xref: /openbmc/linux/tools/perf/builtin-stat.c (revision 82a60352)
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/evlist-hybrid.h"
52 #include "util/evsel.h"
53 #include "util/debug.h"
54 #include "util/color.h"
55 #include "util/stat.h"
56 #include "util/header.h"
57 #include "util/cpumap.h"
58 #include "util/thread_map.h"
59 #include "util/counts.h"
60 #include "util/topdown.h"
61 #include "util/session.h"
62 #include "util/tool.h"
63 #include "util/string2.h"
64 #include "util/metricgroup.h"
65 #include "util/synthetic-events.h"
66 #include "util/target.h"
67 #include "util/time-utils.h"
68 #include "util/top.h"
69 #include "util/affinity.h"
70 #include "util/pfm.h"
71 #include "util/bpf_counter.h"
72 #include "util/iostat.h"
73 #include "util/pmu-hybrid.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 
97 #define DEFAULT_SEPARATOR	" "
98 #define FREEZE_ON_SMI_PATH	"devices/cpu/freeze_on_smi"
99 
100 static void print_counters(struct timespec *ts, int argc, const char **argv);
101 
102 /* Default events used for perf stat -T */
103 static const char *transaction_attrs = {
104 	"task-clock,"
105 	"{"
106 	"instructions,"
107 	"cycles,"
108 	"cpu/cycles-t/,"
109 	"cpu/tx-start/,"
110 	"cpu/el-start/,"
111 	"cpu/cycles-ct/"
112 	"}"
113 };
114 
115 /* More limited version when the CPU does not have all events. */
116 static const char * transaction_limited_attrs = {
117 	"task-clock,"
118 	"{"
119 	"instructions,"
120 	"cycles,"
121 	"cpu/cycles-t/,"
122 	"cpu/tx-start/"
123 	"}"
124 };
125 
126 static const char * topdown_attrs[] = {
127 	"topdown-total-slots",
128 	"topdown-slots-retired",
129 	"topdown-recovery-bubbles",
130 	"topdown-fetch-bubbles",
131 	"topdown-slots-issued",
132 	NULL,
133 };
134 
135 static const char *topdown_metric_attrs[] = {
136 	"slots",
137 	"topdown-retiring",
138 	"topdown-bad-spec",
139 	"topdown-fe-bound",
140 	"topdown-be-bound",
141 	NULL,
142 };
143 
144 static const char *topdown_metric_L2_attrs[] = {
145 	"slots",
146 	"topdown-retiring",
147 	"topdown-bad-spec",
148 	"topdown-fe-bound",
149 	"topdown-be-bound",
150 	"topdown-heavy-ops",
151 	"topdown-br-mispredict",
152 	"topdown-fetch-lat",
153 	"topdown-mem-bound",
154 	NULL,
155 };
156 
157 #define TOPDOWN_MAX_LEVEL			2
158 
159 static const char *smi_cost_attrs = {
160 	"{"
161 	"msr/aperf/,"
162 	"msr/smi/,"
163 	"cycles"
164 	"}"
165 };
166 
167 static struct evlist	*evsel_list;
168 static bool all_counters_use_bpf = true;
169 
170 static struct target target = {
171 	.uid	= UINT_MAX,
172 };
173 
174 #define METRIC_ONLY_LEN 20
175 
176 static volatile pid_t		child_pid			= -1;
177 static int			detailed_run			=  0;
178 static bool			transaction_run;
179 static bool			topdown_run			= false;
180 static bool			smi_cost			= false;
181 static bool			smi_reset			= false;
182 static int			big_num_opt			=  -1;
183 static bool			group				= false;
184 static const char		*pre_cmd			= NULL;
185 static const char		*post_cmd			= NULL;
186 static bool			sync_run			= false;
187 static bool			forever				= false;
188 static bool			force_metric_only		= false;
189 static struct timespec		ref_time;
190 static bool			append_file;
191 static bool			interval_count;
192 static const char		*output_name;
193 static int			output_fd;
194 
195 struct perf_stat {
196 	bool			 record;
197 	struct perf_data	 data;
198 	struct perf_session	*session;
199 	u64			 bytes_written;
200 	struct perf_tool	 tool;
201 	bool			 maps_allocated;
202 	struct perf_cpu_map	*cpus;
203 	struct perf_thread_map *threads;
204 	enum aggr_mode		 aggr_mode;
205 };
206 
207 static struct perf_stat		perf_stat;
208 #define STAT_RECORD		perf_stat.record
209 
210 static volatile int done = 0;
211 
212 static struct perf_stat_config stat_config = {
213 	.aggr_mode		= AGGR_GLOBAL,
214 	.scale			= true,
215 	.unit_width		= 4, /* strlen("unit") */
216 	.run_count		= 1,
217 	.metric_only_len	= METRIC_ONLY_LEN,
218 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
219 	.big_num		= true,
220 	.ctl_fd			= -1,
221 	.ctl_fd_ack		= -1,
222 	.iostat_run		= false,
223 };
224 
225 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
226 {
227 	if (!a->core.cpus && !b->core.cpus)
228 		return true;
229 
230 	if (!a->core.cpus || !b->core.cpus)
231 		return false;
232 
233 	if (a->core.cpus->nr != b->core.cpus->nr)
234 		return false;
235 
236 	for (int i = 0; i < a->core.cpus->nr; i++) {
237 		if (a->core.cpus->map[i] != b->core.cpus->map[i])
238 			return false;
239 	}
240 
241 	return true;
242 }
243 
244 static void evlist__check_cpu_maps(struct evlist *evlist)
245 {
246 	struct evsel *evsel, *pos, *leader;
247 	char buf[1024];
248 
249 	if (evlist__has_hybrid(evlist))
250 		evlist__warn_hybrid_group(evlist);
251 
252 	evlist__for_each_entry(evlist, evsel) {
253 		leader = evsel__leader(evsel);
254 
255 		/* Check that leader matches cpus with each member. */
256 		if (leader == evsel)
257 			continue;
258 		if (cpus_map_matched(leader, evsel))
259 			continue;
260 
261 		/* If there's mismatch disable the group and warn user. */
262 		WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
263 		evsel__group_desc(leader, buf, sizeof(buf));
264 		pr_warning("  %s\n", buf);
265 
266 		if (verbose) {
267 			cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
268 			pr_warning("     %s: %s\n", leader->name, buf);
269 			cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
270 			pr_warning("     %s: %s\n", evsel->name, buf);
271 		}
272 
273 		for_each_group_evsel(pos, leader) {
274 			evsel__set_leader(pos, pos);
275 			pos->core.nr_members = 0;
276 		}
277 		evsel->core.leader->nr_members = 0;
278 	}
279 }
280 
281 static inline void diff_timespec(struct timespec *r, struct timespec *a,
282 				 struct timespec *b)
283 {
284 	r->tv_sec = a->tv_sec - b->tv_sec;
285 	if (a->tv_nsec < b->tv_nsec) {
286 		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
287 		r->tv_sec--;
288 	} else {
289 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
290 	}
291 }
292 
293 static void perf_stat__reset_stats(void)
294 {
295 	int i;
296 
297 	evlist__reset_stats(evsel_list);
298 	perf_stat__reset_shadow_stats();
299 
300 	for (i = 0; i < stat_config.stats_num; i++)
301 		perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
302 }
303 
304 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
305 				     union perf_event *event,
306 				     struct perf_sample *sample __maybe_unused,
307 				     struct machine *machine __maybe_unused)
308 {
309 	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
310 		pr_err("failed to write perf data, error: %m\n");
311 		return -1;
312 	}
313 
314 	perf_stat.bytes_written += event->header.size;
315 	return 0;
316 }
317 
318 static int write_stat_round_event(u64 tm, u64 type)
319 {
320 	return perf_event__synthesize_stat_round(NULL, tm, type,
321 						 process_synthesized_event,
322 						 NULL);
323 }
324 
325 #define WRITE_STAT_ROUND_EVENT(time, interval) \
326 	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
327 
328 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
329 
330 static int evsel__write_stat_event(struct evsel *counter, u32 cpu, u32 thread,
331 				   struct perf_counts_values *count)
332 {
333 	struct perf_sample_id *sid = SID(counter, cpu, thread);
334 
335 	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
336 					   process_synthesized_event, NULL);
337 }
338 
339 static int read_single_counter(struct evsel *counter, int cpu,
340 			       int thread, struct timespec *rs)
341 {
342 	if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
343 		u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
344 		struct perf_counts_values *count =
345 			perf_counts(counter->counts, cpu, thread);
346 		count->ena = count->run = val;
347 		count->val = val;
348 		return 0;
349 	}
350 	return evsel__read_counter(counter, cpu, thread);
351 }
352 
353 /*
354  * Read out the results of a single counter:
355  * do not aggregate counts across CPUs in system-wide mode
356  */
357 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu)
358 {
359 	int nthreads = perf_thread_map__nr(evsel_list->core.threads);
360 	int thread;
361 
362 	if (!counter->supported)
363 		return -ENOENT;
364 
365 	if (counter->core.system_wide)
366 		nthreads = 1;
367 
368 	for (thread = 0; thread < nthreads; thread++) {
369 		struct perf_counts_values *count;
370 
371 		count = perf_counts(counter->counts, cpu, thread);
372 
373 		/*
374 		 * The leader's group read loads data into its group members
375 		 * (via evsel__read_counter()) and sets their count->loaded.
376 		 */
377 		if (!perf_counts__is_loaded(counter->counts, cpu, thread) &&
378 		    read_single_counter(counter, cpu, thread, rs)) {
379 			counter->counts->scaled = -1;
380 			perf_counts(counter->counts, cpu, thread)->ena = 0;
381 			perf_counts(counter->counts, cpu, thread)->run = 0;
382 			return -1;
383 		}
384 
385 		perf_counts__set_loaded(counter->counts, cpu, thread, false);
386 
387 		if (STAT_RECORD) {
388 			if (evsel__write_stat_event(counter, cpu, thread, count)) {
389 				pr_err("failed to write stat event\n");
390 				return -1;
391 			}
392 		}
393 
394 		if (verbose > 1) {
395 			fprintf(stat_config.output,
396 				"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
397 					evsel__name(counter),
398 					cpu,
399 					count->val, count->ena, count->run);
400 		}
401 	}
402 
403 	return 0;
404 }
405 
406 static int read_affinity_counters(struct timespec *rs)
407 {
408 	struct evsel *counter;
409 	struct affinity affinity;
410 	int i, ncpus, cpu;
411 
412 	if (all_counters_use_bpf)
413 		return 0;
414 
415 	if (affinity__setup(&affinity) < 0)
416 		return -1;
417 
418 	ncpus = perf_cpu_map__nr(evsel_list->core.all_cpus);
419 	if (!target__has_cpu(&target) || target__has_per_thread(&target))
420 		ncpus = 1;
421 	evlist__for_each_cpu(evsel_list, i, cpu) {
422 		if (i >= ncpus)
423 			break;
424 		affinity__set(&affinity, cpu);
425 
426 		evlist__for_each_entry(evsel_list, counter) {
427 			if (evsel__cpu_iter_skip(counter, cpu))
428 				continue;
429 			if (evsel__is_bpf(counter))
430 				continue;
431 			if (!counter->err) {
432 				counter->err = read_counter_cpu(counter, rs,
433 								counter->cpu_iter - 1);
434 			}
435 		}
436 	}
437 	affinity__cleanup(&affinity);
438 	return 0;
439 }
440 
441 static int read_bpf_map_counters(void)
442 {
443 	struct evsel *counter;
444 	int err;
445 
446 	evlist__for_each_entry(evsel_list, counter) {
447 		if (!evsel__is_bpf(counter))
448 			continue;
449 
450 		err = bpf_counter__read(counter);
451 		if (err)
452 			return err;
453 	}
454 	return 0;
455 }
456 
457 static void read_counters(struct timespec *rs)
458 {
459 	struct evsel *counter;
460 
461 	if (!stat_config.stop_read_counter) {
462 		if (read_bpf_map_counters() ||
463 		    read_affinity_counters(rs))
464 			return;
465 	}
466 
467 	evlist__for_each_entry(evsel_list, counter) {
468 		if (counter->err)
469 			pr_debug("failed to read counter %s\n", counter->name);
470 		if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
471 			pr_warning("failed to process counter %s\n", counter->name);
472 		counter->err = 0;
473 	}
474 }
475 
476 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
477 {
478 	int i;
479 
480 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
481 	if (!config->stats)
482 		return -1;
483 
484 	config->stats_num = nthreads;
485 
486 	for (i = 0; i < nthreads; i++)
487 		runtime_stat__init(&config->stats[i]);
488 
489 	return 0;
490 }
491 
492 static void runtime_stat_delete(struct perf_stat_config *config)
493 {
494 	int i;
495 
496 	if (!config->stats)
497 		return;
498 
499 	for (i = 0; i < config->stats_num; i++)
500 		runtime_stat__exit(&config->stats[i]);
501 
502 	zfree(&config->stats);
503 }
504 
505 static void runtime_stat_reset(struct perf_stat_config *config)
506 {
507 	int i;
508 
509 	if (!config->stats)
510 		return;
511 
512 	for (i = 0; i < config->stats_num; i++)
513 		perf_stat__reset_shadow_per_stat(&config->stats[i]);
514 }
515 
516 static void process_interval(void)
517 {
518 	struct timespec ts, rs;
519 
520 	clock_gettime(CLOCK_MONOTONIC, &ts);
521 	diff_timespec(&rs, &ts, &ref_time);
522 
523 	perf_stat__reset_shadow_per_stat(&rt_stat);
524 	runtime_stat_reset(&stat_config);
525 	read_counters(&rs);
526 
527 	if (STAT_RECORD) {
528 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
529 			pr_err("failed to write stat round event\n");
530 	}
531 
532 	init_stats(&walltime_nsecs_stats);
533 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
534 	print_counters(&rs, 0, NULL);
535 }
536 
537 static bool handle_interval(unsigned int interval, int *times)
538 {
539 	if (interval) {
540 		process_interval();
541 		if (interval_count && !(--(*times)))
542 			return true;
543 	}
544 	return false;
545 }
546 
547 static int enable_counters(void)
548 {
549 	struct evsel *evsel;
550 	int err;
551 
552 	evlist__for_each_entry(evsel_list, evsel) {
553 		if (!evsel__is_bpf(evsel))
554 			continue;
555 
556 		err = bpf_counter__enable(evsel);
557 		if (err)
558 			return err;
559 	}
560 
561 	if (stat_config.initial_delay < 0) {
562 		pr_info(EVLIST_DISABLED_MSG);
563 		return 0;
564 	}
565 
566 	if (stat_config.initial_delay > 0) {
567 		pr_info(EVLIST_DISABLED_MSG);
568 		usleep(stat_config.initial_delay * USEC_PER_MSEC);
569 	}
570 
571 	/*
572 	 * We need to enable counters only if:
573 	 * - we don't have tracee (attaching to task or cpu)
574 	 * - we have initial delay configured
575 	 */
576 	if (!target__none(&target) || stat_config.initial_delay) {
577 		if (!all_counters_use_bpf)
578 			evlist__enable(evsel_list);
579 		if (stat_config.initial_delay > 0)
580 			pr_info(EVLIST_ENABLED_MSG);
581 	}
582 	return 0;
583 }
584 
585 static void disable_counters(void)
586 {
587 	struct evsel *counter;
588 
589 	/*
590 	 * If we don't have tracee (attaching to task or cpu), counters may
591 	 * still be running. To get accurate group ratios, we must stop groups
592 	 * from counting before reading their constituent counters.
593 	 */
594 	if (!target__none(&target)) {
595 		evlist__for_each_entry(evsel_list, counter)
596 			bpf_counter__disable(counter);
597 		if (!all_counters_use_bpf)
598 			evlist__disable(evsel_list);
599 	}
600 }
601 
602 static volatile int workload_exec_errno;
603 
604 /*
605  * evlist__prepare_workload will send a SIGUSR1
606  * if the fork fails, since we asked by setting its
607  * want_signal to true.
608  */
609 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
610 					void *ucontext __maybe_unused)
611 {
612 	workload_exec_errno = info->si_value.sival_int;
613 }
614 
615 static bool evsel__should_store_id(struct evsel *counter)
616 {
617 	return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
618 }
619 
620 static bool is_target_alive(struct target *_target,
621 			    struct perf_thread_map *threads)
622 {
623 	struct stat st;
624 	int i;
625 
626 	if (!target__has_task(_target))
627 		return true;
628 
629 	for (i = 0; i < threads->nr; i++) {
630 		char path[PATH_MAX];
631 
632 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
633 			  threads->map[i].pid);
634 
635 		if (!stat(path, &st))
636 			return true;
637 	}
638 
639 	return false;
640 }
641 
642 static void process_evlist(struct evlist *evlist, unsigned int interval)
643 {
644 	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
645 
646 	if (evlist__ctlfd_process(evlist, &cmd) > 0) {
647 		switch (cmd) {
648 		case EVLIST_CTL_CMD_ENABLE:
649 			if (interval)
650 				process_interval();
651 			break;
652 		case EVLIST_CTL_CMD_DISABLE:
653 			if (interval)
654 				process_interval();
655 			break;
656 		case EVLIST_CTL_CMD_SNAPSHOT:
657 		case EVLIST_CTL_CMD_ACK:
658 		case EVLIST_CTL_CMD_UNSUPPORTED:
659 		case EVLIST_CTL_CMD_EVLIST:
660 		case EVLIST_CTL_CMD_STOP:
661 		case EVLIST_CTL_CMD_PING:
662 		default:
663 			break;
664 		}
665 	}
666 }
667 
668 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
669 			int *time_to_sleep)
670 {
671 	int tts = *time_to_sleep;
672 	struct timespec time_diff;
673 
674 	diff_timespec(&time_diff, time_stop, time_start);
675 
676 	tts -= time_diff.tv_sec * MSEC_PER_SEC +
677 	       time_diff.tv_nsec / NSEC_PER_MSEC;
678 
679 	if (tts < 0)
680 		tts = 0;
681 
682 	*time_to_sleep = tts;
683 }
684 
685 static int dispatch_events(bool forks, int timeout, int interval, int *times)
686 {
687 	int child_exited = 0, status = 0;
688 	int time_to_sleep, sleep_time;
689 	struct timespec time_start, time_stop;
690 
691 	if (interval)
692 		sleep_time = interval;
693 	else if (timeout)
694 		sleep_time = timeout;
695 	else
696 		sleep_time = 1000;
697 
698 	time_to_sleep = sleep_time;
699 
700 	while (!done) {
701 		if (forks)
702 			child_exited = waitpid(child_pid, &status, WNOHANG);
703 		else
704 			child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
705 
706 		if (child_exited)
707 			break;
708 
709 		clock_gettime(CLOCK_MONOTONIC, &time_start);
710 		if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
711 			if (timeout || handle_interval(interval, times))
712 				break;
713 			time_to_sleep = sleep_time;
714 		} else { /* fd revent */
715 			process_evlist(evsel_list, interval);
716 			clock_gettime(CLOCK_MONOTONIC, &time_stop);
717 			compute_tts(&time_start, &time_stop, &time_to_sleep);
718 		}
719 	}
720 
721 	return status;
722 }
723 
724 enum counter_recovery {
725 	COUNTER_SKIP,
726 	COUNTER_RETRY,
727 	COUNTER_FATAL,
728 };
729 
730 static enum counter_recovery stat_handle_error(struct evsel *counter)
731 {
732 	char msg[BUFSIZ];
733 	/*
734 	 * PPC returns ENXIO for HW counters until 2.6.37
735 	 * (behavior changed with commit b0a873e).
736 	 */
737 	if (errno == EINVAL || errno == ENOSYS ||
738 	    errno == ENOENT || errno == EOPNOTSUPP ||
739 	    errno == ENXIO) {
740 		if (verbose > 0)
741 			ui__warning("%s event is not supported by the kernel.\n",
742 				    evsel__name(counter));
743 		counter->supported = false;
744 		/*
745 		 * errored is a sticky flag that means one of the counter's
746 		 * cpu event had a problem and needs to be reexamined.
747 		 */
748 		counter->errored = true;
749 
750 		if ((evsel__leader(counter) != counter) ||
751 		    !(counter->core.leader->nr_members > 1))
752 			return COUNTER_SKIP;
753 	} else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
754 		if (verbose > 0)
755 			ui__warning("%s\n", msg);
756 		return COUNTER_RETRY;
757 	} else if (target__has_per_thread(&target) &&
758 		   evsel_list->core.threads &&
759 		   evsel_list->core.threads->err_thread != -1) {
760 		/*
761 		 * For global --per-thread case, skip current
762 		 * error thread.
763 		 */
764 		if (!thread_map__remove(evsel_list->core.threads,
765 					evsel_list->core.threads->err_thread)) {
766 			evsel_list->core.threads->err_thread = -1;
767 			return COUNTER_RETRY;
768 		}
769 	}
770 
771 	evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
772 	ui__error("%s\n", msg);
773 
774 	if (child_pid != -1)
775 		kill(child_pid, SIGTERM);
776 	return COUNTER_FATAL;
777 }
778 
779 static int __run_perf_stat(int argc, const char **argv, int run_idx)
780 {
781 	int interval = stat_config.interval;
782 	int times = stat_config.times;
783 	int timeout = stat_config.timeout;
784 	char msg[BUFSIZ];
785 	unsigned long long t0, t1;
786 	struct evsel *counter;
787 	size_t l;
788 	int status = 0;
789 	const bool forks = (argc > 0);
790 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
791 	struct affinity affinity;
792 	int i, cpu, err;
793 	bool second_pass = false;
794 
795 	if (forks) {
796 		if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
797 			perror("failed to prepare workload");
798 			return -1;
799 		}
800 		child_pid = evsel_list->workload.pid;
801 	}
802 
803 	if (group)
804 		evlist__set_leader(evsel_list);
805 
806 	if (affinity__setup(&affinity) < 0)
807 		return -1;
808 
809 	evlist__for_each_entry(evsel_list, counter) {
810 		if (bpf_counter__load(counter, &target))
811 			return -1;
812 		if (!evsel__is_bpf(counter))
813 			all_counters_use_bpf = false;
814 	}
815 
816 	evlist__for_each_cpu (evsel_list, i, cpu) {
817 		/*
818 		 * bperf calls evsel__open_per_cpu() in bperf__load(), so
819 		 * no need to call it again here.
820 		 */
821 		if (target.use_bpf)
822 			break;
823 		affinity__set(&affinity, cpu);
824 
825 		evlist__for_each_entry(evsel_list, counter) {
826 			if (evsel__cpu_iter_skip(counter, cpu))
827 				continue;
828 			if (counter->reset_group || counter->errored)
829 				continue;
830 			if (evsel__is_bpf(counter))
831 				continue;
832 try_again:
833 			if (create_perf_stat_counter(counter, &stat_config, &target,
834 						     counter->cpu_iter - 1) < 0) {
835 
836 				/*
837 				 * Weak group failed. We cannot just undo this here
838 				 * because earlier CPUs might be in group mode, and the kernel
839 				 * doesn't support mixing group and non group reads. Defer
840 				 * it to later.
841 				 * Don't close here because we're in the wrong affinity.
842 				 */
843 				if ((errno == EINVAL || errno == EBADF) &&
844 				    evsel__leader(counter) != counter &&
845 				    counter->weak_group) {
846 					evlist__reset_weak_group(evsel_list, counter, false);
847 					assert(counter->reset_group);
848 					second_pass = true;
849 					continue;
850 				}
851 
852 				switch (stat_handle_error(counter)) {
853 				case COUNTER_FATAL:
854 					return -1;
855 				case COUNTER_RETRY:
856 					goto try_again;
857 				case COUNTER_SKIP:
858 					continue;
859 				default:
860 					break;
861 				}
862 
863 			}
864 			counter->supported = true;
865 		}
866 	}
867 
868 	if (second_pass) {
869 		/*
870 		 * Now redo all the weak group after closing them,
871 		 * and also close errored counters.
872 		 */
873 
874 		evlist__for_each_cpu(evsel_list, i, cpu) {
875 			affinity__set(&affinity, cpu);
876 			/* First close errored or weak retry */
877 			evlist__for_each_entry(evsel_list, counter) {
878 				if (!counter->reset_group && !counter->errored)
879 					continue;
880 				if (evsel__cpu_iter_skip_no_inc(counter, cpu))
881 					continue;
882 				perf_evsel__close_cpu(&counter->core, counter->cpu_iter);
883 			}
884 			/* Now reopen weak */
885 			evlist__for_each_entry(evsel_list, counter) {
886 				if (!counter->reset_group && !counter->errored)
887 					continue;
888 				if (evsel__cpu_iter_skip(counter, cpu))
889 					continue;
890 				if (!counter->reset_group)
891 					continue;
892 try_again_reset:
893 				pr_debug2("reopening weak %s\n", evsel__name(counter));
894 				if (create_perf_stat_counter(counter, &stat_config, &target,
895 							     counter->cpu_iter - 1) < 0) {
896 
897 					switch (stat_handle_error(counter)) {
898 					case COUNTER_FATAL:
899 						return -1;
900 					case COUNTER_RETRY:
901 						goto try_again_reset;
902 					case COUNTER_SKIP:
903 						continue;
904 					default:
905 						break;
906 					}
907 				}
908 				counter->supported = true;
909 			}
910 		}
911 	}
912 	affinity__cleanup(&affinity);
913 
914 	evlist__for_each_entry(evsel_list, counter) {
915 		if (!counter->supported) {
916 			perf_evsel__free_fd(&counter->core);
917 			continue;
918 		}
919 
920 		l = strlen(counter->unit);
921 		if (l > stat_config.unit_width)
922 			stat_config.unit_width = l;
923 
924 		if (evsel__should_store_id(counter) &&
925 		    evsel__store_ids(counter, evsel_list))
926 			return -1;
927 	}
928 
929 	if (evlist__apply_filters(evsel_list, &counter)) {
930 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
931 			counter->filter, evsel__name(counter), errno,
932 			str_error_r(errno, msg, sizeof(msg)));
933 		return -1;
934 	}
935 
936 	if (STAT_RECORD) {
937 		int fd = perf_data__fd(&perf_stat.data);
938 
939 		if (is_pipe) {
940 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
941 		} else {
942 			err = perf_session__write_header(perf_stat.session, evsel_list,
943 							 fd, false);
944 		}
945 
946 		if (err < 0)
947 			return err;
948 
949 		err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
950 							 process_synthesized_event, is_pipe);
951 		if (err < 0)
952 			return err;
953 	}
954 
955 	/*
956 	 * Enable counters and exec the command:
957 	 */
958 	if (forks) {
959 		evlist__start_workload(evsel_list);
960 		err = enable_counters();
961 		if (err)
962 			return -1;
963 
964 		t0 = rdclock();
965 		clock_gettime(CLOCK_MONOTONIC, &ref_time);
966 
967 		if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
968 			status = dispatch_events(forks, timeout, interval, &times);
969 		if (child_pid != -1) {
970 			if (timeout)
971 				kill(child_pid, SIGTERM);
972 			wait4(child_pid, &status, 0, &stat_config.ru_data);
973 		}
974 
975 		if (workload_exec_errno) {
976 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
977 			pr_err("Workload failed: %s\n", emsg);
978 			return -1;
979 		}
980 
981 		if (WIFSIGNALED(status))
982 			psignal(WTERMSIG(status), argv[0]);
983 	} else {
984 		err = enable_counters();
985 		if (err)
986 			return -1;
987 
988 		t0 = rdclock();
989 		clock_gettime(CLOCK_MONOTONIC, &ref_time);
990 
991 		status = dispatch_events(forks, timeout, interval, &times);
992 	}
993 
994 	disable_counters();
995 
996 	t1 = rdclock();
997 
998 	if (stat_config.walltime_run_table)
999 		stat_config.walltime_run[run_idx] = t1 - t0;
1000 
1001 	if (interval && stat_config.summary) {
1002 		stat_config.interval = 0;
1003 		stat_config.stop_read_counter = true;
1004 		init_stats(&walltime_nsecs_stats);
1005 		update_stats(&walltime_nsecs_stats, t1 - t0);
1006 
1007 		if (stat_config.aggr_mode == AGGR_GLOBAL)
1008 			evlist__save_aggr_prev_raw_counts(evsel_list);
1009 
1010 		evlist__copy_prev_raw_counts(evsel_list);
1011 		evlist__reset_prev_raw_counts(evsel_list);
1012 		runtime_stat_reset(&stat_config);
1013 		perf_stat__reset_shadow_per_stat(&rt_stat);
1014 	} else
1015 		update_stats(&walltime_nsecs_stats, t1 - t0);
1016 
1017 	/*
1018 	 * Closing a group leader splits the group, and as we only disable
1019 	 * group leaders, results in remaining events becoming enabled. To
1020 	 * avoid arbitrary skew, we must read all counters before closing any
1021 	 * group leaders.
1022 	 */
1023 	read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
1024 
1025 	/*
1026 	 * We need to keep evsel_list alive, because it's processed
1027 	 * later the evsel_list will be closed after.
1028 	 */
1029 	if (!STAT_RECORD)
1030 		evlist__close(evsel_list);
1031 
1032 	return WEXITSTATUS(status);
1033 }
1034 
1035 static int run_perf_stat(int argc, const char **argv, int run_idx)
1036 {
1037 	int ret;
1038 
1039 	if (pre_cmd) {
1040 		ret = system(pre_cmd);
1041 		if (ret)
1042 			return ret;
1043 	}
1044 
1045 	if (sync_run)
1046 		sync();
1047 
1048 	ret = __run_perf_stat(argc, argv, run_idx);
1049 	if (ret)
1050 		return ret;
1051 
1052 	if (post_cmd) {
1053 		ret = system(post_cmd);
1054 		if (ret)
1055 			return ret;
1056 	}
1057 
1058 	return ret;
1059 }
1060 
1061 static void print_counters(struct timespec *ts, int argc, const char **argv)
1062 {
1063 	/* Do not print anything if we record to the pipe. */
1064 	if (STAT_RECORD && perf_stat.data.is_pipe)
1065 		return;
1066 	if (stat_config.quiet)
1067 		return;
1068 
1069 	evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1070 }
1071 
1072 static volatile int signr = -1;
1073 
1074 static void skip_signal(int signo)
1075 {
1076 	if ((child_pid == -1) || stat_config.interval)
1077 		done = 1;
1078 
1079 	signr = signo;
1080 	/*
1081 	 * render child_pid harmless
1082 	 * won't send SIGTERM to a random
1083 	 * process in case of race condition
1084 	 * and fast PID recycling
1085 	 */
1086 	child_pid = -1;
1087 }
1088 
1089 static void sig_atexit(void)
1090 {
1091 	sigset_t set, oset;
1092 
1093 	/*
1094 	 * avoid race condition with SIGCHLD handler
1095 	 * in skip_signal() which is modifying child_pid
1096 	 * goal is to avoid send SIGTERM to a random
1097 	 * process
1098 	 */
1099 	sigemptyset(&set);
1100 	sigaddset(&set, SIGCHLD);
1101 	sigprocmask(SIG_BLOCK, &set, &oset);
1102 
1103 	if (child_pid != -1)
1104 		kill(child_pid, SIGTERM);
1105 
1106 	sigprocmask(SIG_SETMASK, &oset, NULL);
1107 
1108 	if (signr == -1)
1109 		return;
1110 
1111 	signal(signr, SIG_DFL);
1112 	kill(getpid(), signr);
1113 }
1114 
1115 void perf_stat__set_big_num(int set)
1116 {
1117 	stat_config.big_num = (set != 0);
1118 }
1119 
1120 void perf_stat__set_no_csv_summary(int set)
1121 {
1122 	stat_config.no_csv_summary = (set != 0);
1123 }
1124 
1125 static int stat__set_big_num(const struct option *opt __maybe_unused,
1126 			     const char *s __maybe_unused, int unset)
1127 {
1128 	big_num_opt = unset ? 0 : 1;
1129 	perf_stat__set_big_num(!unset);
1130 	return 0;
1131 }
1132 
1133 static int enable_metric_only(const struct option *opt __maybe_unused,
1134 			      const char *s __maybe_unused, int unset)
1135 {
1136 	force_metric_only = true;
1137 	stat_config.metric_only = !unset;
1138 	return 0;
1139 }
1140 
1141 static int parse_metric_groups(const struct option *opt,
1142 			       const char *str,
1143 			       int unset __maybe_unused)
1144 {
1145 	return metricgroup__parse_groups(opt, str,
1146 					 stat_config.metric_no_group,
1147 					 stat_config.metric_no_merge,
1148 					 &stat_config.metric_events);
1149 }
1150 
1151 static int parse_control_option(const struct option *opt,
1152 				const char *str,
1153 				int unset __maybe_unused)
1154 {
1155 	struct perf_stat_config *config = opt->value;
1156 
1157 	return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1158 }
1159 
1160 static int parse_stat_cgroups(const struct option *opt,
1161 			      const char *str, int unset)
1162 {
1163 	if (stat_config.cgroup_list) {
1164 		pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1165 		return -1;
1166 	}
1167 
1168 	return parse_cgroups(opt, str, unset);
1169 }
1170 
1171 static struct option stat_options[] = {
1172 	OPT_BOOLEAN('T', "transaction", &transaction_run,
1173 		    "hardware transaction statistics"),
1174 	OPT_CALLBACK('e', "event", &evsel_list, "event",
1175 		     "event selector. use 'perf list' to list available events",
1176 		     parse_events_option),
1177 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1178 		     "event filter", parse_filter),
1179 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1180 		    "child tasks do not inherit counters"),
1181 	OPT_STRING('p', "pid", &target.pid, "pid",
1182 		   "stat events on existing process id"),
1183 	OPT_STRING('t', "tid", &target.tid, "tid",
1184 		   "stat events on existing thread id"),
1185 #ifdef HAVE_BPF_SKEL
1186 	OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1187 		   "stat events on existing bpf program id"),
1188 	OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1189 		    "use bpf program to count events"),
1190 	OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1191 		   "path to perf_event_attr map"),
1192 #endif
1193 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1194 		    "system-wide collection from all CPUs"),
1195 	OPT_BOOLEAN('g', "group", &group,
1196 		    "put the counters into a counter group"),
1197 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
1198 		    "Use --no-scale to disable counter scaling for multiplexing"),
1199 	OPT_INCR('v', "verbose", &verbose,
1200 		    "be more verbose (show counter open errors, etc)"),
1201 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
1202 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1203 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1204 		    "display details about each run (only with -r option)"),
1205 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
1206 		    "null run - dont start any counters"),
1207 	OPT_INCR('d', "detailed", &detailed_run,
1208 		    "detailed run - start a lot of events"),
1209 	OPT_BOOLEAN('S', "sync", &sync_run,
1210 		    "call sync() before starting a run"),
1211 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1212 			   "print large numbers with thousands\' separators",
1213 			   stat__set_big_num),
1214 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1215 		    "list of cpus to monitor in system-wide"),
1216 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1217 		    "disable CPU count aggregation", AGGR_NONE),
1218 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1219 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1220 		   "print counts with custom separator"),
1221 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1222 		     "monitor event in cgroup name only", parse_stat_cgroups),
1223 	OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1224 		    "expand events for each cgroup"),
1225 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1226 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1227 	OPT_INTEGER(0, "log-fd", &output_fd,
1228 		    "log output to fd, instead of stderr"),
1229 	OPT_STRING(0, "pre", &pre_cmd, "command",
1230 			"command to run prior to the measured command"),
1231 	OPT_STRING(0, "post", &post_cmd, "command",
1232 			"command to run after to the measured command"),
1233 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1234 		    "print counts at regular interval in ms "
1235 		    "(overhead is possible for values <= 100ms)"),
1236 	OPT_INTEGER(0, "interval-count", &stat_config.times,
1237 		    "print counts for fixed number of times"),
1238 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1239 		    "clear screen in between new interval"),
1240 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1241 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1242 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1243 		     "aggregate counts per processor socket", AGGR_SOCKET),
1244 	OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1245 		     "aggregate counts per processor die", AGGR_DIE),
1246 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1247 		     "aggregate counts per physical processor core", AGGR_CORE),
1248 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1249 		     "aggregate counts per thread", AGGR_THREAD),
1250 	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1251 		     "aggregate counts per numa node", AGGR_NODE),
1252 	OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1253 		    "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1254 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1255 			"Only print computed metrics. No raw values", enable_metric_only),
1256 	OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1257 		       "don't group metric events, impacts multiplexing"),
1258 	OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1259 		       "don't try to share events between metrics in a group"),
1260 	OPT_BOOLEAN(0, "topdown", &topdown_run,
1261 			"measure top-down statistics"),
1262 	OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1263 			"Set the metrics level for the top-down statistics (0: max level)"),
1264 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1265 			"measure SMI cost"),
1266 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1267 		     "monitor specified metrics or metric groups (separated by ,)",
1268 		     parse_metric_groups),
1269 	OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1270 			 "Configure all used events to run in kernel space.",
1271 			 PARSE_OPT_EXCLUSIVE),
1272 	OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1273 			 "Configure all used events to run in user space.",
1274 			 PARSE_OPT_EXCLUSIVE),
1275 	OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1276 		    "Use with 'percore' event qualifier to show the event "
1277 		    "counts of one hardware thread by sum up total hardware "
1278 		    "threads of same physical core"),
1279 	OPT_BOOLEAN(0, "summary", &stat_config.summary,
1280 		       "print summary for interval mode"),
1281 	OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1282 		       "don't print 'summary' for CSV summary output"),
1283 	OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1284 			"don't print output (useful with record)"),
1285 #ifdef HAVE_LIBPFM
1286 	OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1287 		"libpfm4 event selector. use 'perf list' to list available events",
1288 		parse_libpfm_events_option),
1289 #endif
1290 	OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1291 		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1292 		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1293 		     "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1294 		      parse_control_option),
1295 	OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1296 			    "measure I/O performance metrics provided by arch/platform",
1297 			    iostat_parse),
1298 	OPT_END()
1299 };
1300 
1301 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1302 				 struct perf_cpu_map *map, int cpu)
1303 {
1304 	return cpu_map__get_socket(map, cpu, NULL);
1305 }
1306 
1307 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1308 			      struct perf_cpu_map *map, int cpu)
1309 {
1310 	return cpu_map__get_die(map, cpu, NULL);
1311 }
1312 
1313 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1314 			       struct perf_cpu_map *map, int cpu)
1315 {
1316 	return cpu_map__get_core(map, cpu, NULL);
1317 }
1318 
1319 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1320 			       struct perf_cpu_map *map, int cpu)
1321 {
1322 	return cpu_map__get_node(map, cpu, NULL);
1323 }
1324 
1325 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1326 			       aggr_get_id_t get_id, struct perf_cpu_map *map, int idx)
1327 {
1328 	int cpu;
1329 	struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1330 
1331 	if (idx >= map->nr)
1332 		return id;
1333 
1334 	cpu = map->map[idx];
1335 
1336 	if (cpu_map__aggr_cpu_id_is_empty(config->cpus_aggr_map->map[cpu]))
1337 		config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
1338 
1339 	id = config->cpus_aggr_map->map[cpu];
1340 	return id;
1341 }
1342 
1343 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1344 					struct perf_cpu_map *map, int idx)
1345 {
1346 	return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
1347 }
1348 
1349 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1350 					struct perf_cpu_map *map, int idx)
1351 {
1352 	return perf_stat__get_aggr(config, perf_stat__get_die, map, idx);
1353 }
1354 
1355 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1356 				      struct perf_cpu_map *map, int idx)
1357 {
1358 	return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
1359 }
1360 
1361 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1362 				      struct perf_cpu_map *map, int idx)
1363 {
1364 	return perf_stat__get_aggr(config, perf_stat__get_node, map, idx);
1365 }
1366 
1367 static bool term_percore_set(void)
1368 {
1369 	struct evsel *counter;
1370 
1371 	evlist__for_each_entry(evsel_list, counter) {
1372 		if (counter->percore)
1373 			return true;
1374 	}
1375 
1376 	return false;
1377 }
1378 
1379 static int perf_stat_init_aggr_mode(void)
1380 {
1381 	int nr;
1382 
1383 	switch (stat_config.aggr_mode) {
1384 	case AGGR_SOCKET:
1385 		if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1386 			perror("cannot build socket map");
1387 			return -1;
1388 		}
1389 		stat_config.aggr_get_id = perf_stat__get_socket_cached;
1390 		break;
1391 	case AGGR_DIE:
1392 		if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1393 			perror("cannot build die map");
1394 			return -1;
1395 		}
1396 		stat_config.aggr_get_id = perf_stat__get_die_cached;
1397 		break;
1398 	case AGGR_CORE:
1399 		if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1400 			perror("cannot build core map");
1401 			return -1;
1402 		}
1403 		stat_config.aggr_get_id = perf_stat__get_core_cached;
1404 		break;
1405 	case AGGR_NODE:
1406 		if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1407 			perror("cannot build core map");
1408 			return -1;
1409 		}
1410 		stat_config.aggr_get_id = perf_stat__get_node_cached;
1411 		break;
1412 	case AGGR_NONE:
1413 		if (term_percore_set()) {
1414 			if (cpu_map__build_core_map(evsel_list->core.cpus,
1415 						    &stat_config.aggr_map)) {
1416 				perror("cannot build core map");
1417 				return -1;
1418 			}
1419 			stat_config.aggr_get_id = perf_stat__get_core_cached;
1420 		}
1421 		break;
1422 	case AGGR_GLOBAL:
1423 	case AGGR_THREAD:
1424 	case AGGR_UNSET:
1425 	default:
1426 		break;
1427 	}
1428 
1429 	/*
1430 	 * The evsel_list->cpus is the base we operate on,
1431 	 * taking the highest cpu number to be the size of
1432 	 * the aggregation translate cpumap.
1433 	 */
1434 	nr = perf_cpu_map__max(evsel_list->core.cpus);
1435 	stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1436 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1437 }
1438 
1439 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1440 {
1441 	if (map) {
1442 		WARN_ONCE(refcount_read(&map->refcnt) != 0,
1443 			  "cpu_aggr_map refcnt unbalanced\n");
1444 		free(map);
1445 	}
1446 }
1447 
1448 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1449 {
1450 	if (map && refcount_dec_and_test(&map->refcnt))
1451 		cpu_aggr_map__delete(map);
1452 }
1453 
1454 static void perf_stat__exit_aggr_mode(void)
1455 {
1456 	cpu_aggr_map__put(stat_config.aggr_map);
1457 	cpu_aggr_map__put(stat_config.cpus_aggr_map);
1458 	stat_config.aggr_map = NULL;
1459 	stat_config.cpus_aggr_map = NULL;
1460 }
1461 
1462 static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx)
1463 {
1464 	int cpu;
1465 
1466 	if (idx > map->nr)
1467 		return -1;
1468 
1469 	cpu = map->map[idx];
1470 
1471 	if (cpu >= env->nr_cpus_avail)
1472 		return -1;
1473 
1474 	return cpu;
1475 }
1476 
1477 static struct aggr_cpu_id perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data)
1478 {
1479 	struct perf_env *env = data;
1480 	int cpu = perf_env__get_cpu(env, map, idx);
1481 	struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1482 
1483 	if (cpu != -1)
1484 		id.socket = env->cpu[cpu].socket_id;
1485 
1486 	return id;
1487 }
1488 
1489 static struct aggr_cpu_id perf_env__get_die(struct perf_cpu_map *map, int idx, void *data)
1490 {
1491 	struct perf_env *env = data;
1492 	struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1493 	int cpu = perf_env__get_cpu(env, map, idx);
1494 
1495 	if (cpu != -1) {
1496 		/*
1497 		 * die_id is relative to socket, so start
1498 		 * with the socket ID and then add die to
1499 		 * make a unique ID.
1500 		 */
1501 		id.socket = env->cpu[cpu].socket_id;
1502 		id.die = env->cpu[cpu].die_id;
1503 	}
1504 
1505 	return id;
1506 }
1507 
1508 static struct aggr_cpu_id perf_env__get_core(struct perf_cpu_map *map, int idx, void *data)
1509 {
1510 	struct perf_env *env = data;
1511 	struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1512 	int cpu = perf_env__get_cpu(env, map, idx);
1513 
1514 	if (cpu != -1) {
1515 		/*
1516 		 * core_id is relative to socket and die,
1517 		 * we need a global id. So we set
1518 		 * socket, die id and core id
1519 		 */
1520 		id.socket = env->cpu[cpu].socket_id;
1521 		id.die = env->cpu[cpu].die_id;
1522 		id.core = env->cpu[cpu].core_id;
1523 	}
1524 
1525 	return id;
1526 }
1527 
1528 static struct aggr_cpu_id perf_env__get_node(struct perf_cpu_map *map, int idx, void *data)
1529 {
1530 	int cpu = perf_env__get_cpu(data, map, idx);
1531 	struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1532 
1533 	id.node = perf_env__numa_node(data, cpu);
1534 	return id;
1535 }
1536 
1537 static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus,
1538 				      struct cpu_aggr_map **sockp)
1539 {
1540 	return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1541 }
1542 
1543 static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus,
1544 				   struct cpu_aggr_map **diep)
1545 {
1546 	return cpu_map__build_map(cpus, diep, perf_env__get_die, env);
1547 }
1548 
1549 static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus,
1550 				    struct cpu_aggr_map **corep)
1551 {
1552 	return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1553 }
1554 
1555 static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus,
1556 				    struct cpu_aggr_map **nodep)
1557 {
1558 	return cpu_map__build_map(cpus, nodep, perf_env__get_node, env);
1559 }
1560 
1561 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1562 				      struct perf_cpu_map *map, int idx)
1563 {
1564 	return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1565 }
1566 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1567 				   struct perf_cpu_map *map, int idx)
1568 {
1569 	return perf_env__get_die(map, idx, &perf_stat.session->header.env);
1570 }
1571 
1572 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1573 				    struct perf_cpu_map *map, int idx)
1574 {
1575 	return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1576 }
1577 
1578 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1579 				    struct perf_cpu_map *map, int idx)
1580 {
1581 	return perf_env__get_node(map, idx, &perf_stat.session->header.env);
1582 }
1583 
1584 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1585 {
1586 	struct perf_env *env = &st->session->header.env;
1587 
1588 	switch (stat_config.aggr_mode) {
1589 	case AGGR_SOCKET:
1590 		if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1591 			perror("cannot build socket map");
1592 			return -1;
1593 		}
1594 		stat_config.aggr_get_id = perf_stat__get_socket_file;
1595 		break;
1596 	case AGGR_DIE:
1597 		if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1598 			perror("cannot build die map");
1599 			return -1;
1600 		}
1601 		stat_config.aggr_get_id = perf_stat__get_die_file;
1602 		break;
1603 	case AGGR_CORE:
1604 		if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1605 			perror("cannot build core map");
1606 			return -1;
1607 		}
1608 		stat_config.aggr_get_id = perf_stat__get_core_file;
1609 		break;
1610 	case AGGR_NODE:
1611 		if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1612 			perror("cannot build core map");
1613 			return -1;
1614 		}
1615 		stat_config.aggr_get_id = perf_stat__get_node_file;
1616 		break;
1617 	case AGGR_NONE:
1618 	case AGGR_GLOBAL:
1619 	case AGGR_THREAD:
1620 	case AGGR_UNSET:
1621 	default:
1622 		break;
1623 	}
1624 
1625 	return 0;
1626 }
1627 
1628 /*
1629  * Add default attributes, if there were no attributes specified or
1630  * if -d/--detailed, -d -d or -d -d -d is used:
1631  */
1632 static int add_default_attributes(void)
1633 {
1634 	int err;
1635 	struct perf_event_attr default_attrs0[] = {
1636 
1637   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1638   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1639   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1640   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1641 
1642   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1643 };
1644 	struct perf_event_attr frontend_attrs[] = {
1645   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1646 };
1647 	struct perf_event_attr backend_attrs[] = {
1648   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1649 };
1650 	struct perf_event_attr default_attrs1[] = {
1651   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1652   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1653   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1654 
1655 };
1656 	struct perf_event_attr default_sw_attrs[] = {
1657   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1658   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1659   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1660   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1661 };
1662 
1663 /*
1664  * Detailed stats (-d), covering the L1 and last level data caches:
1665  */
1666 	struct perf_event_attr detailed_attrs[] = {
1667 
1668   { .type = PERF_TYPE_HW_CACHE,
1669     .config =
1670 	 PERF_COUNT_HW_CACHE_L1D		<<  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_L1D		<<  0  |
1677 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1678 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1679 
1680   { .type = PERF_TYPE_HW_CACHE,
1681     .config =
1682 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1683 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1684 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1685 
1686   { .type = PERF_TYPE_HW_CACHE,
1687     .config =
1688 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1689 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1690 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1691 };
1692 
1693 /*
1694  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1695  */
1696 	struct perf_event_attr very_detailed_attrs[] = {
1697 
1698   { .type = PERF_TYPE_HW_CACHE,
1699     .config =
1700 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1701 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1702 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1703 
1704   { .type = PERF_TYPE_HW_CACHE,
1705     .config =
1706 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1707 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1708 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1709 
1710   { .type = PERF_TYPE_HW_CACHE,
1711     .config =
1712 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1713 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1714 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1715 
1716   { .type = PERF_TYPE_HW_CACHE,
1717     .config =
1718 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1719 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1720 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1721 
1722   { .type = PERF_TYPE_HW_CACHE,
1723     .config =
1724 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1725 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1726 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1727 
1728   { .type = PERF_TYPE_HW_CACHE,
1729     .config =
1730 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1731 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1732 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1733 
1734 };
1735 
1736 /*
1737  * Very, very detailed stats (-d -d -d), adding prefetch events:
1738  */
1739 	struct perf_event_attr very_very_detailed_attrs[] = {
1740 
1741   { .type = PERF_TYPE_HW_CACHE,
1742     .config =
1743 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1744 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1745 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1746 
1747   { .type = PERF_TYPE_HW_CACHE,
1748     .config =
1749 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1750 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1751 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1752 };
1753 	struct parse_events_error errinfo;
1754 
1755 	/* Set attrs if no event is selected and !null_run: */
1756 	if (stat_config.null_run)
1757 		return 0;
1758 
1759 	bzero(&errinfo, sizeof(errinfo));
1760 	if (transaction_run) {
1761 		/* Handle -T as -M transaction. Once platform specific metrics
1762 		 * support has been added to the json files, all architectures
1763 		 * will use this approach. To determine transaction support
1764 		 * on an architecture test for such a metric name.
1765 		 */
1766 		if (metricgroup__has_metric("transaction")) {
1767 			struct option opt = { .value = &evsel_list };
1768 
1769 			return metricgroup__parse_groups(&opt, "transaction",
1770 							 stat_config.metric_no_group,
1771 							stat_config.metric_no_merge,
1772 							 &stat_config.metric_events);
1773 		}
1774 
1775 		if (pmu_have_event("cpu", "cycles-ct") &&
1776 		    pmu_have_event("cpu", "el-start"))
1777 			err = parse_events(evsel_list, transaction_attrs,
1778 					   &errinfo);
1779 		else
1780 			err = parse_events(evsel_list,
1781 					   transaction_limited_attrs,
1782 					   &errinfo);
1783 		if (err) {
1784 			fprintf(stderr, "Cannot set up transaction events\n");
1785 			parse_events_print_error(&errinfo, transaction_attrs);
1786 			return -1;
1787 		}
1788 		return 0;
1789 	}
1790 
1791 	if (smi_cost) {
1792 		int smi;
1793 
1794 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1795 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1796 			return -1;
1797 		}
1798 
1799 		if (!smi) {
1800 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1801 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1802 				return -1;
1803 			}
1804 			smi_reset = true;
1805 		}
1806 
1807 		if (pmu_have_event("msr", "aperf") &&
1808 		    pmu_have_event("msr", "smi")) {
1809 			if (!force_metric_only)
1810 				stat_config.metric_only = true;
1811 			err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1812 		} else {
1813 			fprintf(stderr, "To measure SMI cost, it needs "
1814 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1815 			parse_events_print_error(&errinfo, smi_cost_attrs);
1816 			return -1;
1817 		}
1818 		if (err) {
1819 			parse_events_print_error(&errinfo, smi_cost_attrs);
1820 			fprintf(stderr, "Cannot set up SMI cost events\n");
1821 			return -1;
1822 		}
1823 		return 0;
1824 	}
1825 
1826 	if (topdown_run) {
1827 		const char **metric_attrs = topdown_metric_attrs;
1828 		unsigned int max_level = 1;
1829 		char *str = NULL;
1830 		bool warn = false;
1831 
1832 		if (!force_metric_only)
1833 			stat_config.metric_only = true;
1834 
1835 		if (pmu_have_event("cpu", topdown_metric_L2_attrs[5])) {
1836 			metric_attrs = topdown_metric_L2_attrs;
1837 			max_level = 2;
1838 		}
1839 
1840 		if (stat_config.topdown_level > max_level) {
1841 			pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1842 			return -1;
1843 		} else if (!stat_config.topdown_level)
1844 			stat_config.topdown_level = max_level;
1845 
1846 		if (topdown_filter_events(metric_attrs, &str, 1) < 0) {
1847 			pr_err("Out of memory\n");
1848 			return -1;
1849 		}
1850 		if (metric_attrs[0] && str) {
1851 			if (!stat_config.interval && !stat_config.metric_only) {
1852 				fprintf(stat_config.output,
1853 					"Topdown accuracy may decrease when measuring long periods.\n"
1854 					"Please print the result regularly, e.g. -I1000\n");
1855 			}
1856 			goto setup_metrics;
1857 		}
1858 
1859 		zfree(&str);
1860 
1861 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1862 		    stat_config.aggr_mode != AGGR_CORE) {
1863 			pr_err("top down event configuration requires --per-core mode\n");
1864 			return -1;
1865 		}
1866 		stat_config.aggr_mode = AGGR_CORE;
1867 		if (nr_cgroups || !target__has_cpu(&target)) {
1868 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1869 			return -1;
1870 		}
1871 
1872 		if (topdown_filter_events(topdown_attrs, &str,
1873 				arch_topdown_check_group(&warn)) < 0) {
1874 			pr_err("Out of memory\n");
1875 			return -1;
1876 		}
1877 		if (topdown_attrs[0] && str) {
1878 			if (warn)
1879 				arch_topdown_group_warn();
1880 setup_metrics:
1881 			err = parse_events(evsel_list, str, &errinfo);
1882 			if (err) {
1883 				fprintf(stderr,
1884 					"Cannot set up top down events %s: %d\n",
1885 					str, err);
1886 				parse_events_print_error(&errinfo, str);
1887 				free(str);
1888 				return -1;
1889 			}
1890 		} else {
1891 			fprintf(stderr, "System does not support topdown\n");
1892 			return -1;
1893 		}
1894 		free(str);
1895 	}
1896 
1897 	if (!evsel_list->core.nr_entries) {
1898 		if (perf_pmu__has_hybrid()) {
1899 			const char *hybrid_str = "cycles,instructions,branches,branch-misses";
1900 
1901 			if (target__has_cpu(&target))
1902 				default_sw_attrs[0].config = PERF_COUNT_SW_CPU_CLOCK;
1903 
1904 			if (evlist__add_default_attrs(evsel_list,
1905 						      default_sw_attrs) < 0) {
1906 				return -1;
1907 			}
1908 
1909 			err = parse_events(evsel_list, hybrid_str, &errinfo);
1910 			if (err) {
1911 				fprintf(stderr,
1912 					"Cannot set up hybrid events %s: %d\n",
1913 					hybrid_str, err);
1914 				parse_events_print_error(&errinfo, hybrid_str);
1915 				return -1;
1916 			}
1917 			return err;
1918 		}
1919 
1920 		if (target__has_cpu(&target))
1921 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1922 
1923 		if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1924 			return -1;
1925 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1926 			if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1927 				return -1;
1928 		}
1929 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1930 			if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1931 				return -1;
1932 		}
1933 		if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1934 			return -1;
1935 
1936 		stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1937 		if (arch_evlist__add_default_attrs(evsel_list) < 0)
1938 			return -1;
1939 	}
1940 
1941 	/* Detailed events get appended to the event list: */
1942 
1943 	if (detailed_run <  1)
1944 		return 0;
1945 
1946 	/* Append detailed run extra attributes: */
1947 	if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1948 		return -1;
1949 
1950 	if (detailed_run < 2)
1951 		return 0;
1952 
1953 	/* Append very detailed run extra attributes: */
1954 	if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1955 		return -1;
1956 
1957 	if (detailed_run < 3)
1958 		return 0;
1959 
1960 	/* Append very, very detailed run extra attributes: */
1961 	return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1962 }
1963 
1964 static const char * const stat_record_usage[] = {
1965 	"perf stat record [<options>]",
1966 	NULL,
1967 };
1968 
1969 static void init_features(struct perf_session *session)
1970 {
1971 	int feat;
1972 
1973 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1974 		perf_header__set_feat(&session->header, feat);
1975 
1976 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1977 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1978 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1979 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1980 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1981 }
1982 
1983 static int __cmd_record(int argc, const char **argv)
1984 {
1985 	struct perf_session *session;
1986 	struct perf_data *data = &perf_stat.data;
1987 
1988 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
1989 			     PARSE_OPT_STOP_AT_NON_OPTION);
1990 
1991 	if (output_name)
1992 		data->path = output_name;
1993 
1994 	if (stat_config.run_count != 1 || forever) {
1995 		pr_err("Cannot use -r option with perf stat record.\n");
1996 		return -1;
1997 	}
1998 
1999 	session = perf_session__new(data, false, NULL);
2000 	if (IS_ERR(session)) {
2001 		pr_err("Perf session creation failed\n");
2002 		return PTR_ERR(session);
2003 	}
2004 
2005 	init_features(session);
2006 
2007 	session->evlist   = evsel_list;
2008 	perf_stat.session = session;
2009 	perf_stat.record  = true;
2010 	return argc;
2011 }
2012 
2013 static int process_stat_round_event(struct perf_session *session,
2014 				    union perf_event *event)
2015 {
2016 	struct perf_record_stat_round *stat_round = &event->stat_round;
2017 	struct evsel *counter;
2018 	struct timespec tsh, *ts = NULL;
2019 	const char **argv = session->header.env.cmdline_argv;
2020 	int argc = session->header.env.nr_cmdline;
2021 
2022 	evlist__for_each_entry(evsel_list, counter)
2023 		perf_stat_process_counter(&stat_config, counter);
2024 
2025 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2026 		update_stats(&walltime_nsecs_stats, stat_round->time);
2027 
2028 	if (stat_config.interval && stat_round->time) {
2029 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2030 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2031 		ts = &tsh;
2032 	}
2033 
2034 	print_counters(ts, argc, argv);
2035 	return 0;
2036 }
2037 
2038 static
2039 int process_stat_config_event(struct perf_session *session,
2040 			      union perf_event *event)
2041 {
2042 	struct perf_tool *tool = session->tool;
2043 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2044 
2045 	perf_event__read_stat_config(&stat_config, &event->stat_config);
2046 
2047 	if (perf_cpu_map__empty(st->cpus)) {
2048 		if (st->aggr_mode != AGGR_UNSET)
2049 			pr_warning("warning: processing task data, aggregation mode not set\n");
2050 		return 0;
2051 	}
2052 
2053 	if (st->aggr_mode != AGGR_UNSET)
2054 		stat_config.aggr_mode = st->aggr_mode;
2055 
2056 	if (perf_stat.data.is_pipe)
2057 		perf_stat_init_aggr_mode();
2058 	else
2059 		perf_stat_init_aggr_mode_file(st);
2060 
2061 	return 0;
2062 }
2063 
2064 static int set_maps(struct perf_stat *st)
2065 {
2066 	if (!st->cpus || !st->threads)
2067 		return 0;
2068 
2069 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2070 		return -EINVAL;
2071 
2072 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2073 
2074 	if (evlist__alloc_stats(evsel_list, true))
2075 		return -ENOMEM;
2076 
2077 	st->maps_allocated = true;
2078 	return 0;
2079 }
2080 
2081 static
2082 int process_thread_map_event(struct perf_session *session,
2083 			     union perf_event *event)
2084 {
2085 	struct perf_tool *tool = session->tool;
2086 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2087 
2088 	if (st->threads) {
2089 		pr_warning("Extra thread map event, ignoring.\n");
2090 		return 0;
2091 	}
2092 
2093 	st->threads = thread_map__new_event(&event->thread_map);
2094 	if (!st->threads)
2095 		return -ENOMEM;
2096 
2097 	return set_maps(st);
2098 }
2099 
2100 static
2101 int process_cpu_map_event(struct perf_session *session,
2102 			  union perf_event *event)
2103 {
2104 	struct perf_tool *tool = session->tool;
2105 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2106 	struct perf_cpu_map *cpus;
2107 
2108 	if (st->cpus) {
2109 		pr_warning("Extra cpu map event, ignoring.\n");
2110 		return 0;
2111 	}
2112 
2113 	cpus = cpu_map__new_data(&event->cpu_map.data);
2114 	if (!cpus)
2115 		return -ENOMEM;
2116 
2117 	st->cpus = cpus;
2118 	return set_maps(st);
2119 }
2120 
2121 static const char * const stat_report_usage[] = {
2122 	"perf stat report [<options>]",
2123 	NULL,
2124 };
2125 
2126 static struct perf_stat perf_stat = {
2127 	.tool = {
2128 		.attr		= perf_event__process_attr,
2129 		.event_update	= perf_event__process_event_update,
2130 		.thread_map	= process_thread_map_event,
2131 		.cpu_map	= process_cpu_map_event,
2132 		.stat_config	= process_stat_config_event,
2133 		.stat		= perf_event__process_stat_event,
2134 		.stat_round	= process_stat_round_event,
2135 	},
2136 	.aggr_mode = AGGR_UNSET,
2137 };
2138 
2139 static int __cmd_report(int argc, const char **argv)
2140 {
2141 	struct perf_session *session;
2142 	const struct option options[] = {
2143 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2144 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2145 		     "aggregate counts per processor socket", AGGR_SOCKET),
2146 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2147 		     "aggregate counts per processor die", AGGR_DIE),
2148 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2149 		     "aggregate counts per physical processor core", AGGR_CORE),
2150 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2151 		     "aggregate counts per numa node", AGGR_NODE),
2152 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2153 		     "disable CPU count aggregation", AGGR_NONE),
2154 	OPT_END()
2155 	};
2156 	struct stat st;
2157 	int ret;
2158 
2159 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2160 
2161 	if (!input_name || !strlen(input_name)) {
2162 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2163 			input_name = "-";
2164 		else
2165 			input_name = "perf.data";
2166 	}
2167 
2168 	perf_stat.data.path = input_name;
2169 	perf_stat.data.mode = PERF_DATA_MODE_READ;
2170 
2171 	session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2172 	if (IS_ERR(session))
2173 		return PTR_ERR(session);
2174 
2175 	perf_stat.session  = session;
2176 	stat_config.output = stderr;
2177 	evsel_list         = session->evlist;
2178 
2179 	ret = perf_session__process_events(session);
2180 	if (ret)
2181 		return ret;
2182 
2183 	perf_session__delete(session);
2184 	return 0;
2185 }
2186 
2187 static void setup_system_wide(int forks)
2188 {
2189 	/*
2190 	 * Make system wide (-a) the default target if
2191 	 * no target was specified and one of following
2192 	 * conditions is met:
2193 	 *
2194 	 *   - there's no workload specified
2195 	 *   - there is workload specified but all requested
2196 	 *     events are system wide events
2197 	 */
2198 	if (!target__none(&target))
2199 		return;
2200 
2201 	if (!forks)
2202 		target.system_wide = true;
2203 	else {
2204 		struct evsel *counter;
2205 
2206 		evlist__for_each_entry(evsel_list, counter) {
2207 			if (!counter->core.system_wide &&
2208 			    strcmp(counter->name, "duration_time")) {
2209 				return;
2210 			}
2211 		}
2212 
2213 		if (evsel_list->core.nr_entries)
2214 			target.system_wide = true;
2215 	}
2216 }
2217 
2218 int cmd_stat(int argc, const char **argv)
2219 {
2220 	const char * const stat_usage[] = {
2221 		"perf stat [<options>] [<command>]",
2222 		NULL
2223 	};
2224 	int status = -EINVAL, run_idx, err;
2225 	const char *mode;
2226 	FILE *output = stderr;
2227 	unsigned int interval, timeout;
2228 	const char * const stat_subcommands[] = { "record", "report" };
2229 	char errbuf[BUFSIZ];
2230 
2231 	setlocale(LC_ALL, "");
2232 
2233 	evsel_list = evlist__new();
2234 	if (evsel_list == NULL)
2235 		return -ENOMEM;
2236 
2237 	parse_events__shrink_config_terms();
2238 
2239 	/* String-parsing callback-based options would segfault when negated */
2240 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2241 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2242 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2243 
2244 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2245 					(const char **) stat_usage,
2246 					PARSE_OPT_STOP_AT_NON_OPTION);
2247 	perf_stat__collect_metric_expr(evsel_list);
2248 	perf_stat__init_shadow_stats();
2249 
2250 	if (stat_config.csv_sep) {
2251 		stat_config.csv_output = true;
2252 		if (!strcmp(stat_config.csv_sep, "\\t"))
2253 			stat_config.csv_sep = "\t";
2254 	} else
2255 		stat_config.csv_sep = DEFAULT_SEPARATOR;
2256 
2257 	if (argc && !strncmp(argv[0], "rec", 3)) {
2258 		argc = __cmd_record(argc, argv);
2259 		if (argc < 0)
2260 			return -1;
2261 	} else if (argc && !strncmp(argv[0], "rep", 3))
2262 		return __cmd_report(argc, argv);
2263 
2264 	interval = stat_config.interval;
2265 	timeout = stat_config.timeout;
2266 
2267 	/*
2268 	 * For record command the -o is already taken care of.
2269 	 */
2270 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2271 		output = NULL;
2272 
2273 	if (output_name && output_fd) {
2274 		fprintf(stderr, "cannot use both --output and --log-fd\n");
2275 		parse_options_usage(stat_usage, stat_options, "o", 1);
2276 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2277 		goto out;
2278 	}
2279 
2280 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2281 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2282 		goto out;
2283 	}
2284 
2285 	if (stat_config.metric_only && stat_config.run_count > 1) {
2286 		fprintf(stderr, "--metric-only is not supported with -r\n");
2287 		goto out;
2288 	}
2289 
2290 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2291 		fprintf(stderr, "--table is only supported with -r\n");
2292 		parse_options_usage(stat_usage, stat_options, "r", 1);
2293 		parse_options_usage(NULL, stat_options, "table", 0);
2294 		goto out;
2295 	}
2296 
2297 	if (output_fd < 0) {
2298 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2299 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2300 		goto out;
2301 	}
2302 
2303 	if (!output && !stat_config.quiet) {
2304 		struct timespec tm;
2305 		mode = append_file ? "a" : "w";
2306 
2307 		output = fopen(output_name, mode);
2308 		if (!output) {
2309 			perror("failed to create output file");
2310 			return -1;
2311 		}
2312 		clock_gettime(CLOCK_REALTIME, &tm);
2313 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2314 	} else if (output_fd > 0) {
2315 		mode = append_file ? "a" : "w";
2316 		output = fdopen(output_fd, mode);
2317 		if (!output) {
2318 			perror("Failed opening logfd");
2319 			return -errno;
2320 		}
2321 	}
2322 
2323 	stat_config.output = output;
2324 
2325 	/*
2326 	 * let the spreadsheet do the pretty-printing
2327 	 */
2328 	if (stat_config.csv_output) {
2329 		/* User explicitly passed -B? */
2330 		if (big_num_opt == 1) {
2331 			fprintf(stderr, "-B option not supported with -x\n");
2332 			parse_options_usage(stat_usage, stat_options, "B", 1);
2333 			parse_options_usage(NULL, stat_options, "x", 1);
2334 			goto out;
2335 		} else /* Nope, so disable big number formatting */
2336 			stat_config.big_num = false;
2337 	} else if (big_num_opt == 0) /* User passed --no-big-num */
2338 		stat_config.big_num = false;
2339 
2340 	err = target__validate(&target);
2341 	if (err) {
2342 		target__strerror(&target, err, errbuf, BUFSIZ);
2343 		pr_warning("%s\n", errbuf);
2344 	}
2345 
2346 	setup_system_wide(argc);
2347 
2348 	/*
2349 	 * Display user/system times only for single
2350 	 * run and when there's specified tracee.
2351 	 */
2352 	if ((stat_config.run_count == 1) && target__none(&target))
2353 		stat_config.ru_display = true;
2354 
2355 	if (stat_config.run_count < 0) {
2356 		pr_err("Run count must be a positive number\n");
2357 		parse_options_usage(stat_usage, stat_options, "r", 1);
2358 		goto out;
2359 	} else if (stat_config.run_count == 0) {
2360 		forever = true;
2361 		stat_config.run_count = 1;
2362 	}
2363 
2364 	if (stat_config.walltime_run_table) {
2365 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2366 		if (!stat_config.walltime_run) {
2367 			pr_err("failed to setup -r option");
2368 			goto out;
2369 		}
2370 	}
2371 
2372 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2373 		!target__has_task(&target)) {
2374 		if (!target.system_wide || target.cpu_list) {
2375 			fprintf(stderr, "The --per-thread option is only "
2376 				"available when monitoring via -p -t -a "
2377 				"options or only --per-thread.\n");
2378 			parse_options_usage(NULL, stat_options, "p", 1);
2379 			parse_options_usage(NULL, stat_options, "t", 1);
2380 			goto out;
2381 		}
2382 	}
2383 
2384 	/*
2385 	 * no_aggr, cgroup are for system-wide only
2386 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2387 	 */
2388 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2389 	      stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2390 	    !target__has_cpu(&target)) {
2391 		fprintf(stderr, "both cgroup and no-aggregation "
2392 			"modes only available in system-wide mode\n");
2393 
2394 		parse_options_usage(stat_usage, stat_options, "G", 1);
2395 		parse_options_usage(NULL, stat_options, "A", 1);
2396 		parse_options_usage(NULL, stat_options, "a", 1);
2397 		goto out;
2398 	}
2399 
2400 	if (stat_config.iostat_run) {
2401 		status = iostat_prepare(evsel_list, &stat_config);
2402 		if (status)
2403 			goto out;
2404 		if (iostat_mode == IOSTAT_LIST) {
2405 			iostat_list(evsel_list, &stat_config);
2406 			goto out;
2407 		} else if (verbose)
2408 			iostat_list(evsel_list, &stat_config);
2409 	}
2410 
2411 	if (add_default_attributes())
2412 		goto out;
2413 
2414 	if (stat_config.cgroup_list) {
2415 		if (nr_cgroups > 0) {
2416 			pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2417 			parse_options_usage(stat_usage, stat_options, "G", 1);
2418 			parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2419 			goto out;
2420 		}
2421 
2422 		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2423 					  &stat_config.metric_events, true) < 0) {
2424 			parse_options_usage(stat_usage, stat_options,
2425 					    "for-each-cgroup", 0);
2426 			goto out;
2427 		}
2428 	}
2429 
2430 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2431 		target.per_thread = true;
2432 
2433 	if (evlist__create_maps(evsel_list, &target) < 0) {
2434 		if (target__has_task(&target)) {
2435 			pr_err("Problems finding threads of monitor\n");
2436 			parse_options_usage(stat_usage, stat_options, "p", 1);
2437 			parse_options_usage(NULL, stat_options, "t", 1);
2438 		} else if (target__has_cpu(&target)) {
2439 			perror("failed to parse CPUs map");
2440 			parse_options_usage(stat_usage, stat_options, "C", 1);
2441 			parse_options_usage(NULL, stat_options, "a", 1);
2442 		}
2443 		goto out;
2444 	}
2445 
2446 	evlist__check_cpu_maps(evsel_list);
2447 
2448 	if (perf_pmu__has_hybrid())
2449 		stat_config.no_merge = true;
2450 
2451 	/*
2452 	 * Initialize thread_map with comm names,
2453 	 * so we could print it out on output.
2454 	 */
2455 	if (stat_config.aggr_mode == AGGR_THREAD) {
2456 		thread_map__read_comms(evsel_list->core.threads);
2457 		if (target.system_wide) {
2458 			if (runtime_stat_new(&stat_config,
2459 				perf_thread_map__nr(evsel_list->core.threads))) {
2460 				goto out;
2461 			}
2462 		}
2463 	}
2464 
2465 	if (stat_config.aggr_mode == AGGR_NODE)
2466 		cpu__setup_cpunode_map();
2467 
2468 	if (stat_config.times && interval)
2469 		interval_count = true;
2470 	else if (stat_config.times && !interval) {
2471 		pr_err("interval-count option should be used together with "
2472 				"interval-print.\n");
2473 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2474 		parse_options_usage(stat_usage, stat_options, "I", 1);
2475 		goto out;
2476 	}
2477 
2478 	if (timeout && timeout < 100) {
2479 		if (timeout < 10) {
2480 			pr_err("timeout must be >= 10ms.\n");
2481 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2482 			goto out;
2483 		} else
2484 			pr_warning("timeout < 100ms. "
2485 				   "The overhead percentage could be high in some cases. "
2486 				   "Please proceed with caution.\n");
2487 	}
2488 	if (timeout && interval) {
2489 		pr_err("timeout option is not supported with interval-print.\n");
2490 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2491 		parse_options_usage(stat_usage, stat_options, "I", 1);
2492 		goto out;
2493 	}
2494 
2495 	if (evlist__alloc_stats(evsel_list, interval))
2496 		goto out;
2497 
2498 	if (perf_stat_init_aggr_mode())
2499 		goto out;
2500 
2501 	/*
2502 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2503 	 * while avoiding that older tools show confusing messages.
2504 	 *
2505 	 * However for pipe sessions we need to keep it zero,
2506 	 * because script's perf_evsel__check_attr is triggered
2507 	 * by attr->sample_type != 0, and we can't run it on
2508 	 * stat sessions.
2509 	 */
2510 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2511 
2512 	/*
2513 	 * We dont want to block the signals - that would cause
2514 	 * child tasks to inherit that and Ctrl-C would not work.
2515 	 * What we want is for Ctrl-C to work in the exec()-ed
2516 	 * task, but being ignored by perf stat itself:
2517 	 */
2518 	atexit(sig_atexit);
2519 	if (!forever)
2520 		signal(SIGINT,  skip_signal);
2521 	signal(SIGCHLD, skip_signal);
2522 	signal(SIGALRM, skip_signal);
2523 	signal(SIGABRT, skip_signal);
2524 
2525 	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2526 		goto out;
2527 
2528 	status = 0;
2529 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2530 		if (stat_config.run_count != 1 && verbose > 0)
2531 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2532 				run_idx + 1);
2533 
2534 		if (run_idx != 0)
2535 			evlist__reset_prev_raw_counts(evsel_list);
2536 
2537 		status = run_perf_stat(argc, argv, run_idx);
2538 		if (forever && status != -1 && !interval) {
2539 			print_counters(NULL, argc, argv);
2540 			perf_stat__reset_stats();
2541 		}
2542 	}
2543 
2544 	if (!forever && status != -1 && (!interval || stat_config.summary))
2545 		print_counters(NULL, argc, argv);
2546 
2547 	evlist__finalize_ctlfd(evsel_list);
2548 
2549 	if (STAT_RECORD) {
2550 		/*
2551 		 * We synthesize the kernel mmap record just so that older tools
2552 		 * don't emit warnings about not being able to resolve symbols
2553 		 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2554 		 * a saner message about no samples being in the perf.data file.
2555 		 *
2556 		 * This also serves to suppress a warning about f_header.data.size == 0
2557 		 * in header.c at the moment 'perf stat record' gets introduced, which
2558 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2559 		 * records, but the need to suppress the kptr_restrict messages in older
2560 		 * tools remain  -acme
2561 		 */
2562 		int fd = perf_data__fd(&perf_stat.data);
2563 
2564 		err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2565 							 process_synthesized_event,
2566 							 &perf_stat.session->machines.host);
2567 		if (err) {
2568 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2569 				   "older tools may produce warnings about this file\n.");
2570 		}
2571 
2572 		if (!interval) {
2573 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2574 				pr_err("failed to write stat round event\n");
2575 		}
2576 
2577 		if (!perf_stat.data.is_pipe) {
2578 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2579 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2580 		}
2581 
2582 		evlist__close(evsel_list);
2583 		perf_session__delete(perf_stat.session);
2584 	}
2585 
2586 	perf_stat__exit_aggr_mode();
2587 	evlist__free_stats(evsel_list);
2588 out:
2589 	if (stat_config.iostat_run)
2590 		iostat_release(evsel_list);
2591 
2592 	zfree(&stat_config.walltime_run);
2593 
2594 	if (smi_cost && smi_reset)
2595 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2596 
2597 	evlist__delete(evsel_list);
2598 
2599 	metricgroup__rblist_exit(&stat_config.metric_events);
2600 	runtime_stat_delete(&stat_config);
2601 	evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2602 
2603 	return status;
2604 }
2605