xref: /linux/tools/perf/builtin-stat.c (revision bdfa82f5b8998a6311a8ef0cf89ad413f5cd9ea4)
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/tool_pmu.h"
50 #include "util/event.h"
51 #include "util/evlist.h"
52 #include "util/evsel.h"
53 #include "util/debug.h"
54 #include "util/color.h"
55 #include "util/stat.h"
56 #include "util/header.h"
57 #include "util/cpumap.h"
58 #include "util/thread_map.h"
59 #include "util/counts.h"
60 #include "util/topdown.h"
61 #include "util/session.h"
62 #include "util/tool.h"
63 #include "util/string2.h"
64 #include "util/metricgroup.h"
65 #include "util/synthetic-events.h"
66 #include "util/target.h"
67 #include "util/time-utils.h"
68 #include "util/top.h"
69 #include "util/affinity.h"
70 #include "util/pfm.h"
71 #include "util/bpf_counter.h"
72 #include "util/iostat.h"
73 #include "util/util.h"
74 #include "util/intel-tpebs.h"
75 #include "asm/bug.h"
76 
77 #include <linux/time64.h>
78 #include <linux/zalloc.h>
79 #include <api/fs/fs.h>
80 #include <errno.h>
81 #include <signal.h>
82 #include <stdlib.h>
83 #include <sys/prctl.h>
84 #include <inttypes.h>
85 #include <locale.h>
86 #include <math.h>
87 #include <sys/types.h>
88 #include <sys/stat.h>
89 #include <sys/wait.h>
90 #include <unistd.h>
91 #include <sys/time.h>
92 #include <sys/resource.h>
93 #include <linux/err.h>
94 
95 #include <linux/ctype.h>
96 #include <perf/evlist.h>
97 #include <internal/threadmap.h>
98 
99 #define DEFAULT_SEPARATOR	" "
100 #define FREEZE_ON_SMI_PATH	"bus/event_source/devices/cpu/freeze_on_smi"
101 
102 static void print_counters(struct timespec *ts, int argc, const char **argv);
103 
104 static struct evlist	*evsel_list;
105 static struct parse_events_option_args parse_events_option_args = {
106 	.evlistp = &evsel_list,
107 };
108 
109 static bool all_counters_use_bpf = true;
110 
111 static struct target target = {
112 	.uid	= UINT_MAX,
113 };
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 /* Options set from the command line. */
153 struct opt_aggr_mode {
154 	bool node, socket, die, cluster, cache, core, thread, no_aggr;
155 };
156 
157 /* Turn command line option into most generic aggregation mode setting. */
158 static enum aggr_mode opt_aggr_mode_to_aggr_mode(struct opt_aggr_mode *opt_mode)
159 {
160 	enum aggr_mode mode = AGGR_GLOBAL;
161 
162 	if (opt_mode->node)
163 		mode = AGGR_NODE;
164 	if (opt_mode->socket)
165 		mode = AGGR_SOCKET;
166 	if (opt_mode->die)
167 		mode = AGGR_DIE;
168 	if (opt_mode->cluster)
169 		mode = AGGR_CLUSTER;
170 	if (opt_mode->cache)
171 		mode = AGGR_CACHE;
172 	if (opt_mode->core)
173 		mode = AGGR_CORE;
174 	if (opt_mode->thread)
175 		mode = AGGR_THREAD;
176 	if (opt_mode->no_aggr)
177 		mode = AGGR_NONE;
178 	return mode;
179 }
180 
181 static void evlist__check_cpu_maps(struct evlist *evlist)
182 {
183 	struct evsel *evsel, *warned_leader = NULL;
184 
185 	evlist__for_each_entry(evlist, evsel) {
186 		struct evsel *leader = evsel__leader(evsel);
187 
188 		/* Check that leader matches cpus with each member. */
189 		if (leader == evsel)
190 			continue;
191 		if (perf_cpu_map__equal(leader->core.cpus, evsel->core.cpus))
192 			continue;
193 
194 		/* If there's mismatch disable the group and warn user. */
195 		if (warned_leader != leader) {
196 			char buf[200];
197 
198 			pr_warning("WARNING: grouped events cpus do not match.\n"
199 				"Events with CPUs not matching the leader will "
200 				"be removed from the group.\n");
201 			evsel__group_desc(leader, buf, sizeof(buf));
202 			pr_warning("  %s\n", buf);
203 			warned_leader = leader;
204 		}
205 		if (verbose > 0) {
206 			char buf[200];
207 
208 			cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
209 			pr_warning("     %s: %s\n", leader->name, buf);
210 			cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
211 			pr_warning("     %s: %s\n", evsel->name, buf);
212 		}
213 
214 		evsel__remove_from_group(evsel, leader);
215 	}
216 }
217 
218 static inline void diff_timespec(struct timespec *r, struct timespec *a,
219 				 struct timespec *b)
220 {
221 	r->tv_sec = a->tv_sec - b->tv_sec;
222 	if (a->tv_nsec < b->tv_nsec) {
223 		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
224 		r->tv_sec--;
225 	} else {
226 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
227 	}
228 }
229 
230 static void perf_stat__reset_stats(void)
231 {
232 	evlist__reset_stats(evsel_list);
233 	perf_stat__reset_shadow_stats();
234 }
235 
236 static int process_synthesized_event(const struct perf_tool *tool __maybe_unused,
237 				     union perf_event *event,
238 				     struct perf_sample *sample __maybe_unused,
239 				     struct machine *machine __maybe_unused)
240 {
241 	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
242 		pr_err("failed to write perf data, error: %m\n");
243 		return -1;
244 	}
245 
246 	perf_stat.bytes_written += event->header.size;
247 	return 0;
248 }
249 
250 static int write_stat_round_event(u64 tm, u64 type)
251 {
252 	return perf_event__synthesize_stat_round(NULL, tm, type,
253 						 process_synthesized_event,
254 						 NULL);
255 }
256 
257 #define WRITE_STAT_ROUND_EVENT(time, interval) \
258 	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
259 
260 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
261 
262 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
263 				   struct perf_counts_values *count)
264 {
265 	struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
266 	struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
267 
268 	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
269 					   process_synthesized_event, NULL);
270 }
271 
272 static int read_single_counter(struct evsel *counter, int cpu_map_idx, int thread)
273 {
274 	int err = evsel__read_counter(counter, cpu_map_idx, thread);
275 
276 	/*
277 	 * Reading user and system time will fail when the process
278 	 * terminates. Use the wait4 values in that case.
279 	 */
280 	if (err && cpu_map_idx == 0 &&
281 	    (evsel__tool_event(counter) == TOOL_PMU__EVENT_USER_TIME ||
282 	     evsel__tool_event(counter) == TOOL_PMU__EVENT_SYSTEM_TIME)) {
283 		u64 val, *start_time;
284 		struct perf_counts_values *count =
285 			perf_counts(counter->counts, cpu_map_idx, thread);
286 
287 		start_time = xyarray__entry(counter->start_times, cpu_map_idx, thread);
288 		if (evsel__tool_event(counter) == TOOL_PMU__EVENT_USER_TIME)
289 			val = ru_stats.ru_utime_usec_stat.mean;
290 		else
291 			val = ru_stats.ru_stime_usec_stat.mean;
292 		count->ena = count->run = *start_time + val;
293 		count->val = val;
294 		return 0;
295 	}
296 	return err;
297 }
298 
299 /*
300  * Read out the results of a single counter:
301  * do not aggregate counts across CPUs in system-wide mode
302  */
303 static int read_counter_cpu(struct evsel *counter, int cpu_map_idx)
304 {
305 	int nthreads = perf_thread_map__nr(evsel_list->core.threads);
306 	int thread;
307 
308 	if (!counter->supported)
309 		return -ENOENT;
310 
311 	for (thread = 0; thread < nthreads; thread++) {
312 		struct perf_counts_values *count;
313 
314 		count = perf_counts(counter->counts, cpu_map_idx, thread);
315 
316 		/*
317 		 * The leader's group read loads data into its group members
318 		 * (via evsel__read_counter()) and sets their count->loaded.
319 		 */
320 		if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
321 		    read_single_counter(counter, cpu_map_idx, thread)) {
322 			counter->counts->scaled = -1;
323 			perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
324 			perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
325 			return -1;
326 		}
327 
328 		perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
329 
330 		if (STAT_RECORD) {
331 			if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
332 				pr_err("failed to write stat event\n");
333 				return -1;
334 			}
335 		}
336 
337 		if (verbose > 1) {
338 			fprintf(stat_config.output,
339 				"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
340 					evsel__name(counter),
341 					perf_cpu_map__cpu(evsel__cpus(counter),
342 							  cpu_map_idx).cpu,
343 					count->val, count->ena, count->run);
344 		}
345 	}
346 
347 	return 0;
348 }
349 
350 static int read_affinity_counters(void)
351 {
352 	struct evlist_cpu_iterator evlist_cpu_itr;
353 	struct affinity saved_affinity, *affinity;
354 
355 	if (all_counters_use_bpf)
356 		return 0;
357 
358 	if (!target__has_cpu(&target) || target__has_per_thread(&target))
359 		affinity = NULL;
360 	else if (affinity__setup(&saved_affinity) < 0)
361 		return -1;
362 	else
363 		affinity = &saved_affinity;
364 
365 	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
366 		struct evsel *counter = evlist_cpu_itr.evsel;
367 
368 		if (evsel__is_bpf(counter))
369 			continue;
370 
371 		if (!counter->err)
372 			counter->err = read_counter_cpu(counter, evlist_cpu_itr.cpu_map_idx);
373 	}
374 	if (affinity)
375 		affinity__cleanup(&saved_affinity);
376 
377 	return 0;
378 }
379 
380 static int read_bpf_map_counters(void)
381 {
382 	struct evsel *counter;
383 	int err;
384 
385 	evlist__for_each_entry(evsel_list, counter) {
386 		if (!evsel__is_bpf(counter))
387 			continue;
388 
389 		err = bpf_counter__read(counter);
390 		if (err)
391 			return err;
392 	}
393 	return 0;
394 }
395 
396 static int read_counters(void)
397 {
398 	if (!stat_config.stop_read_counter) {
399 		if (read_bpf_map_counters() ||
400 		    read_affinity_counters())
401 			return -1;
402 	}
403 	return 0;
404 }
405 
406 static void process_counters(void)
407 {
408 	struct evsel *counter;
409 
410 	evlist__for_each_entry(evsel_list, counter) {
411 		if (counter->err)
412 			pr_debug("failed to read counter %s\n", counter->name);
413 		if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
414 			pr_warning("failed to process counter %s\n", counter->name);
415 		counter->err = 0;
416 	}
417 
418 	perf_stat_merge_counters(&stat_config, evsel_list);
419 	perf_stat_process_percore(&stat_config, evsel_list);
420 }
421 
422 static void process_interval(void)
423 {
424 	struct timespec ts, rs;
425 
426 	clock_gettime(CLOCK_MONOTONIC, &ts);
427 	diff_timespec(&rs, &ts, &ref_time);
428 
429 	evlist__reset_aggr_stats(evsel_list);
430 
431 	if (read_counters() == 0)
432 		process_counters();
433 
434 	if (STAT_RECORD) {
435 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
436 			pr_err("failed to write stat round event\n");
437 	}
438 
439 	init_stats(&walltime_nsecs_stats);
440 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
441 	print_counters(&rs, 0, NULL);
442 }
443 
444 static bool handle_interval(unsigned int interval, int *times)
445 {
446 	if (interval) {
447 		process_interval();
448 		if (interval_count && !(--(*times)))
449 			return true;
450 	}
451 	return false;
452 }
453 
454 static int enable_counters(void)
455 {
456 	struct evsel *evsel;
457 	int err;
458 
459 	evlist__for_each_entry(evsel_list, evsel) {
460 		if (!evsel__is_bpf(evsel))
461 			continue;
462 
463 		err = bpf_counter__enable(evsel);
464 		if (err)
465 			return err;
466 	}
467 
468 	if (!target__enable_on_exec(&target)) {
469 		if (!all_counters_use_bpf)
470 			evlist__enable(evsel_list);
471 	}
472 	return 0;
473 }
474 
475 static void disable_counters(void)
476 {
477 	struct evsel *counter;
478 
479 	/*
480 	 * If we don't have tracee (attaching to task or cpu), counters may
481 	 * still be running. To get accurate group ratios, we must stop groups
482 	 * from counting before reading their constituent counters.
483 	 */
484 	if (!target__none(&target)) {
485 		evlist__for_each_entry(evsel_list, counter)
486 			bpf_counter__disable(counter);
487 		if (!all_counters_use_bpf)
488 			evlist__disable(evsel_list);
489 	}
490 }
491 
492 static volatile sig_atomic_t workload_exec_errno;
493 
494 /*
495  * evlist__prepare_workload will send a SIGUSR1
496  * if the fork fails, since we asked by setting its
497  * want_signal to true.
498  */
499 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
500 					void *ucontext __maybe_unused)
501 {
502 	workload_exec_errno = info->si_value.sival_int;
503 }
504 
505 static bool evsel__should_store_id(struct evsel *counter)
506 {
507 	return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
508 }
509 
510 static bool is_target_alive(struct target *_target,
511 			    struct perf_thread_map *threads)
512 {
513 	struct stat st;
514 	int i;
515 
516 	if (!target__has_task(_target))
517 		return true;
518 
519 	for (i = 0; i < threads->nr; i++) {
520 		char path[PATH_MAX];
521 
522 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
523 			  threads->map[i].pid);
524 
525 		if (!stat(path, &st))
526 			return true;
527 	}
528 
529 	return false;
530 }
531 
532 static void process_evlist(struct evlist *evlist, unsigned int interval)
533 {
534 	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
535 
536 	if (evlist__ctlfd_process(evlist, &cmd) > 0) {
537 		switch (cmd) {
538 		case EVLIST_CTL_CMD_ENABLE:
539 			fallthrough;
540 		case EVLIST_CTL_CMD_DISABLE:
541 			if (interval)
542 				process_interval();
543 			break;
544 		case EVLIST_CTL_CMD_SNAPSHOT:
545 		case EVLIST_CTL_CMD_ACK:
546 		case EVLIST_CTL_CMD_UNSUPPORTED:
547 		case EVLIST_CTL_CMD_EVLIST:
548 		case EVLIST_CTL_CMD_STOP:
549 		case EVLIST_CTL_CMD_PING:
550 		default:
551 			break;
552 		}
553 	}
554 }
555 
556 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
557 			int *time_to_sleep)
558 {
559 	int tts = *time_to_sleep;
560 	struct timespec time_diff;
561 
562 	diff_timespec(&time_diff, time_stop, time_start);
563 
564 	tts -= time_diff.tv_sec * MSEC_PER_SEC +
565 	       time_diff.tv_nsec / NSEC_PER_MSEC;
566 
567 	if (tts < 0)
568 		tts = 0;
569 
570 	*time_to_sleep = tts;
571 }
572 
573 static int dispatch_events(bool forks, int timeout, int interval, int *times)
574 {
575 	int child_exited = 0, status = 0;
576 	int time_to_sleep, sleep_time;
577 	struct timespec time_start, time_stop;
578 
579 	if (interval)
580 		sleep_time = interval;
581 	else if (timeout)
582 		sleep_time = timeout;
583 	else
584 		sleep_time = 1000;
585 
586 	time_to_sleep = sleep_time;
587 
588 	while (!done) {
589 		if (forks)
590 			child_exited = waitpid(child_pid, &status, WNOHANG);
591 		else
592 			child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
593 
594 		if (child_exited)
595 			break;
596 
597 		clock_gettime(CLOCK_MONOTONIC, &time_start);
598 		if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
599 			if (timeout || handle_interval(interval, times))
600 				break;
601 			time_to_sleep = sleep_time;
602 		} else { /* fd revent */
603 			process_evlist(evsel_list, interval);
604 			clock_gettime(CLOCK_MONOTONIC, &time_stop);
605 			compute_tts(&time_start, &time_stop, &time_to_sleep);
606 		}
607 	}
608 
609 	return status;
610 }
611 
612 enum counter_recovery {
613 	COUNTER_SKIP,
614 	COUNTER_RETRY,
615 	COUNTER_FATAL,
616 };
617 
618 static enum counter_recovery stat_handle_error(struct evsel *counter)
619 {
620 	char msg[BUFSIZ];
621 	/*
622 	 * PPC returns ENXIO for HW counters until 2.6.37
623 	 * (behavior changed with commit b0a873e).
624 	 */
625 	if (errno == EINVAL || errno == ENOSYS ||
626 	    errno == ENOENT || errno == ENXIO) {
627 		if (verbose > 0)
628 			ui__warning("%s event is not supported by the kernel.\n",
629 				    evsel__name(counter));
630 		counter->supported = false;
631 		/*
632 		 * errored is a sticky flag that means one of the counter's
633 		 * cpu event had a problem and needs to be reexamined.
634 		 */
635 		counter->errored = true;
636 
637 		if ((evsel__leader(counter) != counter) ||
638 		    !(counter->core.leader->nr_members > 1))
639 			return COUNTER_SKIP;
640 	} else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) {
641 		if (verbose > 0)
642 			ui__warning("%s\n", msg);
643 		return COUNTER_RETRY;
644 	} else if (target__has_per_thread(&target) && errno != EOPNOTSUPP &&
645 		   evsel_list->core.threads &&
646 		   evsel_list->core.threads->err_thread != -1) {
647 		/*
648 		 * For global --per-thread case, skip current
649 		 * error thread.
650 		 */
651 		if (!thread_map__remove(evsel_list->core.threads,
652 					evsel_list->core.threads->err_thread)) {
653 			evsel_list->core.threads->err_thread = -1;
654 			return COUNTER_RETRY;
655 		}
656 	} else if (counter->skippable) {
657 		if (verbose > 0)
658 			ui__warning("skipping event %s that kernel failed to open .\n",
659 				    evsel__name(counter));
660 		counter->supported = false;
661 		counter->errored = true;
662 		return COUNTER_SKIP;
663 	}
664 
665 	if (errno == EOPNOTSUPP) {
666 		if (verbose > 0) {
667 			ui__warning("%s event is not supported by the kernel.\n",
668 				    evsel__name(counter));
669 		}
670 		counter->supported = false;
671 		counter->errored = true;
672 
673 		if ((evsel__leader(counter) != counter) ||
674 		    !(counter->core.leader->nr_members > 1))
675 			return COUNTER_SKIP;
676 	}
677 
678 	evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
679 	ui__error("%s\n", msg);
680 
681 	if (child_pid != -1)
682 		kill(child_pid, SIGTERM);
683 
684 	return COUNTER_FATAL;
685 }
686 
687 static int __run_perf_stat(int argc, const char **argv, int run_idx)
688 {
689 	int interval = stat_config.interval;
690 	int times = stat_config.times;
691 	int timeout = stat_config.timeout;
692 	char msg[BUFSIZ];
693 	unsigned long long t0, t1;
694 	struct evsel *counter;
695 	size_t l;
696 	int status = 0;
697 	const bool forks = (argc > 0);
698 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
699 	struct evlist_cpu_iterator evlist_cpu_itr;
700 	struct affinity saved_affinity, *affinity = NULL;
701 	int err;
702 	bool second_pass = false;
703 
704 	if (forks) {
705 		if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
706 			perror("failed to prepare workload");
707 			return -1;
708 		}
709 		child_pid = evsel_list->workload.pid;
710 	}
711 
712 	if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
713 		if (affinity__setup(&saved_affinity) < 0) {
714 			err = -1;
715 			goto err_out;
716 		}
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 			err = -1;
724 			goto err_out;
725 		}
726 		if (!(evsel__is_bperf(counter)))
727 			all_counters_use_bpf = false;
728 	}
729 
730 	evlist__reset_aggr_stats(evsel_list);
731 
732 	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
733 		counter = evlist_cpu_itr.evsel;
734 
735 		/*
736 		 * bperf calls evsel__open_per_cpu() in bperf__load(), so
737 		 * no need to call it again here.
738 		 */
739 		if (target.use_bpf)
740 			break;
741 
742 		if (counter->reset_group || counter->errored)
743 			continue;
744 		if (evsel__is_bperf(counter))
745 			continue;
746 try_again:
747 		if (create_perf_stat_counter(counter, &stat_config, &target,
748 					     evlist_cpu_itr.cpu_map_idx) < 0) {
749 
750 			/*
751 			 * Weak group failed. We cannot just undo this here
752 			 * because earlier CPUs might be in group mode, and the kernel
753 			 * doesn't support mixing group and non group reads. Defer
754 			 * it to later.
755 			 * Don't close here because we're in the wrong affinity.
756 			 */
757 			if ((errno == EINVAL || errno == EBADF) &&
758 				evsel__leader(counter) != counter &&
759 				counter->weak_group) {
760 				evlist__reset_weak_group(evsel_list, counter, false);
761 				assert(counter->reset_group);
762 				second_pass = true;
763 				continue;
764 			}
765 
766 			switch (stat_handle_error(counter)) {
767 			case COUNTER_FATAL:
768 				err = -1;
769 				goto err_out;
770 			case COUNTER_RETRY:
771 				goto try_again;
772 			case COUNTER_SKIP:
773 				continue;
774 			default:
775 				break;
776 			}
777 
778 		}
779 		counter->supported = true;
780 	}
781 
782 	if (second_pass) {
783 		/*
784 		 * Now redo all the weak group after closing them,
785 		 * and also close errored counters.
786 		 */
787 
788 		/* First close errored or weak retry */
789 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
790 			counter = evlist_cpu_itr.evsel;
791 
792 			if (!counter->reset_group && !counter->errored)
793 				continue;
794 
795 			perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
796 		}
797 		/* Now reopen weak */
798 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
799 			counter = evlist_cpu_itr.evsel;
800 
801 			if (!counter->reset_group)
802 				continue;
803 try_again_reset:
804 			pr_debug2("reopening weak %s\n", evsel__name(counter));
805 			if (create_perf_stat_counter(counter, &stat_config, &target,
806 						     evlist_cpu_itr.cpu_map_idx) < 0) {
807 
808 				switch (stat_handle_error(counter)) {
809 				case COUNTER_FATAL:
810 					err = -1;
811 					goto err_out;
812 				case COUNTER_RETRY:
813 					goto try_again_reset;
814 				case COUNTER_SKIP:
815 					continue;
816 				default:
817 					break;
818 				}
819 			}
820 			counter->supported = true;
821 		}
822 	}
823 	affinity__cleanup(affinity);
824 	affinity = NULL;
825 
826 	evlist__for_each_entry(evsel_list, counter) {
827 		if (!counter->supported) {
828 			perf_evsel__free_fd(&counter->core);
829 			continue;
830 		}
831 
832 		l = strlen(counter->unit);
833 		if (l > stat_config.unit_width)
834 			stat_config.unit_width = l;
835 
836 		if (evsel__should_store_id(counter) &&
837 		    evsel__store_ids(counter, evsel_list)) {
838 			err = -1;
839 			goto err_out;
840 		}
841 	}
842 
843 	if (evlist__apply_filters(evsel_list, &counter, &target)) {
844 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
845 			counter->filter, evsel__name(counter), errno,
846 			str_error_r(errno, msg, sizeof(msg)));
847 		return -1;
848 	}
849 
850 	if (STAT_RECORD) {
851 		int fd = perf_data__fd(&perf_stat.data);
852 
853 		if (is_pipe) {
854 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
855 		} else {
856 			err = perf_session__write_header(perf_stat.session, evsel_list,
857 							 fd, false);
858 		}
859 
860 		if (err < 0)
861 			goto err_out;
862 
863 		err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
864 							 process_synthesized_event, is_pipe);
865 		if (err < 0)
866 			goto err_out;
867 
868 	}
869 
870 	if (target.initial_delay) {
871 		pr_info(EVLIST_DISABLED_MSG);
872 	} else {
873 		err = enable_counters();
874 		if (err) {
875 			err = -1;
876 			goto err_out;
877 		}
878 	}
879 
880 	/* Exec the command, if any */
881 	if (forks)
882 		evlist__start_workload(evsel_list);
883 
884 	if (target.initial_delay > 0) {
885 		usleep(target.initial_delay * USEC_PER_MSEC);
886 		err = enable_counters();
887 		if (err) {
888 			err = -1;
889 			goto err_out;
890 		}
891 
892 		pr_info(EVLIST_ENABLED_MSG);
893 	}
894 
895 	t0 = rdclock();
896 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
897 
898 	if (forks) {
899 		if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
900 			status = dispatch_events(forks, timeout, interval, &times);
901 		if (child_pid != -1) {
902 			if (timeout)
903 				kill(child_pid, SIGTERM);
904 			wait4(child_pid, &status, 0, &stat_config.ru_data);
905 		}
906 
907 		if (workload_exec_errno) {
908 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
909 			pr_err("Workload failed: %s\n", emsg);
910 			err = -1;
911 			goto err_out;
912 		}
913 
914 		if (WIFSIGNALED(status))
915 			psignal(WTERMSIG(status), argv[0]);
916 	} else {
917 		status = dispatch_events(forks, timeout, interval, &times);
918 	}
919 
920 	disable_counters();
921 
922 	t1 = rdclock();
923 
924 	if (stat_config.walltime_run_table)
925 		stat_config.walltime_run[run_idx] = t1 - t0;
926 
927 	if (interval && stat_config.summary) {
928 		stat_config.interval = 0;
929 		stat_config.stop_read_counter = true;
930 		init_stats(&walltime_nsecs_stats);
931 		update_stats(&walltime_nsecs_stats, t1 - t0);
932 
933 		evlist__copy_prev_raw_counts(evsel_list);
934 		evlist__reset_prev_raw_counts(evsel_list);
935 		evlist__reset_aggr_stats(evsel_list);
936 	} else {
937 		update_stats(&walltime_nsecs_stats, t1 - t0);
938 		update_rusage_stats(&ru_stats, &stat_config.ru_data);
939 	}
940 
941 	/*
942 	 * Closing a group leader splits the group, and as we only disable
943 	 * group leaders, results in remaining events becoming enabled. To
944 	 * avoid arbitrary skew, we must read all counters before closing any
945 	 * group leaders.
946 	 */
947 	if (read_counters() == 0)
948 		process_counters();
949 
950 	/*
951 	 * We need to keep evsel_list alive, because it's processed
952 	 * later the evsel_list will be closed after.
953 	 */
954 	if (!STAT_RECORD)
955 		evlist__close(evsel_list);
956 
957 	return WEXITSTATUS(status);
958 
959 err_out:
960 	if (forks)
961 		evlist__cancel_workload(evsel_list);
962 
963 	affinity__cleanup(affinity);
964 	return err;
965 }
966 
967 /*
968  * Returns -1 for fatal errors which signifies to not continue
969  * when in repeat mode.
970  *
971  * Returns < -1 error codes when stat record is used. These
972  * result in the stat information being displayed, but writing
973  * to the file fails and is non fatal.
974  */
975 static int run_perf_stat(int argc, const char **argv, int run_idx)
976 {
977 	int ret;
978 
979 	if (pre_cmd) {
980 		ret = system(pre_cmd);
981 		if (ret)
982 			return ret;
983 	}
984 
985 	if (sync_run)
986 		sync();
987 
988 	ret = __run_perf_stat(argc, argv, run_idx);
989 	if (ret)
990 		return ret;
991 
992 	if (post_cmd) {
993 		ret = system(post_cmd);
994 		if (ret)
995 			return ret;
996 	}
997 
998 	return ret;
999 }
1000 
1001 static void print_counters(struct timespec *ts, int argc, const char **argv)
1002 {
1003 	/* Do not print anything if we record to the pipe. */
1004 	if (STAT_RECORD && perf_stat.data.is_pipe)
1005 		return;
1006 	if (quiet)
1007 		return;
1008 
1009 	evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1010 }
1011 
1012 static volatile sig_atomic_t signr = -1;
1013 
1014 static void skip_signal(int signo)
1015 {
1016 	if ((child_pid == -1) || stat_config.interval)
1017 		done = 1;
1018 
1019 	signr = signo;
1020 	/*
1021 	 * render child_pid harmless
1022 	 * won't send SIGTERM to a random
1023 	 * process in case of race condition
1024 	 * and fast PID recycling
1025 	 */
1026 	child_pid = -1;
1027 }
1028 
1029 static void sig_atexit(void)
1030 {
1031 	sigset_t set, oset;
1032 
1033 	/*
1034 	 * avoid race condition with SIGCHLD handler
1035 	 * in skip_signal() which is modifying child_pid
1036 	 * goal is to avoid send SIGTERM to a random
1037 	 * process
1038 	 */
1039 	sigemptyset(&set);
1040 	sigaddset(&set, SIGCHLD);
1041 	sigprocmask(SIG_BLOCK, &set, &oset);
1042 
1043 	if (child_pid != -1)
1044 		kill(child_pid, SIGTERM);
1045 
1046 	sigprocmask(SIG_SETMASK, &oset, NULL);
1047 
1048 	if (signr == -1)
1049 		return;
1050 
1051 	signal(signr, SIG_DFL);
1052 	kill(getpid(), signr);
1053 }
1054 
1055 static int stat__set_big_num(const struct option *opt __maybe_unused,
1056 			     const char *s __maybe_unused, int unset)
1057 {
1058 	big_num_opt = unset ? 0 : 1;
1059 	perf_stat__set_big_num(!unset);
1060 	return 0;
1061 }
1062 
1063 static int enable_metric_only(const struct option *opt __maybe_unused,
1064 			      const char *s __maybe_unused, int unset)
1065 {
1066 	force_metric_only = true;
1067 	stat_config.metric_only = !unset;
1068 	return 0;
1069 }
1070 
1071 static int append_metric_groups(const struct option *opt __maybe_unused,
1072 			       const char *str,
1073 			       int unset __maybe_unused)
1074 {
1075 	if (metrics) {
1076 		char *tmp;
1077 
1078 		if (asprintf(&tmp, "%s,%s", metrics, str) < 0)
1079 			return -ENOMEM;
1080 		free(metrics);
1081 		metrics = tmp;
1082 	} else {
1083 		metrics = strdup(str);
1084 		if (!metrics)
1085 			return -ENOMEM;
1086 	}
1087 	return 0;
1088 }
1089 
1090 static int parse_control_option(const struct option *opt,
1091 				const char *str,
1092 				int unset __maybe_unused)
1093 {
1094 	struct perf_stat_config *config = opt->value;
1095 
1096 	return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1097 }
1098 
1099 static int parse_stat_cgroups(const struct option *opt,
1100 			      const char *str, int unset)
1101 {
1102 	if (stat_config.cgroup_list) {
1103 		pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1104 		return -1;
1105 	}
1106 
1107 	return parse_cgroups(opt, str, unset);
1108 }
1109 
1110 static int parse_cputype(const struct option *opt,
1111 			     const char *str,
1112 			     int unset __maybe_unused)
1113 {
1114 	const struct perf_pmu *pmu;
1115 	struct evlist *evlist = *(struct evlist **)opt->value;
1116 
1117 	if (!list_empty(&evlist->core.entries)) {
1118 		fprintf(stderr, "Must define cputype before events/metrics\n");
1119 		return -1;
1120 	}
1121 
1122 	pmu = perf_pmus__pmu_for_pmu_filter(str);
1123 	if (!pmu) {
1124 		fprintf(stderr, "--cputype %s is not supported!\n", str);
1125 		return -1;
1126 	}
1127 	parse_events_option_args.pmu_filter = pmu->name;
1128 
1129 	return 0;
1130 }
1131 
1132 static int parse_cache_level(const struct option *opt,
1133 			     const char *str,
1134 			     int unset __maybe_unused)
1135 {
1136 	int level;
1137 	struct opt_aggr_mode *opt_aggr_mode = (struct opt_aggr_mode *)opt->value;
1138 	u32 *aggr_level = (u32 *)opt->data;
1139 
1140 	/*
1141 	 * If no string is specified, aggregate based on the topology of
1142 	 * Last Level Cache (LLC). Since the LLC level can change from
1143 	 * architecture to architecture, set level greater than
1144 	 * MAX_CACHE_LVL which will be interpreted as LLC.
1145 	 */
1146 	if (str == NULL) {
1147 		level = MAX_CACHE_LVL + 1;
1148 		goto out;
1149 	}
1150 
1151 	/*
1152 	 * The format to specify cache level is LX or lX where X is the
1153 	 * cache level.
1154 	 */
1155 	if (strlen(str) != 2 || (str[0] != 'l' && str[0] != 'L')) {
1156 		pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1157 		       MAX_CACHE_LVL,
1158 		       MAX_CACHE_LVL);
1159 		return -EINVAL;
1160 	}
1161 
1162 	level = atoi(&str[1]);
1163 	if (level < 1) {
1164 		pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1165 		       MAX_CACHE_LVL,
1166 		       MAX_CACHE_LVL);
1167 		return -EINVAL;
1168 	}
1169 
1170 	if (level > MAX_CACHE_LVL) {
1171 		pr_err("perf only supports max cache level of %d.\n"
1172 		       "Consider increasing MAX_CACHE_LVL\n", MAX_CACHE_LVL);
1173 		return -EINVAL;
1174 	}
1175 out:
1176 	opt_aggr_mode->cache = true;
1177 	*aggr_level = level;
1178 	return 0;
1179 }
1180 
1181 /**
1182  * Calculate the cache instance ID from the map in
1183  * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1184  * Cache instance ID is the first CPU reported in the shared_cpu_list file.
1185  */
1186 static int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map)
1187 {
1188 	int id;
1189 	struct perf_cpu_map *cpu_map = perf_cpu_map__new(map);
1190 
1191 	/*
1192 	 * If the map contains no CPU, consider the current CPU to
1193 	 * be the first online CPU in the cache domain else use the
1194 	 * first online CPU of the cache domain as the ID.
1195 	 */
1196 	id = perf_cpu_map__min(cpu_map).cpu;
1197 	if (id == -1)
1198 		id = cpu.cpu;
1199 
1200 	/* Free the perf_cpu_map used to find the cache ID */
1201 	perf_cpu_map__put(cpu_map);
1202 
1203 	return id;
1204 }
1205 
1206 /**
1207  * cpu__get_cache_id - Returns 0 if successful in populating the
1208  * cache level and cache id. Cache level is read from
1209  * /sys/devices/system/cpu/cpuX/cache/indexY/level where as cache instance ID
1210  * is the first CPU reported by
1211  * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1212  */
1213 static int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache)
1214 {
1215 	int ret = 0;
1216 	u32 cache_level = stat_config.aggr_level;
1217 	struct cpu_cache_level caches[MAX_CACHE_LVL];
1218 	u32 i = 0, caches_cnt = 0;
1219 
1220 	cache->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1221 	cache->cache = -1;
1222 
1223 	ret = build_caches_for_cpu(cpu.cpu, caches, &caches_cnt);
1224 	if (ret) {
1225 		/*
1226 		 * If caches_cnt is not 0, cpu_cache_level data
1227 		 * was allocated when building the topology.
1228 		 * Free the allocated data before returning.
1229 		 */
1230 		if (caches_cnt)
1231 			goto free_caches;
1232 
1233 		return ret;
1234 	}
1235 
1236 	if (!caches_cnt)
1237 		return -1;
1238 
1239 	/*
1240 	 * Save the data for the highest level if no
1241 	 * level was specified by the user.
1242 	 */
1243 	if (cache_level > MAX_CACHE_LVL) {
1244 		int max_level_index = 0;
1245 
1246 		for (i = 1; i < caches_cnt; ++i) {
1247 			if (caches[i].level > caches[max_level_index].level)
1248 				max_level_index = i;
1249 		}
1250 
1251 		cache->cache_lvl = caches[max_level_index].level;
1252 		cache->cache = cpu__get_cache_id_from_map(cpu, caches[max_level_index].map);
1253 
1254 		/* Reset i to 0 to free entire caches[] */
1255 		i = 0;
1256 		goto free_caches;
1257 	}
1258 
1259 	for (i = 0; i < caches_cnt; ++i) {
1260 		if (caches[i].level == cache_level) {
1261 			cache->cache_lvl = cache_level;
1262 			cache->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1263 		}
1264 
1265 		cpu_cache_level__free(&caches[i]);
1266 	}
1267 
1268 free_caches:
1269 	/*
1270 	 * Free all the allocated cpu_cache_level data.
1271 	 */
1272 	while (i < caches_cnt)
1273 		cpu_cache_level__free(&caches[i++]);
1274 
1275 	return ret;
1276 }
1277 
1278 /**
1279  * aggr_cpu_id__cache - Create an aggr_cpu_id with cache instache ID, cache
1280  * level, die and socket populated with the cache instache ID, cache level,
1281  * die and socket for cpu. The function signature is compatible with
1282  * aggr_cpu_id_get_t.
1283  */
1284 static struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data)
1285 {
1286 	int ret;
1287 	struct aggr_cpu_id id;
1288 	struct perf_cache cache;
1289 
1290 	id = aggr_cpu_id__die(cpu, data);
1291 	if (aggr_cpu_id__is_empty(&id))
1292 		return id;
1293 
1294 	ret = cpu__get_cache_details(cpu, &cache);
1295 	if (ret)
1296 		return id;
1297 
1298 	id.cache_lvl = cache.cache_lvl;
1299 	id.cache = cache.cache;
1300 	return id;
1301 }
1302 
1303 static const char *const aggr_mode__string[] = {
1304 	[AGGR_CORE] = "core",
1305 	[AGGR_CACHE] = "cache",
1306 	[AGGR_CLUSTER] = "cluster",
1307 	[AGGR_DIE] = "die",
1308 	[AGGR_GLOBAL] = "global",
1309 	[AGGR_NODE] = "node",
1310 	[AGGR_NONE] = "none",
1311 	[AGGR_SOCKET] = "socket",
1312 	[AGGR_THREAD] = "thread",
1313 	[AGGR_UNSET] = "unset",
1314 };
1315 
1316 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1317 						struct perf_cpu cpu)
1318 {
1319 	return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1320 }
1321 
1322 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1323 					     struct perf_cpu cpu)
1324 {
1325 	return aggr_cpu_id__die(cpu, /*data=*/NULL);
1326 }
1327 
1328 static struct aggr_cpu_id perf_stat__get_cache_id(struct perf_stat_config *config __maybe_unused,
1329 						  struct perf_cpu cpu)
1330 {
1331 	return aggr_cpu_id__cache(cpu, /*data=*/NULL);
1332 }
1333 
1334 static struct aggr_cpu_id perf_stat__get_cluster(struct perf_stat_config *config __maybe_unused,
1335 						 struct perf_cpu cpu)
1336 {
1337 	return aggr_cpu_id__cluster(cpu, /*data=*/NULL);
1338 }
1339 
1340 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1341 					      struct perf_cpu cpu)
1342 {
1343 	return aggr_cpu_id__core(cpu, /*data=*/NULL);
1344 }
1345 
1346 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1347 					      struct perf_cpu cpu)
1348 {
1349 	return aggr_cpu_id__node(cpu, /*data=*/NULL);
1350 }
1351 
1352 static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused,
1353 						struct perf_cpu cpu)
1354 {
1355 	return aggr_cpu_id__global(cpu, /*data=*/NULL);
1356 }
1357 
1358 static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused,
1359 					     struct perf_cpu cpu)
1360 {
1361 	return aggr_cpu_id__cpu(cpu, /*data=*/NULL);
1362 }
1363 
1364 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1365 					      aggr_get_id_t get_id, struct perf_cpu cpu)
1366 {
1367 	struct aggr_cpu_id id;
1368 
1369 	/* per-process mode - should use global aggr mode */
1370 	if (cpu.cpu == -1)
1371 		return get_id(config, cpu);
1372 
1373 	if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1374 		config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1375 
1376 	id = config->cpus_aggr_map->map[cpu.cpu];
1377 	return id;
1378 }
1379 
1380 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1381 						       struct perf_cpu cpu)
1382 {
1383 	return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1384 }
1385 
1386 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1387 						    struct perf_cpu cpu)
1388 {
1389 	return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1390 }
1391 
1392 static struct aggr_cpu_id perf_stat__get_cluster_cached(struct perf_stat_config *config,
1393 							struct perf_cpu cpu)
1394 {
1395 	return perf_stat__get_aggr(config, perf_stat__get_cluster, cpu);
1396 }
1397 
1398 static struct aggr_cpu_id perf_stat__get_cache_id_cached(struct perf_stat_config *config,
1399 							 struct perf_cpu cpu)
1400 {
1401 	return perf_stat__get_aggr(config, perf_stat__get_cache_id, cpu);
1402 }
1403 
1404 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1405 						     struct perf_cpu cpu)
1406 {
1407 	return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1408 }
1409 
1410 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1411 						     struct perf_cpu cpu)
1412 {
1413 	return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1414 }
1415 
1416 static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config,
1417 						       struct perf_cpu cpu)
1418 {
1419 	return perf_stat__get_aggr(config, perf_stat__get_global, cpu);
1420 }
1421 
1422 static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config,
1423 						    struct perf_cpu cpu)
1424 {
1425 	return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu);
1426 }
1427 
1428 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1429 {
1430 	switch (aggr_mode) {
1431 	case AGGR_SOCKET:
1432 		return aggr_cpu_id__socket;
1433 	case AGGR_DIE:
1434 		return aggr_cpu_id__die;
1435 	case AGGR_CLUSTER:
1436 		return aggr_cpu_id__cluster;
1437 	case AGGR_CACHE:
1438 		return aggr_cpu_id__cache;
1439 	case AGGR_CORE:
1440 		return aggr_cpu_id__core;
1441 	case AGGR_NODE:
1442 		return aggr_cpu_id__node;
1443 	case AGGR_NONE:
1444 		return aggr_cpu_id__cpu;
1445 	case AGGR_GLOBAL:
1446 		return aggr_cpu_id__global;
1447 	case AGGR_THREAD:
1448 	case AGGR_UNSET:
1449 	case AGGR_MAX:
1450 	default:
1451 		return NULL;
1452 	}
1453 }
1454 
1455 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1456 {
1457 	switch (aggr_mode) {
1458 	case AGGR_SOCKET:
1459 		return perf_stat__get_socket_cached;
1460 	case AGGR_DIE:
1461 		return perf_stat__get_die_cached;
1462 	case AGGR_CLUSTER:
1463 		return perf_stat__get_cluster_cached;
1464 	case AGGR_CACHE:
1465 		return perf_stat__get_cache_id_cached;
1466 	case AGGR_CORE:
1467 		return perf_stat__get_core_cached;
1468 	case AGGR_NODE:
1469 		return perf_stat__get_node_cached;
1470 	case AGGR_NONE:
1471 		return perf_stat__get_cpu_cached;
1472 	case AGGR_GLOBAL:
1473 		return perf_stat__get_global_cached;
1474 	case AGGR_THREAD:
1475 	case AGGR_UNSET:
1476 	case AGGR_MAX:
1477 	default:
1478 		return NULL;
1479 	}
1480 }
1481 
1482 static int perf_stat_init_aggr_mode(void)
1483 {
1484 	int nr;
1485 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1486 
1487 	if (get_id) {
1488 		bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1489 		stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1490 							 get_id, /*data=*/NULL, needs_sort);
1491 		if (!stat_config.aggr_map) {
1492 			pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]);
1493 			return -1;
1494 		}
1495 		stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1496 	}
1497 
1498 	if (stat_config.aggr_mode == AGGR_THREAD) {
1499 		nr = perf_thread_map__nr(evsel_list->core.threads);
1500 		stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1501 		if (stat_config.aggr_map == NULL)
1502 			return -ENOMEM;
1503 
1504 		for (int s = 0; s < nr; s++) {
1505 			struct aggr_cpu_id id = aggr_cpu_id__empty();
1506 
1507 			id.thread_idx = s;
1508 			stat_config.aggr_map->map[s] = id;
1509 		}
1510 		return 0;
1511 	}
1512 
1513 	/*
1514 	 * The evsel_list->cpus is the base we operate on,
1515 	 * taking the highest cpu number to be the size of
1516 	 * the aggregation translate cpumap.
1517 	 */
1518 	if (!perf_cpu_map__is_any_cpu_or_is_empty(evsel_list->core.user_requested_cpus))
1519 		nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1520 	else
1521 		nr = 0;
1522 	stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1523 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1524 }
1525 
1526 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1527 {
1528 	free(map);
1529 }
1530 
1531 static void perf_stat__exit_aggr_mode(void)
1532 {
1533 	cpu_aggr_map__delete(stat_config.aggr_map);
1534 	cpu_aggr_map__delete(stat_config.cpus_aggr_map);
1535 	stat_config.aggr_map = NULL;
1536 	stat_config.cpus_aggr_map = NULL;
1537 }
1538 
1539 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1540 {
1541 	struct perf_env *env = data;
1542 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1543 
1544 	if (cpu.cpu != -1)
1545 		id.socket = env->cpu[cpu.cpu].socket_id;
1546 
1547 	return id;
1548 }
1549 
1550 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1551 {
1552 	struct perf_env *env = data;
1553 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1554 
1555 	if (cpu.cpu != -1) {
1556 		/*
1557 		 * die_id is relative to socket, so start
1558 		 * with the socket ID and then add die to
1559 		 * make a unique ID.
1560 		 */
1561 		id.socket = env->cpu[cpu.cpu].socket_id;
1562 		id.die = env->cpu[cpu.cpu].die_id;
1563 	}
1564 
1565 	return id;
1566 }
1567 
1568 static void perf_env__get_cache_id_for_cpu(struct perf_cpu cpu, struct perf_env *env,
1569 					   u32 cache_level, struct aggr_cpu_id *id)
1570 {
1571 	int i;
1572 	int caches_cnt = env->caches_cnt;
1573 	struct cpu_cache_level *caches = env->caches;
1574 
1575 	id->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1576 	id->cache = -1;
1577 
1578 	if (!caches_cnt)
1579 		return;
1580 
1581 	for (i = caches_cnt - 1; i > -1; --i) {
1582 		struct perf_cpu_map *cpu_map;
1583 		int map_contains_cpu;
1584 
1585 		/*
1586 		 * If user has not specified a level, find the fist level with
1587 		 * the cpu in the map. Since building the map is expensive, do
1588 		 * this only if levels match.
1589 		 */
1590 		if (cache_level <= MAX_CACHE_LVL && caches[i].level != cache_level)
1591 			continue;
1592 
1593 		cpu_map = perf_cpu_map__new(caches[i].map);
1594 		map_contains_cpu = perf_cpu_map__idx(cpu_map, cpu);
1595 		perf_cpu_map__put(cpu_map);
1596 
1597 		if (map_contains_cpu != -1) {
1598 			id->cache_lvl = caches[i].level;
1599 			id->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1600 			return;
1601 		}
1602 	}
1603 }
1604 
1605 static struct aggr_cpu_id perf_env__get_cache_aggr_by_cpu(struct perf_cpu cpu,
1606 							  void *data)
1607 {
1608 	struct perf_env *env = data;
1609 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1610 
1611 	if (cpu.cpu != -1) {
1612 		u32 cache_level = (perf_stat.aggr_level) ?: stat_config.aggr_level;
1613 
1614 		id.socket = env->cpu[cpu.cpu].socket_id;
1615 		id.die = env->cpu[cpu.cpu].die_id;
1616 		perf_env__get_cache_id_for_cpu(cpu, env, cache_level, &id);
1617 	}
1618 
1619 	return id;
1620 }
1621 
1622 static struct aggr_cpu_id perf_env__get_cluster_aggr_by_cpu(struct perf_cpu cpu,
1623 							    void *data)
1624 {
1625 	struct perf_env *env = data;
1626 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1627 
1628 	if (cpu.cpu != -1) {
1629 		id.socket = env->cpu[cpu.cpu].socket_id;
1630 		id.die = env->cpu[cpu.cpu].die_id;
1631 		id.cluster = env->cpu[cpu.cpu].cluster_id;
1632 	}
1633 
1634 	return id;
1635 }
1636 
1637 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1638 {
1639 	struct perf_env *env = data;
1640 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1641 
1642 	if (cpu.cpu != -1) {
1643 		/*
1644 		 * core_id is relative to socket, die and cluster, we need a
1645 		 * global id. So we set socket, die id, cluster id and core id.
1646 		 */
1647 		id.socket = env->cpu[cpu.cpu].socket_id;
1648 		id.die = env->cpu[cpu.cpu].die_id;
1649 		id.cluster = env->cpu[cpu.cpu].cluster_id;
1650 		id.core = env->cpu[cpu.cpu].core_id;
1651 	}
1652 
1653 	return id;
1654 }
1655 
1656 static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data)
1657 {
1658 	struct perf_env *env = data;
1659 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1660 
1661 	if (cpu.cpu != -1) {
1662 		/*
1663 		 * core_id is relative to socket and die,
1664 		 * we need a global id. So we set
1665 		 * socket, die id and core id
1666 		 */
1667 		id.socket = env->cpu[cpu.cpu].socket_id;
1668 		id.die = env->cpu[cpu.cpu].die_id;
1669 		id.core = env->cpu[cpu.cpu].core_id;
1670 		id.cpu = cpu;
1671 	}
1672 
1673 	return id;
1674 }
1675 
1676 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1677 {
1678 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1679 
1680 	id.node = perf_env__numa_node(data, cpu);
1681 	return id;
1682 }
1683 
1684 static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused,
1685 							   void *data __maybe_unused)
1686 {
1687 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1688 
1689 	/* it always aggregates to the cpu 0 */
1690 	id.cpu = (struct perf_cpu){ .cpu = 0 };
1691 	return id;
1692 }
1693 
1694 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1695 						     struct perf_cpu cpu)
1696 {
1697 	return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1698 }
1699 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1700 						  struct perf_cpu cpu)
1701 {
1702 	return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1703 }
1704 
1705 static struct aggr_cpu_id perf_stat__get_cluster_file(struct perf_stat_config *config __maybe_unused,
1706 						      struct perf_cpu cpu)
1707 {
1708 	return perf_env__get_cluster_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1709 }
1710 
1711 static struct aggr_cpu_id perf_stat__get_cache_file(struct perf_stat_config *config __maybe_unused,
1712 						    struct perf_cpu cpu)
1713 {
1714 	return perf_env__get_cache_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1715 }
1716 
1717 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1718 						   struct perf_cpu cpu)
1719 {
1720 	return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1721 }
1722 
1723 static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused,
1724 						  struct perf_cpu cpu)
1725 {
1726 	return perf_env__get_cpu_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1727 }
1728 
1729 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1730 						   struct perf_cpu cpu)
1731 {
1732 	return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1733 }
1734 
1735 static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused,
1736 						     struct perf_cpu cpu)
1737 {
1738 	return perf_env__get_global_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1739 }
1740 
1741 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1742 {
1743 	switch (aggr_mode) {
1744 	case AGGR_SOCKET:
1745 		return perf_env__get_socket_aggr_by_cpu;
1746 	case AGGR_DIE:
1747 		return perf_env__get_die_aggr_by_cpu;
1748 	case AGGR_CLUSTER:
1749 		return perf_env__get_cluster_aggr_by_cpu;
1750 	case AGGR_CACHE:
1751 		return perf_env__get_cache_aggr_by_cpu;
1752 	case AGGR_CORE:
1753 		return perf_env__get_core_aggr_by_cpu;
1754 	case AGGR_NODE:
1755 		return perf_env__get_node_aggr_by_cpu;
1756 	case AGGR_GLOBAL:
1757 		return perf_env__get_global_aggr_by_cpu;
1758 	case AGGR_NONE:
1759 		return perf_env__get_cpu_aggr_by_cpu;
1760 	case AGGR_THREAD:
1761 	case AGGR_UNSET:
1762 	case AGGR_MAX:
1763 	default:
1764 		return NULL;
1765 	}
1766 }
1767 
1768 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1769 {
1770 	switch (aggr_mode) {
1771 	case AGGR_SOCKET:
1772 		return perf_stat__get_socket_file;
1773 	case AGGR_DIE:
1774 		return perf_stat__get_die_file;
1775 	case AGGR_CLUSTER:
1776 		return perf_stat__get_cluster_file;
1777 	case AGGR_CACHE:
1778 		return perf_stat__get_cache_file;
1779 	case AGGR_CORE:
1780 		return perf_stat__get_core_file;
1781 	case AGGR_NODE:
1782 		return perf_stat__get_node_file;
1783 	case AGGR_GLOBAL:
1784 		return perf_stat__get_global_file;
1785 	case AGGR_NONE:
1786 		return perf_stat__get_cpu_file;
1787 	case AGGR_THREAD:
1788 	case AGGR_UNSET:
1789 	case AGGR_MAX:
1790 	default:
1791 		return NULL;
1792 	}
1793 }
1794 
1795 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1796 {
1797 	struct perf_env *env = &st->session->header.env;
1798 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1799 	bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1800 
1801 	if (stat_config.aggr_mode == AGGR_THREAD) {
1802 		int nr = perf_thread_map__nr(evsel_list->core.threads);
1803 
1804 		stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1805 		if (stat_config.aggr_map == NULL)
1806 			return -ENOMEM;
1807 
1808 		for (int s = 0; s < nr; s++) {
1809 			struct aggr_cpu_id id = aggr_cpu_id__empty();
1810 
1811 			id.thread_idx = s;
1812 			stat_config.aggr_map->map[s] = id;
1813 		}
1814 		return 0;
1815 	}
1816 
1817 	if (!get_id)
1818 		return 0;
1819 
1820 	stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1821 						 get_id, env, needs_sort);
1822 	if (!stat_config.aggr_map) {
1823 		pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]);
1824 		return -1;
1825 	}
1826 	stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1827 	return 0;
1828 }
1829 
1830 /*
1831  * Add default events, if there were no attributes specified or
1832  * if -d/--detailed, -d -d or -d -d -d is used:
1833  */
1834 static int add_default_events(void)
1835 {
1836 	const char *pmu = parse_events_option_args.pmu_filter ?: "all";
1837 	struct parse_events_error err;
1838 	struct evlist *evlist = evlist__new();
1839 	struct evsel *evsel;
1840 	int ret = 0;
1841 
1842 	if (!evlist)
1843 		return -ENOMEM;
1844 
1845 	parse_events_error__init(&err);
1846 
1847 	/* Set attrs if no event is selected and !null_run: */
1848 	if (stat_config.null_run)
1849 		goto out;
1850 
1851 	if (transaction_run) {
1852 		/* Handle -T as -M transaction. Once platform specific metrics
1853 		 * support has been added to the json files, all architectures
1854 		 * will use this approach. To determine transaction support
1855 		 * on an architecture test for such a metric name.
1856 		 */
1857 		if (!metricgroup__has_metric_or_groups(pmu, "transaction")) {
1858 			pr_err("Missing transaction metrics\n");
1859 			ret = -1;
1860 			goto out;
1861 		}
1862 		ret = metricgroup__parse_groups(evlist, pmu, "transaction",
1863 						stat_config.metric_no_group,
1864 						stat_config.metric_no_merge,
1865 						stat_config.metric_no_threshold,
1866 						stat_config.user_requested_cpu_list,
1867 						stat_config.system_wide,
1868 						stat_config.hardware_aware_grouping,
1869 						&stat_config.metric_events);
1870 		goto out;
1871 	}
1872 
1873 	if (smi_cost) {
1874 		int smi;
1875 
1876 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1877 			pr_err("freeze_on_smi is not supported.\n");
1878 			ret = -1;
1879 			goto out;
1880 		}
1881 
1882 		if (!smi) {
1883 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1884 				pr_err("Failed to set freeze_on_smi.\n");
1885 				ret = -1;
1886 				goto out;
1887 			}
1888 			smi_reset = true;
1889 		}
1890 
1891 		if (!metricgroup__has_metric_or_groups(pmu, "smi")) {
1892 			pr_err("Missing smi metrics\n");
1893 			ret = -1;
1894 			goto out;
1895 		}
1896 
1897 		if (!force_metric_only)
1898 			stat_config.metric_only = true;
1899 
1900 		ret = metricgroup__parse_groups(evlist, pmu, "smi",
1901 						stat_config.metric_no_group,
1902 						stat_config.metric_no_merge,
1903 						stat_config.metric_no_threshold,
1904 						stat_config.user_requested_cpu_list,
1905 						stat_config.system_wide,
1906 						stat_config.hardware_aware_grouping,
1907 						&stat_config.metric_events);
1908 		goto out;
1909 	}
1910 
1911 	if (topdown_run) {
1912 		unsigned int max_level = metricgroups__topdown_max_level();
1913 		char str[] = "TopdownL1";
1914 
1915 		if (!force_metric_only)
1916 			stat_config.metric_only = true;
1917 
1918 		if (!max_level) {
1919 			pr_err("Topdown requested but the topdown metric groups aren't present.\n"
1920 				"(See perf list the metric groups have names like TopdownL1)\n");
1921 			ret = -1;
1922 			goto out;
1923 		}
1924 		if (stat_config.topdown_level > max_level) {
1925 			pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1926 			ret = -1;
1927 			goto out;
1928 		} else if (!stat_config.topdown_level) {
1929 			stat_config.topdown_level = 1;
1930 		}
1931 		if (!stat_config.interval && !stat_config.metric_only) {
1932 			fprintf(stat_config.output,
1933 				"Topdown accuracy may decrease when measuring long periods.\n"
1934 				"Please print the result regularly, e.g. -I1000\n");
1935 		}
1936 		str[8] = stat_config.topdown_level + '0';
1937 		if (metricgroup__parse_groups(evlist,
1938 						pmu, str,
1939 						/*metric_no_group=*/false,
1940 						/*metric_no_merge=*/false,
1941 						/*metric_no_threshold=*/true,
1942 						stat_config.user_requested_cpu_list,
1943 						stat_config.system_wide,
1944 						stat_config.hardware_aware_grouping,
1945 						&stat_config.metric_events) < 0) {
1946 			ret = -1;
1947 			goto out;
1948 		}
1949 	}
1950 
1951 	if (!stat_config.topdown_level)
1952 		stat_config.topdown_level = 1;
1953 
1954 	if (!evlist->core.nr_entries && !evsel_list->core.nr_entries) {
1955 		/* No events so add defaults. */
1956 		if (target__has_cpu(&target))
1957 			ret = parse_events(evlist, "cpu-clock", &err);
1958 		else
1959 			ret = parse_events(evlist, "task-clock", &err);
1960 		if (ret)
1961 			goto out;
1962 
1963 		ret = parse_events(evlist,
1964 				"context-switches,"
1965 				"cpu-migrations,"
1966 				"page-faults,"
1967 				"instructions,"
1968 				"cycles,"
1969 				"stalled-cycles-frontend,"
1970 				"stalled-cycles-backend,"
1971 				"branches,"
1972 				"branch-misses",
1973 				&err);
1974 		if (ret)
1975 			goto out;
1976 
1977 		/*
1978 		 * Add TopdownL1 metrics if they exist. To minimize
1979 		 * multiplexing, don't request threshold computation.
1980 		 */
1981 		if (metricgroup__has_metric_or_groups(pmu, "Default")) {
1982 			struct evlist *metric_evlist = evlist__new();
1983 
1984 			if (!metric_evlist) {
1985 				ret = -ENOMEM;
1986 				goto out;
1987 			}
1988 			if (metricgroup__parse_groups(metric_evlist, pmu, "Default",
1989 							/*metric_no_group=*/false,
1990 							/*metric_no_merge=*/false,
1991 							/*metric_no_threshold=*/true,
1992 							stat_config.user_requested_cpu_list,
1993 							stat_config.system_wide,
1994 							stat_config.hardware_aware_grouping,
1995 							&stat_config.metric_events) < 0) {
1996 				ret = -1;
1997 				goto out;
1998 			}
1999 
2000 			evlist__for_each_entry(metric_evlist, evsel)
2001 				evsel->default_metricgroup = true;
2002 
2003 			evlist__splice_list_tail(evlist, &metric_evlist->core.entries);
2004 			evlist__delete(metric_evlist);
2005 		}
2006 	}
2007 
2008 	/* Detailed events get appended to the event list: */
2009 
2010 	if (!ret && detailed_run >=  1) {
2011 		/*
2012 		 * Detailed stats (-d), covering the L1 and last level data
2013 		 * caches:
2014 		 */
2015 		ret = parse_events(evlist,
2016 				"L1-dcache-loads,"
2017 				"L1-dcache-load-misses,"
2018 				"LLC-loads,"
2019 				"LLC-load-misses",
2020 				&err);
2021 	}
2022 	if (!ret && detailed_run >=  2) {
2023 		/*
2024 		 * Very detailed stats (-d -d), covering the instruction cache
2025 		 * and the TLB caches:
2026 		 */
2027 		ret = parse_events(evlist,
2028 				"L1-icache-loads,"
2029 				"L1-icache-load-misses,"
2030 				"dTLB-loads,"
2031 				"dTLB-load-misses,"
2032 				"iTLB-loads,"
2033 				"iTLB-load-misses",
2034 				&err);
2035 	}
2036 	if (!ret && detailed_run >=  3) {
2037 		/*
2038 		 * Very, very detailed stats (-d -d -d), adding prefetch events:
2039 		 */
2040 		ret = parse_events(evlist,
2041 				"L1-dcache-prefetches,"
2042 				"L1-dcache-prefetch-misses",
2043 				&err);
2044 	}
2045 out:
2046 	if (!ret) {
2047 		evlist__for_each_entry(evlist, evsel) {
2048 			/*
2049 			 * Make at least one event non-skippable so fatal errors are visible.
2050 			 * 'cycles' always used to be default and non-skippable, so use that.
2051 			 */
2052 			if (strcmp("cycles", evsel__name(evsel)))
2053 				evsel->skippable = true;
2054 		}
2055 	}
2056 	parse_events_error__exit(&err);
2057 	evlist__splice_list_tail(evsel_list, &evlist->core.entries);
2058 	evlist__delete(evlist);
2059 	return ret;
2060 }
2061 
2062 static const char * const stat_record_usage[] = {
2063 	"perf stat record [<options>]",
2064 	NULL,
2065 };
2066 
2067 static void init_features(struct perf_session *session)
2068 {
2069 	int feat;
2070 
2071 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2072 		perf_header__set_feat(&session->header, feat);
2073 
2074 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
2075 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2076 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2077 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2078 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2079 }
2080 
2081 static int __cmd_record(const struct option stat_options[], struct opt_aggr_mode *opt_mode,
2082 			int argc, const char **argv)
2083 {
2084 	struct perf_session *session;
2085 	struct perf_data *data = &perf_stat.data;
2086 
2087 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
2088 			     PARSE_OPT_STOP_AT_NON_OPTION);
2089 	stat_config.aggr_mode = opt_aggr_mode_to_aggr_mode(opt_mode);
2090 
2091 	if (output_name)
2092 		data->path = output_name;
2093 
2094 	if (stat_config.run_count != 1 || forever) {
2095 		pr_err("Cannot use -r option with perf stat record.\n");
2096 		return -1;
2097 	}
2098 
2099 	session = perf_session__new(data, NULL);
2100 	if (IS_ERR(session)) {
2101 		pr_err("Perf session creation failed\n");
2102 		return PTR_ERR(session);
2103 	}
2104 
2105 	init_features(session);
2106 
2107 	session->evlist   = evsel_list;
2108 	perf_stat.session = session;
2109 	perf_stat.record  = true;
2110 	return argc;
2111 }
2112 
2113 static int process_stat_round_event(struct perf_session *session,
2114 				    union perf_event *event)
2115 {
2116 	struct perf_record_stat_round *stat_round = &event->stat_round;
2117 	struct timespec tsh, *ts = NULL;
2118 	const char **argv = session->header.env.cmdline_argv;
2119 	int argc = session->header.env.nr_cmdline;
2120 
2121 	process_counters();
2122 
2123 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2124 		update_stats(&walltime_nsecs_stats, stat_round->time);
2125 
2126 	if (stat_config.interval && stat_round->time) {
2127 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2128 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2129 		ts = &tsh;
2130 	}
2131 
2132 	print_counters(ts, argc, argv);
2133 	return 0;
2134 }
2135 
2136 static
2137 int process_stat_config_event(struct perf_session *session,
2138 			      union perf_event *event)
2139 {
2140 	const struct perf_tool *tool = session->tool;
2141 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2142 
2143 	perf_event__read_stat_config(&stat_config, &event->stat_config);
2144 
2145 	if (perf_cpu_map__is_empty(st->cpus)) {
2146 		if (st->aggr_mode != AGGR_UNSET)
2147 			pr_warning("warning: processing task data, aggregation mode not set\n");
2148 	} else if (st->aggr_mode != AGGR_UNSET) {
2149 		stat_config.aggr_mode = st->aggr_mode;
2150 	}
2151 
2152 	if (perf_stat.data.is_pipe)
2153 		perf_stat_init_aggr_mode();
2154 	else
2155 		perf_stat_init_aggr_mode_file(st);
2156 
2157 	if (stat_config.aggr_map) {
2158 		int nr_aggr = stat_config.aggr_map->nr;
2159 
2160 		if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) {
2161 			pr_err("cannot allocate aggr counts\n");
2162 			return -1;
2163 		}
2164 	}
2165 	return 0;
2166 }
2167 
2168 static int set_maps(struct perf_stat *st)
2169 {
2170 	if (!st->cpus || !st->threads)
2171 		return 0;
2172 
2173 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2174 		return -EINVAL;
2175 
2176 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2177 
2178 	if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true))
2179 		return -ENOMEM;
2180 
2181 	st->maps_allocated = true;
2182 	return 0;
2183 }
2184 
2185 static
2186 int process_thread_map_event(struct perf_session *session,
2187 			     union perf_event *event)
2188 {
2189 	const struct perf_tool *tool = session->tool;
2190 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2191 
2192 	if (st->threads) {
2193 		pr_warning("Extra thread map event, ignoring.\n");
2194 		return 0;
2195 	}
2196 
2197 	st->threads = thread_map__new_event(&event->thread_map);
2198 	if (!st->threads)
2199 		return -ENOMEM;
2200 
2201 	return set_maps(st);
2202 }
2203 
2204 static
2205 int process_cpu_map_event(struct perf_session *session,
2206 			  union perf_event *event)
2207 {
2208 	const struct perf_tool *tool = session->tool;
2209 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2210 	struct perf_cpu_map *cpus;
2211 
2212 	if (st->cpus) {
2213 		pr_warning("Extra cpu map event, ignoring.\n");
2214 		return 0;
2215 	}
2216 
2217 	cpus = cpu_map__new_data(&event->cpu_map.data);
2218 	if (!cpus)
2219 		return -ENOMEM;
2220 
2221 	st->cpus = cpus;
2222 	return set_maps(st);
2223 }
2224 
2225 static const char * const stat_report_usage[] = {
2226 	"perf stat report [<options>]",
2227 	NULL,
2228 };
2229 
2230 static struct perf_stat perf_stat = {
2231 	.aggr_mode	= AGGR_UNSET,
2232 	.aggr_level	= 0,
2233 };
2234 
2235 static int __cmd_report(int argc, const char **argv)
2236 {
2237 	struct perf_session *session;
2238 	const struct option options[] = {
2239 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2240 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2241 		     "aggregate counts per processor socket", AGGR_SOCKET),
2242 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2243 		     "aggregate counts per processor die", AGGR_DIE),
2244 	OPT_SET_UINT(0, "per-cluster", &perf_stat.aggr_mode,
2245 		     "aggregate counts perf processor cluster", AGGR_CLUSTER),
2246 	OPT_CALLBACK_OPTARG(0, "per-cache", &perf_stat.aggr_mode, &perf_stat.aggr_level,
2247 			    "cache level",
2248 			    "aggregate count at this cache level (Default: LLC)",
2249 			    parse_cache_level),
2250 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2251 		     "aggregate counts per physical processor core", AGGR_CORE),
2252 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2253 		     "aggregate counts per numa node", AGGR_NODE),
2254 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2255 		     "disable CPU count aggregation", AGGR_NONE),
2256 	OPT_END()
2257 	};
2258 	struct stat st;
2259 	int ret;
2260 
2261 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2262 
2263 	if (!input_name || !strlen(input_name)) {
2264 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2265 			input_name = "-";
2266 		else
2267 			input_name = "perf.data";
2268 	}
2269 
2270 	perf_stat.data.path = input_name;
2271 	perf_stat.data.mode = PERF_DATA_MODE_READ;
2272 
2273 	perf_tool__init(&perf_stat.tool, /*ordered_events=*/false);
2274 	perf_stat.tool.attr		= perf_event__process_attr;
2275 	perf_stat.tool.event_update	= perf_event__process_event_update;
2276 	perf_stat.tool.thread_map	= process_thread_map_event;
2277 	perf_stat.tool.cpu_map		= process_cpu_map_event;
2278 	perf_stat.tool.stat_config	= process_stat_config_event;
2279 	perf_stat.tool.stat		= perf_event__process_stat_event;
2280 	perf_stat.tool.stat_round	= process_stat_round_event;
2281 
2282 	session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2283 	if (IS_ERR(session))
2284 		return PTR_ERR(session);
2285 
2286 	perf_stat.session  = session;
2287 	stat_config.output = stderr;
2288 	evlist__delete(evsel_list);
2289 	evsel_list         = session->evlist;
2290 
2291 	ret = perf_session__process_events(session);
2292 	if (ret)
2293 		return ret;
2294 
2295 	perf_session__delete(session);
2296 	return 0;
2297 }
2298 
2299 static void setup_system_wide(int forks)
2300 {
2301 	/*
2302 	 * Make system wide (-a) the default target if
2303 	 * no target was specified and one of following
2304 	 * conditions is met:
2305 	 *
2306 	 *   - there's no workload specified
2307 	 *   - there is workload specified but all requested
2308 	 *     events are system wide events
2309 	 */
2310 	if (!target__none(&target))
2311 		return;
2312 
2313 	if (!forks)
2314 		target.system_wide = true;
2315 	else {
2316 		struct evsel *counter;
2317 
2318 		evlist__for_each_entry(evsel_list, counter) {
2319 			if (!counter->core.requires_cpu &&
2320 			    !evsel__name_is(counter, "duration_time")) {
2321 				return;
2322 			}
2323 		}
2324 
2325 		if (evsel_list->core.nr_entries)
2326 			target.system_wide = true;
2327 	}
2328 }
2329 
2330 #ifdef HAVE_ARCH_X86_64_SUPPORT
2331 static int parse_tpebs_mode(const struct option *opt, const char *str,
2332 			    int unset __maybe_unused)
2333 {
2334 	enum tpebs_mode *mode = opt->value;
2335 
2336 	if (!strcasecmp("mean", str)) {
2337 		*mode = TPEBS_MODE__MEAN;
2338 		return 0;
2339 	}
2340 	if (!strcasecmp("min", str)) {
2341 		*mode = TPEBS_MODE__MIN;
2342 		return 0;
2343 	}
2344 	if (!strcasecmp("max", str)) {
2345 		*mode = TPEBS_MODE__MAX;
2346 		return 0;
2347 	}
2348 	if (!strcasecmp("last", str)) {
2349 		*mode = TPEBS_MODE__LAST;
2350 		return 0;
2351 	}
2352 	return -1;
2353 }
2354 #endif // HAVE_ARCH_X86_64_SUPPORT
2355 
2356 int cmd_stat(int argc, const char **argv)
2357 {
2358 	struct opt_aggr_mode opt_mode = {};
2359 	struct option stat_options[] = {
2360 		OPT_BOOLEAN('T', "transaction", &transaction_run,
2361 			"hardware transaction statistics"),
2362 		OPT_CALLBACK('e', "event", &parse_events_option_args, "event",
2363 			"event selector. use 'perf list' to list available events",
2364 			parse_events_option),
2365 		OPT_CALLBACK(0, "filter", &evsel_list, "filter",
2366 			"event filter", parse_filter),
2367 		OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
2368 			"child tasks do not inherit counters"),
2369 		OPT_STRING('p', "pid", &target.pid, "pid",
2370 			"stat events on existing process id"),
2371 		OPT_STRING('t', "tid", &target.tid, "tid",
2372 			"stat events on existing thread id"),
2373 #ifdef HAVE_BPF_SKEL
2374 		OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
2375 			"stat events on existing bpf program id"),
2376 		OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
2377 			"use bpf program to count events"),
2378 		OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
2379 			"path to perf_event_attr map"),
2380 #endif
2381 		OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
2382 			"system-wide collection from all CPUs"),
2383 		OPT_BOOLEAN(0, "scale", &stat_config.scale,
2384 			"Use --no-scale to disable counter scaling for multiplexing"),
2385 		OPT_INCR('v', "verbose", &verbose,
2386 			"be more verbose (show counter open errors, etc)"),
2387 		OPT_INTEGER('r', "repeat", &stat_config.run_count,
2388 			"repeat command and print average + stddev (max: 100, forever: 0)"),
2389 		OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
2390 			"display details about each run (only with -r option)"),
2391 		OPT_BOOLEAN('n', "null", &stat_config.null_run,
2392 			"null run - dont start any counters"),
2393 		OPT_INCR('d', "detailed", &detailed_run,
2394 			"detailed run - start a lot of events"),
2395 		OPT_BOOLEAN('S', "sync", &sync_run,
2396 			"call sync() before starting a run"),
2397 		OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
2398 				"print large numbers with thousands\' separators",
2399 				stat__set_big_num),
2400 		OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
2401 			"list of cpus to monitor in system-wide"),
2402 		OPT_BOOLEAN('A', "no-aggr", &opt_mode.no_aggr,
2403 			"disable aggregation across CPUs or PMUs"),
2404 		OPT_BOOLEAN(0, "no-merge", &opt_mode.no_aggr,
2405 			"disable aggregation the same as -A or -no-aggr"),
2406 		OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
2407 			"Merge identical named hybrid events"),
2408 		OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
2409 			"print counts with custom separator"),
2410 		OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
2411 			"print counts in JSON format"),
2412 		OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
2413 			"monitor event in cgroup name only", parse_stat_cgroups),
2414 		OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
2415 			"expand events for each cgroup"),
2416 		OPT_STRING('o', "output", &output_name, "file", "output file name"),
2417 		OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
2418 		OPT_INTEGER(0, "log-fd", &output_fd,
2419 			"log output to fd, instead of stderr"),
2420 		OPT_STRING(0, "pre", &pre_cmd, "command",
2421 			"command to run prior to the measured command"),
2422 		OPT_STRING(0, "post", &post_cmd, "command",
2423 			"command to run after to the measured command"),
2424 		OPT_UINTEGER('I', "interval-print", &stat_config.interval,
2425 			"print counts at regular interval in ms "
2426 			"(overhead is possible for values <= 100ms)"),
2427 		OPT_INTEGER(0, "interval-count", &stat_config.times,
2428 			"print counts for fixed number of times"),
2429 		OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
2430 			"clear screen in between new interval"),
2431 		OPT_UINTEGER(0, "timeout", &stat_config.timeout,
2432 			"stop workload and print counts after a timeout period in ms (>= 10ms)"),
2433 		OPT_BOOLEAN(0, "per-socket", &opt_mode.socket,
2434 			"aggregate counts per processor socket"),
2435 		OPT_BOOLEAN(0, "per-die", &opt_mode.die, "aggregate counts per processor die"),
2436 		OPT_BOOLEAN(0, "per-cluster", &opt_mode.cluster,
2437 			"aggregate counts per processor cluster"),
2438 		OPT_CALLBACK_OPTARG(0, "per-cache", &opt_mode, &stat_config.aggr_level,
2439 				"cache level", "aggregate count at this cache level (Default: LLC)",
2440 				parse_cache_level),
2441 		OPT_BOOLEAN(0, "per-core", &opt_mode.core,
2442 			"aggregate counts per physical processor core"),
2443 		OPT_BOOLEAN(0, "per-thread", &opt_mode.thread, "aggregate counts per thread"),
2444 		OPT_BOOLEAN(0, "per-node", &opt_mode.node, "aggregate counts per numa node"),
2445 		OPT_INTEGER('D', "delay", &target.initial_delay,
2446 			"ms to wait before starting measurement after program start (-1: start with events disabled)"),
2447 		OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
2448 				"Only print computed metrics. No raw values", enable_metric_only),
2449 		OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
2450 			"don't group metric events, impacts multiplexing"),
2451 		OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
2452 			"don't try to share events between metrics in a group"),
2453 		OPT_BOOLEAN(0, "metric-no-threshold", &stat_config.metric_no_threshold,
2454 			"disable adding events for the metric threshold calculation"),
2455 		OPT_BOOLEAN(0, "topdown", &topdown_run,
2456 			"measure top-down statistics"),
2457 #ifdef HAVE_ARCH_X86_64_SUPPORT
2458 		OPT_BOOLEAN(0, "record-tpebs", &tpebs_recording,
2459 			"enable recording for tpebs when retire_latency required"),
2460 		OPT_CALLBACK(0, "tpebs-mode", &tpebs_mode, "tpebs-mode",
2461 			"Mode of TPEBS recording: mean, min or max",
2462 			parse_tpebs_mode),
2463 #endif
2464 		OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
2465 			"Set the metrics level for the top-down statistics (0: max level)"),
2466 		OPT_BOOLEAN(0, "smi-cost", &smi_cost,
2467 			"measure SMI cost"),
2468 		OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
2469 			"monitor specified metrics or metric groups (separated by ,)",
2470 			append_metric_groups),
2471 		OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
2472 				"Configure all used events to run in kernel space.",
2473 				PARSE_OPT_EXCLUSIVE),
2474 		OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
2475 				"Configure all used events to run in user space.",
2476 				PARSE_OPT_EXCLUSIVE),
2477 		OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
2478 			"Use with 'percore' event qualifier to show the event "
2479 			"counts of one hardware thread by sum up total hardware "
2480 			"threads of same physical core"),
2481 		OPT_BOOLEAN(0, "summary", &stat_config.summary,
2482 			"print summary for interval mode"),
2483 		OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
2484 			"don't print 'summary' for CSV summary output"),
2485 		OPT_BOOLEAN(0, "quiet", &quiet,
2486 			"don't print any output, messages or warnings (useful with record)"),
2487 		OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
2488 			"Only enable events on applying cpu with this type "
2489 			"for hybrid platform (e.g. core or atom)",
2490 			parse_cputype),
2491 #ifdef HAVE_LIBPFM
2492 		OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
2493 			"libpfm4 event selector. use 'perf list' to list available events",
2494 			parse_libpfm_events_option),
2495 #endif
2496 		OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
2497 			"Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
2498 			"\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
2499 			"\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
2500 			parse_control_option),
2501 		OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
2502 				"measure I/O performance metrics provided by arch/platform",
2503 				iostat_parse),
2504 		OPT_END()
2505 	};
2506 	const char * const stat_usage[] = {
2507 		"perf stat [<options>] [<command>]",
2508 		NULL
2509 	};
2510 	int status = -EINVAL, run_idx, err;
2511 	const char *mode;
2512 	FILE *output = stderr;
2513 	unsigned int interval, timeout;
2514 	const char * const stat_subcommands[] = { "record", "report" };
2515 	char errbuf[BUFSIZ];
2516 
2517 	setlocale(LC_ALL, "");
2518 
2519 	evsel_list = evlist__new();
2520 	if (evsel_list == NULL)
2521 		return -ENOMEM;
2522 
2523 	parse_events__shrink_config_terms();
2524 
2525 	/* String-parsing callback-based options would segfault when negated */
2526 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2527 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2528 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2529 
2530 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2531 					(const char **) stat_usage,
2532 					PARSE_OPT_STOP_AT_NON_OPTION);
2533 
2534 	stat_config.aggr_mode = opt_aggr_mode_to_aggr_mode(&opt_mode);
2535 
2536 	if (stat_config.csv_sep) {
2537 		stat_config.csv_output = true;
2538 		if (!strcmp(stat_config.csv_sep, "\\t"))
2539 			stat_config.csv_sep = "\t";
2540 	} else
2541 		stat_config.csv_sep = DEFAULT_SEPARATOR;
2542 
2543 	if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2544 		argc = __cmd_record(stat_options, &opt_mode, argc, argv);
2545 		if (argc < 0)
2546 			return -1;
2547 	} else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2548 		return __cmd_report(argc, argv);
2549 
2550 	interval = stat_config.interval;
2551 	timeout = stat_config.timeout;
2552 
2553 	/*
2554 	 * For record command the -o is already taken care of.
2555 	 */
2556 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2557 		output = NULL;
2558 
2559 	if (output_name && output_fd) {
2560 		fprintf(stderr, "cannot use both --output and --log-fd\n");
2561 		parse_options_usage(stat_usage, stat_options, "o", 1);
2562 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2563 		goto out;
2564 	}
2565 
2566 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2567 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2568 		goto out;
2569 	}
2570 
2571 	if (stat_config.metric_only && stat_config.run_count > 1) {
2572 		fprintf(stderr, "--metric-only is not supported with -r\n");
2573 		goto out;
2574 	}
2575 
2576 	if (stat_config.csv_output || (stat_config.metric_only && stat_config.json_output)) {
2577 		/*
2578 		 * Current CSV and metric-only JSON output doesn't display the
2579 		 * metric threshold so don't compute it.
2580 		 */
2581 		stat_config.metric_no_threshold = true;
2582 	}
2583 
2584 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2585 		fprintf(stderr, "--table is only supported with -r\n");
2586 		parse_options_usage(stat_usage, stat_options, "r", 1);
2587 		parse_options_usage(NULL, stat_options, "table", 0);
2588 		goto out;
2589 	}
2590 
2591 	if (output_fd < 0) {
2592 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2593 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2594 		goto out;
2595 	}
2596 
2597 	if (!output && !quiet) {
2598 		struct timespec tm;
2599 		mode = append_file ? "a" : "w";
2600 
2601 		output = fopen(output_name, mode);
2602 		if (!output) {
2603 			perror("failed to create output file");
2604 			return -1;
2605 		}
2606 		if (!stat_config.json_output) {
2607 			clock_gettime(CLOCK_REALTIME, &tm);
2608 			fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2609 		}
2610 	} else if (output_fd > 0) {
2611 		mode = append_file ? "a" : "w";
2612 		output = fdopen(output_fd, mode);
2613 		if (!output) {
2614 			perror("Failed opening logfd");
2615 			return -errno;
2616 		}
2617 	}
2618 
2619 	if (stat_config.interval_clear && !isatty(fileno(output))) {
2620 		fprintf(stderr, "--interval-clear does not work with output\n");
2621 		parse_options_usage(stat_usage, stat_options, "o", 1);
2622 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2623 		parse_options_usage(NULL, stat_options, "interval-clear", 0);
2624 		return -1;
2625 	}
2626 
2627 	stat_config.output = output;
2628 
2629 	/*
2630 	 * let the spreadsheet do the pretty-printing
2631 	 */
2632 	if (stat_config.csv_output) {
2633 		/* User explicitly passed -B? */
2634 		if (big_num_opt == 1) {
2635 			fprintf(stderr, "-B option not supported with -x\n");
2636 			parse_options_usage(stat_usage, stat_options, "B", 1);
2637 			parse_options_usage(NULL, stat_options, "x", 1);
2638 			goto out;
2639 		} else /* Nope, so disable big number formatting */
2640 			stat_config.big_num = false;
2641 	} else if (big_num_opt == 0) /* User passed --no-big-num */
2642 		stat_config.big_num = false;
2643 
2644 	target.inherit = !stat_config.no_inherit;
2645 	err = target__validate(&target);
2646 	if (err) {
2647 		target__strerror(&target, err, errbuf, BUFSIZ);
2648 		pr_warning("%s\n", errbuf);
2649 	}
2650 
2651 	setup_system_wide(argc);
2652 
2653 	/*
2654 	 * Display user/system times only for single
2655 	 * run and when there's specified tracee.
2656 	 */
2657 	if ((stat_config.run_count == 1) && target__none(&target))
2658 		stat_config.ru_display = true;
2659 
2660 	if (stat_config.run_count < 0) {
2661 		pr_err("Run count must be a positive number\n");
2662 		parse_options_usage(stat_usage, stat_options, "r", 1);
2663 		goto out;
2664 	} else if (stat_config.run_count == 0) {
2665 		forever = true;
2666 		stat_config.run_count = 1;
2667 	}
2668 
2669 	if (stat_config.walltime_run_table) {
2670 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2671 		if (!stat_config.walltime_run) {
2672 			pr_err("failed to setup -r option");
2673 			goto out;
2674 		}
2675 	}
2676 
2677 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2678 		!target__has_task(&target)) {
2679 		if (!target.system_wide || target.cpu_list) {
2680 			fprintf(stderr, "The --per-thread option is only "
2681 				"available when monitoring via -p -t -a "
2682 				"options or only --per-thread.\n");
2683 			parse_options_usage(NULL, stat_options, "p", 1);
2684 			parse_options_usage(NULL, stat_options, "t", 1);
2685 			goto out;
2686 		}
2687 	}
2688 
2689 	/*
2690 	 * no_aggr, cgroup are for system-wide only
2691 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2692 	 */
2693 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2694 	      stat_config.aggr_mode != AGGR_THREAD) ||
2695 	     (nr_cgroups || stat_config.cgroup_list)) &&
2696 	    !target__has_cpu(&target)) {
2697 		fprintf(stderr, "both cgroup and no-aggregation "
2698 			"modes only available in system-wide mode\n");
2699 
2700 		parse_options_usage(stat_usage, stat_options, "G", 1);
2701 		parse_options_usage(NULL, stat_options, "A", 1);
2702 		parse_options_usage(NULL, stat_options, "a", 1);
2703 		parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2704 		goto out;
2705 	}
2706 
2707 	if (stat_config.iostat_run) {
2708 		status = iostat_prepare(evsel_list, &stat_config);
2709 		if (status)
2710 			goto out;
2711 		if (iostat_mode == IOSTAT_LIST) {
2712 			iostat_list(evsel_list, &stat_config);
2713 			goto out;
2714 		} else if (verbose > 0)
2715 			iostat_list(evsel_list, &stat_config);
2716 		if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2717 			target.system_wide = true;
2718 	}
2719 
2720 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2721 		target.per_thread = true;
2722 
2723 	stat_config.system_wide = target.system_wide;
2724 	if (target.cpu_list) {
2725 		stat_config.user_requested_cpu_list = strdup(target.cpu_list);
2726 		if (!stat_config.user_requested_cpu_list) {
2727 			status = -ENOMEM;
2728 			goto out;
2729 		}
2730 	}
2731 
2732 	/*
2733 	 * Metric parsing needs to be delayed as metrics may optimize events
2734 	 * knowing the target is system-wide.
2735 	 */
2736 	if (metrics) {
2737 		const char *pmu = parse_events_option_args.pmu_filter ?: "all";
2738 		int ret = metricgroup__parse_groups(evsel_list, pmu, metrics,
2739 						stat_config.metric_no_group,
2740 						stat_config.metric_no_merge,
2741 						stat_config.metric_no_threshold,
2742 						stat_config.user_requested_cpu_list,
2743 						stat_config.system_wide,
2744 						stat_config.hardware_aware_grouping,
2745 						&stat_config.metric_events);
2746 
2747 		zfree(&metrics);
2748 		if (ret) {
2749 			status = ret;
2750 			goto out;
2751 		}
2752 	}
2753 
2754 	if (add_default_events())
2755 		goto out;
2756 
2757 	if (stat_config.cgroup_list) {
2758 		if (nr_cgroups > 0) {
2759 			pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2760 			parse_options_usage(stat_usage, stat_options, "G", 1);
2761 			parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2762 			goto out;
2763 		}
2764 
2765 		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2766 					  &stat_config.metric_events, true) < 0) {
2767 			parse_options_usage(stat_usage, stat_options,
2768 					    "for-each-cgroup", 0);
2769 			goto out;
2770 		}
2771 	}
2772 
2773 	evlist__warn_user_requested_cpus(evsel_list, target.cpu_list);
2774 
2775 	if (evlist__create_maps(evsel_list, &target) < 0) {
2776 		if (target__has_task(&target)) {
2777 			pr_err("Problems finding threads of monitor\n");
2778 			parse_options_usage(stat_usage, stat_options, "p", 1);
2779 			parse_options_usage(NULL, stat_options, "t", 1);
2780 		} else if (target__has_cpu(&target)) {
2781 			perror("failed to parse CPUs map");
2782 			parse_options_usage(stat_usage, stat_options, "C", 1);
2783 			parse_options_usage(NULL, stat_options, "a", 1);
2784 		}
2785 		goto out;
2786 	}
2787 
2788 	evlist__check_cpu_maps(evsel_list);
2789 
2790 	/*
2791 	 * Initialize thread_map with comm names,
2792 	 * so we could print it out on output.
2793 	 */
2794 	if (stat_config.aggr_mode == AGGR_THREAD) {
2795 		thread_map__read_comms(evsel_list->core.threads);
2796 	}
2797 
2798 	if (stat_config.aggr_mode == AGGR_NODE)
2799 		cpu__setup_cpunode_map();
2800 
2801 	if (stat_config.times && interval)
2802 		interval_count = true;
2803 	else if (stat_config.times && !interval) {
2804 		pr_err("interval-count option should be used together with "
2805 				"interval-print.\n");
2806 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2807 		parse_options_usage(stat_usage, stat_options, "I", 1);
2808 		goto out;
2809 	}
2810 
2811 	if (timeout && timeout < 100) {
2812 		if (timeout < 10) {
2813 			pr_err("timeout must be >= 10ms.\n");
2814 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2815 			goto out;
2816 		} else
2817 			pr_warning("timeout < 100ms. "
2818 				   "The overhead percentage could be high in some cases. "
2819 				   "Please proceed with caution.\n");
2820 	}
2821 	if (timeout && interval) {
2822 		pr_err("timeout option is not supported with interval-print.\n");
2823 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2824 		parse_options_usage(stat_usage, stat_options, "I", 1);
2825 		goto out;
2826 	}
2827 
2828 	if (perf_stat_init_aggr_mode())
2829 		goto out;
2830 
2831 	if (evlist__alloc_stats(&stat_config, evsel_list, interval))
2832 		goto out;
2833 
2834 	/*
2835 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2836 	 * while avoiding that older tools show confusing messages.
2837 	 *
2838 	 * However for pipe sessions we need to keep it zero,
2839 	 * because script's perf_evsel__check_attr is triggered
2840 	 * by attr->sample_type != 0, and we can't run it on
2841 	 * stat sessions.
2842 	 */
2843 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2844 
2845 	/*
2846 	 * We dont want to block the signals - that would cause
2847 	 * child tasks to inherit that and Ctrl-C would not work.
2848 	 * What we want is for Ctrl-C to work in the exec()-ed
2849 	 * task, but being ignored by perf stat itself:
2850 	 */
2851 	atexit(sig_atexit);
2852 	if (!forever)
2853 		signal(SIGINT,  skip_signal);
2854 	signal(SIGCHLD, skip_signal);
2855 	signal(SIGALRM, skip_signal);
2856 	signal(SIGABRT, skip_signal);
2857 
2858 	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2859 		goto out;
2860 
2861 	/* Enable ignoring missing threads when -p option is defined. */
2862 	evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2863 	status = 0;
2864 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2865 		if (stat_config.run_count != 1 && verbose > 0)
2866 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2867 				run_idx + 1);
2868 
2869 		if (run_idx != 0)
2870 			evlist__reset_prev_raw_counts(evsel_list);
2871 
2872 		status = run_perf_stat(argc, argv, run_idx);
2873 		if (status == -1)
2874 			break;
2875 
2876 		if (forever && !interval) {
2877 			print_counters(NULL, argc, argv);
2878 			perf_stat__reset_stats();
2879 		}
2880 	}
2881 
2882 	if (!forever && status != -1 && (!interval || stat_config.summary)) {
2883 		if (stat_config.run_count > 1)
2884 			evlist__copy_res_stats(&stat_config, evsel_list);
2885 		print_counters(NULL, argc, argv);
2886 	}
2887 
2888 	evlist__finalize_ctlfd(evsel_list);
2889 
2890 	if (STAT_RECORD) {
2891 		/*
2892 		 * We synthesize the kernel mmap record just so that older tools
2893 		 * don't emit warnings about not being able to resolve symbols
2894 		 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2895 		 * a saner message about no samples being in the perf.data file.
2896 		 *
2897 		 * This also serves to suppress a warning about f_header.data.size == 0
2898 		 * in header.c at the moment 'perf stat record' gets introduced, which
2899 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2900 		 * records, but the need to suppress the kptr_restrict messages in older
2901 		 * tools remain  -acme
2902 		 */
2903 		int fd = perf_data__fd(&perf_stat.data);
2904 
2905 		err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2906 							 process_synthesized_event,
2907 							 &perf_stat.session->machines.host);
2908 		if (err) {
2909 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2910 				   "older tools may produce warnings about this file\n.");
2911 		}
2912 
2913 		if (!interval) {
2914 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2915 				pr_err("failed to write stat round event\n");
2916 		}
2917 
2918 		if (!perf_stat.data.is_pipe) {
2919 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2920 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2921 		}
2922 
2923 		evlist__close(evsel_list);
2924 		perf_session__delete(perf_stat.session);
2925 	}
2926 
2927 	perf_stat__exit_aggr_mode();
2928 	evlist__free_stats(evsel_list);
2929 out:
2930 	if (stat_config.iostat_run)
2931 		iostat_release(evsel_list);
2932 
2933 	zfree(&stat_config.walltime_run);
2934 	zfree(&stat_config.user_requested_cpu_list);
2935 
2936 	if (smi_cost && smi_reset)
2937 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2938 
2939 	evlist__delete(evsel_list);
2940 
2941 	metricgroup__rblist_exit(&stat_config.metric_events);
2942 	evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2943 
2944 	return status;
2945 }
2946