xref: /linux/tools/perf/builtin-stat.c (revision 55d0969c451159cff86949b38c39171cab962069)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * builtin-stat.c
4  *
5  * Builtin stat command: Give a precise performance counters summary
6  * overview about any workload, CPU or specific PID.
7  *
8  * Sample output:
9 
10    $ perf stat ./hackbench 10
11 
12   Time: 0.118
13 
14   Performance counter stats for './hackbench 10':
15 
16        1708.761321 task-clock                #   11.037 CPUs utilized
17             41,190 context-switches          #    0.024 M/sec
18              6,735 CPU-migrations            #    0.004 M/sec
19             17,318 page-faults               #    0.010 M/sec
20      5,205,202,243 cycles                    #    3.046 GHz
21      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
22      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
23      2,603,501,247 instructions              #    0.50  insns per cycle
24                                              #    1.48  stalled cycles per insn
25        484,357,498 branches                  #  283.455 M/sec
26          6,388,934 branch-misses             #    1.32% of all branches
27 
28         0.154822978  seconds time elapsed
29 
30  *
31  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32  *
33  * Improvements and fixes by:
34  *
35  *   Arjan van de Ven <arjan@linux.intel.com>
36  *   Yanmin Zhang <yanmin.zhang@intel.com>
37  *   Wu Fengguang <fengguang.wu@intel.com>
38  *   Mike Galbraith <efault@gmx.de>
39  *   Paul Mackerras <paulus@samba.org>
40  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
41  */
42 
43 #include "builtin.h"
44 #include "util/cgroup.h"
45 #include <subcmd/parse-options.h>
46 #include "util/parse-events.h"
47 #include "util/pmus.h"
48 #include "util/pmu.h"
49 #include "util/event.h"
50 #include "util/evlist.h"
51 #include "util/evsel.h"
52 #include "util/debug.h"
53 #include "util/color.h"
54 #include "util/stat.h"
55 #include "util/header.h"
56 #include "util/cpumap.h"
57 #include "util/thread_map.h"
58 #include "util/counts.h"
59 #include "util/topdown.h"
60 #include "util/session.h"
61 #include "util/tool.h"
62 #include "util/string2.h"
63 #include "util/metricgroup.h"
64 #include "util/synthetic-events.h"
65 #include "util/target.h"
66 #include "util/time-utils.h"
67 #include "util/top.h"
68 #include "util/affinity.h"
69 #include "util/pfm.h"
70 #include "util/bpf_counter.h"
71 #include "util/iostat.h"
72 #include "util/util.h"
73 #include "util/intel-tpebs.h"
74 #include "asm/bug.h"
75 
76 #include <linux/time64.h>
77 #include <linux/zalloc.h>
78 #include <api/fs/fs.h>
79 #include <errno.h>
80 #include <signal.h>
81 #include <stdlib.h>
82 #include <sys/prctl.h>
83 #include <inttypes.h>
84 #include <locale.h>
85 #include <math.h>
86 #include <sys/types.h>
87 #include <sys/stat.h>
88 #include <sys/wait.h>
89 #include <unistd.h>
90 #include <sys/time.h>
91 #include <sys/resource.h>
92 #include <linux/err.h>
93 
94 #include <linux/ctype.h>
95 #include <perf/evlist.h>
96 #include <internal/threadmap.h>
97 
98 #define DEFAULT_SEPARATOR	" "
99 #define FREEZE_ON_SMI_PATH	"devices/cpu/freeze_on_smi"
100 
101 static void print_counters(struct timespec *ts, int argc, const char **argv);
102 
103 static struct evlist	*evsel_list;
104 static struct parse_events_option_args parse_events_option_args = {
105 	.evlistp = &evsel_list,
106 };
107 
108 static bool all_counters_use_bpf = true;
109 
110 static struct target target = {
111 	.uid	= UINT_MAX,
112 };
113 
114 #define METRIC_ONLY_LEN 20
115 
116 static volatile sig_atomic_t	child_pid			= -1;
117 static int			detailed_run			=  0;
118 static bool			transaction_run;
119 static bool			topdown_run			= false;
120 static bool			smi_cost			= false;
121 static bool			smi_reset			= false;
122 static int			big_num_opt			=  -1;
123 static const char		*pre_cmd			= NULL;
124 static const char		*post_cmd			= NULL;
125 static bool			sync_run			= false;
126 static bool			forever				= false;
127 static bool			force_metric_only		= false;
128 static struct timespec		ref_time;
129 static bool			append_file;
130 static bool			interval_count;
131 static const char		*output_name;
132 static int			output_fd;
133 static char			*metrics;
134 
135 struct perf_stat {
136 	bool			 record;
137 	struct perf_data	 data;
138 	struct perf_session	*session;
139 	u64			 bytes_written;
140 	struct perf_tool	 tool;
141 	bool			 maps_allocated;
142 	struct perf_cpu_map	*cpus;
143 	struct perf_thread_map *threads;
144 	enum aggr_mode		 aggr_mode;
145 	u32			 aggr_level;
146 };
147 
148 static struct perf_stat		perf_stat;
149 #define STAT_RECORD		perf_stat.record
150 
151 static volatile sig_atomic_t done = 0;
152 
153 static struct perf_stat_config stat_config = {
154 	.aggr_mode		= AGGR_GLOBAL,
155 	.aggr_level		= MAX_CACHE_LVL + 1,
156 	.scale			= true,
157 	.unit_width		= 4, /* strlen("unit") */
158 	.run_count		= 1,
159 	.metric_only_len	= METRIC_ONLY_LEN,
160 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
161 	.ru_stats		= &ru_stats,
162 	.big_num		= true,
163 	.ctl_fd			= -1,
164 	.ctl_fd_ack		= -1,
165 	.iostat_run		= false,
166 };
167 
168 /* Options set from the command line. */
169 struct opt_aggr_mode {
170 	bool node, socket, die, cluster, cache, core, thread, no_aggr;
171 };
172 
173 /* Turn command line option into most generic aggregation mode setting. */
174 static enum aggr_mode opt_aggr_mode_to_aggr_mode(struct opt_aggr_mode *opt_mode)
175 {
176 	enum aggr_mode mode = AGGR_GLOBAL;
177 
178 	if (opt_mode->node)
179 		mode = AGGR_NODE;
180 	if (opt_mode->socket)
181 		mode = AGGR_SOCKET;
182 	if (opt_mode->die)
183 		mode = AGGR_DIE;
184 	if (opt_mode->cluster)
185 		mode = AGGR_CLUSTER;
186 	if (opt_mode->cache)
187 		mode = AGGR_CACHE;
188 	if (opt_mode->core)
189 		mode = AGGR_CORE;
190 	if (opt_mode->thread)
191 		mode = AGGR_THREAD;
192 	if (opt_mode->no_aggr)
193 		mode = AGGR_NONE;
194 	return mode;
195 }
196 
197 static void evlist__check_cpu_maps(struct evlist *evlist)
198 {
199 	struct evsel *evsel, *warned_leader = NULL;
200 
201 	evlist__for_each_entry(evlist, evsel) {
202 		struct evsel *leader = evsel__leader(evsel);
203 
204 		/* Check that leader matches cpus with each member. */
205 		if (leader == evsel)
206 			continue;
207 		if (perf_cpu_map__equal(leader->core.cpus, evsel->core.cpus))
208 			continue;
209 
210 		/* If there's mismatch disable the group and warn user. */
211 		if (warned_leader != leader) {
212 			char buf[200];
213 
214 			pr_warning("WARNING: grouped events cpus do not match.\n"
215 				"Events with CPUs not matching the leader will "
216 				"be removed from the group.\n");
217 			evsel__group_desc(leader, buf, sizeof(buf));
218 			pr_warning("  %s\n", buf);
219 			warned_leader = leader;
220 		}
221 		if (verbose > 0) {
222 			char buf[200];
223 
224 			cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
225 			pr_warning("     %s: %s\n", leader->name, buf);
226 			cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
227 			pr_warning("     %s: %s\n", evsel->name, buf);
228 		}
229 
230 		evsel__remove_from_group(evsel, leader);
231 	}
232 }
233 
234 static inline void diff_timespec(struct timespec *r, struct timespec *a,
235 				 struct timespec *b)
236 {
237 	r->tv_sec = a->tv_sec - b->tv_sec;
238 	if (a->tv_nsec < b->tv_nsec) {
239 		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
240 		r->tv_sec--;
241 	} else {
242 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
243 	}
244 }
245 
246 static void perf_stat__reset_stats(void)
247 {
248 	evlist__reset_stats(evsel_list);
249 	perf_stat__reset_shadow_stats();
250 }
251 
252 static int process_synthesized_event(const struct perf_tool *tool __maybe_unused,
253 				     union perf_event *event,
254 				     struct perf_sample *sample __maybe_unused,
255 				     struct machine *machine __maybe_unused)
256 {
257 	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
258 		pr_err("failed to write perf data, error: %m\n");
259 		return -1;
260 	}
261 
262 	perf_stat.bytes_written += event->header.size;
263 	return 0;
264 }
265 
266 static int write_stat_round_event(u64 tm, u64 type)
267 {
268 	return perf_event__synthesize_stat_round(NULL, tm, type,
269 						 process_synthesized_event,
270 						 NULL);
271 }
272 
273 #define WRITE_STAT_ROUND_EVENT(time, interval) \
274 	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
275 
276 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
277 
278 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
279 				   struct perf_counts_values *count)
280 {
281 	struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
282 	struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
283 
284 	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
285 					   process_synthesized_event, NULL);
286 }
287 
288 static int read_single_counter(struct evsel *counter, int cpu_map_idx, int thread)
289 {
290 	int err = evsel__read_counter(counter, cpu_map_idx, thread);
291 
292 	/*
293 	 * Reading user and system time will fail when the process
294 	 * terminates. Use the wait4 values in that case.
295 	 */
296 	if (err && cpu_map_idx == 0 &&
297 	    (evsel__tool_event(counter) == PERF_TOOL_USER_TIME ||
298 	     evsel__tool_event(counter) == PERF_TOOL_SYSTEM_TIME)) {
299 		u64 val, *start_time;
300 		struct perf_counts_values *count =
301 			perf_counts(counter->counts, cpu_map_idx, thread);
302 
303 		start_time = xyarray__entry(counter->start_times, cpu_map_idx, thread);
304 		if (evsel__tool_event(counter) == PERF_TOOL_USER_TIME)
305 			val = ru_stats.ru_utime_usec_stat.mean;
306 		else
307 			val = ru_stats.ru_stime_usec_stat.mean;
308 		count->ena = count->run = *start_time + val;
309 		count->val = val;
310 		return 0;
311 	}
312 	return err;
313 }
314 
315 /*
316  * Read out the results of a single counter:
317  * do not aggregate counts across CPUs in system-wide mode
318  */
319 static int read_counter_cpu(struct evsel *counter, int cpu_map_idx)
320 {
321 	int nthreads = perf_thread_map__nr(evsel_list->core.threads);
322 	int thread;
323 
324 	if (!counter->supported)
325 		return -ENOENT;
326 
327 	for (thread = 0; thread < nthreads; thread++) {
328 		struct perf_counts_values *count;
329 
330 		count = perf_counts(counter->counts, cpu_map_idx, thread);
331 
332 		/*
333 		 * The leader's group read loads data into its group members
334 		 * (via evsel__read_counter()) and sets their count->loaded.
335 		 */
336 		if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
337 		    read_single_counter(counter, cpu_map_idx, thread)) {
338 			counter->counts->scaled = -1;
339 			perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
340 			perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
341 			return -1;
342 		}
343 
344 		perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
345 
346 		if (STAT_RECORD) {
347 			if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
348 				pr_err("failed to write stat event\n");
349 				return -1;
350 			}
351 		}
352 
353 		if (verbose > 1) {
354 			fprintf(stat_config.output,
355 				"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
356 					evsel__name(counter),
357 					perf_cpu_map__cpu(evsel__cpus(counter),
358 							  cpu_map_idx).cpu,
359 					count->val, count->ena, count->run);
360 		}
361 	}
362 
363 	return 0;
364 }
365 
366 static int read_affinity_counters(void)
367 {
368 	struct evlist_cpu_iterator evlist_cpu_itr;
369 	struct affinity saved_affinity, *affinity;
370 
371 	if (all_counters_use_bpf)
372 		return 0;
373 
374 	if (!target__has_cpu(&target) || target__has_per_thread(&target))
375 		affinity = NULL;
376 	else if (affinity__setup(&saved_affinity) < 0)
377 		return -1;
378 	else
379 		affinity = &saved_affinity;
380 
381 	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
382 		struct evsel *counter = evlist_cpu_itr.evsel;
383 
384 		if (evsel__is_bpf(counter))
385 			continue;
386 
387 		if (!counter->err)
388 			counter->err = read_counter_cpu(counter, evlist_cpu_itr.cpu_map_idx);
389 	}
390 	if (affinity)
391 		affinity__cleanup(&saved_affinity);
392 
393 	return 0;
394 }
395 
396 static int read_bpf_map_counters(void)
397 {
398 	struct evsel *counter;
399 	int err;
400 
401 	evlist__for_each_entry(evsel_list, counter) {
402 		if (!evsel__is_bpf(counter))
403 			continue;
404 
405 		err = bpf_counter__read(counter);
406 		if (err)
407 			return err;
408 	}
409 	return 0;
410 }
411 
412 static int read_counters(void)
413 {
414 	if (!stat_config.stop_read_counter) {
415 		if (read_bpf_map_counters() ||
416 		    read_affinity_counters())
417 			return -1;
418 	}
419 	return 0;
420 }
421 
422 static void process_counters(void)
423 {
424 	struct evsel *counter;
425 
426 	evlist__for_each_entry(evsel_list, counter) {
427 		if (counter->err)
428 			pr_debug("failed to read counter %s\n", counter->name);
429 		if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
430 			pr_warning("failed to process counter %s\n", counter->name);
431 		counter->err = 0;
432 	}
433 
434 	perf_stat_merge_counters(&stat_config, evsel_list);
435 	perf_stat_process_percore(&stat_config, evsel_list);
436 }
437 
438 static void process_interval(void)
439 {
440 	struct timespec ts, rs;
441 
442 	clock_gettime(CLOCK_MONOTONIC, &ts);
443 	diff_timespec(&rs, &ts, &ref_time);
444 
445 	evlist__reset_aggr_stats(evsel_list);
446 
447 	if (read_counters() == 0)
448 		process_counters();
449 
450 	if (STAT_RECORD) {
451 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
452 			pr_err("failed to write stat round event\n");
453 	}
454 
455 	init_stats(&walltime_nsecs_stats);
456 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
457 	print_counters(&rs, 0, NULL);
458 }
459 
460 static bool handle_interval(unsigned int interval, int *times)
461 {
462 	if (interval) {
463 		process_interval();
464 		if (interval_count && !(--(*times)))
465 			return true;
466 	}
467 	return false;
468 }
469 
470 static int enable_counters(void)
471 {
472 	struct evsel *evsel;
473 	int err;
474 
475 	evlist__for_each_entry(evsel_list, evsel) {
476 		if (!evsel__is_bpf(evsel))
477 			continue;
478 
479 		err = bpf_counter__enable(evsel);
480 		if (err)
481 			return err;
482 	}
483 
484 	if (!target__enable_on_exec(&target)) {
485 		if (!all_counters_use_bpf)
486 			evlist__enable(evsel_list);
487 	}
488 	return 0;
489 }
490 
491 static void disable_counters(void)
492 {
493 	struct evsel *counter;
494 
495 	/*
496 	 * If we don't have tracee (attaching to task or cpu), counters may
497 	 * still be running. To get accurate group ratios, we must stop groups
498 	 * from counting before reading their constituent counters.
499 	 */
500 	if (!target__none(&target)) {
501 		evlist__for_each_entry(evsel_list, counter)
502 			bpf_counter__disable(counter);
503 		if (!all_counters_use_bpf)
504 			evlist__disable(evsel_list);
505 	}
506 }
507 
508 static volatile sig_atomic_t workload_exec_errno;
509 
510 /*
511  * evlist__prepare_workload will send a SIGUSR1
512  * if the fork fails, since we asked by setting its
513  * want_signal to true.
514  */
515 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
516 					void *ucontext __maybe_unused)
517 {
518 	workload_exec_errno = info->si_value.sival_int;
519 }
520 
521 static bool evsel__should_store_id(struct evsel *counter)
522 {
523 	return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
524 }
525 
526 static bool is_target_alive(struct target *_target,
527 			    struct perf_thread_map *threads)
528 {
529 	struct stat st;
530 	int i;
531 
532 	if (!target__has_task(_target))
533 		return true;
534 
535 	for (i = 0; i < threads->nr; i++) {
536 		char path[PATH_MAX];
537 
538 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
539 			  threads->map[i].pid);
540 
541 		if (!stat(path, &st))
542 			return true;
543 	}
544 
545 	return false;
546 }
547 
548 static void process_evlist(struct evlist *evlist, unsigned int interval)
549 {
550 	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
551 
552 	if (evlist__ctlfd_process(evlist, &cmd) > 0) {
553 		switch (cmd) {
554 		case EVLIST_CTL_CMD_ENABLE:
555 			fallthrough;
556 		case EVLIST_CTL_CMD_DISABLE:
557 			if (interval)
558 				process_interval();
559 			break;
560 		case EVLIST_CTL_CMD_SNAPSHOT:
561 		case EVLIST_CTL_CMD_ACK:
562 		case EVLIST_CTL_CMD_UNSUPPORTED:
563 		case EVLIST_CTL_CMD_EVLIST:
564 		case EVLIST_CTL_CMD_STOP:
565 		case EVLIST_CTL_CMD_PING:
566 		default:
567 			break;
568 		}
569 	}
570 }
571 
572 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
573 			int *time_to_sleep)
574 {
575 	int tts = *time_to_sleep;
576 	struct timespec time_diff;
577 
578 	diff_timespec(&time_diff, time_stop, time_start);
579 
580 	tts -= time_diff.tv_sec * MSEC_PER_SEC +
581 	       time_diff.tv_nsec / NSEC_PER_MSEC;
582 
583 	if (tts < 0)
584 		tts = 0;
585 
586 	*time_to_sleep = tts;
587 }
588 
589 static int dispatch_events(bool forks, int timeout, int interval, int *times)
590 {
591 	int child_exited = 0, status = 0;
592 	int time_to_sleep, sleep_time;
593 	struct timespec time_start, time_stop;
594 
595 	if (interval)
596 		sleep_time = interval;
597 	else if (timeout)
598 		sleep_time = timeout;
599 	else
600 		sleep_time = 1000;
601 
602 	time_to_sleep = sleep_time;
603 
604 	while (!done) {
605 		if (forks)
606 			child_exited = waitpid(child_pid, &status, WNOHANG);
607 		else
608 			child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
609 
610 		if (child_exited)
611 			break;
612 
613 		clock_gettime(CLOCK_MONOTONIC, &time_start);
614 		if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
615 			if (timeout || handle_interval(interval, times))
616 				break;
617 			time_to_sleep = sleep_time;
618 		} else { /* fd revent */
619 			process_evlist(evsel_list, interval);
620 			clock_gettime(CLOCK_MONOTONIC, &time_stop);
621 			compute_tts(&time_start, &time_stop, &time_to_sleep);
622 		}
623 	}
624 
625 	return status;
626 }
627 
628 enum counter_recovery {
629 	COUNTER_SKIP,
630 	COUNTER_RETRY,
631 	COUNTER_FATAL,
632 };
633 
634 static enum counter_recovery stat_handle_error(struct evsel *counter)
635 {
636 	char msg[BUFSIZ];
637 	/*
638 	 * PPC returns ENXIO for HW counters until 2.6.37
639 	 * (behavior changed with commit b0a873e).
640 	 */
641 	if (errno == EINVAL || errno == ENOSYS ||
642 	    errno == ENOENT || errno == EOPNOTSUPP ||
643 	    errno == ENXIO) {
644 		if (verbose > 0)
645 			ui__warning("%s event is not supported by the kernel.\n",
646 				    evsel__name(counter));
647 		counter->supported = false;
648 		/*
649 		 * errored is a sticky flag that means one of the counter's
650 		 * cpu event had a problem and needs to be reexamined.
651 		 */
652 		counter->errored = true;
653 
654 		if ((evsel__leader(counter) != counter) ||
655 		    !(counter->core.leader->nr_members > 1))
656 			return COUNTER_SKIP;
657 	} else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) {
658 		if (verbose > 0)
659 			ui__warning("%s\n", msg);
660 		return COUNTER_RETRY;
661 	} else if (target__has_per_thread(&target) &&
662 		   evsel_list->core.threads &&
663 		   evsel_list->core.threads->err_thread != -1) {
664 		/*
665 		 * For global --per-thread case, skip current
666 		 * error thread.
667 		 */
668 		if (!thread_map__remove(evsel_list->core.threads,
669 					evsel_list->core.threads->err_thread)) {
670 			evsel_list->core.threads->err_thread = -1;
671 			return COUNTER_RETRY;
672 		}
673 	} else if (counter->skippable) {
674 		if (verbose > 0)
675 			ui__warning("skipping event %s that kernel failed to open .\n",
676 				    evsel__name(counter));
677 		counter->supported = false;
678 		counter->errored = true;
679 		return COUNTER_SKIP;
680 	}
681 
682 	evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
683 	ui__error("%s\n", msg);
684 
685 	if (child_pid != -1)
686 		kill(child_pid, SIGTERM);
687 
688 	tpebs_delete();
689 
690 	return COUNTER_FATAL;
691 }
692 
693 static int __run_perf_stat(int argc, const char **argv, int run_idx)
694 {
695 	int interval = stat_config.interval;
696 	int times = stat_config.times;
697 	int timeout = stat_config.timeout;
698 	char msg[BUFSIZ];
699 	unsigned long long t0, t1;
700 	struct evsel *counter;
701 	size_t l;
702 	int status = 0;
703 	const bool forks = (argc > 0);
704 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
705 	struct evlist_cpu_iterator evlist_cpu_itr;
706 	struct affinity saved_affinity, *affinity = NULL;
707 	int err;
708 	bool second_pass = false;
709 
710 	if (forks) {
711 		if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
712 			perror("failed to prepare workload");
713 			return -1;
714 		}
715 		child_pid = evsel_list->workload.pid;
716 	}
717 
718 	if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
719 		if (affinity__setup(&saved_affinity) < 0)
720 			return -1;
721 		affinity = &saved_affinity;
722 	}
723 
724 	evlist__for_each_entry(evsel_list, counter) {
725 		counter->reset_group = false;
726 		if (bpf_counter__load(counter, &target))
727 			return -1;
728 		if (!(evsel__is_bperf(counter)))
729 			all_counters_use_bpf = false;
730 	}
731 
732 	evlist__reset_aggr_stats(evsel_list);
733 
734 	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
735 		counter = evlist_cpu_itr.evsel;
736 
737 		/*
738 		 * bperf calls evsel__open_per_cpu() in bperf__load(), so
739 		 * no need to call it again here.
740 		 */
741 		if (target.use_bpf)
742 			break;
743 
744 		if (counter->reset_group || counter->errored)
745 			continue;
746 		if (evsel__is_bperf(counter))
747 			continue;
748 try_again:
749 		if (create_perf_stat_counter(counter, &stat_config, &target,
750 					     evlist_cpu_itr.cpu_map_idx) < 0) {
751 
752 			/*
753 			 * Weak group failed. We cannot just undo this here
754 			 * because earlier CPUs might be in group mode, and the kernel
755 			 * doesn't support mixing group and non group reads. Defer
756 			 * it to later.
757 			 * Don't close here because we're in the wrong affinity.
758 			 */
759 			if ((errno == EINVAL || errno == EBADF) &&
760 				evsel__leader(counter) != counter &&
761 				counter->weak_group) {
762 				evlist__reset_weak_group(evsel_list, counter, false);
763 				assert(counter->reset_group);
764 				second_pass = true;
765 				continue;
766 			}
767 
768 			switch (stat_handle_error(counter)) {
769 			case COUNTER_FATAL:
770 				return -1;
771 			case COUNTER_RETRY:
772 				goto try_again;
773 			case COUNTER_SKIP:
774 				continue;
775 			default:
776 				break;
777 			}
778 
779 		}
780 		counter->supported = true;
781 	}
782 
783 	if (second_pass) {
784 		/*
785 		 * Now redo all the weak group after closing them,
786 		 * and also close errored counters.
787 		 */
788 
789 		/* First close errored or weak retry */
790 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
791 			counter = evlist_cpu_itr.evsel;
792 
793 			if (!counter->reset_group && !counter->errored)
794 				continue;
795 
796 			perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
797 		}
798 		/* Now reopen weak */
799 		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
800 			counter = evlist_cpu_itr.evsel;
801 
802 			if (!counter->reset_group)
803 				continue;
804 try_again_reset:
805 			pr_debug2("reopening weak %s\n", evsel__name(counter));
806 			if (create_perf_stat_counter(counter, &stat_config, &target,
807 						     evlist_cpu_itr.cpu_map_idx) < 0) {
808 
809 				switch (stat_handle_error(counter)) {
810 				case COUNTER_FATAL:
811 					return -1;
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 
825 	evlist__for_each_entry(evsel_list, counter) {
826 		if (!counter->supported) {
827 			perf_evsel__free_fd(&counter->core);
828 			continue;
829 		}
830 
831 		l = strlen(counter->unit);
832 		if (l > stat_config.unit_width)
833 			stat_config.unit_width = l;
834 
835 		if (evsel__should_store_id(counter) &&
836 		    evsel__store_ids(counter, evsel_list))
837 			return -1;
838 	}
839 
840 	if (evlist__apply_filters(evsel_list, &counter, &target)) {
841 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
842 			counter->filter, evsel__name(counter), errno,
843 			str_error_r(errno, msg, sizeof(msg)));
844 		return -1;
845 	}
846 
847 	if (STAT_RECORD) {
848 		int fd = perf_data__fd(&perf_stat.data);
849 
850 		if (is_pipe) {
851 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
852 		} else {
853 			err = perf_session__write_header(perf_stat.session, evsel_list,
854 							 fd, false);
855 		}
856 
857 		if (err < 0)
858 			return err;
859 
860 		err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
861 							 process_synthesized_event, is_pipe);
862 		if (err < 0)
863 			return err;
864 	}
865 
866 	if (target.initial_delay) {
867 		pr_info(EVLIST_DISABLED_MSG);
868 	} else {
869 		err = enable_counters();
870 		if (err)
871 			return -1;
872 	}
873 
874 	/* Exec the command, if any */
875 	if (forks)
876 		evlist__start_workload(evsel_list);
877 
878 	if (target.initial_delay > 0) {
879 		usleep(target.initial_delay * USEC_PER_MSEC);
880 		err = enable_counters();
881 		if (err)
882 			return -1;
883 
884 		pr_info(EVLIST_ENABLED_MSG);
885 	}
886 
887 	t0 = rdclock();
888 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
889 
890 	if (forks) {
891 		if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
892 			status = dispatch_events(forks, timeout, interval, &times);
893 		if (child_pid != -1) {
894 			if (timeout)
895 				kill(child_pid, SIGTERM);
896 			wait4(child_pid, &status, 0, &stat_config.ru_data);
897 		}
898 
899 		if (workload_exec_errno) {
900 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
901 			pr_err("Workload failed: %s\n", emsg);
902 			return -1;
903 		}
904 
905 		if (WIFSIGNALED(status))
906 			psignal(WTERMSIG(status), argv[0]);
907 	} else {
908 		status = dispatch_events(forks, timeout, interval, &times);
909 	}
910 
911 	disable_counters();
912 
913 	t1 = rdclock();
914 
915 	if (stat_config.walltime_run_table)
916 		stat_config.walltime_run[run_idx] = t1 - t0;
917 
918 	if (interval && stat_config.summary) {
919 		stat_config.interval = 0;
920 		stat_config.stop_read_counter = true;
921 		init_stats(&walltime_nsecs_stats);
922 		update_stats(&walltime_nsecs_stats, t1 - t0);
923 
924 		evlist__copy_prev_raw_counts(evsel_list);
925 		evlist__reset_prev_raw_counts(evsel_list);
926 		evlist__reset_aggr_stats(evsel_list);
927 	} else {
928 		update_stats(&walltime_nsecs_stats, t1 - t0);
929 		update_rusage_stats(&ru_stats, &stat_config.ru_data);
930 	}
931 
932 	/*
933 	 * Closing a group leader splits the group, and as we only disable
934 	 * group leaders, results in remaining events becoming enabled. To
935 	 * avoid arbitrary skew, we must read all counters before closing any
936 	 * group leaders.
937 	 */
938 	if (read_counters() == 0)
939 		process_counters();
940 
941 	/*
942 	 * We need to keep evsel_list alive, because it's processed
943 	 * later the evsel_list will be closed after.
944 	 */
945 	if (!STAT_RECORD)
946 		evlist__close(evsel_list);
947 
948 	return WEXITSTATUS(status);
949 }
950 
951 static int run_perf_stat(int argc, const char **argv, int run_idx)
952 {
953 	int ret;
954 
955 	if (pre_cmd) {
956 		ret = system(pre_cmd);
957 		if (ret)
958 			return ret;
959 	}
960 
961 	if (sync_run)
962 		sync();
963 
964 	ret = __run_perf_stat(argc, argv, run_idx);
965 	if (ret)
966 		return ret;
967 
968 	if (post_cmd) {
969 		ret = system(post_cmd);
970 		if (ret)
971 			return ret;
972 	}
973 
974 	return ret;
975 }
976 
977 static void print_counters(struct timespec *ts, int argc, const char **argv)
978 {
979 	/* Do not print anything if we record to the pipe. */
980 	if (STAT_RECORD && perf_stat.data.is_pipe)
981 		return;
982 	if (quiet)
983 		return;
984 
985 	evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
986 }
987 
988 static volatile sig_atomic_t signr = -1;
989 
990 static void skip_signal(int signo)
991 {
992 	if ((child_pid == -1) || stat_config.interval)
993 		done = 1;
994 
995 	signr = signo;
996 	/*
997 	 * render child_pid harmless
998 	 * won't send SIGTERM to a random
999 	 * process in case of race condition
1000 	 * and fast PID recycling
1001 	 */
1002 	child_pid = -1;
1003 }
1004 
1005 static void sig_atexit(void)
1006 {
1007 	sigset_t set, oset;
1008 
1009 	/*
1010 	 * avoid race condition with SIGCHLD handler
1011 	 * in skip_signal() which is modifying child_pid
1012 	 * goal is to avoid send SIGTERM to a random
1013 	 * process
1014 	 */
1015 	sigemptyset(&set);
1016 	sigaddset(&set, SIGCHLD);
1017 	sigprocmask(SIG_BLOCK, &set, &oset);
1018 
1019 	if (child_pid != -1)
1020 		kill(child_pid, SIGTERM);
1021 
1022 	sigprocmask(SIG_SETMASK, &oset, NULL);
1023 
1024 	if (signr == -1)
1025 		return;
1026 
1027 	signal(signr, SIG_DFL);
1028 	kill(getpid(), signr);
1029 }
1030 
1031 void perf_stat__set_big_num(int set)
1032 {
1033 	stat_config.big_num = (set != 0);
1034 }
1035 
1036 void perf_stat__set_no_csv_summary(int set)
1037 {
1038 	stat_config.no_csv_summary = (set != 0);
1039 }
1040 
1041 static int stat__set_big_num(const struct option *opt __maybe_unused,
1042 			     const char *s __maybe_unused, int unset)
1043 {
1044 	big_num_opt = unset ? 0 : 1;
1045 	perf_stat__set_big_num(!unset);
1046 	return 0;
1047 }
1048 
1049 static int enable_metric_only(const struct option *opt __maybe_unused,
1050 			      const char *s __maybe_unused, int unset)
1051 {
1052 	force_metric_only = true;
1053 	stat_config.metric_only = !unset;
1054 	return 0;
1055 }
1056 
1057 static int append_metric_groups(const struct option *opt __maybe_unused,
1058 			       const char *str,
1059 			       int unset __maybe_unused)
1060 {
1061 	if (metrics) {
1062 		char *tmp;
1063 
1064 		if (asprintf(&tmp, "%s,%s", metrics, str) < 0)
1065 			return -ENOMEM;
1066 		free(metrics);
1067 		metrics = tmp;
1068 	} else {
1069 		metrics = strdup(str);
1070 		if (!metrics)
1071 			return -ENOMEM;
1072 	}
1073 	return 0;
1074 }
1075 
1076 static int parse_control_option(const struct option *opt,
1077 				const char *str,
1078 				int unset __maybe_unused)
1079 {
1080 	struct perf_stat_config *config = opt->value;
1081 
1082 	return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1083 }
1084 
1085 static int parse_stat_cgroups(const struct option *opt,
1086 			      const char *str, int unset)
1087 {
1088 	if (stat_config.cgroup_list) {
1089 		pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1090 		return -1;
1091 	}
1092 
1093 	return parse_cgroups(opt, str, unset);
1094 }
1095 
1096 static int parse_cputype(const struct option *opt,
1097 			     const char *str,
1098 			     int unset __maybe_unused)
1099 {
1100 	const struct perf_pmu *pmu;
1101 	struct evlist *evlist = *(struct evlist **)opt->value;
1102 
1103 	if (!list_empty(&evlist->core.entries)) {
1104 		fprintf(stderr, "Must define cputype before events/metrics\n");
1105 		return -1;
1106 	}
1107 
1108 	pmu = perf_pmus__pmu_for_pmu_filter(str);
1109 	if (!pmu) {
1110 		fprintf(stderr, "--cputype %s is not supported!\n", str);
1111 		return -1;
1112 	}
1113 	parse_events_option_args.pmu_filter = pmu->name;
1114 
1115 	return 0;
1116 }
1117 
1118 static int parse_cache_level(const struct option *opt,
1119 			     const char *str,
1120 			     int unset __maybe_unused)
1121 {
1122 	int level;
1123 	struct opt_aggr_mode *opt_aggr_mode = (struct opt_aggr_mode *)opt->value;
1124 	u32 *aggr_level = (u32 *)opt->data;
1125 
1126 	/*
1127 	 * If no string is specified, aggregate based on the topology of
1128 	 * Last Level Cache (LLC). Since the LLC level can change from
1129 	 * architecture to architecture, set level greater than
1130 	 * MAX_CACHE_LVL which will be interpreted as LLC.
1131 	 */
1132 	if (str == NULL) {
1133 		level = MAX_CACHE_LVL + 1;
1134 		goto out;
1135 	}
1136 
1137 	/*
1138 	 * The format to specify cache level is LX or lX where X is the
1139 	 * cache level.
1140 	 */
1141 	if (strlen(str) != 2 || (str[0] != 'l' && str[0] != 'L')) {
1142 		pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1143 		       MAX_CACHE_LVL,
1144 		       MAX_CACHE_LVL);
1145 		return -EINVAL;
1146 	}
1147 
1148 	level = atoi(&str[1]);
1149 	if (level < 1) {
1150 		pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1151 		       MAX_CACHE_LVL,
1152 		       MAX_CACHE_LVL);
1153 		return -EINVAL;
1154 	}
1155 
1156 	if (level > MAX_CACHE_LVL) {
1157 		pr_err("perf only supports max cache level of %d.\n"
1158 		       "Consider increasing MAX_CACHE_LVL\n", MAX_CACHE_LVL);
1159 		return -EINVAL;
1160 	}
1161 out:
1162 	opt_aggr_mode->cache = true;
1163 	*aggr_level = level;
1164 	return 0;
1165 }
1166 
1167 /**
1168  * Calculate the cache instance ID from the map in
1169  * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1170  * Cache instance ID is the first CPU reported in the shared_cpu_list file.
1171  */
1172 static int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map)
1173 {
1174 	int id;
1175 	struct perf_cpu_map *cpu_map = perf_cpu_map__new(map);
1176 
1177 	/*
1178 	 * If the map contains no CPU, consider the current CPU to
1179 	 * be the first online CPU in the cache domain else use the
1180 	 * first online CPU of the cache domain as the ID.
1181 	 */
1182 	id = perf_cpu_map__min(cpu_map).cpu;
1183 	if (id == -1)
1184 		id = cpu.cpu;
1185 
1186 	/* Free the perf_cpu_map used to find the cache ID */
1187 	perf_cpu_map__put(cpu_map);
1188 
1189 	return id;
1190 }
1191 
1192 /**
1193  * cpu__get_cache_id - Returns 0 if successful in populating the
1194  * cache level and cache id. Cache level is read from
1195  * /sys/devices/system/cpu/cpuX/cache/indexY/level where as cache instance ID
1196  * is the first CPU reported by
1197  * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1198  */
1199 static int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache)
1200 {
1201 	int ret = 0;
1202 	u32 cache_level = stat_config.aggr_level;
1203 	struct cpu_cache_level caches[MAX_CACHE_LVL];
1204 	u32 i = 0, caches_cnt = 0;
1205 
1206 	cache->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1207 	cache->cache = -1;
1208 
1209 	ret = build_caches_for_cpu(cpu.cpu, caches, &caches_cnt);
1210 	if (ret) {
1211 		/*
1212 		 * If caches_cnt is not 0, cpu_cache_level data
1213 		 * was allocated when building the topology.
1214 		 * Free the allocated data before returning.
1215 		 */
1216 		if (caches_cnt)
1217 			goto free_caches;
1218 
1219 		return ret;
1220 	}
1221 
1222 	if (!caches_cnt)
1223 		return -1;
1224 
1225 	/*
1226 	 * Save the data for the highest level if no
1227 	 * level was specified by the user.
1228 	 */
1229 	if (cache_level > MAX_CACHE_LVL) {
1230 		int max_level_index = 0;
1231 
1232 		for (i = 1; i < caches_cnt; ++i) {
1233 			if (caches[i].level > caches[max_level_index].level)
1234 				max_level_index = i;
1235 		}
1236 
1237 		cache->cache_lvl = caches[max_level_index].level;
1238 		cache->cache = cpu__get_cache_id_from_map(cpu, caches[max_level_index].map);
1239 
1240 		/* Reset i to 0 to free entire caches[] */
1241 		i = 0;
1242 		goto free_caches;
1243 	}
1244 
1245 	for (i = 0; i < caches_cnt; ++i) {
1246 		if (caches[i].level == cache_level) {
1247 			cache->cache_lvl = cache_level;
1248 			cache->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1249 		}
1250 
1251 		cpu_cache_level__free(&caches[i]);
1252 	}
1253 
1254 free_caches:
1255 	/*
1256 	 * Free all the allocated cpu_cache_level data.
1257 	 */
1258 	while (i < caches_cnt)
1259 		cpu_cache_level__free(&caches[i++]);
1260 
1261 	return ret;
1262 }
1263 
1264 /**
1265  * aggr_cpu_id__cache - Create an aggr_cpu_id with cache instache ID, cache
1266  * level, die and socket populated with the cache instache ID, cache level,
1267  * die and socket for cpu. The function signature is compatible with
1268  * aggr_cpu_id_get_t.
1269  */
1270 static struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data)
1271 {
1272 	int ret;
1273 	struct aggr_cpu_id id;
1274 	struct perf_cache cache;
1275 
1276 	id = aggr_cpu_id__die(cpu, data);
1277 	if (aggr_cpu_id__is_empty(&id))
1278 		return id;
1279 
1280 	ret = cpu__get_cache_details(cpu, &cache);
1281 	if (ret)
1282 		return id;
1283 
1284 	id.cache_lvl = cache.cache_lvl;
1285 	id.cache = cache.cache;
1286 	return id;
1287 }
1288 
1289 static const char *const aggr_mode__string[] = {
1290 	[AGGR_CORE] = "core",
1291 	[AGGR_CACHE] = "cache",
1292 	[AGGR_CLUSTER] = "cluster",
1293 	[AGGR_DIE] = "die",
1294 	[AGGR_GLOBAL] = "global",
1295 	[AGGR_NODE] = "node",
1296 	[AGGR_NONE] = "none",
1297 	[AGGR_SOCKET] = "socket",
1298 	[AGGR_THREAD] = "thread",
1299 	[AGGR_UNSET] = "unset",
1300 };
1301 
1302 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1303 						struct perf_cpu cpu)
1304 {
1305 	return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1306 }
1307 
1308 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1309 					     struct perf_cpu cpu)
1310 {
1311 	return aggr_cpu_id__die(cpu, /*data=*/NULL);
1312 }
1313 
1314 static struct aggr_cpu_id perf_stat__get_cache_id(struct perf_stat_config *config __maybe_unused,
1315 						  struct perf_cpu cpu)
1316 {
1317 	return aggr_cpu_id__cache(cpu, /*data=*/NULL);
1318 }
1319 
1320 static struct aggr_cpu_id perf_stat__get_cluster(struct perf_stat_config *config __maybe_unused,
1321 						 struct perf_cpu cpu)
1322 {
1323 	return aggr_cpu_id__cluster(cpu, /*data=*/NULL);
1324 }
1325 
1326 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1327 					      struct perf_cpu cpu)
1328 {
1329 	return aggr_cpu_id__core(cpu, /*data=*/NULL);
1330 }
1331 
1332 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1333 					      struct perf_cpu cpu)
1334 {
1335 	return aggr_cpu_id__node(cpu, /*data=*/NULL);
1336 }
1337 
1338 static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused,
1339 						struct perf_cpu cpu)
1340 {
1341 	return aggr_cpu_id__global(cpu, /*data=*/NULL);
1342 }
1343 
1344 static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused,
1345 					     struct perf_cpu cpu)
1346 {
1347 	return aggr_cpu_id__cpu(cpu, /*data=*/NULL);
1348 }
1349 
1350 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1351 					      aggr_get_id_t get_id, struct perf_cpu cpu)
1352 {
1353 	struct aggr_cpu_id id;
1354 
1355 	/* per-process mode - should use global aggr mode */
1356 	if (cpu.cpu == -1)
1357 		return get_id(config, cpu);
1358 
1359 	if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1360 		config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1361 
1362 	id = config->cpus_aggr_map->map[cpu.cpu];
1363 	return id;
1364 }
1365 
1366 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1367 						       struct perf_cpu cpu)
1368 {
1369 	return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1370 }
1371 
1372 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1373 						    struct perf_cpu cpu)
1374 {
1375 	return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1376 }
1377 
1378 static struct aggr_cpu_id perf_stat__get_cluster_cached(struct perf_stat_config *config,
1379 							struct perf_cpu cpu)
1380 {
1381 	return perf_stat__get_aggr(config, perf_stat__get_cluster, cpu);
1382 }
1383 
1384 static struct aggr_cpu_id perf_stat__get_cache_id_cached(struct perf_stat_config *config,
1385 							 struct perf_cpu cpu)
1386 {
1387 	return perf_stat__get_aggr(config, perf_stat__get_cache_id, cpu);
1388 }
1389 
1390 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1391 						     struct perf_cpu cpu)
1392 {
1393 	return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1394 }
1395 
1396 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1397 						     struct perf_cpu cpu)
1398 {
1399 	return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1400 }
1401 
1402 static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config,
1403 						       struct perf_cpu cpu)
1404 {
1405 	return perf_stat__get_aggr(config, perf_stat__get_global, cpu);
1406 }
1407 
1408 static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config,
1409 						    struct perf_cpu cpu)
1410 {
1411 	return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu);
1412 }
1413 
1414 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1415 {
1416 	switch (aggr_mode) {
1417 	case AGGR_SOCKET:
1418 		return aggr_cpu_id__socket;
1419 	case AGGR_DIE:
1420 		return aggr_cpu_id__die;
1421 	case AGGR_CLUSTER:
1422 		return aggr_cpu_id__cluster;
1423 	case AGGR_CACHE:
1424 		return aggr_cpu_id__cache;
1425 	case AGGR_CORE:
1426 		return aggr_cpu_id__core;
1427 	case AGGR_NODE:
1428 		return aggr_cpu_id__node;
1429 	case AGGR_NONE:
1430 		return aggr_cpu_id__cpu;
1431 	case AGGR_GLOBAL:
1432 		return aggr_cpu_id__global;
1433 	case AGGR_THREAD:
1434 	case AGGR_UNSET:
1435 	case AGGR_MAX:
1436 	default:
1437 		return NULL;
1438 	}
1439 }
1440 
1441 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1442 {
1443 	switch (aggr_mode) {
1444 	case AGGR_SOCKET:
1445 		return perf_stat__get_socket_cached;
1446 	case AGGR_DIE:
1447 		return perf_stat__get_die_cached;
1448 	case AGGR_CLUSTER:
1449 		return perf_stat__get_cluster_cached;
1450 	case AGGR_CACHE:
1451 		return perf_stat__get_cache_id_cached;
1452 	case AGGR_CORE:
1453 		return perf_stat__get_core_cached;
1454 	case AGGR_NODE:
1455 		return perf_stat__get_node_cached;
1456 	case AGGR_NONE:
1457 		return perf_stat__get_cpu_cached;
1458 	case AGGR_GLOBAL:
1459 		return perf_stat__get_global_cached;
1460 	case AGGR_THREAD:
1461 	case AGGR_UNSET:
1462 	case AGGR_MAX:
1463 	default:
1464 		return NULL;
1465 	}
1466 }
1467 
1468 static int perf_stat_init_aggr_mode(void)
1469 {
1470 	int nr;
1471 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1472 
1473 	if (get_id) {
1474 		bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1475 		stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1476 							 get_id, /*data=*/NULL, needs_sort);
1477 		if (!stat_config.aggr_map) {
1478 			pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]);
1479 			return -1;
1480 		}
1481 		stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1482 	}
1483 
1484 	if (stat_config.aggr_mode == AGGR_THREAD) {
1485 		nr = perf_thread_map__nr(evsel_list->core.threads);
1486 		stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1487 		if (stat_config.aggr_map == NULL)
1488 			return -ENOMEM;
1489 
1490 		for (int s = 0; s < nr; s++) {
1491 			struct aggr_cpu_id id = aggr_cpu_id__empty();
1492 
1493 			id.thread_idx = s;
1494 			stat_config.aggr_map->map[s] = id;
1495 		}
1496 		return 0;
1497 	}
1498 
1499 	/*
1500 	 * The evsel_list->cpus is the base we operate on,
1501 	 * taking the highest cpu number to be the size of
1502 	 * the aggregation translate cpumap.
1503 	 */
1504 	if (!perf_cpu_map__is_any_cpu_or_is_empty(evsel_list->core.user_requested_cpus))
1505 		nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1506 	else
1507 		nr = 0;
1508 	stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1509 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1510 }
1511 
1512 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1513 {
1514 	free(map);
1515 }
1516 
1517 static void perf_stat__exit_aggr_mode(void)
1518 {
1519 	cpu_aggr_map__delete(stat_config.aggr_map);
1520 	cpu_aggr_map__delete(stat_config.cpus_aggr_map);
1521 	stat_config.aggr_map = NULL;
1522 	stat_config.cpus_aggr_map = NULL;
1523 }
1524 
1525 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1526 {
1527 	struct perf_env *env = data;
1528 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1529 
1530 	if (cpu.cpu != -1)
1531 		id.socket = env->cpu[cpu.cpu].socket_id;
1532 
1533 	return id;
1534 }
1535 
1536 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1537 {
1538 	struct perf_env *env = data;
1539 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1540 
1541 	if (cpu.cpu != -1) {
1542 		/*
1543 		 * die_id is relative to socket, so start
1544 		 * with the socket ID and then add die to
1545 		 * make a unique ID.
1546 		 */
1547 		id.socket = env->cpu[cpu.cpu].socket_id;
1548 		id.die = env->cpu[cpu.cpu].die_id;
1549 	}
1550 
1551 	return id;
1552 }
1553 
1554 static void perf_env__get_cache_id_for_cpu(struct perf_cpu cpu, struct perf_env *env,
1555 					   u32 cache_level, struct aggr_cpu_id *id)
1556 {
1557 	int i;
1558 	int caches_cnt = env->caches_cnt;
1559 	struct cpu_cache_level *caches = env->caches;
1560 
1561 	id->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1562 	id->cache = -1;
1563 
1564 	if (!caches_cnt)
1565 		return;
1566 
1567 	for (i = caches_cnt - 1; i > -1; --i) {
1568 		struct perf_cpu_map *cpu_map;
1569 		int map_contains_cpu;
1570 
1571 		/*
1572 		 * If user has not specified a level, find the fist level with
1573 		 * the cpu in the map. Since building the map is expensive, do
1574 		 * this only if levels match.
1575 		 */
1576 		if (cache_level <= MAX_CACHE_LVL && caches[i].level != cache_level)
1577 			continue;
1578 
1579 		cpu_map = perf_cpu_map__new(caches[i].map);
1580 		map_contains_cpu = perf_cpu_map__idx(cpu_map, cpu);
1581 		perf_cpu_map__put(cpu_map);
1582 
1583 		if (map_contains_cpu != -1) {
1584 			id->cache_lvl = caches[i].level;
1585 			id->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1586 			return;
1587 		}
1588 	}
1589 }
1590 
1591 static struct aggr_cpu_id perf_env__get_cache_aggr_by_cpu(struct perf_cpu cpu,
1592 							  void *data)
1593 {
1594 	struct perf_env *env = data;
1595 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1596 
1597 	if (cpu.cpu != -1) {
1598 		u32 cache_level = (perf_stat.aggr_level) ?: stat_config.aggr_level;
1599 
1600 		id.socket = env->cpu[cpu.cpu].socket_id;
1601 		id.die = env->cpu[cpu.cpu].die_id;
1602 		perf_env__get_cache_id_for_cpu(cpu, env, cache_level, &id);
1603 	}
1604 
1605 	return id;
1606 }
1607 
1608 static struct aggr_cpu_id perf_env__get_cluster_aggr_by_cpu(struct perf_cpu cpu,
1609 							    void *data)
1610 {
1611 	struct perf_env *env = data;
1612 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1613 
1614 	if (cpu.cpu != -1) {
1615 		id.socket = env->cpu[cpu.cpu].socket_id;
1616 		id.die = env->cpu[cpu.cpu].die_id;
1617 		id.cluster = env->cpu[cpu.cpu].cluster_id;
1618 	}
1619 
1620 	return id;
1621 }
1622 
1623 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, 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 		/*
1630 		 * core_id is relative to socket, die and cluster, we need a
1631 		 * global id. So we set socket, die id, cluster id and core id.
1632 		 */
1633 		id.socket = env->cpu[cpu.cpu].socket_id;
1634 		id.die = env->cpu[cpu.cpu].die_id;
1635 		id.cluster = env->cpu[cpu.cpu].cluster_id;
1636 		id.core = env->cpu[cpu.cpu].core_id;
1637 	}
1638 
1639 	return id;
1640 }
1641 
1642 static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data)
1643 {
1644 	struct perf_env *env = data;
1645 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1646 
1647 	if (cpu.cpu != -1) {
1648 		/*
1649 		 * core_id is relative to socket and die,
1650 		 * we need a global id. So we set
1651 		 * socket, die id and core id
1652 		 */
1653 		id.socket = env->cpu[cpu.cpu].socket_id;
1654 		id.die = env->cpu[cpu.cpu].die_id;
1655 		id.core = env->cpu[cpu.cpu].core_id;
1656 		id.cpu = cpu;
1657 	}
1658 
1659 	return id;
1660 }
1661 
1662 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1663 {
1664 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1665 
1666 	id.node = perf_env__numa_node(data, cpu);
1667 	return id;
1668 }
1669 
1670 static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused,
1671 							   void *data __maybe_unused)
1672 {
1673 	struct aggr_cpu_id id = aggr_cpu_id__empty();
1674 
1675 	/* it always aggregates to the cpu 0 */
1676 	id.cpu = (struct perf_cpu){ .cpu = 0 };
1677 	return id;
1678 }
1679 
1680 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1681 						     struct perf_cpu cpu)
1682 {
1683 	return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1684 }
1685 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1686 						  struct perf_cpu cpu)
1687 {
1688 	return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1689 }
1690 
1691 static struct aggr_cpu_id perf_stat__get_cluster_file(struct perf_stat_config *config __maybe_unused,
1692 						      struct perf_cpu cpu)
1693 {
1694 	return perf_env__get_cluster_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1695 }
1696 
1697 static struct aggr_cpu_id perf_stat__get_cache_file(struct perf_stat_config *config __maybe_unused,
1698 						    struct perf_cpu cpu)
1699 {
1700 	return perf_env__get_cache_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1701 }
1702 
1703 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1704 						   struct perf_cpu cpu)
1705 {
1706 	return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1707 }
1708 
1709 static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused,
1710 						  struct perf_cpu cpu)
1711 {
1712 	return perf_env__get_cpu_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1713 }
1714 
1715 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1716 						   struct perf_cpu cpu)
1717 {
1718 	return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1719 }
1720 
1721 static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused,
1722 						     struct perf_cpu cpu)
1723 {
1724 	return perf_env__get_global_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1725 }
1726 
1727 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1728 {
1729 	switch (aggr_mode) {
1730 	case AGGR_SOCKET:
1731 		return perf_env__get_socket_aggr_by_cpu;
1732 	case AGGR_DIE:
1733 		return perf_env__get_die_aggr_by_cpu;
1734 	case AGGR_CLUSTER:
1735 		return perf_env__get_cluster_aggr_by_cpu;
1736 	case AGGR_CACHE:
1737 		return perf_env__get_cache_aggr_by_cpu;
1738 	case AGGR_CORE:
1739 		return perf_env__get_core_aggr_by_cpu;
1740 	case AGGR_NODE:
1741 		return perf_env__get_node_aggr_by_cpu;
1742 	case AGGR_GLOBAL:
1743 		return perf_env__get_global_aggr_by_cpu;
1744 	case AGGR_NONE:
1745 		return perf_env__get_cpu_aggr_by_cpu;
1746 	case AGGR_THREAD:
1747 	case AGGR_UNSET:
1748 	case AGGR_MAX:
1749 	default:
1750 		return NULL;
1751 	}
1752 }
1753 
1754 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1755 {
1756 	switch (aggr_mode) {
1757 	case AGGR_SOCKET:
1758 		return perf_stat__get_socket_file;
1759 	case AGGR_DIE:
1760 		return perf_stat__get_die_file;
1761 	case AGGR_CLUSTER:
1762 		return perf_stat__get_cluster_file;
1763 	case AGGR_CACHE:
1764 		return perf_stat__get_cache_file;
1765 	case AGGR_CORE:
1766 		return perf_stat__get_core_file;
1767 	case AGGR_NODE:
1768 		return perf_stat__get_node_file;
1769 	case AGGR_GLOBAL:
1770 		return perf_stat__get_global_file;
1771 	case AGGR_NONE:
1772 		return perf_stat__get_cpu_file;
1773 	case AGGR_THREAD:
1774 	case AGGR_UNSET:
1775 	case AGGR_MAX:
1776 	default:
1777 		return NULL;
1778 	}
1779 }
1780 
1781 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1782 {
1783 	struct perf_env *env = &st->session->header.env;
1784 	aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1785 	bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1786 
1787 	if (stat_config.aggr_mode == AGGR_THREAD) {
1788 		int nr = perf_thread_map__nr(evsel_list->core.threads);
1789 
1790 		stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1791 		if (stat_config.aggr_map == NULL)
1792 			return -ENOMEM;
1793 
1794 		for (int s = 0; s < nr; s++) {
1795 			struct aggr_cpu_id id = aggr_cpu_id__empty();
1796 
1797 			id.thread_idx = s;
1798 			stat_config.aggr_map->map[s] = id;
1799 		}
1800 		return 0;
1801 	}
1802 
1803 	if (!get_id)
1804 		return 0;
1805 
1806 	stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1807 						 get_id, env, needs_sort);
1808 	if (!stat_config.aggr_map) {
1809 		pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]);
1810 		return -1;
1811 	}
1812 	stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1813 	return 0;
1814 }
1815 
1816 /*
1817  * Add default attributes, if there were no attributes specified or
1818  * if -d/--detailed, -d -d or -d -d -d is used:
1819  */
1820 static int add_default_attributes(void)
1821 {
1822 	struct perf_event_attr default_attrs0[] = {
1823 
1824   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1825   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1826   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1827   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1828 
1829   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1830 };
1831 	struct perf_event_attr frontend_attrs[] = {
1832   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1833 };
1834 	struct perf_event_attr backend_attrs[] = {
1835   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1836 };
1837 	struct perf_event_attr default_attrs1[] = {
1838   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1839   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1840   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1841 
1842 };
1843 
1844 /*
1845  * Detailed stats (-d), covering the L1 and last level data caches:
1846  */
1847 	struct perf_event_attr detailed_attrs[] = {
1848 
1849   { .type = PERF_TYPE_HW_CACHE,
1850     .config =
1851 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1852 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1853 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1854 
1855   { .type = PERF_TYPE_HW_CACHE,
1856     .config =
1857 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1858 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1859 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1860 
1861   { .type = PERF_TYPE_HW_CACHE,
1862     .config =
1863 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1864 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1865 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1866 
1867   { .type = PERF_TYPE_HW_CACHE,
1868     .config =
1869 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1870 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1871 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1872 };
1873 
1874 /*
1875  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1876  */
1877 	struct perf_event_attr very_detailed_attrs[] = {
1878 
1879   { .type = PERF_TYPE_HW_CACHE,
1880     .config =
1881 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1882 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1883 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1884 
1885   { .type = PERF_TYPE_HW_CACHE,
1886     .config =
1887 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1888 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1889 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1890 
1891   { .type = PERF_TYPE_HW_CACHE,
1892     .config =
1893 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1894 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1895 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1896 
1897   { .type = PERF_TYPE_HW_CACHE,
1898     .config =
1899 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1900 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1901 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1902 
1903   { .type = PERF_TYPE_HW_CACHE,
1904     .config =
1905 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1906 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1907 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1908 
1909   { .type = PERF_TYPE_HW_CACHE,
1910     .config =
1911 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1912 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1913 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1914 
1915 };
1916 
1917 /*
1918  * Very, very detailed stats (-d -d -d), adding prefetch events:
1919  */
1920 	struct perf_event_attr very_very_detailed_attrs[] = {
1921 
1922   { .type = PERF_TYPE_HW_CACHE,
1923     .config =
1924 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1925 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1926 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1927 
1928   { .type = PERF_TYPE_HW_CACHE,
1929     .config =
1930 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1931 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1932 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1933 };
1934 
1935 	struct perf_event_attr default_null_attrs[] = {};
1936 	const char *pmu = parse_events_option_args.pmu_filter ?: "all";
1937 
1938 	/* Set attrs if no event is selected and !null_run: */
1939 	if (stat_config.null_run)
1940 		return 0;
1941 
1942 	if (transaction_run) {
1943 		/* Handle -T as -M transaction. Once platform specific metrics
1944 		 * support has been added to the json files, all architectures
1945 		 * will use this approach. To determine transaction support
1946 		 * on an architecture test for such a metric name.
1947 		 */
1948 		if (!metricgroup__has_metric(pmu, "transaction")) {
1949 			pr_err("Missing transaction metrics\n");
1950 			return -1;
1951 		}
1952 		return metricgroup__parse_groups(evsel_list, pmu, "transaction",
1953 						stat_config.metric_no_group,
1954 						stat_config.metric_no_merge,
1955 						stat_config.metric_no_threshold,
1956 						stat_config.user_requested_cpu_list,
1957 						stat_config.system_wide,
1958 						stat_config.hardware_aware_grouping,
1959 						&stat_config.metric_events);
1960 	}
1961 
1962 	if (smi_cost) {
1963 		int smi;
1964 
1965 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1966 			pr_err("freeze_on_smi is not supported.\n");
1967 			return -1;
1968 		}
1969 
1970 		if (!smi) {
1971 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1972 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1973 				return -1;
1974 			}
1975 			smi_reset = true;
1976 		}
1977 
1978 		if (!metricgroup__has_metric(pmu, "smi")) {
1979 			pr_err("Missing smi metrics\n");
1980 			return -1;
1981 		}
1982 
1983 		if (!force_metric_only)
1984 			stat_config.metric_only = true;
1985 
1986 		return metricgroup__parse_groups(evsel_list, pmu, "smi",
1987 						stat_config.metric_no_group,
1988 						stat_config.metric_no_merge,
1989 						stat_config.metric_no_threshold,
1990 						stat_config.user_requested_cpu_list,
1991 						stat_config.system_wide,
1992 						stat_config.hardware_aware_grouping,
1993 						&stat_config.metric_events);
1994 	}
1995 
1996 	if (topdown_run) {
1997 		unsigned int max_level = metricgroups__topdown_max_level();
1998 		char str[] = "TopdownL1";
1999 
2000 		if (!force_metric_only)
2001 			stat_config.metric_only = true;
2002 
2003 		if (!max_level) {
2004 			pr_err("Topdown requested but the topdown metric groups aren't present.\n"
2005 				"(See perf list the metric groups have names like TopdownL1)\n");
2006 			return -1;
2007 		}
2008 		if (stat_config.topdown_level > max_level) {
2009 			pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
2010 			return -1;
2011 		} else if (!stat_config.topdown_level)
2012 			stat_config.topdown_level = 1;
2013 
2014 		if (!stat_config.interval && !stat_config.metric_only) {
2015 			fprintf(stat_config.output,
2016 				"Topdown accuracy may decrease when measuring long periods.\n"
2017 				"Please print the result regularly, e.g. -I1000\n");
2018 		}
2019 		str[8] = stat_config.topdown_level + '0';
2020 		if (metricgroup__parse_groups(evsel_list,
2021 						pmu, str,
2022 						/*metric_no_group=*/false,
2023 						/*metric_no_merge=*/false,
2024 						/*metric_no_threshold=*/true,
2025 						stat_config.user_requested_cpu_list,
2026 						stat_config.system_wide,
2027 						stat_config.hardware_aware_grouping,
2028 						&stat_config.metric_events) < 0)
2029 			return -1;
2030 	}
2031 
2032 	if (!stat_config.topdown_level)
2033 		stat_config.topdown_level = 1;
2034 
2035 	if (!evsel_list->core.nr_entries) {
2036 		/* No events so add defaults. */
2037 		if (target__has_cpu(&target))
2038 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2039 
2040 		if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2041 			return -1;
2042 		if (perf_pmus__have_event("cpu", "stalled-cycles-frontend")) {
2043 			if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
2044 				return -1;
2045 		}
2046 		if (perf_pmus__have_event("cpu", "stalled-cycles-backend")) {
2047 			if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
2048 				return -1;
2049 		}
2050 		if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2051 			return -1;
2052 		/*
2053 		 * Add TopdownL1 metrics if they exist. To minimize
2054 		 * multiplexing, don't request threshold computation.
2055 		 */
2056 		if (metricgroup__has_metric(pmu, "Default")) {
2057 			struct evlist *metric_evlist = evlist__new();
2058 			struct evsel *metric_evsel;
2059 
2060 			if (!metric_evlist)
2061 				return -1;
2062 
2063 			if (metricgroup__parse_groups(metric_evlist, pmu, "Default",
2064 							/*metric_no_group=*/false,
2065 							/*metric_no_merge=*/false,
2066 							/*metric_no_threshold=*/true,
2067 							stat_config.user_requested_cpu_list,
2068 							stat_config.system_wide,
2069 							stat_config.hardware_aware_grouping,
2070 							&stat_config.metric_events) < 0)
2071 				return -1;
2072 
2073 			evlist__for_each_entry(metric_evlist, metric_evsel) {
2074 				metric_evsel->skippable = true;
2075 				metric_evsel->default_metricgroup = true;
2076 			}
2077 			evlist__splice_list_tail(evsel_list, &metric_evlist->core.entries);
2078 			evlist__delete(metric_evlist);
2079 		}
2080 
2081 		/* Platform specific attrs */
2082 		if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0)
2083 			return -1;
2084 	}
2085 
2086 	/* Detailed events get appended to the event list: */
2087 
2088 	if (detailed_run <  1)
2089 		return 0;
2090 
2091 	/* Append detailed run extra attributes: */
2092 	if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2093 		return -1;
2094 
2095 	if (detailed_run < 2)
2096 		return 0;
2097 
2098 	/* Append very detailed run extra attributes: */
2099 	if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2100 		return -1;
2101 
2102 	if (detailed_run < 3)
2103 		return 0;
2104 
2105 	/* Append very, very detailed run extra attributes: */
2106 	return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2107 }
2108 
2109 static const char * const stat_record_usage[] = {
2110 	"perf stat record [<options>]",
2111 	NULL,
2112 };
2113 
2114 static void init_features(struct perf_session *session)
2115 {
2116 	int feat;
2117 
2118 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2119 		perf_header__set_feat(&session->header, feat);
2120 
2121 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
2122 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2123 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2124 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2125 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2126 }
2127 
2128 static int __cmd_record(const struct option stat_options[], struct opt_aggr_mode *opt_mode,
2129 			int argc, const char **argv)
2130 {
2131 	struct perf_session *session;
2132 	struct perf_data *data = &perf_stat.data;
2133 
2134 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
2135 			     PARSE_OPT_STOP_AT_NON_OPTION);
2136 	stat_config.aggr_mode = opt_aggr_mode_to_aggr_mode(opt_mode);
2137 
2138 	if (output_name)
2139 		data->path = output_name;
2140 
2141 	if (stat_config.run_count != 1 || forever) {
2142 		pr_err("Cannot use -r option with perf stat record.\n");
2143 		return -1;
2144 	}
2145 
2146 	session = perf_session__new(data, NULL);
2147 	if (IS_ERR(session)) {
2148 		pr_err("Perf session creation failed\n");
2149 		return PTR_ERR(session);
2150 	}
2151 
2152 	init_features(session);
2153 
2154 	session->evlist   = evsel_list;
2155 	perf_stat.session = session;
2156 	perf_stat.record  = true;
2157 	return argc;
2158 }
2159 
2160 static int process_stat_round_event(struct perf_session *session,
2161 				    union perf_event *event)
2162 {
2163 	struct perf_record_stat_round *stat_round = &event->stat_round;
2164 	struct timespec tsh, *ts = NULL;
2165 	const char **argv = session->header.env.cmdline_argv;
2166 	int argc = session->header.env.nr_cmdline;
2167 
2168 	process_counters();
2169 
2170 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2171 		update_stats(&walltime_nsecs_stats, stat_round->time);
2172 
2173 	if (stat_config.interval && stat_round->time) {
2174 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2175 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2176 		ts = &tsh;
2177 	}
2178 
2179 	print_counters(ts, argc, argv);
2180 	return 0;
2181 }
2182 
2183 static
2184 int process_stat_config_event(struct perf_session *session,
2185 			      union perf_event *event)
2186 {
2187 	const struct perf_tool *tool = session->tool;
2188 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2189 
2190 	perf_event__read_stat_config(&stat_config, &event->stat_config);
2191 
2192 	if (perf_cpu_map__is_empty(st->cpus)) {
2193 		if (st->aggr_mode != AGGR_UNSET)
2194 			pr_warning("warning: processing task data, aggregation mode not set\n");
2195 	} else if (st->aggr_mode != AGGR_UNSET) {
2196 		stat_config.aggr_mode = st->aggr_mode;
2197 	}
2198 
2199 	if (perf_stat.data.is_pipe)
2200 		perf_stat_init_aggr_mode();
2201 	else
2202 		perf_stat_init_aggr_mode_file(st);
2203 
2204 	if (stat_config.aggr_map) {
2205 		int nr_aggr = stat_config.aggr_map->nr;
2206 
2207 		if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) {
2208 			pr_err("cannot allocate aggr counts\n");
2209 			return -1;
2210 		}
2211 	}
2212 	return 0;
2213 }
2214 
2215 static int set_maps(struct perf_stat *st)
2216 {
2217 	if (!st->cpus || !st->threads)
2218 		return 0;
2219 
2220 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2221 		return -EINVAL;
2222 
2223 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2224 
2225 	if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true))
2226 		return -ENOMEM;
2227 
2228 	st->maps_allocated = true;
2229 	return 0;
2230 }
2231 
2232 static
2233 int process_thread_map_event(struct perf_session *session,
2234 			     union perf_event *event)
2235 {
2236 	const struct perf_tool *tool = session->tool;
2237 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2238 
2239 	if (st->threads) {
2240 		pr_warning("Extra thread map event, ignoring.\n");
2241 		return 0;
2242 	}
2243 
2244 	st->threads = thread_map__new_event(&event->thread_map);
2245 	if (!st->threads)
2246 		return -ENOMEM;
2247 
2248 	return set_maps(st);
2249 }
2250 
2251 static
2252 int process_cpu_map_event(struct perf_session *session,
2253 			  union perf_event *event)
2254 {
2255 	const struct perf_tool *tool = session->tool;
2256 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2257 	struct perf_cpu_map *cpus;
2258 
2259 	if (st->cpus) {
2260 		pr_warning("Extra cpu map event, ignoring.\n");
2261 		return 0;
2262 	}
2263 
2264 	cpus = cpu_map__new_data(&event->cpu_map.data);
2265 	if (!cpus)
2266 		return -ENOMEM;
2267 
2268 	st->cpus = cpus;
2269 	return set_maps(st);
2270 }
2271 
2272 static const char * const stat_report_usage[] = {
2273 	"perf stat report [<options>]",
2274 	NULL,
2275 };
2276 
2277 static struct perf_stat perf_stat = {
2278 	.aggr_mode	= AGGR_UNSET,
2279 	.aggr_level	= 0,
2280 };
2281 
2282 static int __cmd_report(int argc, const char **argv)
2283 {
2284 	struct perf_session *session;
2285 	const struct option options[] = {
2286 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2287 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2288 		     "aggregate counts per processor socket", AGGR_SOCKET),
2289 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2290 		     "aggregate counts per processor die", AGGR_DIE),
2291 	OPT_SET_UINT(0, "per-cluster", &perf_stat.aggr_mode,
2292 		     "aggregate counts perf processor cluster", AGGR_CLUSTER),
2293 	OPT_CALLBACK_OPTARG(0, "per-cache", &perf_stat.aggr_mode, &perf_stat.aggr_level,
2294 			    "cache level",
2295 			    "aggregate count at this cache level (Default: LLC)",
2296 			    parse_cache_level),
2297 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2298 		     "aggregate counts per physical processor core", AGGR_CORE),
2299 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2300 		     "aggregate counts per numa node", AGGR_NODE),
2301 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2302 		     "disable CPU count aggregation", AGGR_NONE),
2303 	OPT_END()
2304 	};
2305 	struct stat st;
2306 	int ret;
2307 
2308 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2309 
2310 	if (!input_name || !strlen(input_name)) {
2311 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2312 			input_name = "-";
2313 		else
2314 			input_name = "perf.data";
2315 	}
2316 
2317 	perf_stat.data.path = input_name;
2318 	perf_stat.data.mode = PERF_DATA_MODE_READ;
2319 
2320 	perf_tool__init(&perf_stat.tool, /*ordered_events=*/false);
2321 	perf_stat.tool.attr		= perf_event__process_attr;
2322 	perf_stat.tool.event_update	= perf_event__process_event_update;
2323 	perf_stat.tool.thread_map	= process_thread_map_event;
2324 	perf_stat.tool.cpu_map		= process_cpu_map_event;
2325 	perf_stat.tool.stat_config	= process_stat_config_event;
2326 	perf_stat.tool.stat		= perf_event__process_stat_event;
2327 	perf_stat.tool.stat_round	= process_stat_round_event;
2328 
2329 	session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2330 	if (IS_ERR(session))
2331 		return PTR_ERR(session);
2332 
2333 	perf_stat.session  = session;
2334 	stat_config.output = stderr;
2335 	evlist__delete(evsel_list);
2336 	evsel_list         = session->evlist;
2337 
2338 	ret = perf_session__process_events(session);
2339 	if (ret)
2340 		return ret;
2341 
2342 	perf_session__delete(session);
2343 	return 0;
2344 }
2345 
2346 static void setup_system_wide(int forks)
2347 {
2348 	/*
2349 	 * Make system wide (-a) the default target if
2350 	 * no target was specified and one of following
2351 	 * conditions is met:
2352 	 *
2353 	 *   - there's no workload specified
2354 	 *   - there is workload specified but all requested
2355 	 *     events are system wide events
2356 	 */
2357 	if (!target__none(&target))
2358 		return;
2359 
2360 	if (!forks)
2361 		target.system_wide = true;
2362 	else {
2363 		struct evsel *counter;
2364 
2365 		evlist__for_each_entry(evsel_list, counter) {
2366 			if (!counter->core.requires_cpu &&
2367 			    !evsel__name_is(counter, "duration_time")) {
2368 				return;
2369 			}
2370 		}
2371 
2372 		if (evsel_list->core.nr_entries)
2373 			target.system_wide = true;
2374 	}
2375 }
2376 
2377 int cmd_stat(int argc, const char **argv)
2378 {
2379 	struct opt_aggr_mode opt_mode = {};
2380 	struct option stat_options[] = {
2381 		OPT_BOOLEAN('T', "transaction", &transaction_run,
2382 			"hardware transaction statistics"),
2383 		OPT_CALLBACK('e', "event", &parse_events_option_args, "event",
2384 			"event selector. use 'perf list' to list available events",
2385 			parse_events_option),
2386 		OPT_CALLBACK(0, "filter", &evsel_list, "filter",
2387 			"event filter", parse_filter),
2388 		OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
2389 			"child tasks do not inherit counters"),
2390 		OPT_STRING('p', "pid", &target.pid, "pid",
2391 			"stat events on existing process id"),
2392 		OPT_STRING('t', "tid", &target.tid, "tid",
2393 			"stat events on existing thread id"),
2394 #ifdef HAVE_BPF_SKEL
2395 		OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
2396 			"stat events on existing bpf program id"),
2397 		OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
2398 			"use bpf program to count events"),
2399 		OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
2400 			"path to perf_event_attr map"),
2401 #endif
2402 		OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
2403 			"system-wide collection from all CPUs"),
2404 		OPT_BOOLEAN(0, "scale", &stat_config.scale,
2405 			"Use --no-scale to disable counter scaling for multiplexing"),
2406 		OPT_INCR('v', "verbose", &verbose,
2407 			"be more verbose (show counter open errors, etc)"),
2408 		OPT_INTEGER('r', "repeat", &stat_config.run_count,
2409 			"repeat command and print average + stddev (max: 100, forever: 0)"),
2410 		OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
2411 			"display details about each run (only with -r option)"),
2412 		OPT_BOOLEAN('n', "null", &stat_config.null_run,
2413 			"null run - dont start any counters"),
2414 		OPT_INCR('d', "detailed", &detailed_run,
2415 			"detailed run - start a lot of events"),
2416 		OPT_BOOLEAN('S', "sync", &sync_run,
2417 			"call sync() before starting a run"),
2418 		OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
2419 				"print large numbers with thousands\' separators",
2420 				stat__set_big_num),
2421 		OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
2422 			"list of cpus to monitor in system-wide"),
2423 		OPT_BOOLEAN('A', "no-aggr", &opt_mode.no_aggr,
2424 			"disable aggregation across CPUs or PMUs"),
2425 		OPT_BOOLEAN(0, "no-merge", &opt_mode.no_aggr,
2426 			"disable aggregation the same as -A or -no-aggr"),
2427 		OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
2428 			"Merge identical named hybrid events"),
2429 		OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
2430 			"print counts with custom separator"),
2431 		OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
2432 			"print counts in JSON format"),
2433 		OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
2434 			"monitor event in cgroup name only", parse_stat_cgroups),
2435 		OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
2436 			"expand events for each cgroup"),
2437 		OPT_STRING('o', "output", &output_name, "file", "output file name"),
2438 		OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
2439 		OPT_INTEGER(0, "log-fd", &output_fd,
2440 			"log output to fd, instead of stderr"),
2441 		OPT_STRING(0, "pre", &pre_cmd, "command",
2442 			"command to run prior to the measured command"),
2443 		OPT_STRING(0, "post", &post_cmd, "command",
2444 			"command to run after to the measured command"),
2445 		OPT_UINTEGER('I', "interval-print", &stat_config.interval,
2446 			"print counts at regular interval in ms "
2447 			"(overhead is possible for values <= 100ms)"),
2448 		OPT_INTEGER(0, "interval-count", &stat_config.times,
2449 			"print counts for fixed number of times"),
2450 		OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
2451 			"clear screen in between new interval"),
2452 		OPT_UINTEGER(0, "timeout", &stat_config.timeout,
2453 			"stop workload and print counts after a timeout period in ms (>= 10ms)"),
2454 		OPT_BOOLEAN(0, "per-socket", &opt_mode.socket,
2455 			"aggregate counts per processor socket"),
2456 		OPT_BOOLEAN(0, "per-die", &opt_mode.die, "aggregate counts per processor die"),
2457 		OPT_BOOLEAN(0, "per-cluster", &opt_mode.cluster,
2458 			"aggregate counts per processor cluster"),
2459 		OPT_CALLBACK_OPTARG(0, "per-cache", &opt_mode, &stat_config.aggr_level,
2460 				"cache level", "aggregate count at this cache level (Default: LLC)",
2461 				parse_cache_level),
2462 		OPT_BOOLEAN(0, "per-core", &opt_mode.core,
2463 			"aggregate counts per physical processor core"),
2464 		OPT_BOOLEAN(0, "per-thread", &opt_mode.thread, "aggregate counts per thread"),
2465 		OPT_BOOLEAN(0, "per-node", &opt_mode.node, "aggregate counts per numa node"),
2466 		OPT_INTEGER('D', "delay", &target.initial_delay,
2467 			"ms to wait before starting measurement after program start (-1: start with events disabled)"),
2468 		OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
2469 				"Only print computed metrics. No raw values", enable_metric_only),
2470 		OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
2471 			"don't group metric events, impacts multiplexing"),
2472 		OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
2473 			"don't try to share events between metrics in a group"),
2474 		OPT_BOOLEAN(0, "metric-no-threshold", &stat_config.metric_no_threshold,
2475 			"disable adding events for the metric threshold calculation"),
2476 		OPT_BOOLEAN(0, "topdown", &topdown_run,
2477 			"measure top-down statistics"),
2478 #ifdef HAVE_ARCH_X86_64_SUPPORT
2479 		OPT_BOOLEAN(0, "record-tpebs", &tpebs_recording,
2480 			"enable recording for tpebs when retire_latency required"),
2481 #endif
2482 		OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
2483 			"Set the metrics level for the top-down statistics (0: max level)"),
2484 		OPT_BOOLEAN(0, "smi-cost", &smi_cost,
2485 			"measure SMI cost"),
2486 		OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
2487 			"monitor specified metrics or metric groups (separated by ,)",
2488 			append_metric_groups),
2489 		OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
2490 				"Configure all used events to run in kernel space.",
2491 				PARSE_OPT_EXCLUSIVE),
2492 		OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
2493 				"Configure all used events to run in user space.",
2494 				PARSE_OPT_EXCLUSIVE),
2495 		OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
2496 			"Use with 'percore' event qualifier to show the event "
2497 			"counts of one hardware thread by sum up total hardware "
2498 			"threads of same physical core"),
2499 		OPT_BOOLEAN(0, "summary", &stat_config.summary,
2500 			"print summary for interval mode"),
2501 		OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
2502 			"don't print 'summary' for CSV summary output"),
2503 		OPT_BOOLEAN(0, "quiet", &quiet,
2504 			"don't print any output, messages or warnings (useful with record)"),
2505 		OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
2506 			"Only enable events on applying cpu with this type "
2507 			"for hybrid platform (e.g. core or atom)",
2508 			parse_cputype),
2509 #ifdef HAVE_LIBPFM
2510 		OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
2511 			"libpfm4 event selector. use 'perf list' to list available events",
2512 			parse_libpfm_events_option),
2513 #endif
2514 		OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
2515 			"Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
2516 			"\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
2517 			"\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
2518 			parse_control_option),
2519 		OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
2520 				"measure I/O performance metrics provided by arch/platform",
2521 				iostat_parse),
2522 		OPT_END()
2523 	};
2524 	const char * const stat_usage[] = {
2525 		"perf stat [<options>] [<command>]",
2526 		NULL
2527 	};
2528 	int status = -EINVAL, run_idx, err;
2529 	const char *mode;
2530 	FILE *output = stderr;
2531 	unsigned int interval, timeout;
2532 	const char * const stat_subcommands[] = { "record", "report" };
2533 	char errbuf[BUFSIZ];
2534 
2535 	setlocale(LC_ALL, "");
2536 
2537 	evsel_list = evlist__new();
2538 	if (evsel_list == NULL)
2539 		return -ENOMEM;
2540 
2541 	parse_events__shrink_config_terms();
2542 
2543 	/* String-parsing callback-based options would segfault when negated */
2544 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2545 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2546 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2547 
2548 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2549 					(const char **) stat_usage,
2550 					PARSE_OPT_STOP_AT_NON_OPTION);
2551 
2552 	stat_config.aggr_mode = opt_aggr_mode_to_aggr_mode(&opt_mode);
2553 
2554 	if (stat_config.csv_sep) {
2555 		stat_config.csv_output = true;
2556 		if (!strcmp(stat_config.csv_sep, "\\t"))
2557 			stat_config.csv_sep = "\t";
2558 	} else
2559 		stat_config.csv_sep = DEFAULT_SEPARATOR;
2560 
2561 	if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2562 		argc = __cmd_record(stat_options, &opt_mode, argc, argv);
2563 		if (argc < 0)
2564 			return -1;
2565 	} else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2566 		return __cmd_report(argc, argv);
2567 
2568 	interval = stat_config.interval;
2569 	timeout = stat_config.timeout;
2570 
2571 	/*
2572 	 * For record command the -o is already taken care of.
2573 	 */
2574 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2575 		output = NULL;
2576 
2577 	if (output_name && output_fd) {
2578 		fprintf(stderr, "cannot use both --output and --log-fd\n");
2579 		parse_options_usage(stat_usage, stat_options, "o", 1);
2580 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2581 		goto out;
2582 	}
2583 
2584 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2585 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2586 		goto out;
2587 	}
2588 
2589 	if (stat_config.metric_only && stat_config.run_count > 1) {
2590 		fprintf(stderr, "--metric-only is not supported with -r\n");
2591 		goto out;
2592 	}
2593 
2594 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2595 		fprintf(stderr, "--table is only supported with -r\n");
2596 		parse_options_usage(stat_usage, stat_options, "r", 1);
2597 		parse_options_usage(NULL, stat_options, "table", 0);
2598 		goto out;
2599 	}
2600 
2601 	if (output_fd < 0) {
2602 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2603 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2604 		goto out;
2605 	}
2606 
2607 	if (!output && !quiet) {
2608 		struct timespec tm;
2609 		mode = append_file ? "a" : "w";
2610 
2611 		output = fopen(output_name, mode);
2612 		if (!output) {
2613 			perror("failed to create output file");
2614 			return -1;
2615 		}
2616 		if (!stat_config.json_output) {
2617 			clock_gettime(CLOCK_REALTIME, &tm);
2618 			fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2619 		}
2620 	} else if (output_fd > 0) {
2621 		mode = append_file ? "a" : "w";
2622 		output = fdopen(output_fd, mode);
2623 		if (!output) {
2624 			perror("Failed opening logfd");
2625 			return -errno;
2626 		}
2627 	}
2628 
2629 	if (stat_config.interval_clear && !isatty(fileno(output))) {
2630 		fprintf(stderr, "--interval-clear does not work with output\n");
2631 		parse_options_usage(stat_usage, stat_options, "o", 1);
2632 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2633 		parse_options_usage(NULL, stat_options, "interval-clear", 0);
2634 		return -1;
2635 	}
2636 
2637 	stat_config.output = output;
2638 
2639 	/*
2640 	 * let the spreadsheet do the pretty-printing
2641 	 */
2642 	if (stat_config.csv_output) {
2643 		/* User explicitly passed -B? */
2644 		if (big_num_opt == 1) {
2645 			fprintf(stderr, "-B option not supported with -x\n");
2646 			parse_options_usage(stat_usage, stat_options, "B", 1);
2647 			parse_options_usage(NULL, stat_options, "x", 1);
2648 			goto out;
2649 		} else /* Nope, so disable big number formatting */
2650 			stat_config.big_num = false;
2651 	} else if (big_num_opt == 0) /* User passed --no-big-num */
2652 		stat_config.big_num = false;
2653 
2654 	err = target__validate(&target);
2655 	if (err) {
2656 		target__strerror(&target, err, errbuf, BUFSIZ);
2657 		pr_warning("%s\n", errbuf);
2658 	}
2659 
2660 	setup_system_wide(argc);
2661 
2662 	/*
2663 	 * Display user/system times only for single
2664 	 * run and when there's specified tracee.
2665 	 */
2666 	if ((stat_config.run_count == 1) && target__none(&target))
2667 		stat_config.ru_display = true;
2668 
2669 	if (stat_config.run_count < 0) {
2670 		pr_err("Run count must be a positive number\n");
2671 		parse_options_usage(stat_usage, stat_options, "r", 1);
2672 		goto out;
2673 	} else if (stat_config.run_count == 0) {
2674 		forever = true;
2675 		stat_config.run_count = 1;
2676 	}
2677 
2678 	if (stat_config.walltime_run_table) {
2679 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2680 		if (!stat_config.walltime_run) {
2681 			pr_err("failed to setup -r option");
2682 			goto out;
2683 		}
2684 	}
2685 
2686 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2687 		!target__has_task(&target)) {
2688 		if (!target.system_wide || target.cpu_list) {
2689 			fprintf(stderr, "The --per-thread option is only "
2690 				"available when monitoring via -p -t -a "
2691 				"options or only --per-thread.\n");
2692 			parse_options_usage(NULL, stat_options, "p", 1);
2693 			parse_options_usage(NULL, stat_options, "t", 1);
2694 			goto out;
2695 		}
2696 	}
2697 
2698 	/*
2699 	 * no_aggr, cgroup are for system-wide only
2700 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2701 	 */
2702 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2703 	      stat_config.aggr_mode != AGGR_THREAD) ||
2704 	     (nr_cgroups || stat_config.cgroup_list)) &&
2705 	    !target__has_cpu(&target)) {
2706 		fprintf(stderr, "both cgroup and no-aggregation "
2707 			"modes only available in system-wide mode\n");
2708 
2709 		parse_options_usage(stat_usage, stat_options, "G", 1);
2710 		parse_options_usage(NULL, stat_options, "A", 1);
2711 		parse_options_usage(NULL, stat_options, "a", 1);
2712 		parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2713 		goto out;
2714 	}
2715 
2716 	if (stat_config.iostat_run) {
2717 		status = iostat_prepare(evsel_list, &stat_config);
2718 		if (status)
2719 			goto out;
2720 		if (iostat_mode == IOSTAT_LIST) {
2721 			iostat_list(evsel_list, &stat_config);
2722 			goto out;
2723 		} else if (verbose > 0)
2724 			iostat_list(evsel_list, &stat_config);
2725 		if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2726 			target.system_wide = true;
2727 	}
2728 
2729 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2730 		target.per_thread = true;
2731 
2732 	stat_config.system_wide = target.system_wide;
2733 	if (target.cpu_list) {
2734 		stat_config.user_requested_cpu_list = strdup(target.cpu_list);
2735 		if (!stat_config.user_requested_cpu_list) {
2736 			status = -ENOMEM;
2737 			goto out;
2738 		}
2739 	}
2740 
2741 	/*
2742 	 * Metric parsing needs to be delayed as metrics may optimize events
2743 	 * knowing the target is system-wide.
2744 	 */
2745 	if (metrics) {
2746 		const char *pmu = parse_events_option_args.pmu_filter ?: "all";
2747 		int ret = metricgroup__parse_groups(evsel_list, pmu, metrics,
2748 						stat_config.metric_no_group,
2749 						stat_config.metric_no_merge,
2750 						stat_config.metric_no_threshold,
2751 						stat_config.user_requested_cpu_list,
2752 						stat_config.system_wide,
2753 						stat_config.hardware_aware_grouping,
2754 						&stat_config.metric_events);
2755 
2756 		zfree(&metrics);
2757 		if (ret) {
2758 			status = ret;
2759 			goto out;
2760 		}
2761 	}
2762 
2763 	if (add_default_attributes())
2764 		goto out;
2765 
2766 	if (stat_config.cgroup_list) {
2767 		if (nr_cgroups > 0) {
2768 			pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2769 			parse_options_usage(stat_usage, stat_options, "G", 1);
2770 			parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2771 			goto out;
2772 		}
2773 
2774 		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2775 					  &stat_config.metric_events, true) < 0) {
2776 			parse_options_usage(stat_usage, stat_options,
2777 					    "for-each-cgroup", 0);
2778 			goto out;
2779 		}
2780 	}
2781 
2782 	evlist__warn_user_requested_cpus(evsel_list, target.cpu_list);
2783 
2784 	if (evlist__create_maps(evsel_list, &target) < 0) {
2785 		if (target__has_task(&target)) {
2786 			pr_err("Problems finding threads of monitor\n");
2787 			parse_options_usage(stat_usage, stat_options, "p", 1);
2788 			parse_options_usage(NULL, stat_options, "t", 1);
2789 		} else if (target__has_cpu(&target)) {
2790 			perror("failed to parse CPUs map");
2791 			parse_options_usage(stat_usage, stat_options, "C", 1);
2792 			parse_options_usage(NULL, stat_options, "a", 1);
2793 		}
2794 		goto out;
2795 	}
2796 
2797 	evlist__check_cpu_maps(evsel_list);
2798 
2799 	/*
2800 	 * Initialize thread_map with comm names,
2801 	 * so we could print it out on output.
2802 	 */
2803 	if (stat_config.aggr_mode == AGGR_THREAD) {
2804 		thread_map__read_comms(evsel_list->core.threads);
2805 	}
2806 
2807 	if (stat_config.aggr_mode == AGGR_NODE)
2808 		cpu__setup_cpunode_map();
2809 
2810 	if (stat_config.times && interval)
2811 		interval_count = true;
2812 	else if (stat_config.times && !interval) {
2813 		pr_err("interval-count option should be used together with "
2814 				"interval-print.\n");
2815 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2816 		parse_options_usage(stat_usage, stat_options, "I", 1);
2817 		goto out;
2818 	}
2819 
2820 	if (timeout && timeout < 100) {
2821 		if (timeout < 10) {
2822 			pr_err("timeout must be >= 10ms.\n");
2823 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2824 			goto out;
2825 		} else
2826 			pr_warning("timeout < 100ms. "
2827 				   "The overhead percentage could be high in some cases. "
2828 				   "Please proceed with caution.\n");
2829 	}
2830 	if (timeout && interval) {
2831 		pr_err("timeout option is not supported with interval-print.\n");
2832 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2833 		parse_options_usage(stat_usage, stat_options, "I", 1);
2834 		goto out;
2835 	}
2836 
2837 	if (perf_stat_init_aggr_mode())
2838 		goto out;
2839 
2840 	if (evlist__alloc_stats(&stat_config, evsel_list, interval))
2841 		goto out;
2842 
2843 	/*
2844 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2845 	 * while avoiding that older tools show confusing messages.
2846 	 *
2847 	 * However for pipe sessions we need to keep it zero,
2848 	 * because script's perf_evsel__check_attr is triggered
2849 	 * by attr->sample_type != 0, and we can't run it on
2850 	 * stat sessions.
2851 	 */
2852 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2853 
2854 	/*
2855 	 * We dont want to block the signals - that would cause
2856 	 * child tasks to inherit that and Ctrl-C would not work.
2857 	 * What we want is for Ctrl-C to work in the exec()-ed
2858 	 * task, but being ignored by perf stat itself:
2859 	 */
2860 	atexit(sig_atexit);
2861 	if (!forever)
2862 		signal(SIGINT,  skip_signal);
2863 	signal(SIGCHLD, skip_signal);
2864 	signal(SIGALRM, skip_signal);
2865 	signal(SIGABRT, skip_signal);
2866 
2867 	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2868 		goto out;
2869 
2870 	/* Enable ignoring missing threads when -p option is defined. */
2871 	evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2872 	status = 0;
2873 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2874 		if (stat_config.run_count != 1 && verbose > 0)
2875 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2876 				run_idx + 1);
2877 
2878 		if (run_idx != 0)
2879 			evlist__reset_prev_raw_counts(evsel_list);
2880 
2881 		status = run_perf_stat(argc, argv, run_idx);
2882 		if (forever && status != -1 && !interval) {
2883 			print_counters(NULL, argc, argv);
2884 			perf_stat__reset_stats();
2885 		}
2886 	}
2887 
2888 	if (!forever && status != -1 && (!interval || stat_config.summary)) {
2889 		if (stat_config.run_count > 1)
2890 			evlist__copy_res_stats(&stat_config, evsel_list);
2891 		print_counters(NULL, argc, argv);
2892 	}
2893 
2894 	evlist__finalize_ctlfd(evsel_list);
2895 
2896 	if (STAT_RECORD) {
2897 		/*
2898 		 * We synthesize the kernel mmap record just so that older tools
2899 		 * don't emit warnings about not being able to resolve symbols
2900 		 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2901 		 * a saner message about no samples being in the perf.data file.
2902 		 *
2903 		 * This also serves to suppress a warning about f_header.data.size == 0
2904 		 * in header.c at the moment 'perf stat record' gets introduced, which
2905 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2906 		 * records, but the need to suppress the kptr_restrict messages in older
2907 		 * tools remain  -acme
2908 		 */
2909 		int fd = perf_data__fd(&perf_stat.data);
2910 
2911 		err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2912 							 process_synthesized_event,
2913 							 &perf_stat.session->machines.host);
2914 		if (err) {
2915 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2916 				   "older tools may produce warnings about this file\n.");
2917 		}
2918 
2919 		if (!interval) {
2920 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2921 				pr_err("failed to write stat round event\n");
2922 		}
2923 
2924 		if (!perf_stat.data.is_pipe) {
2925 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2926 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2927 		}
2928 
2929 		evlist__close(evsel_list);
2930 		perf_session__delete(perf_stat.session);
2931 	}
2932 
2933 	perf_stat__exit_aggr_mode();
2934 	evlist__free_stats(evsel_list);
2935 out:
2936 	if (stat_config.iostat_run)
2937 		iostat_release(evsel_list);
2938 
2939 	zfree(&stat_config.walltime_run);
2940 	zfree(&stat_config.user_requested_cpu_list);
2941 
2942 	if (smi_cost && smi_reset)
2943 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2944 
2945 	evlist__delete(evsel_list);
2946 
2947 	metricgroup__rblist_exit(&stat_config.metric_events);
2948 	evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2949 
2950 	return status;
2951 }
2952