347 
2348 	return set_maps(st);
2349 }
2350 
2351 static
2352 int process_cpu_map_event(struct perf_session *session,
2353 			  union perf_event *event)
2354 {
2355 	struct perf_tool *tool = session->tool;
2356 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2357 	struct perf_cpu_map *cpus;
2358 
2359 	if (st->cpus) {
2360 		pr_warning("Extra cpu map event, ignoring.\n");
2361 		return 0;
2362 	}
2363 
2364 	cpus = cpu_map__new_data(&event->cpu_map.data);
2365 	if (!cpus)
2366 		return -ENOMEM;
2367 
2368 	st->cpus = cpus;
2369 	return set_maps(st);
2370 }
2371 
2372 static const char * const stat_report_usage[] = {
2373 	"perf stat report [<options>]",
2374 	NULL,
2375 };
2376 
2377 static struct perf_stat perf_stat = {
2378 	.tool = {
2379 		.attr		= perf_event__process_attr,
2380 		.event_update	= perf_event__process_event_update,
2381 		.thread_map	= process_thread_map_event,
2382 		.cpu_map	= process_cpu_map_event,
2383 		.stat_config	= process_stat_config_event,
2384 		.stat		= perf_event__process_stat_event,
2385 		.stat_round	= process_stat_round_event,
2386 	},
2387 	.aggr_mode	= AGGR_UNSET,
2388 	.aggr_level	= 0,
2389 };
2390 
2391 static int __cmd_report(int argc, const char **argv)
2392 {
2393 	struct perf_session *session;
2394 	const struct option options[] = {
2395 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2396 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2397 		     "aggregate counts per processor socket", AGGR_SOCKET),
2398 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2399 		     "aggregate counts per processor die", AGGR_DIE),
2400 	OPT_CALLBACK_OPTARG(0, "per-cache", &perf_stat.aggr_mode, &perf_stat.aggr_level,
2401 			    "cache level",
2402 			    "aggregate count at this cache level (Default: LLC)",
2403 			    parse_cache_level),
2404 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2405 		     "aggregate counts per physical processor core", AGGR_CORE),
2406 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2407 		     "aggregate counts per numa node", AGGR_NODE),
2408 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2409 		     "disable CPU count aggregation", AGGR_NONE),
2410 	OPT_END()
2411 	};
2412 	struct stat st;
2413 	int ret;
2414 
2415 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2416 
2417 	if (!input_name || !strlen(input_name)) {
2418 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2419 			input_name = "-";
2420 		else
2421 			input_name = "perf.data";
2422 	}
2423 
2424 	perf_stat.data.path = input_name;
2425 	perf_stat.data.mode = PERF_DATA_MODE_READ;
2426 
2427 	session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2428 	if (IS_ERR(session))
2429 		return PTR_ERR(session);
2430 
2431 	perf_stat.session  = session;
2432 	stat_config.output = stderr;
2433 	evlist__delete(evsel_list);
2434 	evsel_list         = session->evlist;
2435 
2436 	ret = perf_session__process_events(session);
2437 	if (ret)
2438 		return ret;
2439 
2440 	perf_session__delete(session);
2441 	return 0;
2442 }
2443 
2444 static void setup_system_wide(int forks)
2445 {
2446 	/*
2447 	 * Make system wide (-a) the default target if
2448 	 * no target was specified and one of following
2449 	 * conditions is met:
2450 	 *
2451 	 *   - there's no workload specified
2452 	 *   - there is workload specified but all requested
2453 	 *     events are system wide events
2454 	 */
2455 	if (!target__none(&target))
2456 		return;
2457 
2458 	if (!forks)
2459 		target.system_wide = true;
2460 	else {
2461 		struct evsel *counter;
2462 
2463 		evlist__for_each_entry(evsel_list, counter) {
2464 			if (!counter->core.requires_cpu &&
2465 			    !evsel__name_is(counter, "duration_time")) {
2466 				return;
2467 			}
2468 		}
2469 
2470 		if (evsel_list->core.nr_entries)
2471 			target.system_wide = true;
2472 	}
2473 }
2474 
2475 int cmd_stat(int argc, const char **argv)
2476 {
2477 	const char * const stat_usage[] = {
2478 		"perf stat [<options>] [<command>]",
2479 		NULL
2480 	};
2481 	int status = -EINVAL, run_idx, err;
2482 	const char *mode;
2483 	FILE *output = stderr;
2484 	unsigned int interval, timeout;
2485 	const char * const stat_subcommands[] = { "record", "report" };
2486 	char errbuf[BUFSIZ];
2487 
2488 	setlocale(LC_ALL, "");
2489 
2490 	evsel_list = evlist__new();
2491 	if (evsel_list == NULL)
2492 		return -ENOMEM;
2493 
2494 	parse_events__shrink_config_terms();
2495 
2496 	/* String-parsing callback-based options would segfault when negated */
2497 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2498 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2499 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2500 
2501 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2502 					(const char **) stat_usage,
2503 					PARSE_OPT_STOP_AT_NON_OPTION);
2504 
2505 	if (stat_config.csv_sep) {
2506 		stat_config.csv_output = true;
2507 		if (!strcmp(stat_config.csv_sep, "\\t"))
2508 			stat_config.csv_sep = "\t";
2509 	} else
2510 		stat_config.csv_sep = DEFAULT_SEPARATOR;
2511 
2512 	if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2513 		argc = __cmd_record(argc, argv);
2514 		if (argc < 0)
2515 			return -1;
2516 	} else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2517 		return __cmd_report(argc, argv);
2518 
2519 	interval = stat_config.interval;
2520 	timeout = stat_config.timeout;
2521 
2522 	/*
2523 	 * For record command the -o is already taken care of.
2524 	 */
2525 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2526 		output = NULL;
2527 
2528 	if (output_name && output_fd) {
2529 		fprintf(stderr, "cannot use both --output and --log-fd\n");
2530 		parse_options_usage(stat_usage, stat_options, "o", 1);
2531 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2532 		goto out;
2533 	}
2534 
2535 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2536 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2537 		goto out;
2538 	}
2539 
2540 	if (stat_config.metric_only && stat_config.run_count > 1) {
2541 		fprintf(stderr, "--metric-only is not supported with -r\n");
2542 		goto out;
2543 	}
2544 
2545 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2546 		fprintf(stderr, "--table is only supported with -r\n");
2547 		parse_options_usage(stat_usage, stat_options, "r", 1);
2548 		parse_options_usage(NULL, stat_options, "table", 0);
2549 		goto out;
2550 	}
2551 
2552 	if (output_fd < 0) {
2553 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2554 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2555 		goto out;
2556 	}
2557 
2558 	if (!output && !quiet) {
2559 		struct timespec tm;
2560 		mode = append_file ? "a" : "w";
2561 
2562 		output = fopen(output_name, mode);
2563 		if (!output) {
2564 			perror("failed to create output file");
2565 			return -1;
2566 		}
2567 		if (!stat_config.json_output) {
2568 			clock_gettime(CLOCK_REALTIME, &tm);
2569 			fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2570 		}
2571 	} else if (output_fd > 0) {
2572 		mode = append_file ? "a" : "w";
2573 		output = fdopen(output_fd, mode);
2574 		if (!output) {
2575 			perror("Failed opening logfd");
2576 			return -errno;
2577 		}
2578 	}
2579 
2580 	if (stat_config.interval_clear && !isatty(fileno(output))) {
2581 		fprintf(stderr, "--interval-clear does not work with output\n");
2582 		parse_options_usage(stat_usage, stat_options, "o", 1);
2583 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2584 		parse_options_usage(NULL, stat_options, "interval-clear", 0);
2585 		return -1;
2586 	}
2587 
2588 	stat_config.output = output;
2589 
2590 	/*
2591 	 * let the spreadsheet do the pretty-printing
2592 	 */
2593 	if (stat_config.csv_output) {
2594 		/* User explicitly passed -B? */
2595 		if (big_num_opt == 1) {
2596 			fprintf(stderr, "-B option not supported with -x\n");
2597 			parse_options_usage(stat_usage, stat_options, "B", 1);
2598 			parse_options_usage(NULL, stat_options, "x", 1);
2599 			goto out;
2600 		} else /* Nope, so disable big number formatting */
2601 			stat_config.big_num = false;
2602 	} else if (big_num_opt == 0) /* User passed --no-big-num */
2603 		stat_config.big_num = false;
2604 
2605 	err = target__validate(&target);
2606 	if (err) {
2607 		target__strerror(&target, err, errbuf, BUFSIZ);
2608 		pr_warning("%s\n", errbuf);
2609 	}
2610 
2611 	setup_system_wide(argc);
2612 
2613 	/*
2614 	 * Display user/system times only for single
2615 	 * run and when there's specified tracee.
2616 	 */
2617 	if ((stat_config.run_count == 1) && target__none(&target))
2618 		stat_config.ru_display = true;
2619 
2620 	if (stat_config.run_count < 0) {
2621 		pr_err("Run count must be a positive number\n");
2622 		parse_options_usage(stat_usage, stat_options, "r", 1);
2623 		goto out;
2624 	} else if (stat_config.run_count == 0) {
2625 		forever = true;
2626 		stat_config.run_count = 1;
2627 	}
2628 
2629 	if (stat_config.walltime_run_table) {
2630 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2631 		if (!stat_config.walltime_run) {
2632 			pr_err("failed to setup -r option");
2633 			goto out;
2634 		}
2635 	}
2636 
2637 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2638 		!target__has_task(&target)) {
2639 		if (!target.system_wide || target.cpu_list) {
2640 			fprintf(stderr, "The --per-thread option is only "
2641 				"available when monitoring via -p -t -a "
2642 				"options or only --per-thread.\n");
2643 			parse_options_usage(NULL, stat_options, "p", 1);
2644 			parse_options_usage(NULL, stat_options, "t", 1);
2645 			goto out;
2646 		}
2647 	}
2648 
2649 	/*
2650 	 * no_aggr, cgroup are for system-wide only
2651 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2652 	 */
2653 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2654 	      stat_config.aggr_mode != AGGR_THREAD) ||
2655 	     (nr_cgroups || stat_config.cgroup_list)) &&
2656 	    !target__has_cpu(&target)) {
2657 		fprintf(stderr, "both cgroup and no-aggregation "
2658 			"modes only available in system-wide mode\n");
2659 
2660 		parse_options_usage(stat_usage, stat_options, "G", 1);
2661 		parse_options_usage(NULL, stat_options, "A", 1);
2662 		parse_options_usage(NULL, stat_options, "a", 1);
2663 		parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2664 		goto out;
2665 	}
2666 
2667 	if (stat_config.iostat_run) {
2668 		status = iostat_prepare(evsel_list, &stat_config);
2669 		if (status)
2670 			goto out;
2671 		if (iostat_mode == IOSTAT_LIST) {
2672 			iostat_list(evsel_list, &stat_config);
2673 			goto out;
2674 		} else if (verbose > 0)
2675 			iostat_list(evsel_list, &stat_config);
2676 		if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2677 			target.system_wide = true;
2678 	}
2679 
2680 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2681 		target.per_thread = true;
2682 
2683 	stat_config.system_wide = target.system_wide;
2684 	if (target.cpu_list) {
2685 		stat_config.user_requested_cpu_list = strdup(target.cpu_list);
2686 		if (!stat_config.user_requested_cpu_list) {
2687 			status = -ENOMEM;
2688 			goto out;
2689 		}
2690 	}
2691 
2692 	/*
2693 	 * Metric parsing needs to be delayed as metrics may optimize events
2694 	 * knowing the target is system-wide.
2695 	 */
2696 	if (metrics) {
2697 		const char *pmu = parse_events_option_args.pmu_filter ?: "all";
2698 
2699 		metricgroup__parse_groups(evsel_list, pmu, metrics,
2700 					stat_config.metric_no_group,
2701 					stat_config.metric_no_merge,
2702 					stat_config.metric_no_threshold,
2703 					stat_config.user_requested_cpu_list,
2704 					stat_config.system_wide,
2705 					&stat_config.metric_events);
2706 		zfree(&metrics);
2707 	}
2708 
2709 	if (add_default_attributes())
2710 		goto out;
2711 
2712 	if (stat_config.cgroup_list) {
2713 		if (nr_cgroups > 0) {
2714 			pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2715 			parse_options_usage(stat_usage, stat_options, "G", 1);
2716 			parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2717 			goto out;
2718 		}
2719 
2720 		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2721 					  &stat_config.metric_events, true) < 0) {
2722 			parse_options_usage(stat_usage, stat_options,
2723 					    "for-each-cgroup", 0);
2724 			goto out;
2725 		}
2726 	}
2727 
2728 	evlist__warn_user_requested_cpus(evsel_list, target.cpu_list);
2729 
2730 	if (evlist__create_maps(evsel_list, &target) < 0) {
2731 		if (target__has_task(&target)) {
2732 			pr_err("Problems finding threads of monitor\n");
2733 			parse_options_usage(stat_usage, stat_options, "p", 1);
2734 			parse_options_usage(NULL, stat_options, "t", 1);
2735 		} else if (target__has_cpu(&target)) {
2736 			perror("failed to parse CPUs map");
2737 			parse_options_usage(stat_usage, stat_options, "C", 1);
2738 			parse_options_usage(NULL, stat_options, "a", 1);
2739 		}
2740 		goto out;
2741 	}
2742 
2743 	evlist__check_cpu_maps(evsel_list);
2744 
2745 	/*
2746 	 * Initialize thread_map with comm names,
2747 	 * so we could print it out on output.
2748 	 */
2749 	if (stat_config.aggr_mode == AGGR_THREAD) {
2750 		thread_map__read_comms(evsel_list->core.threads);
2751 	}
2752 
2753 	if (stat_config.aggr_mode == AGGR_NODE)
2754 		cpu__setup_cpunode_map();
2755 
2756 	if (stat_config.times && interval)
2757 		interval_count = true;
2758 	else if (stat_config.times && !interval) {
2759 		pr_err("interval-count option should be used together with "
2760 				"interval-print.\n");
2761 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2762 		parse_options_usage(stat_usage, stat_options, "I", 1);
2763 		goto out;
2764 	}
2765 
2766 	if (timeout && timeout < 100) {
2767 		if (timeout < 10) {
2768 			pr_err("timeout must be >= 10ms.\n");
2769 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2770 			goto out;
2771 		} else
2772 			pr_warning("timeout < 100ms. "
2773 				   "The overhead percentage could be high in some cases. "
2774 				   "Please proceed with caution.\n");
2775 	}
2776 	if (timeout && interval) {
2777 		pr_err("timeout option is not supported with interval-print.\n");
2778 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2779 		parse_options_usage(stat_usage, stat_options, "I", 1);
2780 		goto out;
2781 	}
2782 
2783 	if (perf_stat_init_aggr_mode())
2784 		goto out;
2785 
2786 	if (evlist__alloc_stats(&stat_config, evsel_list, interval))
2787 		goto out;
2788 
2789 	/*
2790 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2791 	 * while avoiding that older tools show confusing messages.
2792 	 *
2793 	 * However for pipe sessions we need to keep it zero,
2794 	 * because script's perf_evsel__check_attr is triggered
2795 	 * by attr->sample_type != 0, and we can't run it on
2796 	 * stat sessions.
2797 	 */
2798 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2799 
2800 	/*
2801 	 * We dont want to block the signals - that would cause
2802 	 * child tasks to inherit that and Ctrl-C would not work.
2803 	 * What we want is for Ctrl-C to work in the exec()-ed
2804 	 * task, but being ignored by perf stat itself:
2805 	 */
2806 	atexit(sig_atexit);
2807 	if (!forever)
2808 		signal(SIGINT,  skip_signal);
2809 	signal(SIGCHLD, skip_signal);
2810 	signal(SIGALRM, skip_signal);
2811 	signal(SIGABRT, skip_signal);
2812 
2813 	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2814 		goto out;
2815 
2816 	/* Enable ignoring missing threads when -p option is defined. */
2817 	evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2818 	status = 0;
2819 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2820 		if (stat_config.run_count != 1 && verbose > 0)
2821 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2822 				run_idx + 1);
2823 
2824 		if (run_idx != 0)
2825 			evlist__reset_prev_raw_counts(evsel_list);
2826 
2827 		status = run_perf_stat(argc, argv, run_idx);
2828 		if (forever && status != -1 && !interval) {
2829 			print_counters(NULL, argc, argv);
2830 			perf_stat__reset_stats();
2831 		}
2832 	}
2833 
2834 	if (!forever && status != -1 && (!interval || stat_config.summary)) {
2835 		if (stat_config.run_count > 1)
2836 			evlist__copy_res_stats(&stat_config, evsel_list);
2837 		print_counters(NULL, argc, argv);
2838 	}
2839 
2840 	evlist__finalize_ctlfd(evsel_list);
2841 
2842 	if (STAT_RECORD) {
2843 		/*
2844 		 * We synthesize the kernel mmap record just so that older tools
2845 		 * don't emit warnings about not being able to resolve symbols
2846 		 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2847 		 * a saner message about no samples being in the perf.data file.
2848 		 *
2849 		 * This also serves to suppress a warning about f_header.data.size == 0
2850 		 * in header.c at the moment 'perf stat record' gets introduced, which
2851 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2852 		 * records, but the need to suppress the kptr_restrict messages in older
2853 		 * tools remain  -acme
2854 		 */
2855 		int fd = perf_data__fd(&perf_stat.data);
2856 
2857 		err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2858 							 process_synthesized_event,
2859 							 &perf_stat.session->machines.host);
2860 		if (err) {
2861 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2862 				   "older tools may produce warnings about this file\n.");
2863 		}
2864 
2865 		if (!interval) {
2866 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2867 				pr_err("failed to write stat round event\n");
2868 		}
2869 
2870 		if (!perf_stat.data.is_pipe) {
2871 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2872 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2873 		}
2874 
2875 		evlist__close(evsel_list);
2876 		perf_session__delete(perf_stat.session);
2877 	}
2878 
2879 	perf_stat__exit_aggr_mode();
2880 	evlist__free_stats(evsel_list);
2881 out:
2882 	if (stat_config.iostat_run)
2883 		iostat_release(evsel_list);
2884 
2885 	zfree(&stat_config.walltime_run);
2886 	zfree(&stat_config.user_requested_cpu_list);
2887 
2888 	if (smi_cost && smi_reset)
2889 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2890 
2891 	evlist__delete(evsel_list);
2892 
2893 	metricgroup__rblist_exit(&stat_config.metric_events);
2894 	evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2895 
2896 	return status;
2897 }
2898