xref: /linux/tools/perf/builtin-stat.c (revision c79c3c34f75d72a066e292b10aa50fc758c97c89)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * builtin-stat.c
4  *
5  * Builtin stat command: Give a precise performance counters summary
6  * overview about any workload, CPU or specific PID.
7  *
8  * Sample output:
9 
10    $ perf stat ./hackbench 10
11 
12   Time: 0.118
13 
14   Performance counter stats for './hackbench 10':
15 
16        1708.761321 task-clock                #   11.037 CPUs utilized
17             41,190 context-switches          #    0.024 M/sec
18              6,735 CPU-migrations            #    0.004 M/sec
19             17,318 page-faults               #    0.010 M/sec
20      5,205,202,243 cycles                    #    3.046 GHz
21      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
22      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
23      2,603,501,247 instructions              #    0.50  insns per cycle
24                                              #    1.48  stalled cycles per insn
25        484,357,498 branches                  #  283.455 M/sec
26          6,388,934 branch-misses             #    1.32% of all branches
27 
28         0.154822978  seconds time elapsed
29 
30  *
31  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32  *
33  * Improvements and fixes by:
34  *
35  *   Arjan van de Ven <arjan@linux.intel.com>
36  *   Yanmin Zhang <yanmin.zhang@intel.com>
37  *   Wu Fengguang <fengguang.wu@intel.com>
38  *   Mike Galbraith <efault@gmx.de>
39  *   Paul Mackerras <paulus@samba.org>
40  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
41  */
42 
43 #include "builtin.h"
44 #include "perf.h"
45 #include "util/cgroup.h"
46 #include <subcmd/parse-options.h>
47 #include "util/parse-events.h"
48 #include "util/pmu.h"
49 #include "util/event.h"
50 #include "util/evlist.h"
51 #include "util/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 "asm/bug.h"
72 
73 #include <linux/time64.h>
74 #include <linux/zalloc.h>
75 #include <api/fs/fs.h>
76 #include <errno.h>
77 #include <signal.h>
78 #include <stdlib.h>
79 #include <sys/prctl.h>
80 #include <inttypes.h>
81 #include <locale.h>
82 #include <math.h>
83 #include <sys/types.h>
84 #include <sys/stat.h>
85 #include <sys/wait.h>
86 #include <unistd.h>
87 #include <sys/time.h>
88 #include <sys/resource.h>
89 #include <linux/err.h>
90 
91 #include <linux/ctype.h>
92 #include <perf/evlist.h>
93 
94 #define DEFAULT_SEPARATOR	" "
95 #define FREEZE_ON_SMI_PATH	"devices/cpu/freeze_on_smi"
96 
97 static void print_counters(struct timespec *ts, int argc, const char **argv);
98 
99 /* Default events used for perf stat -T */
100 static const char *transaction_attrs = {
101 	"task-clock,"
102 	"{"
103 	"instructions,"
104 	"cycles,"
105 	"cpu/cycles-t/,"
106 	"cpu/tx-start/,"
107 	"cpu/el-start/,"
108 	"cpu/cycles-ct/"
109 	"}"
110 };
111 
112 /* More limited version when the CPU does not have all events. */
113 static const char * transaction_limited_attrs = {
114 	"task-clock,"
115 	"{"
116 	"instructions,"
117 	"cycles,"
118 	"cpu/cycles-t/,"
119 	"cpu/tx-start/"
120 	"}"
121 };
122 
123 static const char * topdown_attrs[] = {
124 	"topdown-total-slots",
125 	"topdown-slots-retired",
126 	"topdown-recovery-bubbles",
127 	"topdown-fetch-bubbles",
128 	"topdown-slots-issued",
129 	NULL,
130 };
131 
132 static const char *topdown_metric_attrs[] = {
133 	"slots",
134 	"topdown-retiring",
135 	"topdown-bad-spec",
136 	"topdown-fe-bound",
137 	"topdown-be-bound",
138 	NULL,
139 };
140 
141 static const char *topdown_metric_L2_attrs[] = {
142 	"slots",
143 	"topdown-retiring",
144 	"topdown-bad-spec",
145 	"topdown-fe-bound",
146 	"topdown-be-bound",
147 	"topdown-heavy-ops",
148 	"topdown-br-mispredict",
149 	"topdown-fetch-lat",
150 	"topdown-mem-bound",
151 	NULL,
152 };
153 
154 static const char *smi_cost_attrs = {
155 	"{"
156 	"msr/aperf/,"
157 	"msr/smi/,"
158 	"cycles"
159 	"}"
160 };
161 
162 static struct evlist	*evsel_list;
163 
164 static struct target target = {
165 	.uid	= UINT_MAX,
166 };
167 
168 #define METRIC_ONLY_LEN 20
169 
170 static volatile pid_t		child_pid			= -1;
171 static int			detailed_run			=  0;
172 static bool			transaction_run;
173 static bool			topdown_run			= false;
174 static bool			smi_cost			= false;
175 static bool			smi_reset			= false;
176 static int			big_num_opt			=  -1;
177 static bool			group				= false;
178 static const char		*pre_cmd			= NULL;
179 static const char		*post_cmd			= NULL;
180 static bool			sync_run			= false;
181 static bool			forever				= false;
182 static bool			force_metric_only		= false;
183 static struct timespec		ref_time;
184 static bool			append_file;
185 static bool			interval_count;
186 static const char		*output_name;
187 static int			output_fd;
188 
189 struct perf_stat {
190 	bool			 record;
191 	struct perf_data	 data;
192 	struct perf_session	*session;
193 	u64			 bytes_written;
194 	struct perf_tool	 tool;
195 	bool			 maps_allocated;
196 	struct perf_cpu_map	*cpus;
197 	struct perf_thread_map *threads;
198 	enum aggr_mode		 aggr_mode;
199 };
200 
201 static struct perf_stat		perf_stat;
202 #define STAT_RECORD		perf_stat.record
203 
204 static volatile int done = 0;
205 
206 static struct perf_stat_config stat_config = {
207 	.aggr_mode		= AGGR_GLOBAL,
208 	.scale			= true,
209 	.unit_width		= 4, /* strlen("unit") */
210 	.run_count		= 1,
211 	.metric_only_len	= METRIC_ONLY_LEN,
212 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
213 	.big_num		= true,
214 	.ctl_fd			= -1,
215 	.ctl_fd_ack		= -1
216 };
217 
218 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
219 {
220 	if (!a->core.cpus && !b->core.cpus)
221 		return true;
222 
223 	if (!a->core.cpus || !b->core.cpus)
224 		return false;
225 
226 	if (a->core.cpus->nr != b->core.cpus->nr)
227 		return false;
228 
229 	for (int i = 0; i < a->core.cpus->nr; i++) {
230 		if (a->core.cpus->map[i] != b->core.cpus->map[i])
231 			return false;
232 	}
233 
234 	return true;
235 }
236 
237 static void evlist__check_cpu_maps(struct evlist *evlist)
238 {
239 	struct evsel *evsel, *pos, *leader;
240 	char buf[1024];
241 
242 	evlist__for_each_entry(evlist, evsel) {
243 		leader = evsel->leader;
244 
245 		/* Check that leader matches cpus with each member. */
246 		if (leader == evsel)
247 			continue;
248 		if (cpus_map_matched(leader, evsel))
249 			continue;
250 
251 		/* If there's mismatch disable the group and warn user. */
252 		WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
253 		evsel__group_desc(leader, buf, sizeof(buf));
254 		pr_warning("  %s\n", buf);
255 
256 		if (verbose) {
257 			cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
258 			pr_warning("     %s: %s\n", leader->name, buf);
259 			cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
260 			pr_warning("     %s: %s\n", evsel->name, buf);
261 		}
262 
263 		for_each_group_evsel(pos, leader) {
264 			pos->leader = pos;
265 			pos->core.nr_members = 0;
266 		}
267 		evsel->leader->core.nr_members = 0;
268 	}
269 }
270 
271 static inline void diff_timespec(struct timespec *r, struct timespec *a,
272 				 struct timespec *b)
273 {
274 	r->tv_sec = a->tv_sec - b->tv_sec;
275 	if (a->tv_nsec < b->tv_nsec) {
276 		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
277 		r->tv_sec--;
278 	} else {
279 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
280 	}
281 }
282 
283 static void perf_stat__reset_stats(void)
284 {
285 	int i;
286 
287 	evlist__reset_stats(evsel_list);
288 	perf_stat__reset_shadow_stats();
289 
290 	for (i = 0; i < stat_config.stats_num; i++)
291 		perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
292 }
293 
294 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
295 				     union perf_event *event,
296 				     struct perf_sample *sample __maybe_unused,
297 				     struct machine *machine __maybe_unused)
298 {
299 	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
300 		pr_err("failed to write perf data, error: %m\n");
301 		return -1;
302 	}
303 
304 	perf_stat.bytes_written += event->header.size;
305 	return 0;
306 }
307 
308 static int write_stat_round_event(u64 tm, u64 type)
309 {
310 	return perf_event__synthesize_stat_round(NULL, tm, type,
311 						 process_synthesized_event,
312 						 NULL);
313 }
314 
315 #define WRITE_STAT_ROUND_EVENT(time, interval) \
316 	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
317 
318 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
319 
320 static int evsel__write_stat_event(struct evsel *counter, u32 cpu, u32 thread,
321 				   struct perf_counts_values *count)
322 {
323 	struct perf_sample_id *sid = SID(counter, cpu, thread);
324 
325 	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
326 					   process_synthesized_event, NULL);
327 }
328 
329 static int read_single_counter(struct evsel *counter, int cpu,
330 			       int thread, struct timespec *rs)
331 {
332 	if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
333 		u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
334 		struct perf_counts_values *count =
335 			perf_counts(counter->counts, cpu, thread);
336 		count->ena = count->run = val;
337 		count->val = val;
338 		return 0;
339 	}
340 	return evsel__read_counter(counter, cpu, thread);
341 }
342 
343 /*
344  * Read out the results of a single counter:
345  * do not aggregate counts across CPUs in system-wide mode
346  */
347 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu)
348 {
349 	int nthreads = perf_thread_map__nr(evsel_list->core.threads);
350 	int thread;
351 
352 	if (!counter->supported)
353 		return -ENOENT;
354 
355 	if (counter->core.system_wide)
356 		nthreads = 1;
357 
358 	for (thread = 0; thread < nthreads; thread++) {
359 		struct perf_counts_values *count;
360 
361 		count = perf_counts(counter->counts, cpu, thread);
362 
363 		/*
364 		 * The leader's group read loads data into its group members
365 		 * (via evsel__read_counter()) and sets their count->loaded.
366 		 */
367 		if (!perf_counts__is_loaded(counter->counts, cpu, thread) &&
368 		    read_single_counter(counter, cpu, thread, rs)) {
369 			counter->counts->scaled = -1;
370 			perf_counts(counter->counts, cpu, thread)->ena = 0;
371 			perf_counts(counter->counts, cpu, thread)->run = 0;
372 			return -1;
373 		}
374 
375 		perf_counts__set_loaded(counter->counts, cpu, thread, false);
376 
377 		if (STAT_RECORD) {
378 			if (evsel__write_stat_event(counter, cpu, thread, count)) {
379 				pr_err("failed to write stat event\n");
380 				return -1;
381 			}
382 		}
383 
384 		if (verbose > 1) {
385 			fprintf(stat_config.output,
386 				"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
387 					evsel__name(counter),
388 					cpu,
389 					count->val, count->ena, count->run);
390 		}
391 	}
392 
393 	return 0;
394 }
395 
396 static int read_affinity_counters(struct timespec *rs)
397 {
398 	struct evsel *counter;
399 	struct affinity affinity;
400 	int i, ncpus, cpu;
401 
402 	if (affinity__setup(&affinity) < 0)
403 		return -1;
404 
405 	ncpus = perf_cpu_map__nr(evsel_list->core.all_cpus);
406 	if (!target__has_cpu(&target) || target__has_per_thread(&target))
407 		ncpus = 1;
408 	evlist__for_each_cpu(evsel_list, i, cpu) {
409 		if (i >= ncpus)
410 			break;
411 		affinity__set(&affinity, cpu);
412 
413 		evlist__for_each_entry(evsel_list, counter) {
414 			if (evsel__cpu_iter_skip(counter, cpu))
415 				continue;
416 			if (!counter->err) {
417 				counter->err = read_counter_cpu(counter, rs,
418 								counter->cpu_iter - 1);
419 			}
420 		}
421 	}
422 	affinity__cleanup(&affinity);
423 	return 0;
424 }
425 
426 static int read_bpf_map_counters(void)
427 {
428 	struct evsel *counter;
429 	int err;
430 
431 	evlist__for_each_entry(evsel_list, counter) {
432 		err = bpf_counter__read(counter);
433 		if (err)
434 			return err;
435 	}
436 	return 0;
437 }
438 
439 static void read_counters(struct timespec *rs)
440 {
441 	struct evsel *counter;
442 	int err;
443 
444 	if (!stat_config.stop_read_counter) {
445 		if (target__has_bpf(&target))
446 			err = read_bpf_map_counters();
447 		else
448 			err = read_affinity_counters(rs);
449 		if (err < 0)
450 			return;
451 	}
452 
453 	evlist__for_each_entry(evsel_list, counter) {
454 		if (counter->err)
455 			pr_debug("failed to read counter %s\n", counter->name);
456 		if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
457 			pr_warning("failed to process counter %s\n", counter->name);
458 		counter->err = 0;
459 	}
460 }
461 
462 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
463 {
464 	int i;
465 
466 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
467 	if (!config->stats)
468 		return -1;
469 
470 	config->stats_num = nthreads;
471 
472 	for (i = 0; i < nthreads; i++)
473 		runtime_stat__init(&config->stats[i]);
474 
475 	return 0;
476 }
477 
478 static void runtime_stat_delete(struct perf_stat_config *config)
479 {
480 	int i;
481 
482 	if (!config->stats)
483 		return;
484 
485 	for (i = 0; i < config->stats_num; i++)
486 		runtime_stat__exit(&config->stats[i]);
487 
488 	zfree(&config->stats);
489 }
490 
491 static void runtime_stat_reset(struct perf_stat_config *config)
492 {
493 	int i;
494 
495 	if (!config->stats)
496 		return;
497 
498 	for (i = 0; i < config->stats_num; i++)
499 		perf_stat__reset_shadow_per_stat(&config->stats[i]);
500 }
501 
502 static void process_interval(void)
503 {
504 	struct timespec ts, rs;
505 
506 	clock_gettime(CLOCK_MONOTONIC, &ts);
507 	diff_timespec(&rs, &ts, &ref_time);
508 
509 	perf_stat__reset_shadow_per_stat(&rt_stat);
510 	runtime_stat_reset(&stat_config);
511 	read_counters(&rs);
512 
513 	if (STAT_RECORD) {
514 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
515 			pr_err("failed to write stat round event\n");
516 	}
517 
518 	init_stats(&walltime_nsecs_stats);
519 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
520 	print_counters(&rs, 0, NULL);
521 }
522 
523 static bool handle_interval(unsigned int interval, int *times)
524 {
525 	if (interval) {
526 		process_interval();
527 		if (interval_count && !(--(*times)))
528 			return true;
529 	}
530 	return false;
531 }
532 
533 static int enable_counters(void)
534 {
535 	struct evsel *evsel;
536 	int err;
537 
538 	if (target__has_bpf(&target)) {
539 		evlist__for_each_entry(evsel_list, evsel) {
540 			err = bpf_counter__enable(evsel);
541 			if (err)
542 				return err;
543 		}
544 	}
545 
546 	if (stat_config.initial_delay < 0) {
547 		pr_info(EVLIST_DISABLED_MSG);
548 		return 0;
549 	}
550 
551 	if (stat_config.initial_delay > 0) {
552 		pr_info(EVLIST_DISABLED_MSG);
553 		usleep(stat_config.initial_delay * USEC_PER_MSEC);
554 	}
555 
556 	/*
557 	 * We need to enable counters only if:
558 	 * - we don't have tracee (attaching to task or cpu)
559 	 * - we have initial delay configured
560 	 */
561 	if (!target__none(&target) || stat_config.initial_delay) {
562 		evlist__enable(evsel_list);
563 		if (stat_config.initial_delay > 0)
564 			pr_info(EVLIST_ENABLED_MSG);
565 	}
566 	return 0;
567 }
568 
569 static void disable_counters(void)
570 {
571 	/*
572 	 * If we don't have tracee (attaching to task or cpu), counters may
573 	 * still be running. To get accurate group ratios, we must stop groups
574 	 * from counting before reading their constituent counters.
575 	 */
576 	if (!target__none(&target))
577 		evlist__disable(evsel_list);
578 }
579 
580 static volatile int workload_exec_errno;
581 
582 /*
583  * evlist__prepare_workload will send a SIGUSR1
584  * if the fork fails, since we asked by setting its
585  * want_signal to true.
586  */
587 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
588 					void *ucontext __maybe_unused)
589 {
590 	workload_exec_errno = info->si_value.sival_int;
591 }
592 
593 static bool evsel__should_store_id(struct evsel *counter)
594 {
595 	return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
596 }
597 
598 static bool is_target_alive(struct target *_target,
599 			    struct perf_thread_map *threads)
600 {
601 	struct stat st;
602 	int i;
603 
604 	if (!target__has_task(_target))
605 		return true;
606 
607 	for (i = 0; i < threads->nr; i++) {
608 		char path[PATH_MAX];
609 
610 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
611 			  threads->map[i].pid);
612 
613 		if (!stat(path, &st))
614 			return true;
615 	}
616 
617 	return false;
618 }
619 
620 static void process_evlist(struct evlist *evlist, unsigned int interval)
621 {
622 	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
623 
624 	if (evlist__ctlfd_process(evlist, &cmd) > 0) {
625 		switch (cmd) {
626 		case EVLIST_CTL_CMD_ENABLE:
627 			if (interval)
628 				process_interval();
629 			break;
630 		case EVLIST_CTL_CMD_DISABLE:
631 			if (interval)
632 				process_interval();
633 			break;
634 		case EVLIST_CTL_CMD_SNAPSHOT:
635 		case EVLIST_CTL_CMD_ACK:
636 		case EVLIST_CTL_CMD_UNSUPPORTED:
637 		case EVLIST_CTL_CMD_EVLIST:
638 		case EVLIST_CTL_CMD_STOP:
639 		case EVLIST_CTL_CMD_PING:
640 		default:
641 			break;
642 		}
643 	}
644 }
645 
646 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
647 			int *time_to_sleep)
648 {
649 	int tts = *time_to_sleep;
650 	struct timespec time_diff;
651 
652 	diff_timespec(&time_diff, time_stop, time_start);
653 
654 	tts -= time_diff.tv_sec * MSEC_PER_SEC +
655 	       time_diff.tv_nsec / NSEC_PER_MSEC;
656 
657 	if (tts < 0)
658 		tts = 0;
659 
660 	*time_to_sleep = tts;
661 }
662 
663 static int dispatch_events(bool forks, int timeout, int interval, int *times)
664 {
665 	int child_exited = 0, status = 0;
666 	int time_to_sleep, sleep_time;
667 	struct timespec time_start, time_stop;
668 
669 	if (interval)
670 		sleep_time = interval;
671 	else if (timeout)
672 		sleep_time = timeout;
673 	else
674 		sleep_time = 1000;
675 
676 	time_to_sleep = sleep_time;
677 
678 	while (!done) {
679 		if (forks)
680 			child_exited = waitpid(child_pid, &status, WNOHANG);
681 		else
682 			child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
683 
684 		if (child_exited)
685 			break;
686 
687 		clock_gettime(CLOCK_MONOTONIC, &time_start);
688 		if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
689 			if (timeout || handle_interval(interval, times))
690 				break;
691 			time_to_sleep = sleep_time;
692 		} else { /* fd revent */
693 			process_evlist(evsel_list, interval);
694 			clock_gettime(CLOCK_MONOTONIC, &time_stop);
695 			compute_tts(&time_start, &time_stop, &time_to_sleep);
696 		}
697 	}
698 
699 	return status;
700 }
701 
702 enum counter_recovery {
703 	COUNTER_SKIP,
704 	COUNTER_RETRY,
705 	COUNTER_FATAL,
706 };
707 
708 static enum counter_recovery stat_handle_error(struct evsel *counter)
709 {
710 	char msg[BUFSIZ];
711 	/*
712 	 * PPC returns ENXIO for HW counters until 2.6.37
713 	 * (behavior changed with commit b0a873e).
714 	 */
715 	if (errno == EINVAL || errno == ENOSYS ||
716 	    errno == ENOENT || errno == EOPNOTSUPP ||
717 	    errno == ENXIO) {
718 		if (verbose > 0)
719 			ui__warning("%s event is not supported by the kernel.\n",
720 				    evsel__name(counter));
721 		counter->supported = false;
722 		/*
723 		 * errored is a sticky flag that means one of the counter's
724 		 * cpu event had a problem and needs to be reexamined.
725 		 */
726 		counter->errored = true;
727 
728 		if ((counter->leader != counter) ||
729 		    !(counter->leader->core.nr_members > 1))
730 			return COUNTER_SKIP;
731 	} else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
732 		if (verbose > 0)
733 			ui__warning("%s\n", msg);
734 		return COUNTER_RETRY;
735 	} else if (target__has_per_thread(&target) &&
736 		   evsel_list->core.threads &&
737 		   evsel_list->core.threads->err_thread != -1) {
738 		/*
739 		 * For global --per-thread case, skip current
740 		 * error thread.
741 		 */
742 		if (!thread_map__remove(evsel_list->core.threads,
743 					evsel_list->core.threads->err_thread)) {
744 			evsel_list->core.threads->err_thread = -1;
745 			return COUNTER_RETRY;
746 		}
747 	}
748 
749 	evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
750 	ui__error("%s\n", msg);
751 
752 	if (child_pid != -1)
753 		kill(child_pid, SIGTERM);
754 	return COUNTER_FATAL;
755 }
756 
757 static int __run_perf_stat(int argc, const char **argv, int run_idx)
758 {
759 	int interval = stat_config.interval;
760 	int times = stat_config.times;
761 	int timeout = stat_config.timeout;
762 	char msg[BUFSIZ];
763 	unsigned long long t0, t1;
764 	struct evsel *counter;
765 	size_t l;
766 	int status = 0;
767 	const bool forks = (argc > 0);
768 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
769 	struct affinity affinity;
770 	int i, cpu, err;
771 	bool second_pass = false;
772 
773 	if (forks) {
774 		if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
775 			perror("failed to prepare workload");
776 			return -1;
777 		}
778 		child_pid = evsel_list->workload.pid;
779 	}
780 
781 	if (group)
782 		evlist__set_leader(evsel_list);
783 
784 	if (affinity__setup(&affinity) < 0)
785 		return -1;
786 
787 	if (target__has_bpf(&target)) {
788 		evlist__for_each_entry(evsel_list, counter) {
789 			if (bpf_counter__load(counter, &target))
790 				return -1;
791 		}
792 	}
793 
794 	evlist__for_each_cpu (evsel_list, i, cpu) {
795 		affinity__set(&affinity, cpu);
796 
797 		evlist__for_each_entry(evsel_list, counter) {
798 			if (evsel__cpu_iter_skip(counter, cpu))
799 				continue;
800 			if (counter->reset_group || counter->errored)
801 				continue;
802 try_again:
803 			if (create_perf_stat_counter(counter, &stat_config, &target,
804 						     counter->cpu_iter - 1) < 0) {
805 
806 				/*
807 				 * Weak group failed. We cannot just undo this here
808 				 * because earlier CPUs might be in group mode, and the kernel
809 				 * doesn't support mixing group and non group reads. Defer
810 				 * it to later.
811 				 * Don't close here because we're in the wrong affinity.
812 				 */
813 				if ((errno == EINVAL || errno == EBADF) &&
814 				    counter->leader != counter &&
815 				    counter->weak_group) {
816 					evlist__reset_weak_group(evsel_list, counter, false);
817 					assert(counter->reset_group);
818 					second_pass = true;
819 					continue;
820 				}
821 
822 				switch (stat_handle_error(counter)) {
823 				case COUNTER_FATAL:
824 					return -1;
825 				case COUNTER_RETRY:
826 					goto try_again;
827 				case COUNTER_SKIP:
828 					continue;
829 				default:
830 					break;
831 				}
832 
833 			}
834 			counter->supported = true;
835 		}
836 	}
837 
838 	if (second_pass) {
839 		/*
840 		 * Now redo all the weak group after closing them,
841 		 * and also close errored counters.
842 		 */
843 
844 		evlist__for_each_cpu(evsel_list, i, cpu) {
845 			affinity__set(&affinity, cpu);
846 			/* First close errored or weak retry */
847 			evlist__for_each_entry(evsel_list, counter) {
848 				if (!counter->reset_group && !counter->errored)
849 					continue;
850 				if (evsel__cpu_iter_skip_no_inc(counter, cpu))
851 					continue;
852 				perf_evsel__close_cpu(&counter->core, counter->cpu_iter);
853 			}
854 			/* Now reopen weak */
855 			evlist__for_each_entry(evsel_list, counter) {
856 				if (!counter->reset_group && !counter->errored)
857 					continue;
858 				if (evsel__cpu_iter_skip(counter, cpu))
859 					continue;
860 				if (!counter->reset_group)
861 					continue;
862 try_again_reset:
863 				pr_debug2("reopening weak %s\n", evsel__name(counter));
864 				if (create_perf_stat_counter(counter, &stat_config, &target,
865 							     counter->cpu_iter - 1) < 0) {
866 
867 					switch (stat_handle_error(counter)) {
868 					case COUNTER_FATAL:
869 						return -1;
870 					case COUNTER_RETRY:
871 						goto try_again_reset;
872 					case COUNTER_SKIP:
873 						continue;
874 					default:
875 						break;
876 					}
877 				}
878 				counter->supported = true;
879 			}
880 		}
881 	}
882 	affinity__cleanup(&affinity);
883 
884 	evlist__for_each_entry(evsel_list, counter) {
885 		if (!counter->supported) {
886 			perf_evsel__free_fd(&counter->core);
887 			continue;
888 		}
889 
890 		l = strlen(counter->unit);
891 		if (l > stat_config.unit_width)
892 			stat_config.unit_width = l;
893 
894 		if (evsel__should_store_id(counter) &&
895 		    evsel__store_ids(counter, evsel_list))
896 			return -1;
897 	}
898 
899 	if (evlist__apply_filters(evsel_list, &counter)) {
900 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
901 			counter->filter, evsel__name(counter), errno,
902 			str_error_r(errno, msg, sizeof(msg)));
903 		return -1;
904 	}
905 
906 	if (STAT_RECORD) {
907 		int fd = perf_data__fd(&perf_stat.data);
908 
909 		if (is_pipe) {
910 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
911 		} else {
912 			err = perf_session__write_header(perf_stat.session, evsel_list,
913 							 fd, false);
914 		}
915 
916 		if (err < 0)
917 			return err;
918 
919 		err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
920 							 process_synthesized_event, is_pipe);
921 		if (err < 0)
922 			return err;
923 	}
924 
925 	/*
926 	 * Enable counters and exec the command:
927 	 */
928 	t0 = rdclock();
929 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
930 
931 	if (forks) {
932 		evlist__start_workload(evsel_list);
933 		err = enable_counters();
934 		if (err)
935 			return -1;
936 
937 		if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
938 			status = dispatch_events(forks, timeout, interval, &times);
939 		if (child_pid != -1) {
940 			if (timeout)
941 				kill(child_pid, SIGTERM);
942 			wait4(child_pid, &status, 0, &stat_config.ru_data);
943 		}
944 
945 		if (workload_exec_errno) {
946 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
947 			pr_err("Workload failed: %s\n", emsg);
948 			return -1;
949 		}
950 
951 		if (WIFSIGNALED(status))
952 			psignal(WTERMSIG(status), argv[0]);
953 	} else {
954 		err = enable_counters();
955 		if (err)
956 			return -1;
957 		status = dispatch_events(forks, timeout, interval, &times);
958 	}
959 
960 	disable_counters();
961 
962 	t1 = rdclock();
963 
964 	if (stat_config.walltime_run_table)
965 		stat_config.walltime_run[run_idx] = t1 - t0;
966 
967 	if (interval && stat_config.summary) {
968 		stat_config.interval = 0;
969 		stat_config.stop_read_counter = true;
970 		init_stats(&walltime_nsecs_stats);
971 		update_stats(&walltime_nsecs_stats, t1 - t0);
972 
973 		if (stat_config.aggr_mode == AGGR_GLOBAL)
974 			evlist__save_aggr_prev_raw_counts(evsel_list);
975 
976 		evlist__copy_prev_raw_counts(evsel_list);
977 		evlist__reset_prev_raw_counts(evsel_list);
978 		runtime_stat_reset(&stat_config);
979 		perf_stat__reset_shadow_per_stat(&rt_stat);
980 	} else
981 		update_stats(&walltime_nsecs_stats, t1 - t0);
982 
983 	/*
984 	 * Closing a group leader splits the group, and as we only disable
985 	 * group leaders, results in remaining events becoming enabled. To
986 	 * avoid arbitrary skew, we must read all counters before closing any
987 	 * group leaders.
988 	 */
989 	read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
990 
991 	/*
992 	 * We need to keep evsel_list alive, because it's processed
993 	 * later the evsel_list will be closed after.
994 	 */
995 	if (!STAT_RECORD)
996 		evlist__close(evsel_list);
997 
998 	return WEXITSTATUS(status);
999 }
1000 
1001 static int run_perf_stat(int argc, const char **argv, int run_idx)
1002 {
1003 	int ret;
1004 
1005 	if (pre_cmd) {
1006 		ret = system(pre_cmd);
1007 		if (ret)
1008 			return ret;
1009 	}
1010 
1011 	if (sync_run)
1012 		sync();
1013 
1014 	ret = __run_perf_stat(argc, argv, run_idx);
1015 	if (ret)
1016 		return ret;
1017 
1018 	if (post_cmd) {
1019 		ret = system(post_cmd);
1020 		if (ret)
1021 			return ret;
1022 	}
1023 
1024 	return ret;
1025 }
1026 
1027 static void print_counters(struct timespec *ts, int argc, const char **argv)
1028 {
1029 	/* Do not print anything if we record to the pipe. */
1030 	if (STAT_RECORD && perf_stat.data.is_pipe)
1031 		return;
1032 	if (stat_config.quiet)
1033 		return;
1034 
1035 	evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1036 }
1037 
1038 static volatile int signr = -1;
1039 
1040 static void skip_signal(int signo)
1041 {
1042 	if ((child_pid == -1) || stat_config.interval)
1043 		done = 1;
1044 
1045 	signr = signo;
1046 	/*
1047 	 * render child_pid harmless
1048 	 * won't send SIGTERM to a random
1049 	 * process in case of race condition
1050 	 * and fast PID recycling
1051 	 */
1052 	child_pid = -1;
1053 }
1054 
1055 static void sig_atexit(void)
1056 {
1057 	sigset_t set, oset;
1058 
1059 	/*
1060 	 * avoid race condition with SIGCHLD handler
1061 	 * in skip_signal() which is modifying child_pid
1062 	 * goal is to avoid send SIGTERM to a random
1063 	 * process
1064 	 */
1065 	sigemptyset(&set);
1066 	sigaddset(&set, SIGCHLD);
1067 	sigprocmask(SIG_BLOCK, &set, &oset);
1068 
1069 	if (child_pid != -1)
1070 		kill(child_pid, SIGTERM);
1071 
1072 	sigprocmask(SIG_SETMASK, &oset, NULL);
1073 
1074 	if (signr == -1)
1075 		return;
1076 
1077 	signal(signr, SIG_DFL);
1078 	kill(getpid(), signr);
1079 }
1080 
1081 void perf_stat__set_big_num(int set)
1082 {
1083 	stat_config.big_num = (set != 0);
1084 }
1085 
1086 static int stat__set_big_num(const struct option *opt __maybe_unused,
1087 			     const char *s __maybe_unused, int unset)
1088 {
1089 	big_num_opt = unset ? 0 : 1;
1090 	perf_stat__set_big_num(!unset);
1091 	return 0;
1092 }
1093 
1094 static int enable_metric_only(const struct option *opt __maybe_unused,
1095 			      const char *s __maybe_unused, int unset)
1096 {
1097 	force_metric_only = true;
1098 	stat_config.metric_only = !unset;
1099 	return 0;
1100 }
1101 
1102 static int parse_metric_groups(const struct option *opt,
1103 			       const char *str,
1104 			       int unset __maybe_unused)
1105 {
1106 	return metricgroup__parse_groups(opt, str,
1107 					 stat_config.metric_no_group,
1108 					 stat_config.metric_no_merge,
1109 					 &stat_config.metric_events);
1110 }
1111 
1112 static int parse_control_option(const struct option *opt,
1113 				const char *str,
1114 				int unset __maybe_unused)
1115 {
1116 	struct perf_stat_config *config = opt->value;
1117 
1118 	return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1119 }
1120 
1121 static int parse_stat_cgroups(const struct option *opt,
1122 			      const char *str, int unset)
1123 {
1124 	if (stat_config.cgroup_list) {
1125 		pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1126 		return -1;
1127 	}
1128 
1129 	return parse_cgroups(opt, str, unset);
1130 }
1131 
1132 static struct option stat_options[] = {
1133 	OPT_BOOLEAN('T', "transaction", &transaction_run,
1134 		    "hardware transaction statistics"),
1135 	OPT_CALLBACK('e', "event", &evsel_list, "event",
1136 		     "event selector. use 'perf list' to list available events",
1137 		     parse_events_option),
1138 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1139 		     "event filter", parse_filter),
1140 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1141 		    "child tasks do not inherit counters"),
1142 	OPT_STRING('p', "pid", &target.pid, "pid",
1143 		   "stat events on existing process id"),
1144 	OPT_STRING('t', "tid", &target.tid, "tid",
1145 		   "stat events on existing thread id"),
1146 #ifdef HAVE_BPF_SKEL
1147 	OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1148 		   "stat events on existing bpf program id"),
1149 #endif
1150 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1151 		    "system-wide collection from all CPUs"),
1152 	OPT_BOOLEAN('g', "group", &group,
1153 		    "put the counters into a counter group"),
1154 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
1155 		    "Use --no-scale to disable counter scaling for multiplexing"),
1156 	OPT_INCR('v', "verbose", &verbose,
1157 		    "be more verbose (show counter open errors, etc)"),
1158 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
1159 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1160 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1161 		    "display details about each run (only with -r option)"),
1162 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
1163 		    "null run - dont start any counters"),
1164 	OPT_INCR('d', "detailed", &detailed_run,
1165 		    "detailed run - start a lot of events"),
1166 	OPT_BOOLEAN('S', "sync", &sync_run,
1167 		    "call sync() before starting a run"),
1168 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1169 			   "print large numbers with thousands\' separators",
1170 			   stat__set_big_num),
1171 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1172 		    "list of cpus to monitor in system-wide"),
1173 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1174 		    "disable CPU count aggregation", AGGR_NONE),
1175 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1176 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1177 		   "print counts with custom separator"),
1178 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1179 		     "monitor event in cgroup name only", parse_stat_cgroups),
1180 	OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1181 		    "expand events for each cgroup"),
1182 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1183 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1184 	OPT_INTEGER(0, "log-fd", &output_fd,
1185 		    "log output to fd, instead of stderr"),
1186 	OPT_STRING(0, "pre", &pre_cmd, "command",
1187 			"command to run prior to the measured command"),
1188 	OPT_STRING(0, "post", &post_cmd, "command",
1189 			"command to run after to the measured command"),
1190 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1191 		    "print counts at regular interval in ms "
1192 		    "(overhead is possible for values <= 100ms)"),
1193 	OPT_INTEGER(0, "interval-count", &stat_config.times,
1194 		    "print counts for fixed number of times"),
1195 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1196 		    "clear screen in between new interval"),
1197 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1198 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1199 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1200 		     "aggregate counts per processor socket", AGGR_SOCKET),
1201 	OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1202 		     "aggregate counts per processor die", AGGR_DIE),
1203 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1204 		     "aggregate counts per physical processor core", AGGR_CORE),
1205 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1206 		     "aggregate counts per thread", AGGR_THREAD),
1207 	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1208 		     "aggregate counts per numa node", AGGR_NODE),
1209 	OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1210 		    "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1211 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1212 			"Only print computed metrics. No raw values", enable_metric_only),
1213 	OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1214 		       "don't group metric events, impacts multiplexing"),
1215 	OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1216 		       "don't try to share events between metrics in a group"),
1217 	OPT_BOOLEAN(0, "topdown", &topdown_run,
1218 			"measure top-down statistics"),
1219 	OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1220 			"Set the metrics level for the top-down statistics (0: max level)"),
1221 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1222 			"measure SMI cost"),
1223 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1224 		     "monitor specified metrics or metric groups (separated by ,)",
1225 		     parse_metric_groups),
1226 	OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1227 			 "Configure all used events to run in kernel space.",
1228 			 PARSE_OPT_EXCLUSIVE),
1229 	OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1230 			 "Configure all used events to run in user space.",
1231 			 PARSE_OPT_EXCLUSIVE),
1232 	OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1233 		    "Use with 'percore' event qualifier to show the event "
1234 		    "counts of one hardware thread by sum up total hardware "
1235 		    "threads of same physical core"),
1236 	OPT_BOOLEAN(0, "summary", &stat_config.summary,
1237 		       "print summary for interval mode"),
1238 	OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1239 			"don't print output (useful with record)"),
1240 #ifdef HAVE_LIBPFM
1241 	OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1242 		"libpfm4 event selector. use 'perf list' to list available events",
1243 		parse_libpfm_events_option),
1244 #endif
1245 	OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1246 		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1247 		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1248 		     "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1249 		      parse_control_option),
1250 	OPT_END()
1251 };
1252 
1253 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1254 				 struct perf_cpu_map *map, int cpu)
1255 {
1256 	return cpu_map__get_socket(map, cpu, NULL);
1257 }
1258 
1259 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1260 			      struct perf_cpu_map *map, int cpu)
1261 {
1262 	return cpu_map__get_die(map, cpu, NULL);
1263 }
1264 
1265 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1266 			       struct perf_cpu_map *map, int cpu)
1267 {
1268 	return cpu_map__get_core(map, cpu, NULL);
1269 }
1270 
1271 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1272 			       struct perf_cpu_map *map, int cpu)
1273 {
1274 	return cpu_map__get_node(map, cpu, NULL);
1275 }
1276 
1277 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1278 			       aggr_get_id_t get_id, struct perf_cpu_map *map, int idx)
1279 {
1280 	int cpu;
1281 	struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1282 
1283 	if (idx >= map->nr)
1284 		return id;
1285 
1286 	cpu = map->map[idx];
1287 
1288 	if (cpu_map__aggr_cpu_id_is_empty(config->cpus_aggr_map->map[cpu]))
1289 		config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
1290 
1291 	id = config->cpus_aggr_map->map[cpu];
1292 	return id;
1293 }
1294 
1295 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1296 					struct perf_cpu_map *map, int idx)
1297 {
1298 	return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
1299 }
1300 
1301 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1302 					struct perf_cpu_map *map, int idx)
1303 {
1304 	return perf_stat__get_aggr(config, perf_stat__get_die, map, idx);
1305 }
1306 
1307 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1308 				      struct perf_cpu_map *map, int idx)
1309 {
1310 	return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
1311 }
1312 
1313 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1314 				      struct perf_cpu_map *map, int idx)
1315 {
1316 	return perf_stat__get_aggr(config, perf_stat__get_node, map, idx);
1317 }
1318 
1319 static bool term_percore_set(void)
1320 {
1321 	struct evsel *counter;
1322 
1323 	evlist__for_each_entry(evsel_list, counter) {
1324 		if (counter->percore)
1325 			return true;
1326 	}
1327 
1328 	return false;
1329 }
1330 
1331 static int perf_stat_init_aggr_mode(void)
1332 {
1333 	int nr;
1334 
1335 	switch (stat_config.aggr_mode) {
1336 	case AGGR_SOCKET:
1337 		if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1338 			perror("cannot build socket map");
1339 			return -1;
1340 		}
1341 		stat_config.aggr_get_id = perf_stat__get_socket_cached;
1342 		break;
1343 	case AGGR_DIE:
1344 		if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1345 			perror("cannot build die map");
1346 			return -1;
1347 		}
1348 		stat_config.aggr_get_id = perf_stat__get_die_cached;
1349 		break;
1350 	case AGGR_CORE:
1351 		if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1352 			perror("cannot build core map");
1353 			return -1;
1354 		}
1355 		stat_config.aggr_get_id = perf_stat__get_core_cached;
1356 		break;
1357 	case AGGR_NODE:
1358 		if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1359 			perror("cannot build core map");
1360 			return -1;
1361 		}
1362 		stat_config.aggr_get_id = perf_stat__get_node_cached;
1363 		break;
1364 	case AGGR_NONE:
1365 		if (term_percore_set()) {
1366 			if (cpu_map__build_core_map(evsel_list->core.cpus,
1367 						    &stat_config.aggr_map)) {
1368 				perror("cannot build core map");
1369 				return -1;
1370 			}
1371 			stat_config.aggr_get_id = perf_stat__get_core_cached;
1372 		}
1373 		break;
1374 	case AGGR_GLOBAL:
1375 	case AGGR_THREAD:
1376 	case AGGR_UNSET:
1377 	default:
1378 		break;
1379 	}
1380 
1381 	/*
1382 	 * The evsel_list->cpus is the base we operate on,
1383 	 * taking the highest cpu number to be the size of
1384 	 * the aggregation translate cpumap.
1385 	 */
1386 	nr = perf_cpu_map__max(evsel_list->core.cpus);
1387 	stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1388 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1389 }
1390 
1391 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1392 {
1393 	if (map) {
1394 		WARN_ONCE(refcount_read(&map->refcnt) != 0,
1395 			  "cpu_aggr_map refcnt unbalanced\n");
1396 		free(map);
1397 	}
1398 }
1399 
1400 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1401 {
1402 	if (map && refcount_dec_and_test(&map->refcnt))
1403 		cpu_aggr_map__delete(map);
1404 }
1405 
1406 static void perf_stat__exit_aggr_mode(void)
1407 {
1408 	cpu_aggr_map__put(stat_config.aggr_map);
1409 	cpu_aggr_map__put(stat_config.cpus_aggr_map);
1410 	stat_config.aggr_map = NULL;
1411 	stat_config.cpus_aggr_map = NULL;
1412 }
1413 
1414 static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx)
1415 {
1416 	int cpu;
1417 
1418 	if (idx > map->nr)
1419 		return -1;
1420 
1421 	cpu = map->map[idx];
1422 
1423 	if (cpu >= env->nr_cpus_avail)
1424 		return -1;
1425 
1426 	return cpu;
1427 }
1428 
1429 static struct aggr_cpu_id perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data)
1430 {
1431 	struct perf_env *env = data;
1432 	int cpu = perf_env__get_cpu(env, map, idx);
1433 	struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1434 
1435 	if (cpu != -1)
1436 		id.socket = env->cpu[cpu].socket_id;
1437 
1438 	return id;
1439 }
1440 
1441 static struct aggr_cpu_id perf_env__get_die(struct perf_cpu_map *map, int idx, void *data)
1442 {
1443 	struct perf_env *env = data;
1444 	struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1445 	int cpu = perf_env__get_cpu(env, map, idx);
1446 
1447 	if (cpu != -1) {
1448 		/*
1449 		 * die_id is relative to socket, so start
1450 		 * with the socket ID and then add die to
1451 		 * make a unique ID.
1452 		 */
1453 		id.socket = env->cpu[cpu].socket_id;
1454 		id.die = env->cpu[cpu].die_id;
1455 	}
1456 
1457 	return id;
1458 }
1459 
1460 static struct aggr_cpu_id perf_env__get_core(struct perf_cpu_map *map, int idx, void *data)
1461 {
1462 	struct perf_env *env = data;
1463 	struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1464 	int cpu = perf_env__get_cpu(env, map, idx);
1465 
1466 	if (cpu != -1) {
1467 		/*
1468 		 * core_id is relative to socket and die,
1469 		 * we need a global id. So we set
1470 		 * socket, die id and core id
1471 		 */
1472 		id.socket = env->cpu[cpu].socket_id;
1473 		id.die = env->cpu[cpu].die_id;
1474 		id.core = env->cpu[cpu].core_id;
1475 	}
1476 
1477 	return id;
1478 }
1479 
1480 static struct aggr_cpu_id perf_env__get_node(struct perf_cpu_map *map, int idx, void *data)
1481 {
1482 	int cpu = perf_env__get_cpu(data, map, idx);
1483 	struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1484 
1485 	id.node = perf_env__numa_node(data, cpu);
1486 	return id;
1487 }
1488 
1489 static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus,
1490 				      struct cpu_aggr_map **sockp)
1491 {
1492 	return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1493 }
1494 
1495 static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus,
1496 				   struct cpu_aggr_map **diep)
1497 {
1498 	return cpu_map__build_map(cpus, diep, perf_env__get_die, env);
1499 }
1500 
1501 static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus,
1502 				    struct cpu_aggr_map **corep)
1503 {
1504 	return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1505 }
1506 
1507 static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus,
1508 				    struct cpu_aggr_map **nodep)
1509 {
1510 	return cpu_map__build_map(cpus, nodep, perf_env__get_node, env);
1511 }
1512 
1513 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1514 				      struct perf_cpu_map *map, int idx)
1515 {
1516 	return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1517 }
1518 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1519 				   struct perf_cpu_map *map, int idx)
1520 {
1521 	return perf_env__get_die(map, idx, &perf_stat.session->header.env);
1522 }
1523 
1524 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1525 				    struct perf_cpu_map *map, int idx)
1526 {
1527 	return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1528 }
1529 
1530 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1531 				    struct perf_cpu_map *map, int idx)
1532 {
1533 	return perf_env__get_node(map, idx, &perf_stat.session->header.env);
1534 }
1535 
1536 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1537 {
1538 	struct perf_env *env = &st->session->header.env;
1539 
1540 	switch (stat_config.aggr_mode) {
1541 	case AGGR_SOCKET:
1542 		if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1543 			perror("cannot build socket map");
1544 			return -1;
1545 		}
1546 		stat_config.aggr_get_id = perf_stat__get_socket_file;
1547 		break;
1548 	case AGGR_DIE:
1549 		if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1550 			perror("cannot build die map");
1551 			return -1;
1552 		}
1553 		stat_config.aggr_get_id = perf_stat__get_die_file;
1554 		break;
1555 	case AGGR_CORE:
1556 		if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1557 			perror("cannot build core map");
1558 			return -1;
1559 		}
1560 		stat_config.aggr_get_id = perf_stat__get_core_file;
1561 		break;
1562 	case AGGR_NODE:
1563 		if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1564 			perror("cannot build core map");
1565 			return -1;
1566 		}
1567 		stat_config.aggr_get_id = perf_stat__get_node_file;
1568 		break;
1569 	case AGGR_NONE:
1570 	case AGGR_GLOBAL:
1571 	case AGGR_THREAD:
1572 	case AGGR_UNSET:
1573 	default:
1574 		break;
1575 	}
1576 
1577 	return 0;
1578 }
1579 
1580 /*
1581  * Add default attributes, if there were no attributes specified or
1582  * if -d/--detailed, -d -d or -d -d -d is used:
1583  */
1584 static int add_default_attributes(void)
1585 {
1586 	int err;
1587 	struct perf_event_attr default_attrs0[] = {
1588 
1589   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1590   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1591   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1592   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1593 
1594   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1595 };
1596 	struct perf_event_attr frontend_attrs[] = {
1597   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1598 };
1599 	struct perf_event_attr backend_attrs[] = {
1600   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1601 };
1602 	struct perf_event_attr default_attrs1[] = {
1603   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1604   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1605   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1606 
1607 };
1608 
1609 /*
1610  * Detailed stats (-d), covering the L1 and last level data caches:
1611  */
1612 	struct perf_event_attr detailed_attrs[] = {
1613 
1614   { .type = PERF_TYPE_HW_CACHE,
1615     .config =
1616 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1617 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1618 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1619 
1620   { .type = PERF_TYPE_HW_CACHE,
1621     .config =
1622 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1623 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1624 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1625 
1626   { .type = PERF_TYPE_HW_CACHE,
1627     .config =
1628 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1629 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1630 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1631 
1632   { .type = PERF_TYPE_HW_CACHE,
1633     .config =
1634 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1635 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1636 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1637 };
1638 
1639 /*
1640  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1641  */
1642 	struct perf_event_attr very_detailed_attrs[] = {
1643 
1644   { .type = PERF_TYPE_HW_CACHE,
1645     .config =
1646 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1647 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1648 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1649 
1650   { .type = PERF_TYPE_HW_CACHE,
1651     .config =
1652 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1653 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1654 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1655 
1656   { .type = PERF_TYPE_HW_CACHE,
1657     .config =
1658 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1659 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1660 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1661 
1662   { .type = PERF_TYPE_HW_CACHE,
1663     .config =
1664 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1665 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1666 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1667 
1668   { .type = PERF_TYPE_HW_CACHE,
1669     .config =
1670 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1671 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1672 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1673 
1674   { .type = PERF_TYPE_HW_CACHE,
1675     .config =
1676 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1677 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1678 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1679 
1680 };
1681 
1682 /*
1683  * Very, very detailed stats (-d -d -d), adding prefetch events:
1684  */
1685 	struct perf_event_attr very_very_detailed_attrs[] = {
1686 
1687   { .type = PERF_TYPE_HW_CACHE,
1688     .config =
1689 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1690 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1691 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1692 
1693   { .type = PERF_TYPE_HW_CACHE,
1694     .config =
1695 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1696 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1697 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1698 };
1699 	struct parse_events_error errinfo;
1700 
1701 	/* Set attrs if no event is selected and !null_run: */
1702 	if (stat_config.null_run)
1703 		return 0;
1704 
1705 	bzero(&errinfo, sizeof(errinfo));
1706 	if (transaction_run) {
1707 		/* Handle -T as -M transaction. Once platform specific metrics
1708 		 * support has been added to the json files, all archictures
1709 		 * will use this approach. To determine transaction support
1710 		 * on an architecture test for such a metric name.
1711 		 */
1712 		if (metricgroup__has_metric("transaction")) {
1713 			struct option opt = { .value = &evsel_list };
1714 
1715 			return metricgroup__parse_groups(&opt, "transaction",
1716 							 stat_config.metric_no_group,
1717 							stat_config.metric_no_merge,
1718 							 &stat_config.metric_events);
1719 		}
1720 
1721 		if (pmu_have_event("cpu", "cycles-ct") &&
1722 		    pmu_have_event("cpu", "el-start"))
1723 			err = parse_events(evsel_list, transaction_attrs,
1724 					   &errinfo);
1725 		else
1726 			err = parse_events(evsel_list,
1727 					   transaction_limited_attrs,
1728 					   &errinfo);
1729 		if (err) {
1730 			fprintf(stderr, "Cannot set up transaction events\n");
1731 			parse_events_print_error(&errinfo, transaction_attrs);
1732 			return -1;
1733 		}
1734 		return 0;
1735 	}
1736 
1737 	if (smi_cost) {
1738 		int smi;
1739 
1740 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1741 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1742 			return -1;
1743 		}
1744 
1745 		if (!smi) {
1746 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1747 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1748 				return -1;
1749 			}
1750 			smi_reset = true;
1751 		}
1752 
1753 		if (pmu_have_event("msr", "aperf") &&
1754 		    pmu_have_event("msr", "smi")) {
1755 			if (!force_metric_only)
1756 				stat_config.metric_only = true;
1757 			err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1758 		} else {
1759 			fprintf(stderr, "To measure SMI cost, it needs "
1760 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1761 			parse_events_print_error(&errinfo, smi_cost_attrs);
1762 			return -1;
1763 		}
1764 		if (err) {
1765 			parse_events_print_error(&errinfo, smi_cost_attrs);
1766 			fprintf(stderr, "Cannot set up SMI cost events\n");
1767 			return -1;
1768 		}
1769 		return 0;
1770 	}
1771 
1772 	if (topdown_run) {
1773 		const char **metric_attrs = topdown_metric_attrs;
1774 		unsigned int max_level = 1;
1775 		char *str = NULL;
1776 		bool warn = false;
1777 
1778 		if (!force_metric_only)
1779 			stat_config.metric_only = true;
1780 
1781 		if (pmu_have_event("cpu", topdown_metric_L2_attrs[5])) {
1782 			metric_attrs = topdown_metric_L2_attrs;
1783 			max_level = 2;
1784 		}
1785 
1786 		if (stat_config.topdown_level > max_level) {
1787 			pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1788 			return -1;
1789 		} else if (!stat_config.topdown_level)
1790 			stat_config.topdown_level = max_level;
1791 
1792 		if (topdown_filter_events(metric_attrs, &str, 1) < 0) {
1793 			pr_err("Out of memory\n");
1794 			return -1;
1795 		}
1796 		if (metric_attrs[0] && str) {
1797 			if (!stat_config.interval && !stat_config.metric_only) {
1798 				fprintf(stat_config.output,
1799 					"Topdown accuracy may decrease when measuring long periods.\n"
1800 					"Please print the result regularly, e.g. -I1000\n");
1801 			}
1802 			goto setup_metrics;
1803 		}
1804 
1805 		zfree(&str);
1806 
1807 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1808 		    stat_config.aggr_mode != AGGR_CORE) {
1809 			pr_err("top down event configuration requires --per-core mode\n");
1810 			return -1;
1811 		}
1812 		stat_config.aggr_mode = AGGR_CORE;
1813 		if (nr_cgroups || !target__has_cpu(&target)) {
1814 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1815 			return -1;
1816 		}
1817 
1818 		if (topdown_filter_events(topdown_attrs, &str,
1819 				arch_topdown_check_group(&warn)) < 0) {
1820 			pr_err("Out of memory\n");
1821 			return -1;
1822 		}
1823 		if (topdown_attrs[0] && str) {
1824 			if (warn)
1825 				arch_topdown_group_warn();
1826 setup_metrics:
1827 			err = parse_events(evsel_list, str, &errinfo);
1828 			if (err) {
1829 				fprintf(stderr,
1830 					"Cannot set up top down events %s: %d\n",
1831 					str, err);
1832 				parse_events_print_error(&errinfo, str);
1833 				free(str);
1834 				return -1;
1835 			}
1836 		} else {
1837 			fprintf(stderr, "System does not support topdown\n");
1838 			return -1;
1839 		}
1840 		free(str);
1841 	}
1842 
1843 	if (!evsel_list->core.nr_entries) {
1844 		if (target__has_cpu(&target))
1845 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1846 
1847 		if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1848 			return -1;
1849 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1850 			if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1851 				return -1;
1852 		}
1853 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1854 			if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1855 				return -1;
1856 		}
1857 		if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1858 			return -1;
1859 
1860 		if (arch_evlist__add_default_attrs(evsel_list) < 0)
1861 			return -1;
1862 	}
1863 
1864 	/* Detailed events get appended to the event list: */
1865 
1866 	if (detailed_run <  1)
1867 		return 0;
1868 
1869 	/* Append detailed run extra attributes: */
1870 	if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1871 		return -1;
1872 
1873 	if (detailed_run < 2)
1874 		return 0;
1875 
1876 	/* Append very detailed run extra attributes: */
1877 	if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1878 		return -1;
1879 
1880 	if (detailed_run < 3)
1881 		return 0;
1882 
1883 	/* Append very, very detailed run extra attributes: */
1884 	return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1885 }
1886 
1887 static const char * const stat_record_usage[] = {
1888 	"perf stat record [<options>]",
1889 	NULL,
1890 };
1891 
1892 static void init_features(struct perf_session *session)
1893 {
1894 	int feat;
1895 
1896 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1897 		perf_header__set_feat(&session->header, feat);
1898 
1899 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1900 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1901 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1902 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1903 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1904 }
1905 
1906 static int __cmd_record(int argc, const char **argv)
1907 {
1908 	struct perf_session *session;
1909 	struct perf_data *data = &perf_stat.data;
1910 
1911 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
1912 			     PARSE_OPT_STOP_AT_NON_OPTION);
1913 
1914 	if (output_name)
1915 		data->path = output_name;
1916 
1917 	if (stat_config.run_count != 1 || forever) {
1918 		pr_err("Cannot use -r option with perf stat record.\n");
1919 		return -1;
1920 	}
1921 
1922 	session = perf_session__new(data, false, NULL);
1923 	if (IS_ERR(session)) {
1924 		pr_err("Perf session creation failed\n");
1925 		return PTR_ERR(session);
1926 	}
1927 
1928 	init_features(session);
1929 
1930 	session->evlist   = evsel_list;
1931 	perf_stat.session = session;
1932 	perf_stat.record  = true;
1933 	return argc;
1934 }
1935 
1936 static int process_stat_round_event(struct perf_session *session,
1937 				    union perf_event *event)
1938 {
1939 	struct perf_record_stat_round *stat_round = &event->stat_round;
1940 	struct evsel *counter;
1941 	struct timespec tsh, *ts = NULL;
1942 	const char **argv = session->header.env.cmdline_argv;
1943 	int argc = session->header.env.nr_cmdline;
1944 
1945 	evlist__for_each_entry(evsel_list, counter)
1946 		perf_stat_process_counter(&stat_config, counter);
1947 
1948 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1949 		update_stats(&walltime_nsecs_stats, stat_round->time);
1950 
1951 	if (stat_config.interval && stat_round->time) {
1952 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
1953 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1954 		ts = &tsh;
1955 	}
1956 
1957 	print_counters(ts, argc, argv);
1958 	return 0;
1959 }
1960 
1961 static
1962 int process_stat_config_event(struct perf_session *session,
1963 			      union perf_event *event)
1964 {
1965 	struct perf_tool *tool = session->tool;
1966 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1967 
1968 	perf_event__read_stat_config(&stat_config, &event->stat_config);
1969 
1970 	if (perf_cpu_map__empty(st->cpus)) {
1971 		if (st->aggr_mode != AGGR_UNSET)
1972 			pr_warning("warning: processing task data, aggregation mode not set\n");
1973 		return 0;
1974 	}
1975 
1976 	if (st->aggr_mode != AGGR_UNSET)
1977 		stat_config.aggr_mode = st->aggr_mode;
1978 
1979 	if (perf_stat.data.is_pipe)
1980 		perf_stat_init_aggr_mode();
1981 	else
1982 		perf_stat_init_aggr_mode_file(st);
1983 
1984 	return 0;
1985 }
1986 
1987 static int set_maps(struct perf_stat *st)
1988 {
1989 	if (!st->cpus || !st->threads)
1990 		return 0;
1991 
1992 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1993 		return -EINVAL;
1994 
1995 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
1996 
1997 	if (evlist__alloc_stats(evsel_list, true))
1998 		return -ENOMEM;
1999 
2000 	st->maps_allocated = true;
2001 	return 0;
2002 }
2003 
2004 static
2005 int process_thread_map_event(struct perf_session *session,
2006 			     union perf_event *event)
2007 {
2008 	struct perf_tool *tool = session->tool;
2009 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2010 
2011 	if (st->threads) {
2012 		pr_warning("Extra thread map event, ignoring.\n");
2013 		return 0;
2014 	}
2015 
2016 	st->threads = thread_map__new_event(&event->thread_map);
2017 	if (!st->threads)
2018 		return -ENOMEM;
2019 
2020 	return set_maps(st);
2021 }
2022 
2023 static
2024 int process_cpu_map_event(struct perf_session *session,
2025 			  union perf_event *event)
2026 {
2027 	struct perf_tool *tool = session->tool;
2028 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2029 	struct perf_cpu_map *cpus;
2030 
2031 	if (st->cpus) {
2032 		pr_warning("Extra cpu map event, ignoring.\n");
2033 		return 0;
2034 	}
2035 
2036 	cpus = cpu_map__new_data(&event->cpu_map.data);
2037 	if (!cpus)
2038 		return -ENOMEM;
2039 
2040 	st->cpus = cpus;
2041 	return set_maps(st);
2042 }
2043 
2044 static const char * const stat_report_usage[] = {
2045 	"perf stat report [<options>]",
2046 	NULL,
2047 };
2048 
2049 static struct perf_stat perf_stat = {
2050 	.tool = {
2051 		.attr		= perf_event__process_attr,
2052 		.event_update	= perf_event__process_event_update,
2053 		.thread_map	= process_thread_map_event,
2054 		.cpu_map	= process_cpu_map_event,
2055 		.stat_config	= process_stat_config_event,
2056 		.stat		= perf_event__process_stat_event,
2057 		.stat_round	= process_stat_round_event,
2058 	},
2059 	.aggr_mode = AGGR_UNSET,
2060 };
2061 
2062 static int __cmd_report(int argc, const char **argv)
2063 {
2064 	struct perf_session *session;
2065 	const struct option options[] = {
2066 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2067 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2068 		     "aggregate counts per processor socket", AGGR_SOCKET),
2069 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2070 		     "aggregate counts per processor die", AGGR_DIE),
2071 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2072 		     "aggregate counts per physical processor core", AGGR_CORE),
2073 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2074 		     "aggregate counts per numa node", AGGR_NODE),
2075 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2076 		     "disable CPU count aggregation", AGGR_NONE),
2077 	OPT_END()
2078 	};
2079 	struct stat st;
2080 	int ret;
2081 
2082 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2083 
2084 	if (!input_name || !strlen(input_name)) {
2085 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2086 			input_name = "-";
2087 		else
2088 			input_name = "perf.data";
2089 	}
2090 
2091 	perf_stat.data.path = input_name;
2092 	perf_stat.data.mode = PERF_DATA_MODE_READ;
2093 
2094 	session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2095 	if (IS_ERR(session))
2096 		return PTR_ERR(session);
2097 
2098 	perf_stat.session  = session;
2099 	stat_config.output = stderr;
2100 	evsel_list         = session->evlist;
2101 
2102 	ret = perf_session__process_events(session);
2103 	if (ret)
2104 		return ret;
2105 
2106 	perf_session__delete(session);
2107 	return 0;
2108 }
2109 
2110 static void setup_system_wide(int forks)
2111 {
2112 	/*
2113 	 * Make system wide (-a) the default target if
2114 	 * no target was specified and one of following
2115 	 * conditions is met:
2116 	 *
2117 	 *   - there's no workload specified
2118 	 *   - there is workload specified but all requested
2119 	 *     events are system wide events
2120 	 */
2121 	if (!target__none(&target))
2122 		return;
2123 
2124 	if (!forks)
2125 		target.system_wide = true;
2126 	else {
2127 		struct evsel *counter;
2128 
2129 		evlist__for_each_entry(evsel_list, counter) {
2130 			if (!counter->core.system_wide &&
2131 			    strcmp(counter->name, "duration_time")) {
2132 				return;
2133 			}
2134 		}
2135 
2136 		if (evsel_list->core.nr_entries)
2137 			target.system_wide = true;
2138 	}
2139 }
2140 
2141 int cmd_stat(int argc, const char **argv)
2142 {
2143 	const char * const stat_usage[] = {
2144 		"perf stat [<options>] [<command>]",
2145 		NULL
2146 	};
2147 	int status = -EINVAL, run_idx, err;
2148 	const char *mode;
2149 	FILE *output = stderr;
2150 	unsigned int interval, timeout;
2151 	const char * const stat_subcommands[] = { "record", "report" };
2152 	char errbuf[BUFSIZ];
2153 
2154 	setlocale(LC_ALL, "");
2155 
2156 	evsel_list = evlist__new();
2157 	if (evsel_list == NULL)
2158 		return -ENOMEM;
2159 
2160 	parse_events__shrink_config_terms();
2161 
2162 	/* String-parsing callback-based options would segfault when negated */
2163 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2164 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2165 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2166 
2167 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2168 					(const char **) stat_usage,
2169 					PARSE_OPT_STOP_AT_NON_OPTION);
2170 	perf_stat__collect_metric_expr(evsel_list);
2171 	perf_stat__init_shadow_stats();
2172 
2173 	if (stat_config.csv_sep) {
2174 		stat_config.csv_output = true;
2175 		if (!strcmp(stat_config.csv_sep, "\\t"))
2176 			stat_config.csv_sep = "\t";
2177 	} else
2178 		stat_config.csv_sep = DEFAULT_SEPARATOR;
2179 
2180 	if (argc && !strncmp(argv[0], "rec", 3)) {
2181 		argc = __cmd_record(argc, argv);
2182 		if (argc < 0)
2183 			return -1;
2184 	} else if (argc && !strncmp(argv[0], "rep", 3))
2185 		return __cmd_report(argc, argv);
2186 
2187 	interval = stat_config.interval;
2188 	timeout = stat_config.timeout;
2189 
2190 	/*
2191 	 * For record command the -o is already taken care of.
2192 	 */
2193 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2194 		output = NULL;
2195 
2196 	if (output_name && output_fd) {
2197 		fprintf(stderr, "cannot use both --output and --log-fd\n");
2198 		parse_options_usage(stat_usage, stat_options, "o", 1);
2199 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2200 		goto out;
2201 	}
2202 
2203 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2204 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2205 		goto out;
2206 	}
2207 
2208 	if (stat_config.metric_only && stat_config.run_count > 1) {
2209 		fprintf(stderr, "--metric-only is not supported with -r\n");
2210 		goto out;
2211 	}
2212 
2213 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2214 		fprintf(stderr, "--table is only supported with -r\n");
2215 		parse_options_usage(stat_usage, stat_options, "r", 1);
2216 		parse_options_usage(NULL, stat_options, "table", 0);
2217 		goto out;
2218 	}
2219 
2220 	if (output_fd < 0) {
2221 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2222 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2223 		goto out;
2224 	}
2225 
2226 	if (!output && !stat_config.quiet) {
2227 		struct timespec tm;
2228 		mode = append_file ? "a" : "w";
2229 
2230 		output = fopen(output_name, mode);
2231 		if (!output) {
2232 			perror("failed to create output file");
2233 			return -1;
2234 		}
2235 		clock_gettime(CLOCK_REALTIME, &tm);
2236 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2237 	} else if (output_fd > 0) {
2238 		mode = append_file ? "a" : "w";
2239 		output = fdopen(output_fd, mode);
2240 		if (!output) {
2241 			perror("Failed opening logfd");
2242 			return -errno;
2243 		}
2244 	}
2245 
2246 	stat_config.output = output;
2247 
2248 	/*
2249 	 * let the spreadsheet do the pretty-printing
2250 	 */
2251 	if (stat_config.csv_output) {
2252 		/* User explicitly passed -B? */
2253 		if (big_num_opt == 1) {
2254 			fprintf(stderr, "-B option not supported with -x\n");
2255 			parse_options_usage(stat_usage, stat_options, "B", 1);
2256 			parse_options_usage(NULL, stat_options, "x", 1);
2257 			goto out;
2258 		} else /* Nope, so disable big number formatting */
2259 			stat_config.big_num = false;
2260 	} else if (big_num_opt == 0) /* User passed --no-big-num */
2261 		stat_config.big_num = false;
2262 
2263 	err = target__validate(&target);
2264 	if (err) {
2265 		target__strerror(&target, err, errbuf, BUFSIZ);
2266 		pr_warning("%s\n", errbuf);
2267 	}
2268 
2269 	setup_system_wide(argc);
2270 
2271 	/*
2272 	 * Display user/system times only for single
2273 	 * run and when there's specified tracee.
2274 	 */
2275 	if ((stat_config.run_count == 1) && target__none(&target))
2276 		stat_config.ru_display = true;
2277 
2278 	if (stat_config.run_count < 0) {
2279 		pr_err("Run count must be a positive number\n");
2280 		parse_options_usage(stat_usage, stat_options, "r", 1);
2281 		goto out;
2282 	} else if (stat_config.run_count == 0) {
2283 		forever = true;
2284 		stat_config.run_count = 1;
2285 	}
2286 
2287 	if (stat_config.walltime_run_table) {
2288 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2289 		if (!stat_config.walltime_run) {
2290 			pr_err("failed to setup -r option");
2291 			goto out;
2292 		}
2293 	}
2294 
2295 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2296 		!target__has_task(&target)) {
2297 		if (!target.system_wide || target.cpu_list) {
2298 			fprintf(stderr, "The --per-thread option is only "
2299 				"available when monitoring via -p -t -a "
2300 				"options or only --per-thread.\n");
2301 			parse_options_usage(NULL, stat_options, "p", 1);
2302 			parse_options_usage(NULL, stat_options, "t", 1);
2303 			goto out;
2304 		}
2305 	}
2306 
2307 	/*
2308 	 * no_aggr, cgroup are for system-wide only
2309 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2310 	 */
2311 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2312 	      stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2313 	    !target__has_cpu(&target)) {
2314 		fprintf(stderr, "both cgroup and no-aggregation "
2315 			"modes only available in system-wide mode\n");
2316 
2317 		parse_options_usage(stat_usage, stat_options, "G", 1);
2318 		parse_options_usage(NULL, stat_options, "A", 1);
2319 		parse_options_usage(NULL, stat_options, "a", 1);
2320 		goto out;
2321 	}
2322 
2323 	if (add_default_attributes())
2324 		goto out;
2325 
2326 	if (stat_config.cgroup_list) {
2327 		if (nr_cgroups > 0) {
2328 			pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2329 			parse_options_usage(stat_usage, stat_options, "G", 1);
2330 			parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2331 			goto out;
2332 		}
2333 
2334 		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2335 					  &stat_config.metric_events, true) < 0) {
2336 			parse_options_usage(stat_usage, stat_options,
2337 					    "for-each-cgroup", 0);
2338 			goto out;
2339 		}
2340 	}
2341 
2342 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2343 		target.per_thread = true;
2344 
2345 	if (evlist__create_maps(evsel_list, &target) < 0) {
2346 		if (target__has_task(&target)) {
2347 			pr_err("Problems finding threads of monitor\n");
2348 			parse_options_usage(stat_usage, stat_options, "p", 1);
2349 			parse_options_usage(NULL, stat_options, "t", 1);
2350 		} else if (target__has_cpu(&target)) {
2351 			perror("failed to parse CPUs map");
2352 			parse_options_usage(stat_usage, stat_options, "C", 1);
2353 			parse_options_usage(NULL, stat_options, "a", 1);
2354 		}
2355 		goto out;
2356 	}
2357 
2358 	evlist__check_cpu_maps(evsel_list);
2359 
2360 	/*
2361 	 * Initialize thread_map with comm names,
2362 	 * so we could print it out on output.
2363 	 */
2364 	if (stat_config.aggr_mode == AGGR_THREAD) {
2365 		thread_map__read_comms(evsel_list->core.threads);
2366 		if (target.system_wide) {
2367 			if (runtime_stat_new(&stat_config,
2368 				perf_thread_map__nr(evsel_list->core.threads))) {
2369 				goto out;
2370 			}
2371 		}
2372 	}
2373 
2374 	if (stat_config.aggr_mode == AGGR_NODE)
2375 		cpu__setup_cpunode_map();
2376 
2377 	if (stat_config.times && interval)
2378 		interval_count = true;
2379 	else if (stat_config.times && !interval) {
2380 		pr_err("interval-count option should be used together with "
2381 				"interval-print.\n");
2382 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2383 		parse_options_usage(stat_usage, stat_options, "I", 1);
2384 		goto out;
2385 	}
2386 
2387 	if (timeout && timeout < 100) {
2388 		if (timeout < 10) {
2389 			pr_err("timeout must be >= 10ms.\n");
2390 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2391 			goto out;
2392 		} else
2393 			pr_warning("timeout < 100ms. "
2394 				   "The overhead percentage could be high in some cases. "
2395 				   "Please proceed with caution.\n");
2396 	}
2397 	if (timeout && interval) {
2398 		pr_err("timeout option is not supported with interval-print.\n");
2399 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2400 		parse_options_usage(stat_usage, stat_options, "I", 1);
2401 		goto out;
2402 	}
2403 
2404 	if (evlist__alloc_stats(evsel_list, interval))
2405 		goto out;
2406 
2407 	if (perf_stat_init_aggr_mode())
2408 		goto out;
2409 
2410 	/*
2411 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2412 	 * while avoiding that older tools show confusing messages.
2413 	 *
2414 	 * However for pipe sessions we need to keep it zero,
2415 	 * because script's perf_evsel__check_attr is triggered
2416 	 * by attr->sample_type != 0, and we can't run it on
2417 	 * stat sessions.
2418 	 */
2419 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2420 
2421 	/*
2422 	 * We dont want to block the signals - that would cause
2423 	 * child tasks to inherit that and Ctrl-C would not work.
2424 	 * What we want is for Ctrl-C to work in the exec()-ed
2425 	 * task, but being ignored by perf stat itself:
2426 	 */
2427 	atexit(sig_atexit);
2428 	if (!forever)
2429 		signal(SIGINT,  skip_signal);
2430 	signal(SIGCHLD, skip_signal);
2431 	signal(SIGALRM, skip_signal);
2432 	signal(SIGABRT, skip_signal);
2433 
2434 	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2435 		goto out;
2436 
2437 	status = 0;
2438 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2439 		if (stat_config.run_count != 1 && verbose > 0)
2440 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2441 				run_idx + 1);
2442 
2443 		if (run_idx != 0)
2444 			evlist__reset_prev_raw_counts(evsel_list);
2445 
2446 		status = run_perf_stat(argc, argv, run_idx);
2447 		if (forever && status != -1 && !interval) {
2448 			print_counters(NULL, argc, argv);
2449 			perf_stat__reset_stats();
2450 		}
2451 	}
2452 
2453 	if (!forever && status != -1 && (!interval || stat_config.summary))
2454 		print_counters(NULL, argc, argv);
2455 
2456 	evlist__finalize_ctlfd(evsel_list);
2457 
2458 	if (STAT_RECORD) {
2459 		/*
2460 		 * We synthesize the kernel mmap record just so that older tools
2461 		 * don't emit warnings about not being able to resolve symbols
2462 		 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2463 		 * a saner message about no samples being in the perf.data file.
2464 		 *
2465 		 * This also serves to suppress a warning about f_header.data.size == 0
2466 		 * in header.c at the moment 'perf stat record' gets introduced, which
2467 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2468 		 * records, but the need to suppress the kptr_restrict messages in older
2469 		 * tools remain  -acme
2470 		 */
2471 		int fd = perf_data__fd(&perf_stat.data);
2472 
2473 		err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2474 							 process_synthesized_event,
2475 							 &perf_stat.session->machines.host);
2476 		if (err) {
2477 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2478 				   "older tools may produce warnings about this file\n.");
2479 		}
2480 
2481 		if (!interval) {
2482 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2483 				pr_err("failed to write stat round event\n");
2484 		}
2485 
2486 		if (!perf_stat.data.is_pipe) {
2487 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2488 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2489 		}
2490 
2491 		evlist__close(evsel_list);
2492 		perf_session__delete(perf_stat.session);
2493 	}
2494 
2495 	perf_stat__exit_aggr_mode();
2496 	evlist__free_stats(evsel_list);
2497 out:
2498 	zfree(&stat_config.walltime_run);
2499 
2500 	if (smi_cost && smi_reset)
2501 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2502 
2503 	evlist__delete(evsel_list);
2504 
2505 	metricgroup__rblist_exit(&stat_config.metric_events);
2506 	runtime_stat_delete(&stat_config);
2507 	evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2508 
2509 	return status;
2510 }
2511