xref: /openbmc/linux/tools/perf/builtin-stat.c (revision 89df62c3)
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__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
729 		counter = evlist_cpu_itr.evsel;
730 
731 		/*
732 		 * bperf calls evsel__open_per_cpu() in bperf__load(), so
733 		 * no need to call it again here.
734 		 */
735 		if (target.use_bpf)
736 			break;
737 
738 		if (counter->reset_group || counter->errored)
739 			continue;
740 		if (evsel__is_bperf(counter))
741 			continue;
742 try_again:
743 		if (create_perf_stat_counter(counter, &stat_config, &target,
744 					     evlist_cpu_itr.cpu_map_idx) < 0) {
745 
746 			/*
747 			 * Weak group failed. We cannot just undo this here
748 			 * because earlier CPUs might be in group mode, and the kernel
749 			 * doesn't support mixing group and non group reads. Defer
750 			 * it to later.
751 			 * Don't close here because we're in the wrong affinity.
752 			 */
753 			if ((errno == EINVAL || errno == EBADF) &&
754 				evsel__leader(counter) != counter &&
755 				counter->weak_group) {
756 				evlist__reset_weak_group(evsel_list, counter, false);
757 				assert(counter->reset_group);
758 				second_pass = true;
759 				continue;
760 			}
761 
762 			switch (stat_handle_error(counter)) {
763 			case COUNTER_FATAL:
764 				return -1;
765 			case COUNTER_RETRY:
766 				goto try_again;
767 			case COUNTER_SKIP:
768 				continue;
769 			default:
770 				break;
771 			}
772 
773 		}
774 		counter->supported = true;
775 	}
776 
777 	if (second_pass) {
778 		/*
779 		 * Now redo all the weak group after closing them,
780 		 * and also close errored counters.
781 		 */
782 
783 		/* First close errored or weak retry */
784 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
785 			counter = evlist_cpu_itr.evsel;
786 
787 			if (!counter->reset_group && !counter->errored)
788 				continue;
789 
790 			perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
791 		}
792 		/* Now reopen weak */
793 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
794 			counter = evlist_cpu_itr.evsel;
795 
796 			if (!counter->reset_group)
797 				continue;
798 try_again_reset:
799 			pr_debug2("reopening weak %s\n", evsel__name(counter));
800 			if (create_perf_stat_counter(counter, &stat_config, &target,
801 						     evlist_cpu_itr.cpu_map_idx) < 0) {
802 
803 				switch (stat_handle_error(counter)) {
804 				case COUNTER_FATAL:
805 					return -1;
806 				case COUNTER_RETRY:
807 					goto try_again_reset;
808 				case COUNTER_SKIP:
809 					continue;
810 				default:
811 					break;
812 				}
813 			}
814 			counter->supported = true;
815 		}
816 	}
817 	affinity__cleanup(affinity);
818 
819 	evlist__for_each_entry(evsel_list, counter) {
820 		if (!counter->supported) {
821 			perf_evsel__free_fd(&counter->core);
822 			continue;
823 		}
824 
825 		l = strlen(counter->unit);
826 		if (l > stat_config.unit_width)
827 			stat_config.unit_width = l;
828 
829 		if (evsel__should_store_id(counter) &&
830 		    evsel__store_ids(counter, evsel_list))
831 			return -1;
832 	}
833 
834 	if (evlist__apply_filters(evsel_list, &counter)) {
835 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
836 			counter->filter, evsel__name(counter), errno,
837 			str_error_r(errno, msg, sizeof(msg)));
838 		return -1;
839 	}
840 
841 	if (STAT_RECORD) {
842 		int fd = perf_data__fd(&perf_stat.data);
843 
844 		if (is_pipe) {
845 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
846 		} else {
847 			err = perf_session__write_header(perf_stat.session, evsel_list,
848 							 fd, false);
849 		}
850 
851 		if (err < 0)
852 			return err;
853 
854 		err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
855 							 process_synthesized_event, is_pipe);
856 		if (err < 0)
857 			return err;
858 	}
859 
860 	if (target.initial_delay) {
861 		pr_info(EVLIST_DISABLED_MSG);
862 	} else {
863 		err = enable_counters();
864 		if (err)
865 			return -1;
866 	}
867 
868 	/* Exec the command, if any */
869 	if (forks)
870 		evlist__start_workload(evsel_list);
871 
872 	if (target.initial_delay > 0) {
873 		usleep(target.initial_delay * USEC_PER_MSEC);
874 		err = enable_counters();
875 		if (err)
876 			return -1;
877 
878 		pr_info(EVLIST_ENABLED_MSG);
879 	}
880 
881 	t0 = rdclock();
882 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
883 
884 	if (forks) {
885 		if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
886 			status = dispatch_events(forks, timeout, interval, &times);
887 		if (child_pid != -1) {
888 			if (timeout)
889 				kill(child_pid, SIGTERM);
890 			wait4(child_pid, &status, 0, &stat_config.ru_data);
891 		}
892 
893 		if (workload_exec_errno) {
894 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
895 			pr_err("Workload failed: %s\n", emsg);
896 			return -1;
897 		}
898 
899 		if (WIFSIGNALED(status))
900 			psignal(WTERMSIG(status), argv[0]);
901 	} else {
902 		status = dispatch_events(forks, timeout, interval, &times);
903 	}
904 
905 	disable_counters();
906 
907 	t1 = rdclock();
908 
909 	if (stat_config.walltime_run_table)
910 		stat_config.walltime_run[run_idx] = t1 - t0;
911 
912 	if (interval && stat_config.summary) {
913 		stat_config.interval = 0;
914 		stat_config.stop_read_counter = true;
915 		init_stats(&walltime_nsecs_stats);
916 		update_stats(&walltime_nsecs_stats, t1 - t0);
917 
918 		evlist__copy_prev_raw_counts(evsel_list);
919 		evlist__reset_prev_raw_counts(evsel_list);
920 		evlist__reset_aggr_stats(evsel_list);
921 	} else {
922 		update_stats(&walltime_nsecs_stats, t1 - t0);
923 		update_rusage_stats(&ru_stats, &stat_config.ru_data);
924 	}
925 
926 	/*
927 	 * Closing a group leader splits the group, and as we only disable
928 	 * group leaders, results in remaining events becoming enabled. To
929 	 * avoid arbitrary skew, we must read all counters before closing any
930 	 * group leaders.
931 	 */
932 	if (read_counters(&(struct timespec) { .tv_nsec = t1-t0 }) == 0)
933 		process_counters();
934 
935 	/*
936 	 * We need to keep evsel_list alive, because it's processed
937 	 * later the evsel_list will be closed after.
938 	 */
939 	if (!STAT_RECORD)
940 		evlist__close(evsel_list);
941 
942 	return WEXITSTATUS(status);
943 }
944 
945 static int run_perf_stat(int argc, const char **argv, int run_idx)
946 {
947 	int ret;
948 
949 	if (pre_cmd) {
950 		ret = system(pre_cmd);
951 		if (ret)
952 			return ret;
953 	}
954 
955 	if (sync_run)
956 		sync();
957 
958 	ret = __run_perf_stat(argc, argv, run_idx);
959 	if (ret)
960 		return ret;
961 
962 	if (post_cmd) {
963 		ret = system(post_cmd);
964 		if (ret)
965 			return ret;
966 	}
967 
968 	return ret;
969 }
970 
971 static void print_counters(struct timespec *ts, int argc, const char **argv)
972 {
973 	/* Do not print anything if we record to the pipe. */
974 	if (STAT_RECORD && perf_stat.data.is_pipe)
975 		return;
976 	if (quiet)
977 		return;
978 
979 	evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
980 }
981 
982 static volatile sig_atomic_t signr = -1;
983 
984 static void skip_signal(int signo)
985 {
986 	if ((child_pid == -1) || stat_config.interval)
987 		done = 1;
988 
989 	signr = signo;
990 	/*
991 	 * render child_pid harmless
992 	 * won't send SIGTERM to a random
993 	 * process in case of race condition
994 	 * and fast PID recycling
995 	 */
996 	child_pid = -1;
997 }
998 
999 static void sig_atexit(void)
1000 {
1001 	sigset_t set, oset;
1002 
1003 	/*
1004 	 * avoid race condition with SIGCHLD handler
1005 	 * in skip_signal() which is modifying child_pid
1006 	 * goal is to avoid send SIGTERM to a random
1007 	 * process
1008 	 */
1009 	sigemptyset(&set);
1010 	sigaddset(&set, SIGCHLD);
1011 	sigprocmask(SIG_BLOCK, &set, &oset);
1012 
1013 	if (child_pid != -1)
1014 		kill(child_pid, SIGTERM);
1015 
1016 	sigprocmask(SIG_SETMASK, &oset, NULL);
1017 
1018 	if (signr == -1)
1019 		return;
1020 
1021 	signal(signr, SIG_DFL);
1022 	kill(getpid(), signr);
1023 }
1024 
1025 void perf_stat__set_big_num(int set)
1026 {
1027 	stat_config.big_num = (set != 0);
1028 }
1029 
1030 void perf_stat__set_no_csv_summary(int set)
1031 {
1032 	stat_config.no_csv_summary = (set != 0);
1033 }
1034 
1035 static int stat__set_big_num(const struct option *opt __maybe_unused,
1036 			     const char *s __maybe_unused, int unset)
1037 {
1038 	big_num_opt = unset ? 0 : 1;
1039 	perf_stat__set_big_num(!unset);
1040 	return 0;
1041 }
1042 
1043 static int enable_metric_only(const struct option *opt __maybe_unused,
1044 			      const char *s __maybe_unused, int unset)
1045 {
1046 	force_metric_only = true;
1047 	stat_config.metric_only = !unset;
1048 	return 0;
1049 }
1050 
1051 static int append_metric_groups(const struct option *opt __maybe_unused,
1052 			       const char *str,
1053 			       int unset __maybe_unused)
1054 {
1055 	if (metrics) {
1056 		char *tmp;
1057 
1058 		if (asprintf(&tmp, "%s,%s", metrics, str) < 0)
1059 			return -ENOMEM;
1060 		free(metrics);
1061 		metrics = tmp;
1062 	} else {
1063 		metrics = strdup(str);
1064 		if (!metrics)
1065 			return -ENOMEM;
1066 	}
1067 	return 0;
1068 }
1069 
1070 static int parse_control_option(const struct option *opt,
1071 				const char *str,
1072 				int unset __maybe_unused)
1073 {
1074 	struct perf_stat_config *config = opt->value;
1075 
1076 	return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1077 }
1078 
1079 static int parse_stat_cgroups(const struct option *opt,
1080 			      const char *str, int unset)
1081 {
1082 	if (stat_config.cgroup_list) {
1083 		pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1084 		return -1;
1085 	}
1086 
1087 	return parse_cgroups(opt, str, unset);
1088 }
1089 
1090 static int parse_cputype(const struct option *opt,
1091 			     const char *str,
1092 			     int unset __maybe_unused)
1093 {
1094 	const struct perf_pmu *pmu;
1095 	struct evlist *evlist = *(struct evlist **)opt->value;
1096 
1097 	if (!list_empty(&evlist->core.entries)) {
1098 		fprintf(stderr, "Must define cputype before events/metrics\n");
1099 		return -1;
1100 	}
1101 
1102 	pmu = perf_pmus__pmu_for_pmu_filter(str);
1103 	if (!pmu) {
1104 		fprintf(stderr, "--cputype %s is not supported!\n", str);
1105 		return -1;
1106 	}
1107 	parse_events_option_args.pmu_filter = pmu->name;
1108 
1109 	return 0;
1110 }
1111 
1112 static int parse_cache_level(const struct option *opt,
1113 			     const char *str,
1114 			     int unset __maybe_unused)
1115 {
1116 	int level;
1117 	u32 *aggr_mode = (u32 *)opt->value;
1118 	u32 *aggr_level = (u32 *)opt->data;
1119 
1120 	/*
1121 	 * If no string is specified, aggregate based on the topology of
1122 	 * Last Level Cache (LLC). Since the LLC level can change from
1123 	 * architecture to architecture, set level greater than
1124 	 * MAX_CACHE_LVL which will be interpreted as LLC.
1125 	 */
1126 	if (str == NULL) {
1127 		level = MAX_CACHE_LVL + 1;
1128 		goto out;
1129 	}
1130 
1131 	/*
1132 	 * The format to specify cache level is LX or lX where X is the
1133 	 * cache level.
1134 	 */
1135 	if (strlen(str) != 2 || (str[0] != 'l' && str[0] != 'L')) {
1136 		pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1137 		       MAX_CACHE_LVL,
1138 		       MAX_CACHE_LVL);
1139 		return -EINVAL;
1140 	}
1141 
1142 	level = atoi(&str[1]);
1143 	if (level < 1) {
1144 		pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1145 		       MAX_CACHE_LVL,
1146 		       MAX_CACHE_LVL);
1147 		return -EINVAL;
1148 	}
1149 
1150 	if (level > MAX_CACHE_LVL) {
1151 		pr_err("perf only supports max cache level of %d.\n"
1152 		       "Consider increasing MAX_CACHE_LVL\n", MAX_CACHE_LVL);
1153 		return -EINVAL;
1154 	}
1155 out:
1156 	*aggr_mode = AGGR_CACHE;
1157 	*aggr_level = level;
1158 	return 0;
1159 }
1160 
1161 static struct option stat_options[] = {
1162 	OPT_BOOLEAN('T', "transaction", &transaction_run,
1163 		    "hardware transaction statistics"),
1164 	OPT_CALLBACK('e', "event", &parse_events_option_args, "event",
1165 		     "event selector. use 'perf list' to list available events",
1166 		     parse_events_option),
1167 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1168 		     "event filter", parse_filter),
1169 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1170 		    "child tasks do not inherit counters"),
1171 	OPT_STRING('p', "pid", &target.pid, "pid",
1172 		   "stat events on existing process id"),
1173 	OPT_STRING('t', "tid", &target.tid, "tid",
1174 		   "stat events on existing thread id"),
1175 #ifdef HAVE_BPF_SKEL
1176 	OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1177 		   "stat events on existing bpf program id"),
1178 	OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1179 		    "use bpf program to count events"),
1180 	OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1181 		   "path to perf_event_attr map"),
1182 #endif
1183 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1184 		    "system-wide collection from all CPUs"),
1185 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
1186 		    "Use --no-scale to disable counter scaling for multiplexing"),
1187 	OPT_INCR('v', "verbose", &verbose,
1188 		    "be more verbose (show counter open errors, etc)"),
1189 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
1190 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1191 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1192 		    "display details about each run (only with -r option)"),
1193 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
1194 		    "null run - dont start any counters"),
1195 	OPT_INCR('d', "detailed", &detailed_run,
1196 		    "detailed run - start a lot of events"),
1197 	OPT_BOOLEAN('S', "sync", &sync_run,
1198 		    "call sync() before starting a run"),
1199 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1200 			   "print large numbers with thousands\' separators",
1201 			   stat__set_big_num),
1202 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1203 		    "list of cpus to monitor in system-wide"),
1204 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1205 		    "disable CPU count aggregation", AGGR_NONE),
1206 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1207 	OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
1208 		    "Merge identical named hybrid events"),
1209 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1210 		   "print counts with custom separator"),
1211 	OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
1212 		   "print counts in JSON format"),
1213 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1214 		     "monitor event in cgroup name only", parse_stat_cgroups),
1215 	OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1216 		    "expand events for each cgroup"),
1217 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1218 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1219 	OPT_INTEGER(0, "log-fd", &output_fd,
1220 		    "log output to fd, instead of stderr"),
1221 	OPT_STRING(0, "pre", &pre_cmd, "command",
1222 			"command to run prior to the measured command"),
1223 	OPT_STRING(0, "post", &post_cmd, "command",
1224 			"command to run after to the measured command"),
1225 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1226 		    "print counts at regular interval in ms "
1227 		    "(overhead is possible for values <= 100ms)"),
1228 	OPT_INTEGER(0, "interval-count", &stat_config.times,
1229 		    "print counts for fixed number of times"),
1230 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1231 		    "clear screen in between new interval"),
1232 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1233 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1234 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1235 		     "aggregate counts per processor socket", AGGR_SOCKET),
1236 	OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1237 		     "aggregate counts per processor die", AGGR_DIE),
1238 	OPT_CALLBACK_OPTARG(0, "per-cache", &stat_config.aggr_mode, &stat_config.aggr_level,
1239 			    "cache level", "aggregate count at this cache level (Default: LLC)",
1240 			    parse_cache_level),
1241 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1242 		     "aggregate counts per physical processor core", AGGR_CORE),
1243 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1244 		     "aggregate counts per thread", AGGR_THREAD),
1245 	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1246 		     "aggregate counts per numa node", AGGR_NODE),
1247 	OPT_INTEGER('D', "delay", &target.initial_delay,
1248 		    "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1249 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1250 			"Only print computed metrics. No raw values", enable_metric_only),
1251 	OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1252 		       "don't group metric events, impacts multiplexing"),
1253 	OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1254 		       "don't try to share events between metrics in a group"),
1255 	OPT_BOOLEAN(0, "metric-no-threshold", &stat_config.metric_no_threshold,
1256 		       "don't try to share events between metrics in a group  "),
1257 	OPT_BOOLEAN(0, "topdown", &topdown_run,
1258 			"measure top-down statistics"),
1259 	OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1260 			"Set the metrics level for the top-down statistics (0: max level)"),
1261 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1262 			"measure SMI cost"),
1263 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1264 		     "monitor specified metrics or metric groups (separated by ,)",
1265 		     append_metric_groups),
1266 	OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1267 			 "Configure all used events to run in kernel space.",
1268 			 PARSE_OPT_EXCLUSIVE),
1269 	OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1270 			 "Configure all used events to run in user space.",
1271 			 PARSE_OPT_EXCLUSIVE),
1272 	OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1273 		    "Use with 'percore' event qualifier to show the event "
1274 		    "counts of one hardware thread by sum up total hardware "
1275 		    "threads of same physical core"),
1276 	OPT_BOOLEAN(0, "summary", &stat_config.summary,
1277 		       "print summary for interval mode"),
1278 	OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1279 		       "don't print 'summary' for CSV summary output"),
1280 	OPT_BOOLEAN(0, "quiet", &quiet,
1281 			"don't print any output, messages or warnings (useful with record)"),
1282 	OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1283 		     "Only enable events on applying cpu with this type "
1284 		     "for hybrid platform (e.g. core or atom)",
1285 		     parse_cputype),
1286 #ifdef HAVE_LIBPFM
1287 	OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1288 		"libpfm4 event selector. use 'perf list' to list available events",
1289 		parse_libpfm_events_option),
1290 #endif
1291 	OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1292 		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1293 		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1294 		     "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1295 		      parse_control_option),
1296 	OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1297 			    "measure I/O performance metrics provided by arch/platform",
1298 			    iostat_parse),
1299 	OPT_END()
1300 };
1301 
1302 /**
1303  * Calculate the cache instance ID from the map in
1304  * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1305  * Cache instance ID is the first CPU reported in the shared_cpu_list file.
1306  */
1307 static int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map)
1308 {
1309 	int id;
1310 	struct perf_cpu_map *cpu_map = perf_cpu_map__new(map);
1311 
1312 	/*
1313 	 * If the map contains no CPU, consider the current CPU to
1314 	 * be the first online CPU in the cache domain else use the
1315 	 * first online CPU of the cache domain as the ID.
1316 	 */
1317 	if (perf_cpu_map__empty(cpu_map))
1318 		id = cpu.cpu;
1319 	else
1320 		id = perf_cpu_map__cpu(cpu_map, 0).cpu;
1321 
1322 	/* Free the perf_cpu_map used to find the cache ID */
1323 	perf_cpu_map__put(cpu_map);
1324 
1325 	return id;
1326 }
1327 
1328 /**
1329  * cpu__get_cache_id - Returns 0 if successful in populating the
1330  * cache level and cache id. Cache level is read from
1331  * /sys/devices/system/cpu/cpuX/cache/indexY/level where as cache instance ID
1332  * is the first CPU reported by
1333  * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1334  */
1335 static int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache)
1336 {
1337 	int ret = 0;
1338 	u32 cache_level = stat_config.aggr_level;
1339 	struct cpu_cache_level caches[MAX_CACHE_LVL];
1340 	u32 i = 0, caches_cnt = 0;
1341 
1342 	cache->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1343 	cache->cache = -1;
1344 
1345 	ret = build_caches_for_cpu(cpu.cpu, caches, &caches_cnt);
1346 	if (ret) {
1347 		/*
1348 		 * If caches_cnt is not 0, cpu_cache_level data
1349 		 * was allocated when building the topology.
1350 		 * Free the allocated data before returning.
1351 		 */
1352 		if (caches_cnt)
1353 			goto free_caches;
1354 
1355 		return ret;
1356 	}
1357 
1358 	if (!caches_cnt)
1359 		return -1;
1360 
1361 	/*
1362 	 * Save the data for the highest level if no
1363 	 * level was specified by the user.
1364 	 */
1365 	if (cache_level > MAX_CACHE_LVL) {
1366 		int max_level_index = 0;
1367 
1368 		for (i = 1; i < caches_cnt; ++i) {
1369 			if (caches[i].level > caches[max_level_index].level)
1370 				max_level_index = i;
1371 		}
1372 
1373 		cache->cache_lvl = caches[max_level_index].level;
1374 		cache->cache = cpu__get_cache_id_from_map(cpu, caches[max_level_index].map);
1375 
1376 		/* Reset i to 0 to free entire caches[] */
1377 		i = 0;
1378 		goto free_caches;
1379 	}
1380 
1381 	for (i = 0; i < caches_cnt; ++i) {
1382 		if (caches[i].level == cache_level) {
1383 			cache->cache_lvl = cache_level;
1384 			cache->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1385 		}
1386 
1387 		cpu_cache_level__free(&caches[i]);
1388 	}
1389 
1390 free_caches:
1391 	/*
1392 	 * Free all the allocated cpu_cache_level data.
1393 	 */
1394 	while (i < caches_cnt)
1395 		cpu_cache_level__free(&caches[i++]);
1396 
1397 	return ret;
1398 }
1399 
1400 /**
1401  * aggr_cpu_id__cache - Create an aggr_cpu_id with cache instache ID, cache
1402  * level, die and socket populated with the cache instache ID, cache level,
1403  * die and socket for cpu. The function signature is compatible with
1404  * aggr_cpu_id_get_t.
1405  */
1406 static struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data)
1407 {
1408 	int ret;
1409 	struct aggr_cpu_id id;
1410 	struct perf_cache cache;
1411 
1412 	id = aggr_cpu_id__die(cpu, data);
1413 	if (aggr_cpu_id__is_empty(&id))
1414 		return id;
1415 
1416 	ret = cpu__get_cache_details(cpu, &cache);
1417 	if (ret)
1418 		return id;
1419 
1420 	id.cache_lvl = cache.cache_lvl;
1421 	id.cache = cache.cache;
1422 	return id;
1423 }
1424 
1425 static const char *const aggr_mode__string[] = {
1426 	[AGGR_CORE] = "core",
1427 	[AGGR_CACHE] = "cache",
1428 	[AGGR_DIE] = "die",
1429 	[AGGR_GLOBAL] = "global",
1430 	[AGGR_NODE] = "node",
1431 	[AGGR_NONE] = "none",
1432 	[AGGR_SOCKET] = "socket",
1433 	[AGGR_THREAD] = "thread",
1434 	[AGGR_UNSET] = "unset",
1435 };
1436 
1437 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1438 						struct perf_cpu cpu)
1439 {
1440 	return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1441 }
1442 
1443 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1444 					     struct perf_cpu cpu)
1445 {
1446 	return aggr_cpu_id__die(cpu, /*data=*/NULL);
1447 }
1448 
1449 static struct aggr_cpu_id perf_stat__get_cache_id(struct perf_stat_config *config __maybe_unused,
1450 						  struct perf_cpu cpu)
1451 {
1452 	return aggr_cpu_id__cache(cpu, /*data=*/NULL);
1453 }
1454 
1455 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1456 					      struct perf_cpu cpu)
1457 {
1458 	return aggr_cpu_id__core(cpu, /*data=*/NULL);
1459 }
1460 
1461 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1462 					      struct perf_cpu cpu)
1463 {
1464 	return aggr_cpu_id__node(cpu, /*data=*/NULL);
1465 }
1466 
1467 static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused,
1468 						struct perf_cpu cpu)
1469 {
1470 	return aggr_cpu_id__global(cpu, /*data=*/NULL);
1471 }
1472 
1473 static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused,
1474 					     struct perf_cpu cpu)
1475 {
1476 	return aggr_cpu_id__cpu(cpu, /*data=*/NULL);
1477 }
1478 
1479 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1480 					      aggr_get_id_t get_id, struct perf_cpu cpu)
1481 {
1482 	struct aggr_cpu_id id;
1483 
1484 	/* per-process mode - should use global aggr mode */
1485 	if (cpu.cpu == -1)
1486 		return get_id(config, cpu);
1487 
1488 	if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1489 		config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1490 
1491 	id = config->cpus_aggr_map->map[cpu.cpu];
1492 	return id;
1493 }
1494 
1495 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1496 						       struct perf_cpu cpu)
1497 {
1498 	return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1499 }
1500 
1501 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1502 						    struct perf_cpu cpu)
1503 {
1504 	return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1505 }
1506 
1507 static struct aggr_cpu_id perf_stat__get_cache_id_cached(struct perf_stat_config *config,
1508 							 struct perf_cpu cpu)
1509 {
1510 	return perf_stat__get_aggr(config, perf_stat__get_cache_id, cpu);
1511 }
1512 
1513 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1514 						     struct perf_cpu cpu)
1515 {
1516 	return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1517 }
1518 
1519 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1520 						     struct perf_cpu cpu)
1521 {
1522 	return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1523 }
1524 
1525 static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config,
1526 						       struct perf_cpu cpu)
1527 {
1528 	return perf_stat__get_aggr(config, perf_stat__get_global, cpu);
1529 }
1530 
1531 static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config,
1532 						    struct perf_cpu cpu)
1533 {
1534 	return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu);
1535 }
1536 
1537 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1538 {
1539 	switch (aggr_mode) {
1540 	case AGGR_SOCKET:
1541 		return aggr_cpu_id__socket;
1542 	case AGGR_DIE:
1543 		return aggr_cpu_id__die;
1544 	case AGGR_CACHE:
1545 		return aggr_cpu_id__cache;
1546 	case AGGR_CORE:
1547 		return aggr_cpu_id__core;
1548 	case AGGR_NODE:
1549 		return aggr_cpu_id__node;
1550 	case AGGR_NONE:
1551 		return aggr_cpu_id__cpu;
1552 	case AGGR_GLOBAL:
1553 		return aggr_cpu_id__global;
1554 	case AGGR_THREAD:
1555 	case AGGR_UNSET:
1556 	case AGGR_MAX:
1557 	default:
1558 		return NULL;
1559 	}
1560 }
1561 
1562 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1563 {
1564 	switch (aggr_mode) {
1565 	case AGGR_SOCKET:
1566 		return perf_stat__get_socket_cached;
1567 	case AGGR_DIE:
1568 		return perf_stat__get_die_cached;
1569 	case AGGR_CACHE:
1570 		return perf_stat__get_cache_id_cached;
1571 	case AGGR_CORE:
1572 		return perf_stat__get_core_cached;
1573 	case AGGR_NODE:
1574 		return perf_stat__get_node_cached;
1575 	case AGGR_NONE:
1576 		return perf_stat__get_cpu_cached;
1577 	case AGGR_GLOBAL:
1578 		return perf_stat__get_global_cached;
1579 	case AGGR_THREAD:
1580 	case AGGR_UNSET:
1581 	case AGGR_MAX:
1582 	default:
1583 		return NULL;
1584 	}
1585 }
1586 
1587 static int perf_stat_init_aggr_mode(void)
1588 {
1589 	int nr;
1590 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1591 
1592 	if (get_id) {
1593 		bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1594 		stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1595 							 get_id, /*data=*/NULL, needs_sort);
1596 		if (!stat_config.aggr_map) {
1597 			pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1598 			return -1;
1599 		}
1600 		stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1601 	}
1602 
1603 	if (stat_config.aggr_mode == AGGR_THREAD) {
1604 		nr = perf_thread_map__nr(evsel_list->core.threads);
1605 		stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1606 		if (stat_config.aggr_map == NULL)
1607 			return -ENOMEM;
1608 
1609 		for (int s = 0; s < nr; s++) {
1610 			struct aggr_cpu_id id = aggr_cpu_id__empty();
1611 
1612 			id.thread_idx = s;
1613 			stat_config.aggr_map->map[s] = id;
1614 		}
1615 		return 0;
1616 	}
1617 
1618 	/*
1619 	 * The evsel_list->cpus is the base we operate on,
1620 	 * taking the highest cpu number to be the size of
1621 	 * the aggregation translate cpumap.
1622 	 */
1623 	if (evsel_list->core.user_requested_cpus)
1624 		nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1625 	else
1626 		nr = 0;
1627 	stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1628 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1629 }
1630 
1631 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1632 {
1633 	if (map) {
1634 		WARN_ONCE(refcount_read(&map->refcnt) != 0,
1635 			  "cpu_aggr_map refcnt unbalanced\n");
1636 		free(map);
1637 	}
1638 }
1639 
1640 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1641 {
1642 	if (map && refcount_dec_and_test(&map->refcnt))
1643 		cpu_aggr_map__delete(map);
1644 }
1645 
1646 static void perf_stat__exit_aggr_mode(void)
1647 {
1648 	cpu_aggr_map__put(stat_config.aggr_map);
1649 	cpu_aggr_map__put(stat_config.cpus_aggr_map);
1650 	stat_config.aggr_map = NULL;
1651 	stat_config.cpus_aggr_map = NULL;
1652 }
1653 
1654 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1655 {
1656 	struct perf_env *env = data;
1657 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1658 
1659 	if (cpu.cpu != -1)
1660 		id.socket = env->cpu[cpu.cpu].socket_id;
1661 
1662 	return id;
1663 }
1664 
1665 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1666 {
1667 	struct perf_env *env = data;
1668 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1669 
1670 	if (cpu.cpu != -1) {
1671 		/*
1672 		 * die_id is relative to socket, so start
1673 		 * with the socket ID and then add die to
1674 		 * make a unique ID.
1675 		 */
1676 		id.socket = env->cpu[cpu.cpu].socket_id;
1677 		id.die = env->cpu[cpu.cpu].die_id;
1678 	}
1679 
1680 	return id;
1681 }
1682 
1683 static void perf_env__get_cache_id_for_cpu(struct perf_cpu cpu, struct perf_env *env,
1684 					   u32 cache_level, struct aggr_cpu_id *id)
1685 {
1686 	int i;
1687 	int caches_cnt = env->caches_cnt;
1688 	struct cpu_cache_level *caches = env->caches;
1689 
1690 	id->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1691 	id->cache = -1;
1692 
1693 	if (!caches_cnt)
1694 		return;
1695 
1696 	for (i = caches_cnt - 1; i > -1; --i) {
1697 		struct perf_cpu_map *cpu_map;
1698 		int map_contains_cpu;
1699 
1700 		/*
1701 		 * If user has not specified a level, find the fist level with
1702 		 * the cpu in the map. Since building the map is expensive, do
1703 		 * this only if levels match.
1704 		 */
1705 		if (cache_level <= MAX_CACHE_LVL && caches[i].level != cache_level)
1706 			continue;
1707 
1708 		cpu_map = perf_cpu_map__new(caches[i].map);
1709 		map_contains_cpu = perf_cpu_map__idx(cpu_map, cpu);
1710 		perf_cpu_map__put(cpu_map);
1711 
1712 		if (map_contains_cpu != -1) {
1713 			id->cache_lvl = caches[i].level;
1714 			id->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1715 			return;
1716 		}
1717 	}
1718 }
1719 
1720 static struct aggr_cpu_id perf_env__get_cache_aggr_by_cpu(struct perf_cpu cpu,
1721 							  void *data)
1722 {
1723 	struct perf_env *env = data;
1724 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1725 
1726 	if (cpu.cpu != -1) {
1727 		u32 cache_level = (perf_stat.aggr_level) ?: stat_config.aggr_level;
1728 
1729 		id.socket = env->cpu[cpu.cpu].socket_id;
1730 		id.die = env->cpu[cpu.cpu].die_id;
1731 		perf_env__get_cache_id_for_cpu(cpu, env, cache_level, &id);
1732 	}
1733 
1734 	return id;
1735 }
1736 
1737 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1738 {
1739 	struct perf_env *env = data;
1740 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1741 
1742 	if (cpu.cpu != -1) {
1743 		/*
1744 		 * core_id is relative to socket and die,
1745 		 * we need a global id. So we set
1746 		 * socket, die id and core id
1747 		 */
1748 		id.socket = env->cpu[cpu.cpu].socket_id;
1749 		id.die = env->cpu[cpu.cpu].die_id;
1750 		id.core = env->cpu[cpu.cpu].core_id;
1751 	}
1752 
1753 	return id;
1754 }
1755 
1756 static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data)
1757 {
1758 	struct perf_env *env = data;
1759 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1760 
1761 	if (cpu.cpu != -1) {
1762 		/*
1763 		 * core_id is relative to socket and die,
1764 		 * we need a global id. So we set
1765 		 * socket, die id and core id
1766 		 */
1767 		id.socket = env->cpu[cpu.cpu].socket_id;
1768 		id.die = env->cpu[cpu.cpu].die_id;
1769 		id.core = env->cpu[cpu.cpu].core_id;
1770 		id.cpu = cpu;
1771 	}
1772 
1773 	return id;
1774 }
1775 
1776 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1777 {
1778 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1779 
1780 	id.node = perf_env__numa_node(data, cpu);
1781 	return id;
1782 }
1783 
1784 static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused,
1785 							   void *data __maybe_unused)
1786 {
1787 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1788 
1789 	/* it always aggregates to the cpu 0 */
1790 	id.cpu = (struct perf_cpu){ .cpu = 0 };
1791 	return id;
1792 }
1793 
1794 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1795 						     struct perf_cpu cpu)
1796 {
1797 	return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1798 }
1799 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1800 						  struct perf_cpu cpu)
1801 {
1802 	return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1803 }
1804 
1805 static struct aggr_cpu_id perf_stat__get_cache_file(struct perf_stat_config *config __maybe_unused,
1806 						    struct perf_cpu cpu)
1807 {
1808 	return perf_env__get_cache_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1809 }
1810 
1811 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1812 						   struct perf_cpu cpu)
1813 {
1814 	return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1815 }
1816 
1817 static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused,
1818 						  struct perf_cpu cpu)
1819 {
1820 	return perf_env__get_cpu_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1821 }
1822 
1823 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1824 						   struct perf_cpu cpu)
1825 {
1826 	return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1827 }
1828 
1829 static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused,
1830 						     struct perf_cpu cpu)
1831 {
1832 	return perf_env__get_global_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1833 }
1834 
1835 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1836 {
1837 	switch (aggr_mode) {
1838 	case AGGR_SOCKET:
1839 		return perf_env__get_socket_aggr_by_cpu;
1840 	case AGGR_DIE:
1841 		return perf_env__get_die_aggr_by_cpu;
1842 	case AGGR_CACHE:
1843 		return perf_env__get_cache_aggr_by_cpu;
1844 	case AGGR_CORE:
1845 		return perf_env__get_core_aggr_by_cpu;
1846 	case AGGR_NODE:
1847 		return perf_env__get_node_aggr_by_cpu;
1848 	case AGGR_GLOBAL:
1849 		return perf_env__get_global_aggr_by_cpu;
1850 	case AGGR_NONE:
1851 		return perf_env__get_cpu_aggr_by_cpu;
1852 	case AGGR_THREAD:
1853 	case AGGR_UNSET:
1854 	case AGGR_MAX:
1855 	default:
1856 		return NULL;
1857 	}
1858 }
1859 
1860 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1861 {
1862 	switch (aggr_mode) {
1863 	case AGGR_SOCKET:
1864 		return perf_stat__get_socket_file;
1865 	case AGGR_DIE:
1866 		return perf_stat__get_die_file;
1867 	case AGGR_CACHE:
1868 		return perf_stat__get_cache_file;
1869 	case AGGR_CORE:
1870 		return perf_stat__get_core_file;
1871 	case AGGR_NODE:
1872 		return perf_stat__get_node_file;
1873 	case AGGR_GLOBAL:
1874 		return perf_stat__get_global_file;
1875 	case AGGR_NONE:
1876 		return perf_stat__get_cpu_file;
1877 	case AGGR_THREAD:
1878 	case AGGR_UNSET:
1879 	case AGGR_MAX:
1880 	default:
1881 		return NULL;
1882 	}
1883 }
1884 
1885 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1886 {
1887 	struct perf_env *env = &st->session->header.env;
1888 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1889 	bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1890 
1891 	if (stat_config.aggr_mode == AGGR_THREAD) {
1892 		int nr = perf_thread_map__nr(evsel_list->core.threads);
1893 
1894 		stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1895 		if (stat_config.aggr_map == NULL)
1896 			return -ENOMEM;
1897 
1898 		for (int s = 0; s < nr; s++) {
1899 			struct aggr_cpu_id id = aggr_cpu_id__empty();
1900 
1901 			id.thread_idx = s;
1902 			stat_config.aggr_map->map[s] = id;
1903 		}
1904 		return 0;
1905 	}
1906 
1907 	if (!get_id)
1908 		return 0;
1909 
1910 	stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1911 						 get_id, env, needs_sort);
1912 	if (!stat_config.aggr_map) {
1913 		pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1914 		return -1;
1915 	}
1916 	stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1917 	return 0;
1918 }
1919 
1920 /*
1921  * Add default attributes, if there were no attributes specified or
1922  * if -d/--detailed, -d -d or -d -d -d is used:
1923  */
1924 static int add_default_attributes(void)
1925 {
1926 	struct perf_event_attr default_attrs0[] = {
1927 
1928   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1929   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1930   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1931   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1932 
1933   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1934 };
1935 	struct perf_event_attr frontend_attrs[] = {
1936   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1937 };
1938 	struct perf_event_attr backend_attrs[] = {
1939   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1940 };
1941 	struct perf_event_attr default_attrs1[] = {
1942   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1943   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1944   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1945 
1946 };
1947 
1948 /*
1949  * Detailed stats (-d), covering the L1 and last level data caches:
1950  */
1951 	struct perf_event_attr detailed_attrs[] = {
1952 
1953   { .type = PERF_TYPE_HW_CACHE,
1954     .config =
1955 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1956 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1957 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1958 
1959   { .type = PERF_TYPE_HW_CACHE,
1960     .config =
1961 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1962 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1963 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1964 
1965   { .type = PERF_TYPE_HW_CACHE,
1966     .config =
1967 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1968 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1969 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1970 
1971   { .type = PERF_TYPE_HW_CACHE,
1972     .config =
1973 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1974 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1975 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1976 };
1977 
1978 /*
1979  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1980  */
1981 	struct perf_event_attr very_detailed_attrs[] = {
1982 
1983   { .type = PERF_TYPE_HW_CACHE,
1984     .config =
1985 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1986 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1987 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1988 
1989   { .type = PERF_TYPE_HW_CACHE,
1990     .config =
1991 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1992 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1993 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1994 
1995   { .type = PERF_TYPE_HW_CACHE,
1996     .config =
1997 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1998 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1999 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
2000 
2001   { .type = PERF_TYPE_HW_CACHE,
2002     .config =
2003 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
2004 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
2005 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
2006 
2007   { .type = PERF_TYPE_HW_CACHE,
2008     .config =
2009 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
2010 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
2011 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
2012 
2013   { .type = PERF_TYPE_HW_CACHE,
2014     .config =
2015 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
2016 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
2017 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
2018 
2019 };
2020 
2021 /*
2022  * Very, very detailed stats (-d -d -d), adding prefetch events:
2023  */
2024 	struct perf_event_attr very_very_detailed_attrs[] = {
2025 
2026   { .type = PERF_TYPE_HW_CACHE,
2027     .config =
2028 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
2029 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
2030 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
2031 
2032   { .type = PERF_TYPE_HW_CACHE,
2033     .config =
2034 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
2035 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
2036 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
2037 };
2038 
2039 	struct perf_event_attr default_null_attrs[] = {};
2040 	const char *pmu = parse_events_option_args.pmu_filter ?: "all";
2041 
2042 	/* Set attrs if no event is selected and !null_run: */
2043 	if (stat_config.null_run)
2044 		return 0;
2045 
2046 	if (transaction_run) {
2047 		/* Handle -T as -M transaction. Once platform specific metrics
2048 		 * support has been added to the json files, all architectures
2049 		 * will use this approach. To determine transaction support
2050 		 * on an architecture test for such a metric name.
2051 		 */
2052 		if (!metricgroup__has_metric(pmu, "transaction")) {
2053 			pr_err("Missing transaction metrics");
2054 			return -1;
2055 		}
2056 		return metricgroup__parse_groups(evsel_list, pmu, "transaction",
2057 						stat_config.metric_no_group,
2058 						stat_config.metric_no_merge,
2059 						stat_config.metric_no_threshold,
2060 						stat_config.user_requested_cpu_list,
2061 						stat_config.system_wide,
2062 						&stat_config.metric_events);
2063 	}
2064 
2065 	if (smi_cost) {
2066 		int smi;
2067 
2068 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2069 			pr_err("freeze_on_smi is not supported.");
2070 			return -1;
2071 		}
2072 
2073 		if (!smi) {
2074 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2075 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
2076 				return -1;
2077 			}
2078 			smi_reset = true;
2079 		}
2080 
2081 		if (!metricgroup__has_metric(pmu, "smi")) {
2082 			pr_err("Missing smi metrics");
2083 			return -1;
2084 		}
2085 
2086 		if (!force_metric_only)
2087 			stat_config.metric_only = true;
2088 
2089 		return metricgroup__parse_groups(evsel_list, pmu, "smi",
2090 						stat_config.metric_no_group,
2091 						stat_config.metric_no_merge,
2092 						stat_config.metric_no_threshold,
2093 						stat_config.user_requested_cpu_list,
2094 						stat_config.system_wide,
2095 						&stat_config.metric_events);
2096 	}
2097 
2098 	if (topdown_run) {
2099 		unsigned int max_level = metricgroups__topdown_max_level();
2100 		char str[] = "TopdownL1";
2101 
2102 		if (!force_metric_only)
2103 			stat_config.metric_only = true;
2104 
2105 		if (!max_level) {
2106 			pr_err("Topdown requested but the topdown metric groups aren't present.\n"
2107 				"(See perf list the metric groups have names like TopdownL1)");
2108 			return -1;
2109 		}
2110 		if (stat_config.topdown_level > max_level) {
2111 			pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
2112 			return -1;
2113 		} else if (!stat_config.topdown_level)
2114 			stat_config.topdown_level = 1;
2115 
2116 		if (!stat_config.interval && !stat_config.metric_only) {
2117 			fprintf(stat_config.output,
2118 				"Topdown accuracy may decrease when measuring long periods.\n"
2119 				"Please print the result regularly, e.g. -I1000\n");
2120 		}
2121 		str[8] = stat_config.topdown_level + '0';
2122 		if (metricgroup__parse_groups(evsel_list,
2123 						pmu, str,
2124 						/*metric_no_group=*/false,
2125 						/*metric_no_merge=*/false,
2126 						/*metric_no_threshold=*/true,
2127 						stat_config.user_requested_cpu_list,
2128 						stat_config.system_wide,
2129 						&stat_config.metric_events) < 0)
2130 			return -1;
2131 	}
2132 
2133 	if (!stat_config.topdown_level)
2134 		stat_config.topdown_level = 1;
2135 
2136 	if (!evsel_list->core.nr_entries) {
2137 		/* No events so add defaults. */
2138 		if (target__has_cpu(&target))
2139 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2140 
2141 		if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2142 			return -1;
2143 		if (perf_pmus__have_event("cpu", "stalled-cycles-frontend")) {
2144 			if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
2145 				return -1;
2146 		}
2147 		if (perf_pmus__have_event("cpu", "stalled-cycles-backend")) {
2148 			if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
2149 				return -1;
2150 		}
2151 		if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2152 			return -1;
2153 		/*
2154 		 * Add TopdownL1 metrics if they exist. To minimize
2155 		 * multiplexing, don't request threshold computation.
2156 		 */
2157 		if (metricgroup__has_metric(pmu, "TopdownL1")) {
2158 			struct evlist *metric_evlist = evlist__new();
2159 			struct evsel *metric_evsel;
2160 
2161 			if (!metric_evlist)
2162 				return -1;
2163 
2164 			if (metricgroup__parse_groups(metric_evlist, pmu, "TopdownL1",
2165 							/*metric_no_group=*/false,
2166 							/*metric_no_merge=*/false,
2167 							/*metric_no_threshold=*/true,
2168 							stat_config.user_requested_cpu_list,
2169 							stat_config.system_wide,
2170 							&stat_config.metric_events) < 0)
2171 				return -1;
2172 
2173 			evlist__for_each_entry(metric_evlist, metric_evsel) {
2174 				metric_evsel->skippable = true;
2175 			}
2176 			evlist__splice_list_tail(evsel_list, &metric_evlist->core.entries);
2177 			evlist__delete(metric_evlist);
2178 		}
2179 
2180 		/* Platform specific attrs */
2181 		if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0)
2182 			return -1;
2183 	}
2184 
2185 	/* Detailed events get appended to the event list: */
2186 
2187 	if (detailed_run <  1)
2188 		return 0;
2189 
2190 	/* Append detailed run extra attributes: */
2191 	if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2192 		return -1;
2193 
2194 	if (detailed_run < 2)
2195 		return 0;
2196 
2197 	/* Append very detailed run extra attributes: */
2198 	if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2199 		return -1;
2200 
2201 	if (detailed_run < 3)
2202 		return 0;
2203 
2204 	/* Append very, very detailed run extra attributes: */
2205 	return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2206 }
2207 
2208 static const char * const stat_record_usage[] = {
2209 	"perf stat record [<options>]",
2210 	NULL,
2211 };
2212 
2213 static void init_features(struct perf_session *session)
2214 {
2215 	int feat;
2216 
2217 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2218 		perf_header__set_feat(&session->header, feat);
2219 
2220 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
2221 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2222 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2223 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2224 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2225 }
2226 
2227 static int __cmd_record(int argc, const char **argv)
2228 {
2229 	struct perf_session *session;
2230 	struct perf_data *data = &perf_stat.data;
2231 
2232 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
2233 			     PARSE_OPT_STOP_AT_NON_OPTION);
2234 
2235 	if (output_name)
2236 		data->path = output_name;
2237 
2238 	if (stat_config.run_count != 1 || forever) {
2239 		pr_err("Cannot use -r option with perf stat record.\n");
2240 		return -1;
2241 	}
2242 
2243 	session = perf_session__new(data, NULL);
2244 	if (IS_ERR(session)) {
2245 		pr_err("Perf session creation failed\n");
2246 		return PTR_ERR(session);
2247 	}
2248 
2249 	init_features(session);
2250 
2251 	session->evlist   = evsel_list;
2252 	perf_stat.session = session;
2253 	perf_stat.record  = true;
2254 	return argc;
2255 }
2256 
2257 static int process_stat_round_event(struct perf_session *session,
2258 				    union perf_event *event)
2259 {
2260 	struct perf_record_stat_round *stat_round = &event->stat_round;
2261 	struct timespec tsh, *ts = NULL;
2262 	const char **argv = session->header.env.cmdline_argv;
2263 	int argc = session->header.env.nr_cmdline;
2264 
2265 	process_counters();
2266 
2267 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2268 		update_stats(&walltime_nsecs_stats, stat_round->time);
2269 
2270 	if (stat_config.interval && stat_round->time) {
2271 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2272 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2273 		ts = &tsh;
2274 	}
2275 
2276 	print_counters(ts, argc, argv);
2277 	return 0;
2278 }
2279 
2280 static
2281 int process_stat_config_event(struct perf_session *session,
2282 			      union perf_event *event)
2283 {
2284 	struct perf_tool *tool = session->tool;
2285 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2286 
2287 	perf_event__read_stat_config(&stat_config, &event->stat_config);
2288 
2289 	if (perf_cpu_map__empty(st->cpus)) {
2290 		if (st->aggr_mode != AGGR_UNSET)
2291 			pr_warning("warning: processing task data, aggregation mode not set\n");
2292 	} else if (st->aggr_mode != AGGR_UNSET) {
2293 		stat_config.aggr_mode = st->aggr_mode;
2294 	}
2295 
2296 	if (perf_stat.data.is_pipe)
2297 		perf_stat_init_aggr_mode();
2298 	else
2299 		perf_stat_init_aggr_mode_file(st);
2300 
2301 	if (stat_config.aggr_map) {
2302 		int nr_aggr = stat_config.aggr_map->nr;
2303 
2304 		if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) {
2305 			pr_err("cannot allocate aggr counts\n");
2306 			return -1;
2307 		}
2308 	}
2309 	return 0;
2310 }
2311 
2312 static int set_maps(struct perf_stat *st)
2313 {
2314 	if (!st->cpus || !st->threads)
2315 		return 0;
2316 
2317 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2318 		return -EINVAL;
2319 
2320 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2321 
2322 	if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true))
2323 		return -ENOMEM;
2324 
2325 	st->maps_allocated = true;
2326 	return 0;
2327 }
2328 
2329 static
2330 int process_thread_map_event(struct perf_session *session,
2331 			     union perf_event *event)
2332 {
2333 	struct perf_tool *tool = session->tool;
2334 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2335 
2336 	if (st->threads) {
2337 		pr_warning("Extra thread map event, ignoring.\n");
2338 		return 0;
2339 	}
2340 
2341 	st->threads = thread_map__new_event(&event->thread_map);
2342 	if (!st->threads)
2343 		return -ENOMEM;
2344 
2345 	return set_maps(st);
2346 }
2347 
2348 static
2349 int process_cpu_map_event(struct perf_session *session,
2350 			  union perf_event *event)
2351 {
2352 	struct perf_tool *tool = session->tool;
2353 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2354 	struct perf_cpu_map *cpus;
2355 
2356 	if (st->cpus) {
2357 		pr_warning("Extra cpu map event, ignoring.\n");
2358 		return 0;
2359 	}
2360 
2361 	cpus = cpu_map__new_data(&event->cpu_map.data);
2362 	if (!cpus)
2363 		return -ENOMEM;
2364 
2365 	st->cpus = cpus;
2366 	return set_maps(st);
2367 }
2368 
2369 static const char * const stat_report_usage[] = {
2370 	"perf stat report [<options>]",
2371 	NULL,
2372 };
2373 
2374 static struct perf_stat perf_stat = {
2375 	.tool = {
2376 		.attr		= perf_event__process_attr,
2377 		.event_update	= perf_event__process_event_update,
2378 		.thread_map	= process_thread_map_event,
2379 		.cpu_map	= process_cpu_map_event,
2380 		.stat_config	= process_stat_config_event,
2381 		.stat		= perf_event__process_stat_event,
2382 		.stat_round	= process_stat_round_event,
2383 	},
2384 	.aggr_mode	= AGGR_UNSET,
2385 	.aggr_level	= 0,
2386 };
2387 
2388 static int __cmd_report(int argc, const char **argv)
2389 {
2390 	struct perf_session *session;
2391 	const struct option options[] = {
2392 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2393 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2394 		     "aggregate counts per processor socket", AGGR_SOCKET),
2395 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2396 		     "aggregate counts per processor die", AGGR_DIE),
2397 	OPT_CALLBACK_OPTARG(0, "per-cache", &perf_stat.aggr_mode, &perf_stat.aggr_level,
2398 			    "cache level",
2399 			    "aggregate count at this cache level (Default: LLC)",
2400 			    parse_cache_level),
2401 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2402 		     "aggregate counts per physical processor core", AGGR_CORE),
2403 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2404 		     "aggregate counts per numa node", AGGR_NODE),
2405 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2406 		     "disable CPU count aggregation", AGGR_NONE),
2407 	OPT_END()
2408 	};
2409 	struct stat st;
2410 	int ret;
2411 
2412 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2413 
2414 	if (!input_name || !strlen(input_name)) {
2415 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2416 			input_name = "-";
2417 		else
2418 			input_name = "perf.data";
2419 	}
2420 
2421 	perf_stat.data.path = input_name;
2422 	perf_stat.data.mode = PERF_DATA_MODE_READ;
2423 
2424 	session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2425 	if (IS_ERR(session))
2426 		return PTR_ERR(session);
2427 
2428 	perf_stat.session  = session;
2429 	stat_config.output = stderr;
2430 	evsel_list         = session->evlist;
2431 
2432 	ret = perf_session__process_events(session);
2433 	if (ret)
2434 		return ret;
2435 
2436 	perf_session__delete(session);
2437 	return 0;
2438 }
2439 
2440 static void setup_system_wide(int forks)
2441 {
2442 	/*
2443 	 * Make system wide (-a) the default target if
2444 	 * no target was specified and one of following
2445 	 * conditions is met:
2446 	 *
2447 	 *   - there's no workload specified
2448 	 *   - there is workload specified but all requested
2449 	 *     events are system wide events
2450 	 */
2451 	if (!target__none(&target))
2452 		return;
2453 
2454 	if (!forks)
2455 		target.system_wide = true;
2456 	else {
2457 		struct evsel *counter;
2458 
2459 		evlist__for_each_entry(evsel_list, counter) {
2460 			if (!counter->core.requires_cpu &&
2461 			    !evsel__name_is(counter, "duration_time")) {
2462 				return;
2463 			}
2464 		}
2465 
2466 		if (evsel_list->core.nr_entries)
2467 			target.system_wide = true;
2468 	}
2469 }
2470 
2471 int cmd_stat(int argc, const char **argv)
2472 {
2473 	const char * const stat_usage[] = {
2474 		"perf stat [<options>] [<command>]",
2475 		NULL
2476 	};
2477 	int status = -EINVAL, run_idx, err;
2478 	const char *mode;
2479 	FILE *output = stderr;
2480 	unsigned int interval, timeout;
2481 	const char * const stat_subcommands[] = { "record", "report" };
2482 	char errbuf[BUFSIZ];
2483 
2484 	setlocale(LC_ALL, "");
2485 
2486 	evsel_list = evlist__new();
2487 	if (evsel_list == NULL)
2488 		return -ENOMEM;
2489 
2490 	parse_events__shrink_config_terms();
2491 
2492 	/* String-parsing callback-based options would segfault when negated */
2493 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2494 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2495 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2496 
2497 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2498 					(const char **) stat_usage,
2499 					PARSE_OPT_STOP_AT_NON_OPTION);
2500 
2501 	if (stat_config.csv_sep) {
2502 		stat_config.csv_output = true;
2503 		if (!strcmp(stat_config.csv_sep, "\\t"))
2504 			stat_config.csv_sep = "\t";
2505 	} else
2506 		stat_config.csv_sep = DEFAULT_SEPARATOR;
2507 
2508 	if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2509 		argc = __cmd_record(argc, argv);
2510 		if (argc < 0)
2511 			return -1;
2512 	} else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2513 		return __cmd_report(argc, argv);
2514 
2515 	interval = stat_config.interval;
2516 	timeout = stat_config.timeout;
2517 
2518 	/*
2519 	 * For record command the -o is already taken care of.
2520 	 */
2521 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2522 		output = NULL;
2523 
2524 	if (output_name && output_fd) {
2525 		fprintf(stderr, "cannot use both --output and --log-fd\n");
2526 		parse_options_usage(stat_usage, stat_options, "o", 1);
2527 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2528 		goto out;
2529 	}
2530 
2531 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2532 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2533 		goto out;
2534 	}
2535 
2536 	if (stat_config.metric_only && stat_config.run_count > 1) {
2537 		fprintf(stderr, "--metric-only is not supported with -r\n");
2538 		goto out;
2539 	}
2540 
2541 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2542 		fprintf(stderr, "--table is only supported with -r\n");
2543 		parse_options_usage(stat_usage, stat_options, "r", 1);
2544 		parse_options_usage(NULL, stat_options, "table", 0);
2545 		goto out;
2546 	}
2547 
2548 	if (output_fd < 0) {
2549 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2550 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2551 		goto out;
2552 	}
2553 
2554 	if (!output && !quiet) {
2555 		struct timespec tm;
2556 		mode = append_file ? "a" : "w";
2557 
2558 		output = fopen(output_name, mode);
2559 		if (!output) {
2560 			perror("failed to create output file");
2561 			return -1;
2562 		}
2563 		if (!stat_config.json_output) {
2564 			clock_gettime(CLOCK_REALTIME, &tm);
2565 			fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2566 		}
2567 	} else if (output_fd > 0) {
2568 		mode = append_file ? "a" : "w";
2569 		output = fdopen(output_fd, mode);
2570 		if (!output) {
2571 			perror("Failed opening logfd");
2572 			return -errno;
2573 		}
2574 	}
2575 
2576 	if (stat_config.interval_clear && !isatty(fileno(output))) {
2577 		fprintf(stderr, "--interval-clear does not work with output\n");
2578 		parse_options_usage(stat_usage, stat_options, "o", 1);
2579 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2580 		parse_options_usage(NULL, stat_options, "interval-clear", 0);
2581 		return -1;
2582 	}
2583 
2584 	stat_config.output = output;
2585 
2586 	/*
2587 	 * let the spreadsheet do the pretty-printing
2588 	 */
2589 	if (stat_config.csv_output) {
2590 		/* User explicitly passed -B? */
2591 		if (big_num_opt == 1) {
2592 			fprintf(stderr, "-B option not supported with -x\n");
2593 			parse_options_usage(stat_usage, stat_options, "B", 1);
2594 			parse_options_usage(NULL, stat_options, "x", 1);
2595 			goto out;
2596 		} else /* Nope, so disable big number formatting */
2597 			stat_config.big_num = false;
2598 	} else if (big_num_opt == 0) /* User passed --no-big-num */
2599 		stat_config.big_num = false;
2600 
2601 	err = target__validate(&target);
2602 	if (err) {
2603 		target__strerror(&target, err, errbuf, BUFSIZ);
2604 		pr_warning("%s\n", errbuf);
2605 	}
2606 
2607 	setup_system_wide(argc);
2608 
2609 	/*
2610 	 * Display user/system times only for single
2611 	 * run and when there's specified tracee.
2612 	 */
2613 	if ((stat_config.run_count == 1) && target__none(&target))
2614 		stat_config.ru_display = true;
2615 
2616 	if (stat_config.run_count < 0) {
2617 		pr_err("Run count must be a positive number\n");
2618 		parse_options_usage(stat_usage, stat_options, "r", 1);
2619 		goto out;
2620 	} else if (stat_config.run_count == 0) {
2621 		forever = true;
2622 		stat_config.run_count = 1;
2623 	}
2624 
2625 	if (stat_config.walltime_run_table) {
2626 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2627 		if (!stat_config.walltime_run) {
2628 			pr_err("failed to setup -r option");
2629 			goto out;
2630 		}
2631 	}
2632 
2633 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2634 		!target__has_task(&target)) {
2635 		if (!target.system_wide || target.cpu_list) {
2636 			fprintf(stderr, "The --per-thread option is only "
2637 				"available when monitoring via -p -t -a "
2638 				"options or only --per-thread.\n");
2639 			parse_options_usage(NULL, stat_options, "p", 1);
2640 			parse_options_usage(NULL, stat_options, "t", 1);
2641 			goto out;
2642 		}
2643 	}
2644 
2645 	/*
2646 	 * no_aggr, cgroup are for system-wide only
2647 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2648 	 */
2649 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2650 	      stat_config.aggr_mode != AGGR_THREAD) ||
2651 	     (nr_cgroups || stat_config.cgroup_list)) &&
2652 	    !target__has_cpu(&target)) {
2653 		fprintf(stderr, "both cgroup and no-aggregation "
2654 			"modes only available in system-wide mode\n");
2655 
2656 		parse_options_usage(stat_usage, stat_options, "G", 1);
2657 		parse_options_usage(NULL, stat_options, "A", 1);
2658 		parse_options_usage(NULL, stat_options, "a", 1);
2659 		parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2660 		goto out;
2661 	}
2662 
2663 	if (stat_config.iostat_run) {
2664 		status = iostat_prepare(evsel_list, &stat_config);
2665 		if (status)
2666 			goto out;
2667 		if (iostat_mode == IOSTAT_LIST) {
2668 			iostat_list(evsel_list, &stat_config);
2669 			goto out;
2670 		} else if (verbose > 0)
2671 			iostat_list(evsel_list, &stat_config);
2672 		if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2673 			target.system_wide = true;
2674 	}
2675 
2676 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2677 		target.per_thread = true;
2678 
2679 	stat_config.system_wide = target.system_wide;
2680 	if (target.cpu_list) {
2681 		stat_config.user_requested_cpu_list = strdup(target.cpu_list);
2682 		if (!stat_config.user_requested_cpu_list) {
2683 			status = -ENOMEM;
2684 			goto out;
2685 		}
2686 	}
2687 
2688 	/*
2689 	 * Metric parsing needs to be delayed as metrics may optimize events
2690 	 * knowing the target is system-wide.
2691 	 */
2692 	if (metrics) {
2693 		const char *pmu = parse_events_option_args.pmu_filter ?: "all";
2694 
2695 		metricgroup__parse_groups(evsel_list, pmu, metrics,
2696 					stat_config.metric_no_group,
2697 					stat_config.metric_no_merge,
2698 					stat_config.metric_no_threshold,
2699 					stat_config.user_requested_cpu_list,
2700 					stat_config.system_wide,
2701 					&stat_config.metric_events);
2702 		zfree(&metrics);
2703 	}
2704 
2705 	if (add_default_attributes())
2706 		goto out;
2707 
2708 	if (stat_config.cgroup_list) {
2709 		if (nr_cgroups > 0) {
2710 			pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2711 			parse_options_usage(stat_usage, stat_options, "G", 1);
2712 			parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2713 			goto out;
2714 		}
2715 
2716 		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2717 					  &stat_config.metric_events, true) < 0) {
2718 			parse_options_usage(stat_usage, stat_options,
2719 					    "for-each-cgroup", 0);
2720 			goto out;
2721 		}
2722 	}
2723 
2724 	evlist__warn_user_requested_cpus(evsel_list, target.cpu_list);
2725 
2726 	if (evlist__create_maps(evsel_list, &target) < 0) {
2727 		if (target__has_task(&target)) {
2728 			pr_err("Problems finding threads of monitor\n");
2729 			parse_options_usage(stat_usage, stat_options, "p", 1);
2730 			parse_options_usage(NULL, stat_options, "t", 1);
2731 		} else if (target__has_cpu(&target)) {
2732 			perror("failed to parse CPUs map");
2733 			parse_options_usage(stat_usage, stat_options, "C", 1);
2734 			parse_options_usage(NULL, stat_options, "a", 1);
2735 		}
2736 		goto out;
2737 	}
2738 
2739 	evlist__check_cpu_maps(evsel_list);
2740 
2741 	/*
2742 	 * Initialize thread_map with comm names,
2743 	 * so we could print it out on output.
2744 	 */
2745 	if (stat_config.aggr_mode == AGGR_THREAD) {
2746 		thread_map__read_comms(evsel_list->core.threads);
2747 	}
2748 
2749 	if (stat_config.aggr_mode == AGGR_NODE)
2750 		cpu__setup_cpunode_map();
2751 
2752 	if (stat_config.times && interval)
2753 		interval_count = true;
2754 	else if (stat_config.times && !interval) {
2755 		pr_err("interval-count option should be used together with "
2756 				"interval-print.\n");
2757 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2758 		parse_options_usage(stat_usage, stat_options, "I", 1);
2759 		goto out;
2760 	}
2761 
2762 	if (timeout && timeout < 100) {
2763 		if (timeout < 10) {
2764 			pr_err("timeout must be >= 10ms.\n");
2765 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2766 			goto out;
2767 		} else
2768 			pr_warning("timeout < 100ms. "
2769 				   "The overhead percentage could be high in some cases. "
2770 				   "Please proceed with caution.\n");
2771 	}
2772 	if (timeout && interval) {
2773 		pr_err("timeout option is not supported with interval-print.\n");
2774 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2775 		parse_options_usage(stat_usage, stat_options, "I", 1);
2776 		goto out;
2777 	}
2778 
2779 	if (perf_stat_init_aggr_mode())
2780 		goto out;
2781 
2782 	if (evlist__alloc_stats(&stat_config, evsel_list, interval))
2783 		goto out;
2784 
2785 	/*
2786 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2787 	 * while avoiding that older tools show confusing messages.
2788 	 *
2789 	 * However for pipe sessions we need to keep it zero,
2790 	 * because script's perf_evsel__check_attr is triggered
2791 	 * by attr->sample_type != 0, and we can't run it on
2792 	 * stat sessions.
2793 	 */
2794 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2795 
2796 	/*
2797 	 * We dont want to block the signals - that would cause
2798 	 * child tasks to inherit that and Ctrl-C would not work.
2799 	 * What we want is for Ctrl-C to work in the exec()-ed
2800 	 * task, but being ignored by perf stat itself:
2801 	 */
2802 	atexit(sig_atexit);
2803 	if (!forever)
2804 		signal(SIGINT,  skip_signal);
2805 	signal(SIGCHLD, skip_signal);
2806 	signal(SIGALRM, skip_signal);
2807 	signal(SIGABRT, skip_signal);
2808 
2809 	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2810 		goto out;
2811 
2812 	/* Enable ignoring missing threads when -p option is defined. */
2813 	evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2814 	status = 0;
2815 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2816 		if (stat_config.run_count != 1 && verbose > 0)
2817 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2818 				run_idx + 1);
2819 
2820 		if (run_idx != 0)
2821 			evlist__reset_prev_raw_counts(evsel_list);
2822 
2823 		status = run_perf_stat(argc, argv, run_idx);
2824 		if (forever && status != -1 && !interval) {
2825 			print_counters(NULL, argc, argv);
2826 			perf_stat__reset_stats();
2827 		}
2828 	}
2829 
2830 	if (!forever && status != -1 && (!interval || stat_config.summary))
2831 		print_counters(NULL, argc, argv);
2832 
2833 	evlist__finalize_ctlfd(evsel_list);
2834 
2835 	if (STAT_RECORD) {
2836 		/*
2837 		 * We synthesize the kernel mmap record just so that older tools
2838 		 * don't emit warnings about not being able to resolve symbols
2839 		 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2840 		 * a saner message about no samples being in the perf.data file.
2841 		 *
2842 		 * This also serves to suppress a warning about f_header.data.size == 0
2843 		 * in header.c at the moment 'perf stat record' gets introduced, which
2844 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2845 		 * records, but the need to suppress the kptr_restrict messages in older
2846 		 * tools remain  -acme
2847 		 */
2848 		int fd = perf_data__fd(&perf_stat.data);
2849 
2850 		err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2851 							 process_synthesized_event,
2852 							 &perf_stat.session->machines.host);
2853 		if (err) {
2854 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2855 				   "older tools may produce warnings about this file\n.");
2856 		}
2857 
2858 		if (!interval) {
2859 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2860 				pr_err("failed to write stat round event\n");
2861 		}
2862 
2863 		if (!perf_stat.data.is_pipe) {
2864 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2865 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2866 		}
2867 
2868 		evlist__close(evsel_list);
2869 		perf_session__delete(perf_stat.session);
2870 	}
2871 
2872 	perf_stat__exit_aggr_mode();
2873 	evlist__free_stats(evsel_list);
2874 out:
2875 	if (stat_config.iostat_run)
2876 		iostat_release(evsel_list);
2877 
2878 	zfree(&stat_config.walltime_run);
2879 	zfree(&stat_config.user_requested_cpu_list);
2880 
2881 	if (smi_cost && smi_reset)
2882 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2883 
2884 	evlist__delete(evsel_list);
2885 
2886 	metricgroup__rblist_exit(&stat_config.metric_events);
2887 	evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2888 
2889 	return status;
2890 }
2891