xref: /openbmc/linux/tools/perf/builtin-stat.c (revision 29c37341)
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/group.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 "asm/bug.h"
71 
72 #include <linux/time64.h>
73 #include <linux/zalloc.h>
74 #include <api/fs/fs.h>
75 #include <errno.h>
76 #include <signal.h>
77 #include <stdlib.h>
78 #include <sys/prctl.h>
79 #include <inttypes.h>
80 #include <locale.h>
81 #include <math.h>
82 #include <sys/types.h>
83 #include <sys/stat.h>
84 #include <sys/wait.h>
85 #include <unistd.h>
86 #include <sys/time.h>
87 #include <sys/resource.h>
88 #include <linux/err.h>
89 
90 #include <linux/ctype.h>
91 #include <perf/evlist.h>
92 
93 #define DEFAULT_SEPARATOR	" "
94 #define FREEZE_ON_SMI_PATH	"devices/cpu/freeze_on_smi"
95 
96 static void print_counters(struct timespec *ts, int argc, const char **argv);
97 
98 /* Default events used for perf stat -T */
99 static const char *transaction_attrs = {
100 	"task-clock,"
101 	"{"
102 	"instructions,"
103 	"cycles,"
104 	"cpu/cycles-t/,"
105 	"cpu/tx-start/,"
106 	"cpu/el-start/,"
107 	"cpu/cycles-ct/"
108 	"}"
109 };
110 
111 /* More limited version when the CPU does not have all events. */
112 static const char * transaction_limited_attrs = {
113 	"task-clock,"
114 	"{"
115 	"instructions,"
116 	"cycles,"
117 	"cpu/cycles-t/,"
118 	"cpu/tx-start/"
119 	"}"
120 };
121 
122 static const char * topdown_attrs[] = {
123 	"topdown-total-slots",
124 	"topdown-slots-retired",
125 	"topdown-recovery-bubbles",
126 	"topdown-fetch-bubbles",
127 	"topdown-slots-issued",
128 	NULL,
129 };
130 
131 static const char *smi_cost_attrs = {
132 	"{"
133 	"msr/aperf/,"
134 	"msr/smi/,"
135 	"cycles"
136 	"}"
137 };
138 
139 static struct evlist	*evsel_list;
140 
141 static struct target target = {
142 	.uid	= UINT_MAX,
143 };
144 
145 #define METRIC_ONLY_LEN 20
146 
147 static volatile pid_t		child_pid			= -1;
148 static int			detailed_run			=  0;
149 static bool			transaction_run;
150 static bool			topdown_run			= false;
151 static bool			smi_cost			= false;
152 static bool			smi_reset			= false;
153 static int			big_num_opt			=  -1;
154 static bool			group				= false;
155 static const char		*pre_cmd			= NULL;
156 static const char		*post_cmd			= NULL;
157 static bool			sync_run			= false;
158 static bool			forever				= false;
159 static bool			force_metric_only		= false;
160 static struct timespec		ref_time;
161 static bool			append_file;
162 static bool			interval_count;
163 static const char		*output_name;
164 static int			output_fd;
165 
166 struct perf_stat {
167 	bool			 record;
168 	struct perf_data	 data;
169 	struct perf_session	*session;
170 	u64			 bytes_written;
171 	struct perf_tool	 tool;
172 	bool			 maps_allocated;
173 	struct perf_cpu_map	*cpus;
174 	struct perf_thread_map *threads;
175 	enum aggr_mode		 aggr_mode;
176 };
177 
178 static struct perf_stat		perf_stat;
179 #define STAT_RECORD		perf_stat.record
180 
181 static volatile int done = 0;
182 
183 static struct perf_stat_config stat_config = {
184 	.aggr_mode		= AGGR_GLOBAL,
185 	.scale			= true,
186 	.unit_width		= 4, /* strlen("unit") */
187 	.run_count		= 1,
188 	.metric_only_len	= METRIC_ONLY_LEN,
189 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
190 	.big_num		= true,
191 	.ctl_fd			= -1,
192 	.ctl_fd_ack		= -1
193 };
194 
195 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
196 {
197 	if (!a->core.cpus && !b->core.cpus)
198 		return true;
199 
200 	if (!a->core.cpus || !b->core.cpus)
201 		return false;
202 
203 	if (a->core.cpus->nr != b->core.cpus->nr)
204 		return false;
205 
206 	for (int i = 0; i < a->core.cpus->nr; i++) {
207 		if (a->core.cpus->map[i] != b->core.cpus->map[i])
208 			return false;
209 	}
210 
211 	return true;
212 }
213 
214 static void evlist__check_cpu_maps(struct evlist *evlist)
215 {
216 	struct evsel *evsel, *pos, *leader;
217 	char buf[1024];
218 
219 	evlist__for_each_entry(evlist, evsel) {
220 		leader = evsel->leader;
221 
222 		/* Check that leader matches cpus with each member. */
223 		if (leader == evsel)
224 			continue;
225 		if (cpus_map_matched(leader, evsel))
226 			continue;
227 
228 		/* If there's mismatch disable the group and warn user. */
229 		WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
230 		evsel__group_desc(leader, buf, sizeof(buf));
231 		pr_warning("  %s\n", buf);
232 
233 		if (verbose) {
234 			cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
235 			pr_warning("     %s: %s\n", leader->name, buf);
236 			cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
237 			pr_warning("     %s: %s\n", evsel->name, buf);
238 		}
239 
240 		for_each_group_evsel(pos, leader) {
241 			pos->leader = pos;
242 			pos->core.nr_members = 0;
243 		}
244 		evsel->leader->core.nr_members = 0;
245 	}
246 }
247 
248 static inline void diff_timespec(struct timespec *r, struct timespec *a,
249 				 struct timespec *b)
250 {
251 	r->tv_sec = a->tv_sec - b->tv_sec;
252 	if (a->tv_nsec < b->tv_nsec) {
253 		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
254 		r->tv_sec--;
255 	} else {
256 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
257 	}
258 }
259 
260 static void perf_stat__reset_stats(void)
261 {
262 	int i;
263 
264 	perf_evlist__reset_stats(evsel_list);
265 	perf_stat__reset_shadow_stats();
266 
267 	for (i = 0; i < stat_config.stats_num; i++)
268 		perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
269 }
270 
271 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
272 				     union perf_event *event,
273 				     struct perf_sample *sample __maybe_unused,
274 				     struct machine *machine __maybe_unused)
275 {
276 	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
277 		pr_err("failed to write perf data, error: %m\n");
278 		return -1;
279 	}
280 
281 	perf_stat.bytes_written += event->header.size;
282 	return 0;
283 }
284 
285 static int write_stat_round_event(u64 tm, u64 type)
286 {
287 	return perf_event__synthesize_stat_round(NULL, tm, type,
288 						 process_synthesized_event,
289 						 NULL);
290 }
291 
292 #define WRITE_STAT_ROUND_EVENT(time, interval) \
293 	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
294 
295 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
296 
297 static int evsel__write_stat_event(struct evsel *counter, u32 cpu, u32 thread,
298 				   struct perf_counts_values *count)
299 {
300 	struct perf_sample_id *sid = SID(counter, cpu, thread);
301 
302 	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
303 					   process_synthesized_event, NULL);
304 }
305 
306 static int read_single_counter(struct evsel *counter, int cpu,
307 			       int thread, struct timespec *rs)
308 {
309 	if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
310 		u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
311 		struct perf_counts_values *count =
312 			perf_counts(counter->counts, cpu, thread);
313 		count->ena = count->run = val;
314 		count->val = val;
315 		return 0;
316 	}
317 	return evsel__read_counter(counter, cpu, thread);
318 }
319 
320 /*
321  * Read out the results of a single counter:
322  * do not aggregate counts across CPUs in system-wide mode
323  */
324 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu)
325 {
326 	int nthreads = perf_thread_map__nr(evsel_list->core.threads);
327 	int thread;
328 
329 	if (!counter->supported)
330 		return -ENOENT;
331 
332 	if (counter->core.system_wide)
333 		nthreads = 1;
334 
335 	for (thread = 0; thread < nthreads; thread++) {
336 		struct perf_counts_values *count;
337 
338 		count = perf_counts(counter->counts, cpu, thread);
339 
340 		/*
341 		 * The leader's group read loads data into its group members
342 		 * (via evsel__read_counter()) and sets their count->loaded.
343 		 */
344 		if (!perf_counts__is_loaded(counter->counts, cpu, thread) &&
345 		    read_single_counter(counter, cpu, thread, rs)) {
346 			counter->counts->scaled = -1;
347 			perf_counts(counter->counts, cpu, thread)->ena = 0;
348 			perf_counts(counter->counts, cpu, thread)->run = 0;
349 			return -1;
350 		}
351 
352 		perf_counts__set_loaded(counter->counts, cpu, thread, false);
353 
354 		if (STAT_RECORD) {
355 			if (evsel__write_stat_event(counter, cpu, thread, count)) {
356 				pr_err("failed to write stat event\n");
357 				return -1;
358 			}
359 		}
360 
361 		if (verbose > 1) {
362 			fprintf(stat_config.output,
363 				"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
364 					evsel__name(counter),
365 					cpu,
366 					count->val, count->ena, count->run);
367 		}
368 	}
369 
370 	return 0;
371 }
372 
373 static int read_affinity_counters(struct timespec *rs)
374 {
375 	struct evsel *counter;
376 	struct affinity affinity;
377 	int i, ncpus, cpu;
378 
379 	if (affinity__setup(&affinity) < 0)
380 		return -1;
381 
382 	ncpus = perf_cpu_map__nr(evsel_list->core.all_cpus);
383 	if (!target__has_cpu(&target) || target__has_per_thread(&target))
384 		ncpus = 1;
385 	evlist__for_each_cpu(evsel_list, i, cpu) {
386 		if (i >= ncpus)
387 			break;
388 		affinity__set(&affinity, cpu);
389 
390 		evlist__for_each_entry(evsel_list, counter) {
391 			if (evsel__cpu_iter_skip(counter, cpu))
392 				continue;
393 			if (!counter->err) {
394 				counter->err = read_counter_cpu(counter, rs,
395 								counter->cpu_iter - 1);
396 			}
397 		}
398 	}
399 	affinity__cleanup(&affinity);
400 	return 0;
401 }
402 
403 static void read_counters(struct timespec *rs)
404 {
405 	struct evsel *counter;
406 
407 	if (!stat_config.summary && (read_affinity_counters(rs) < 0))
408 		return;
409 
410 	evlist__for_each_entry(evsel_list, counter) {
411 		if (counter->err)
412 			pr_debug("failed to read counter %s\n", counter->name);
413 		if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
414 			pr_warning("failed to process counter %s\n", counter->name);
415 		counter->err = 0;
416 	}
417 }
418 
419 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
420 {
421 	int i;
422 
423 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
424 	if (!config->stats)
425 		return -1;
426 
427 	config->stats_num = nthreads;
428 
429 	for (i = 0; i < nthreads; i++)
430 		runtime_stat__init(&config->stats[i]);
431 
432 	return 0;
433 }
434 
435 static void runtime_stat_delete(struct perf_stat_config *config)
436 {
437 	int i;
438 
439 	if (!config->stats)
440 		return;
441 
442 	for (i = 0; i < config->stats_num; i++)
443 		runtime_stat__exit(&config->stats[i]);
444 
445 	zfree(&config->stats);
446 }
447 
448 static void runtime_stat_reset(struct perf_stat_config *config)
449 {
450 	int i;
451 
452 	if (!config->stats)
453 		return;
454 
455 	for (i = 0; i < config->stats_num; i++)
456 		perf_stat__reset_shadow_per_stat(&config->stats[i]);
457 }
458 
459 static void process_interval(void)
460 {
461 	struct timespec ts, rs;
462 
463 	clock_gettime(CLOCK_MONOTONIC, &ts);
464 	diff_timespec(&rs, &ts, &ref_time);
465 
466 	perf_stat__reset_shadow_per_stat(&rt_stat);
467 	runtime_stat_reset(&stat_config);
468 	read_counters(&rs);
469 
470 	if (STAT_RECORD) {
471 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
472 			pr_err("failed to write stat round event\n");
473 	}
474 
475 	init_stats(&walltime_nsecs_stats);
476 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
477 	print_counters(&rs, 0, NULL);
478 }
479 
480 static bool handle_interval(unsigned int interval, int *times)
481 {
482 	if (interval) {
483 		process_interval();
484 		if (interval_count && !(--(*times)))
485 			return true;
486 	}
487 	return false;
488 }
489 
490 static void enable_counters(void)
491 {
492 	if (stat_config.initial_delay < 0) {
493 		pr_info(EVLIST_DISABLED_MSG);
494 		return;
495 	}
496 
497 	if (stat_config.initial_delay > 0) {
498 		pr_info(EVLIST_DISABLED_MSG);
499 		usleep(stat_config.initial_delay * USEC_PER_MSEC);
500 	}
501 
502 	/*
503 	 * We need to enable counters only if:
504 	 * - we don't have tracee (attaching to task or cpu)
505 	 * - we have initial delay configured
506 	 */
507 	if (!target__none(&target) || stat_config.initial_delay) {
508 		evlist__enable(evsel_list);
509 		if (stat_config.initial_delay > 0)
510 			pr_info(EVLIST_ENABLED_MSG);
511 	}
512 }
513 
514 static void disable_counters(void)
515 {
516 	/*
517 	 * If we don't have tracee (attaching to task or cpu), counters may
518 	 * still be running. To get accurate group ratios, we must stop groups
519 	 * from counting before reading their constituent counters.
520 	 */
521 	if (!target__none(&target))
522 		evlist__disable(evsel_list);
523 }
524 
525 static volatile int workload_exec_errno;
526 
527 /*
528  * perf_evlist__prepare_workload will send a SIGUSR1
529  * if the fork fails, since we asked by setting its
530  * want_signal to true.
531  */
532 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
533 					void *ucontext __maybe_unused)
534 {
535 	workload_exec_errno = info->si_value.sival_int;
536 }
537 
538 static bool evsel__should_store_id(struct evsel *counter)
539 {
540 	return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
541 }
542 
543 static bool is_target_alive(struct target *_target,
544 			    struct perf_thread_map *threads)
545 {
546 	struct stat st;
547 	int i;
548 
549 	if (!target__has_task(_target))
550 		return true;
551 
552 	for (i = 0; i < threads->nr; i++) {
553 		char path[PATH_MAX];
554 
555 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
556 			  threads->map[i].pid);
557 
558 		if (!stat(path, &st))
559 			return true;
560 	}
561 
562 	return false;
563 }
564 
565 static void process_evlist(struct evlist *evlist, unsigned int interval)
566 {
567 	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
568 
569 	if (evlist__ctlfd_process(evlist, &cmd) > 0) {
570 		switch (cmd) {
571 		case EVLIST_CTL_CMD_ENABLE:
572 			pr_info(EVLIST_ENABLED_MSG);
573 			if (interval)
574 				process_interval();
575 			break;
576 		case EVLIST_CTL_CMD_DISABLE:
577 			if (interval)
578 				process_interval();
579 			pr_info(EVLIST_DISABLED_MSG);
580 			break;
581 		case EVLIST_CTL_CMD_ACK:
582 		case EVLIST_CTL_CMD_UNSUPPORTED:
583 		default:
584 			break;
585 		}
586 	}
587 }
588 
589 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
590 			int *time_to_sleep)
591 {
592 	int tts = *time_to_sleep;
593 	struct timespec time_diff;
594 
595 	diff_timespec(&time_diff, time_stop, time_start);
596 
597 	tts -= time_diff.tv_sec * MSEC_PER_SEC +
598 	       time_diff.tv_nsec / NSEC_PER_MSEC;
599 
600 	if (tts < 0)
601 		tts = 0;
602 
603 	*time_to_sleep = tts;
604 }
605 
606 static int dispatch_events(bool forks, int timeout, int interval, int *times)
607 {
608 	int child_exited = 0, status = 0;
609 	int time_to_sleep, sleep_time;
610 	struct timespec time_start, time_stop;
611 
612 	if (interval)
613 		sleep_time = interval;
614 	else if (timeout)
615 		sleep_time = timeout;
616 	else
617 		sleep_time = 1000;
618 
619 	time_to_sleep = sleep_time;
620 
621 	while (!done) {
622 		if (forks)
623 			child_exited = waitpid(child_pid, &status, WNOHANG);
624 		else
625 			child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
626 
627 		if (child_exited)
628 			break;
629 
630 		clock_gettime(CLOCK_MONOTONIC, &time_start);
631 		if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
632 			if (timeout || handle_interval(interval, times))
633 				break;
634 			time_to_sleep = sleep_time;
635 		} else { /* fd revent */
636 			process_evlist(evsel_list, interval);
637 			clock_gettime(CLOCK_MONOTONIC, &time_stop);
638 			compute_tts(&time_start, &time_stop, &time_to_sleep);
639 		}
640 	}
641 
642 	return status;
643 }
644 
645 enum counter_recovery {
646 	COUNTER_SKIP,
647 	COUNTER_RETRY,
648 	COUNTER_FATAL,
649 };
650 
651 static enum counter_recovery stat_handle_error(struct evsel *counter)
652 {
653 	char msg[BUFSIZ];
654 	/*
655 	 * PPC returns ENXIO for HW counters until 2.6.37
656 	 * (behavior changed with commit b0a873e).
657 	 */
658 	if (errno == EINVAL || errno == ENOSYS ||
659 	    errno == ENOENT || errno == EOPNOTSUPP ||
660 	    errno == ENXIO) {
661 		if (verbose > 0)
662 			ui__warning("%s event is not supported by the kernel.\n",
663 				    evsel__name(counter));
664 		counter->supported = false;
665 		/*
666 		 * errored is a sticky flag that means one of the counter's
667 		 * cpu event had a problem and needs to be reexamined.
668 		 */
669 		counter->errored = true;
670 
671 		if ((counter->leader != counter) ||
672 		    !(counter->leader->core.nr_members > 1))
673 			return COUNTER_SKIP;
674 	} else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
675 		if (verbose > 0)
676 			ui__warning("%s\n", msg);
677 		return COUNTER_RETRY;
678 	} else if (target__has_per_thread(&target) &&
679 		   evsel_list->core.threads &&
680 		   evsel_list->core.threads->err_thread != -1) {
681 		/*
682 		 * For global --per-thread case, skip current
683 		 * error thread.
684 		 */
685 		if (!thread_map__remove(evsel_list->core.threads,
686 					evsel_list->core.threads->err_thread)) {
687 			evsel_list->core.threads->err_thread = -1;
688 			return COUNTER_RETRY;
689 		}
690 	}
691 
692 	evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
693 	ui__error("%s\n", msg);
694 
695 	if (child_pid != -1)
696 		kill(child_pid, SIGTERM);
697 	return COUNTER_FATAL;
698 }
699 
700 static int __run_perf_stat(int argc, const char **argv, int run_idx)
701 {
702 	int interval = stat_config.interval;
703 	int times = stat_config.times;
704 	int timeout = stat_config.timeout;
705 	char msg[BUFSIZ];
706 	unsigned long long t0, t1;
707 	struct evsel *counter;
708 	size_t l;
709 	int status = 0;
710 	const bool forks = (argc > 0);
711 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
712 	struct affinity affinity;
713 	int i, cpu;
714 	bool second_pass = false;
715 
716 	if (forks) {
717 		if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
718 						  workload_exec_failed_signal) < 0) {
719 			perror("failed to prepare workload");
720 			return -1;
721 		}
722 		child_pid = evsel_list->workload.pid;
723 	}
724 
725 	if (group)
726 		perf_evlist__set_leader(evsel_list);
727 
728 	if (affinity__setup(&affinity) < 0)
729 		return -1;
730 
731 	evlist__for_each_cpu (evsel_list, i, cpu) {
732 		affinity__set(&affinity, cpu);
733 
734 		evlist__for_each_entry(evsel_list, counter) {
735 			if (evsel__cpu_iter_skip(counter, cpu))
736 				continue;
737 			if (counter->reset_group || counter->errored)
738 				continue;
739 try_again:
740 			if (create_perf_stat_counter(counter, &stat_config, &target,
741 						     counter->cpu_iter - 1) < 0) {
742 
743 				/*
744 				 * Weak group failed. We cannot just undo this here
745 				 * because earlier CPUs might be in group mode, and the kernel
746 				 * doesn't support mixing group and non group reads. Defer
747 				 * it to later.
748 				 * Don't close here because we're in the wrong affinity.
749 				 */
750 				if ((errno == EINVAL || errno == EBADF) &&
751 				    counter->leader != counter &&
752 				    counter->weak_group) {
753 					perf_evlist__reset_weak_group(evsel_list, counter, false);
754 					assert(counter->reset_group);
755 					second_pass = true;
756 					continue;
757 				}
758 
759 				switch (stat_handle_error(counter)) {
760 				case COUNTER_FATAL:
761 					return -1;
762 				case COUNTER_RETRY:
763 					goto try_again;
764 				case COUNTER_SKIP:
765 					continue;
766 				default:
767 					break;
768 				}
769 
770 			}
771 			counter->supported = true;
772 		}
773 	}
774 
775 	if (second_pass) {
776 		/*
777 		 * Now redo all the weak group after closing them,
778 		 * and also close errored counters.
779 		 */
780 
781 		evlist__for_each_cpu(evsel_list, i, cpu) {
782 			affinity__set(&affinity, cpu);
783 			/* First close errored or weak retry */
784 			evlist__for_each_entry(evsel_list, counter) {
785 				if (!counter->reset_group && !counter->errored)
786 					continue;
787 				if (evsel__cpu_iter_skip_no_inc(counter, cpu))
788 					continue;
789 				perf_evsel__close_cpu(&counter->core, counter->cpu_iter);
790 			}
791 			/* Now reopen weak */
792 			evlist__for_each_entry(evsel_list, counter) {
793 				if (!counter->reset_group && !counter->errored)
794 					continue;
795 				if (evsel__cpu_iter_skip(counter, cpu))
796 					continue;
797 				if (!counter->reset_group)
798 					continue;
799 try_again_reset:
800 				pr_debug2("reopening weak %s\n", evsel__name(counter));
801 				if (create_perf_stat_counter(counter, &stat_config, &target,
802 							     counter->cpu_iter - 1) < 0) {
803 
804 					switch (stat_handle_error(counter)) {
805 					case COUNTER_FATAL:
806 						return -1;
807 					case COUNTER_RETRY:
808 						goto try_again_reset;
809 					case COUNTER_SKIP:
810 						continue;
811 					default:
812 						break;
813 					}
814 				}
815 				counter->supported = true;
816 			}
817 		}
818 	}
819 	affinity__cleanup(&affinity);
820 
821 	evlist__for_each_entry(evsel_list, counter) {
822 		if (!counter->supported) {
823 			perf_evsel__free_fd(&counter->core);
824 			continue;
825 		}
826 
827 		l = strlen(counter->unit);
828 		if (l > stat_config.unit_width)
829 			stat_config.unit_width = l;
830 
831 		if (evsel__should_store_id(counter) &&
832 		    evsel__store_ids(counter, evsel_list))
833 			return -1;
834 	}
835 
836 	if (perf_evlist__apply_filters(evsel_list, &counter)) {
837 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
838 			counter->filter, evsel__name(counter), errno,
839 			str_error_r(errno, msg, sizeof(msg)));
840 		return -1;
841 	}
842 
843 	if (STAT_RECORD) {
844 		int err, fd = perf_data__fd(&perf_stat.data);
845 
846 		if (is_pipe) {
847 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
848 		} else {
849 			err = perf_session__write_header(perf_stat.session, evsel_list,
850 							 fd, false);
851 		}
852 
853 		if (err < 0)
854 			return err;
855 
856 		err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
857 							 process_synthesized_event, is_pipe);
858 		if (err < 0)
859 			return err;
860 	}
861 
862 	/*
863 	 * Enable counters and exec the command:
864 	 */
865 	t0 = rdclock();
866 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
867 
868 	if (forks) {
869 		perf_evlist__start_workload(evsel_list);
870 		enable_counters();
871 
872 		if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
873 			status = dispatch_events(forks, timeout, interval, &times);
874 		if (child_pid != -1) {
875 			if (timeout)
876 				kill(child_pid, SIGTERM);
877 			wait4(child_pid, &status, 0, &stat_config.ru_data);
878 		}
879 
880 		if (workload_exec_errno) {
881 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
882 			pr_err("Workload failed: %s\n", emsg);
883 			return -1;
884 		}
885 
886 		if (WIFSIGNALED(status))
887 			psignal(WTERMSIG(status), argv[0]);
888 	} else {
889 		enable_counters();
890 		status = dispatch_events(forks, timeout, interval, &times);
891 	}
892 
893 	disable_counters();
894 
895 	t1 = rdclock();
896 
897 	if (stat_config.walltime_run_table)
898 		stat_config.walltime_run[run_idx] = t1 - t0;
899 
900 	if (interval) {
901 		stat_config.interval = 0;
902 		stat_config.summary = true;
903 		init_stats(&walltime_nsecs_stats);
904 		update_stats(&walltime_nsecs_stats, t1 - t0);
905 
906 		if (stat_config.aggr_mode == AGGR_GLOBAL)
907 			perf_evlist__save_aggr_prev_raw_counts(evsel_list);
908 
909 		perf_evlist__copy_prev_raw_counts(evsel_list);
910 		perf_evlist__reset_prev_raw_counts(evsel_list);
911 		runtime_stat_reset(&stat_config);
912 		perf_stat__reset_shadow_per_stat(&rt_stat);
913 	} else
914 		update_stats(&walltime_nsecs_stats, t1 - t0);
915 
916 	/*
917 	 * Closing a group leader splits the group, and as we only disable
918 	 * group leaders, results in remaining events becoming enabled. To
919 	 * avoid arbitrary skew, we must read all counters before closing any
920 	 * group leaders.
921 	 */
922 	read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
923 
924 	/*
925 	 * We need to keep evsel_list alive, because it's processed
926 	 * later the evsel_list will be closed after.
927 	 */
928 	if (!STAT_RECORD)
929 		evlist__close(evsel_list);
930 
931 	return WEXITSTATUS(status);
932 }
933 
934 static int run_perf_stat(int argc, const char **argv, int run_idx)
935 {
936 	int ret;
937 
938 	if (pre_cmd) {
939 		ret = system(pre_cmd);
940 		if (ret)
941 			return ret;
942 	}
943 
944 	if (sync_run)
945 		sync();
946 
947 	ret = __run_perf_stat(argc, argv, run_idx);
948 	if (ret)
949 		return ret;
950 
951 	if (post_cmd) {
952 		ret = system(post_cmd);
953 		if (ret)
954 			return ret;
955 	}
956 
957 	return ret;
958 }
959 
960 static void print_counters(struct timespec *ts, int argc, const char **argv)
961 {
962 	/* Do not print anything if we record to the pipe. */
963 	if (STAT_RECORD && perf_stat.data.is_pipe)
964 		return;
965 
966 	perf_evlist__print_counters(evsel_list, &stat_config, &target,
967 				    ts, argc, argv);
968 }
969 
970 static volatile int signr = -1;
971 
972 static void skip_signal(int signo)
973 {
974 	if ((child_pid == -1) || stat_config.interval)
975 		done = 1;
976 
977 	signr = signo;
978 	/*
979 	 * render child_pid harmless
980 	 * won't send SIGTERM to a random
981 	 * process in case of race condition
982 	 * and fast PID recycling
983 	 */
984 	child_pid = -1;
985 }
986 
987 static void sig_atexit(void)
988 {
989 	sigset_t set, oset;
990 
991 	/*
992 	 * avoid race condition with SIGCHLD handler
993 	 * in skip_signal() which is modifying child_pid
994 	 * goal is to avoid send SIGTERM to a random
995 	 * process
996 	 */
997 	sigemptyset(&set);
998 	sigaddset(&set, SIGCHLD);
999 	sigprocmask(SIG_BLOCK, &set, &oset);
1000 
1001 	if (child_pid != -1)
1002 		kill(child_pid, SIGTERM);
1003 
1004 	sigprocmask(SIG_SETMASK, &oset, NULL);
1005 
1006 	if (signr == -1)
1007 		return;
1008 
1009 	signal(signr, SIG_DFL);
1010 	kill(getpid(), signr);
1011 }
1012 
1013 void perf_stat__set_big_num(int set)
1014 {
1015 	stat_config.big_num = (set != 0);
1016 }
1017 
1018 static int stat__set_big_num(const struct option *opt __maybe_unused,
1019 			     const char *s __maybe_unused, int unset)
1020 {
1021 	big_num_opt = unset ? 0 : 1;
1022 	perf_stat__set_big_num(!unset);
1023 	return 0;
1024 }
1025 
1026 static int enable_metric_only(const struct option *opt __maybe_unused,
1027 			      const char *s __maybe_unused, int unset)
1028 {
1029 	force_metric_only = true;
1030 	stat_config.metric_only = !unset;
1031 	return 0;
1032 }
1033 
1034 static int parse_metric_groups(const struct option *opt,
1035 			       const char *str,
1036 			       int unset __maybe_unused)
1037 {
1038 	return metricgroup__parse_groups(opt, str,
1039 					 stat_config.metric_no_group,
1040 					 stat_config.metric_no_merge,
1041 					 &stat_config.metric_events);
1042 }
1043 
1044 static int parse_control_option(const struct option *opt,
1045 				const char *str,
1046 				int unset __maybe_unused)
1047 {
1048 	char *comma = NULL, *endptr = NULL;
1049 	struct perf_stat_config *config = (struct perf_stat_config *)opt->value;
1050 
1051 	if (strncmp(str, "fd:", 3))
1052 		return -EINVAL;
1053 
1054 	config->ctl_fd = strtoul(&str[3], &endptr, 0);
1055 	if (endptr == &str[3])
1056 		return -EINVAL;
1057 
1058 	comma = strchr(str, ',');
1059 	if (comma) {
1060 		if (endptr != comma)
1061 			return -EINVAL;
1062 
1063 		config->ctl_fd_ack = strtoul(comma + 1, &endptr, 0);
1064 		if (endptr == comma + 1 || *endptr != '\0')
1065 			return -EINVAL;
1066 	}
1067 
1068 	return 0;
1069 }
1070 
1071 static struct option stat_options[] = {
1072 	OPT_BOOLEAN('T', "transaction", &transaction_run,
1073 		    "hardware transaction statistics"),
1074 	OPT_CALLBACK('e', "event", &evsel_list, "event",
1075 		     "event selector. use 'perf list' to list available events",
1076 		     parse_events_option),
1077 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1078 		     "event filter", parse_filter),
1079 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1080 		    "child tasks do not inherit counters"),
1081 	OPT_STRING('p', "pid", &target.pid, "pid",
1082 		   "stat events on existing process id"),
1083 	OPT_STRING('t', "tid", &target.tid, "tid",
1084 		   "stat events on existing thread id"),
1085 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1086 		    "system-wide collection from all CPUs"),
1087 	OPT_BOOLEAN('g', "group", &group,
1088 		    "put the counters into a counter group"),
1089 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
1090 		    "Use --no-scale to disable counter scaling for multiplexing"),
1091 	OPT_INCR('v', "verbose", &verbose,
1092 		    "be more verbose (show counter open errors, etc)"),
1093 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
1094 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
1095 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1096 		    "display details about each run (only with -r option)"),
1097 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
1098 		    "null run - dont start any counters"),
1099 	OPT_INCR('d', "detailed", &detailed_run,
1100 		    "detailed run - start a lot of events"),
1101 	OPT_BOOLEAN('S', "sync", &sync_run,
1102 		    "call sync() before starting a run"),
1103 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1104 			   "print large numbers with thousands\' separators",
1105 			   stat__set_big_num),
1106 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1107 		    "list of cpus to monitor in system-wide"),
1108 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1109 		    "disable CPU count aggregation", AGGR_NONE),
1110 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1111 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1112 		   "print counts with custom separator"),
1113 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1114 		     "monitor event in cgroup name only", parse_cgroups),
1115 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1116 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1117 	OPT_INTEGER(0, "log-fd", &output_fd,
1118 		    "log output to fd, instead of stderr"),
1119 	OPT_STRING(0, "pre", &pre_cmd, "command",
1120 			"command to run prior to the measured command"),
1121 	OPT_STRING(0, "post", &post_cmd, "command",
1122 			"command to run after to the measured command"),
1123 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1124 		    "print counts at regular interval in ms "
1125 		    "(overhead is possible for values <= 100ms)"),
1126 	OPT_INTEGER(0, "interval-count", &stat_config.times,
1127 		    "print counts for fixed number of times"),
1128 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1129 		    "clear screen in between new interval"),
1130 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1131 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1132 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1133 		     "aggregate counts per processor socket", AGGR_SOCKET),
1134 	OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1135 		     "aggregate counts per processor die", AGGR_DIE),
1136 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1137 		     "aggregate counts per physical processor core", AGGR_CORE),
1138 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1139 		     "aggregate counts per thread", AGGR_THREAD),
1140 	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1141 		     "aggregate counts per numa node", AGGR_NODE),
1142 	OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1143 		    "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1144 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1145 			"Only print computed metrics. No raw values", enable_metric_only),
1146 	OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1147 		       "don't group metric events, impacts multiplexing"),
1148 	OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1149 		       "don't try to share events between metrics in a group"),
1150 	OPT_BOOLEAN(0, "topdown", &topdown_run,
1151 			"measure topdown level 1 statistics"),
1152 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1153 			"measure SMI cost"),
1154 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1155 		     "monitor specified metrics or metric groups (separated by ,)",
1156 		     parse_metric_groups),
1157 	OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1158 			 "Configure all used events to run in kernel space.",
1159 			 PARSE_OPT_EXCLUSIVE),
1160 	OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1161 			 "Configure all used events to run in user space.",
1162 			 PARSE_OPT_EXCLUSIVE),
1163 	OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1164 		    "Use with 'percore' event qualifier to show the event "
1165 		    "counts of one hardware thread by sum up total hardware "
1166 		    "threads of same physical core"),
1167 #ifdef HAVE_LIBPFM
1168 	OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1169 		"libpfm4 event selector. use 'perf list' to list available events",
1170 		parse_libpfm_events_option),
1171 #endif
1172 	OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd]",
1173 		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1174 		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.",
1175 		      parse_control_option),
1176 	OPT_END()
1177 };
1178 
1179 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1180 				 struct perf_cpu_map *map, int cpu)
1181 {
1182 	return cpu_map__get_socket(map, cpu, NULL);
1183 }
1184 
1185 static int perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1186 			      struct perf_cpu_map *map, int cpu)
1187 {
1188 	return cpu_map__get_die(map, cpu, NULL);
1189 }
1190 
1191 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1192 			       struct perf_cpu_map *map, int cpu)
1193 {
1194 	return cpu_map__get_core(map, cpu, NULL);
1195 }
1196 
1197 static int perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1198 			       struct perf_cpu_map *map, int cpu)
1199 {
1200 	return cpu_map__get_node(map, cpu, NULL);
1201 }
1202 
1203 static int perf_stat__get_aggr(struct perf_stat_config *config,
1204 			       aggr_get_id_t get_id, struct perf_cpu_map *map, int idx)
1205 {
1206 	int cpu;
1207 
1208 	if (idx >= map->nr)
1209 		return -1;
1210 
1211 	cpu = map->map[idx];
1212 
1213 	if (config->cpus_aggr_map->map[cpu] == -1)
1214 		config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
1215 
1216 	return config->cpus_aggr_map->map[cpu];
1217 }
1218 
1219 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
1220 					struct perf_cpu_map *map, int idx)
1221 {
1222 	return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
1223 }
1224 
1225 static int perf_stat__get_die_cached(struct perf_stat_config *config,
1226 					struct perf_cpu_map *map, int idx)
1227 {
1228 	return perf_stat__get_aggr(config, perf_stat__get_die, map, idx);
1229 }
1230 
1231 static int perf_stat__get_core_cached(struct perf_stat_config *config,
1232 				      struct perf_cpu_map *map, int idx)
1233 {
1234 	return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
1235 }
1236 
1237 static int perf_stat__get_node_cached(struct perf_stat_config *config,
1238 				      struct perf_cpu_map *map, int idx)
1239 {
1240 	return perf_stat__get_aggr(config, perf_stat__get_node, map, idx);
1241 }
1242 
1243 static bool term_percore_set(void)
1244 {
1245 	struct evsel *counter;
1246 
1247 	evlist__for_each_entry(evsel_list, counter) {
1248 		if (counter->percore)
1249 			return true;
1250 	}
1251 
1252 	return false;
1253 }
1254 
1255 static int perf_stat_init_aggr_mode(void)
1256 {
1257 	int nr;
1258 
1259 	switch (stat_config.aggr_mode) {
1260 	case AGGR_SOCKET:
1261 		if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1262 			perror("cannot build socket map");
1263 			return -1;
1264 		}
1265 		stat_config.aggr_get_id = perf_stat__get_socket_cached;
1266 		break;
1267 	case AGGR_DIE:
1268 		if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1269 			perror("cannot build die map");
1270 			return -1;
1271 		}
1272 		stat_config.aggr_get_id = perf_stat__get_die_cached;
1273 		break;
1274 	case AGGR_CORE:
1275 		if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1276 			perror("cannot build core map");
1277 			return -1;
1278 		}
1279 		stat_config.aggr_get_id = perf_stat__get_core_cached;
1280 		break;
1281 	case AGGR_NODE:
1282 		if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1283 			perror("cannot build core map");
1284 			return -1;
1285 		}
1286 		stat_config.aggr_get_id = perf_stat__get_node_cached;
1287 		break;
1288 	case AGGR_NONE:
1289 		if (term_percore_set()) {
1290 			if (cpu_map__build_core_map(evsel_list->core.cpus,
1291 						    &stat_config.aggr_map)) {
1292 				perror("cannot build core map");
1293 				return -1;
1294 			}
1295 			stat_config.aggr_get_id = perf_stat__get_core_cached;
1296 		}
1297 		break;
1298 	case AGGR_GLOBAL:
1299 	case AGGR_THREAD:
1300 	case AGGR_UNSET:
1301 	default:
1302 		break;
1303 	}
1304 
1305 	/*
1306 	 * The evsel_list->cpus is the base we operate on,
1307 	 * taking the highest cpu number to be the size of
1308 	 * the aggregation translate cpumap.
1309 	 */
1310 	nr = perf_cpu_map__max(evsel_list->core.cpus);
1311 	stat_config.cpus_aggr_map = perf_cpu_map__empty_new(nr + 1);
1312 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1313 }
1314 
1315 static void perf_stat__exit_aggr_mode(void)
1316 {
1317 	perf_cpu_map__put(stat_config.aggr_map);
1318 	perf_cpu_map__put(stat_config.cpus_aggr_map);
1319 	stat_config.aggr_map = NULL;
1320 	stat_config.cpus_aggr_map = NULL;
1321 }
1322 
1323 static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx)
1324 {
1325 	int cpu;
1326 
1327 	if (idx > map->nr)
1328 		return -1;
1329 
1330 	cpu = map->map[idx];
1331 
1332 	if (cpu >= env->nr_cpus_avail)
1333 		return -1;
1334 
1335 	return cpu;
1336 }
1337 
1338 static int perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data)
1339 {
1340 	struct perf_env *env = data;
1341 	int cpu = perf_env__get_cpu(env, map, idx);
1342 
1343 	return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1344 }
1345 
1346 static int perf_env__get_die(struct perf_cpu_map *map, int idx, void *data)
1347 {
1348 	struct perf_env *env = data;
1349 	int die_id = -1, cpu = perf_env__get_cpu(env, map, idx);
1350 
1351 	if (cpu != -1) {
1352 		/*
1353 		 * Encode socket in bit range 15:8
1354 		 * die_id is relative to socket,
1355 		 * we need a global id. So we combine
1356 		 * socket + die id
1357 		 */
1358 		if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
1359 			return -1;
1360 
1361 		if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
1362 			return -1;
1363 
1364 		die_id = (env->cpu[cpu].socket_id << 8) | (env->cpu[cpu].die_id & 0xff);
1365 	}
1366 
1367 	return die_id;
1368 }
1369 
1370 static int perf_env__get_core(struct perf_cpu_map *map, int idx, void *data)
1371 {
1372 	struct perf_env *env = data;
1373 	int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1374 
1375 	if (cpu != -1) {
1376 		/*
1377 		 * Encode socket in bit range 31:24
1378 		 * encode die id in bit range 23:16
1379 		 * core_id is relative to socket and die,
1380 		 * we need a global id. So we combine
1381 		 * socket + die id + core id
1382 		 */
1383 		if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
1384 			return -1;
1385 
1386 		if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
1387 			return -1;
1388 
1389 		if (WARN_ONCE(env->cpu[cpu].core_id >> 16, "The core id number is too big.\n"))
1390 			return -1;
1391 
1392 		core = (env->cpu[cpu].socket_id << 24) |
1393 		       (env->cpu[cpu].die_id << 16) |
1394 		       (env->cpu[cpu].core_id & 0xffff);
1395 	}
1396 
1397 	return core;
1398 }
1399 
1400 static int perf_env__get_node(struct perf_cpu_map *map, int idx, void *data)
1401 {
1402 	int cpu = perf_env__get_cpu(data, map, idx);
1403 
1404 	return perf_env__numa_node(data, cpu);
1405 }
1406 
1407 static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus,
1408 				      struct perf_cpu_map **sockp)
1409 {
1410 	return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1411 }
1412 
1413 static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus,
1414 				   struct perf_cpu_map **diep)
1415 {
1416 	return cpu_map__build_map(cpus, diep, perf_env__get_die, env);
1417 }
1418 
1419 static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus,
1420 				    struct perf_cpu_map **corep)
1421 {
1422 	return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1423 }
1424 
1425 static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus,
1426 				    struct perf_cpu_map **nodep)
1427 {
1428 	return cpu_map__build_map(cpus, nodep, perf_env__get_node, env);
1429 }
1430 
1431 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1432 				      struct perf_cpu_map *map, int idx)
1433 {
1434 	return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1435 }
1436 static int perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1437 				   struct perf_cpu_map *map, int idx)
1438 {
1439 	return perf_env__get_die(map, idx, &perf_stat.session->header.env);
1440 }
1441 
1442 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1443 				    struct perf_cpu_map *map, int idx)
1444 {
1445 	return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1446 }
1447 
1448 static int perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1449 				    struct perf_cpu_map *map, int idx)
1450 {
1451 	return perf_env__get_node(map, idx, &perf_stat.session->header.env);
1452 }
1453 
1454 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1455 {
1456 	struct perf_env *env = &st->session->header.env;
1457 
1458 	switch (stat_config.aggr_mode) {
1459 	case AGGR_SOCKET:
1460 		if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1461 			perror("cannot build socket map");
1462 			return -1;
1463 		}
1464 		stat_config.aggr_get_id = perf_stat__get_socket_file;
1465 		break;
1466 	case AGGR_DIE:
1467 		if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1468 			perror("cannot build die map");
1469 			return -1;
1470 		}
1471 		stat_config.aggr_get_id = perf_stat__get_die_file;
1472 		break;
1473 	case AGGR_CORE:
1474 		if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1475 			perror("cannot build core map");
1476 			return -1;
1477 		}
1478 		stat_config.aggr_get_id = perf_stat__get_core_file;
1479 		break;
1480 	case AGGR_NODE:
1481 		if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1482 			perror("cannot build core map");
1483 			return -1;
1484 		}
1485 		stat_config.aggr_get_id = perf_stat__get_node_file;
1486 		break;
1487 	case AGGR_NONE:
1488 	case AGGR_GLOBAL:
1489 	case AGGR_THREAD:
1490 	case AGGR_UNSET:
1491 	default:
1492 		break;
1493 	}
1494 
1495 	return 0;
1496 }
1497 
1498 static int topdown_filter_events(const char **attr, char **str, bool use_group)
1499 {
1500 	int off = 0;
1501 	int i;
1502 	int len = 0;
1503 	char *s;
1504 
1505 	for (i = 0; attr[i]; i++) {
1506 		if (pmu_have_event("cpu", attr[i])) {
1507 			len += strlen(attr[i]) + 1;
1508 			attr[i - off] = attr[i];
1509 		} else
1510 			off++;
1511 	}
1512 	attr[i - off] = NULL;
1513 
1514 	*str = malloc(len + 1 + 2);
1515 	if (!*str)
1516 		return -1;
1517 	s = *str;
1518 	if (i - off == 0) {
1519 		*s = 0;
1520 		return 0;
1521 	}
1522 	if (use_group)
1523 		*s++ = '{';
1524 	for (i = 0; attr[i]; i++) {
1525 		strcpy(s, attr[i]);
1526 		s += strlen(s);
1527 		*s++ = ',';
1528 	}
1529 	if (use_group) {
1530 		s[-1] = '}';
1531 		*s = 0;
1532 	} else
1533 		s[-1] = 0;
1534 	return 0;
1535 }
1536 
1537 __weak bool arch_topdown_check_group(bool *warn)
1538 {
1539 	*warn = false;
1540 	return false;
1541 }
1542 
1543 __weak void arch_topdown_group_warn(void)
1544 {
1545 }
1546 
1547 /*
1548  * Add default attributes, if there were no attributes specified or
1549  * if -d/--detailed, -d -d or -d -d -d is used:
1550  */
1551 static int add_default_attributes(void)
1552 {
1553 	int err;
1554 	struct perf_event_attr default_attrs0[] = {
1555 
1556   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1557   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1558   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1559   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1560 
1561   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1562 };
1563 	struct perf_event_attr frontend_attrs[] = {
1564   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1565 };
1566 	struct perf_event_attr backend_attrs[] = {
1567   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1568 };
1569 	struct perf_event_attr default_attrs1[] = {
1570   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1571   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1572   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1573 
1574 };
1575 
1576 /*
1577  * Detailed stats (-d), covering the L1 and last level data caches:
1578  */
1579 	struct perf_event_attr detailed_attrs[] = {
1580 
1581   { .type = PERF_TYPE_HW_CACHE,
1582     .config =
1583 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1584 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1585 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1586 
1587   { .type = PERF_TYPE_HW_CACHE,
1588     .config =
1589 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1590 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1591 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1592 
1593   { .type = PERF_TYPE_HW_CACHE,
1594     .config =
1595 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1596 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1597 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1598 
1599   { .type = PERF_TYPE_HW_CACHE,
1600     .config =
1601 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1602 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1603 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1604 };
1605 
1606 /*
1607  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1608  */
1609 	struct perf_event_attr very_detailed_attrs[] = {
1610 
1611   { .type = PERF_TYPE_HW_CACHE,
1612     .config =
1613 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1614 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1615 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1616 
1617   { .type = PERF_TYPE_HW_CACHE,
1618     .config =
1619 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1620 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1621 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1622 
1623   { .type = PERF_TYPE_HW_CACHE,
1624     .config =
1625 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1626 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1627 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1628 
1629   { .type = PERF_TYPE_HW_CACHE,
1630     .config =
1631 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1632 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1633 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1634 
1635   { .type = PERF_TYPE_HW_CACHE,
1636     .config =
1637 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1638 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1639 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1640 
1641   { .type = PERF_TYPE_HW_CACHE,
1642     .config =
1643 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1644 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1645 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1646 
1647 };
1648 
1649 /*
1650  * Very, very detailed stats (-d -d -d), adding prefetch events:
1651  */
1652 	struct perf_event_attr very_very_detailed_attrs[] = {
1653 
1654   { .type = PERF_TYPE_HW_CACHE,
1655     .config =
1656 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1657 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1658 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1659 
1660   { .type = PERF_TYPE_HW_CACHE,
1661     .config =
1662 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1663 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1664 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1665 };
1666 	struct parse_events_error errinfo;
1667 
1668 	/* Set attrs if no event is selected and !null_run: */
1669 	if (stat_config.null_run)
1670 		return 0;
1671 
1672 	bzero(&errinfo, sizeof(errinfo));
1673 	if (transaction_run) {
1674 		/* Handle -T as -M transaction. Once platform specific metrics
1675 		 * support has been added to the json files, all archictures
1676 		 * will use this approach. To determine transaction support
1677 		 * on an architecture test for such a metric name.
1678 		 */
1679 		if (metricgroup__has_metric("transaction")) {
1680 			struct option opt = { .value = &evsel_list };
1681 
1682 			return metricgroup__parse_groups(&opt, "transaction",
1683 							 stat_config.metric_no_group,
1684 							stat_config.metric_no_merge,
1685 							 &stat_config.metric_events);
1686 		}
1687 
1688 		if (pmu_have_event("cpu", "cycles-ct") &&
1689 		    pmu_have_event("cpu", "el-start"))
1690 			err = parse_events(evsel_list, transaction_attrs,
1691 					   &errinfo);
1692 		else
1693 			err = parse_events(evsel_list,
1694 					   transaction_limited_attrs,
1695 					   &errinfo);
1696 		if (err) {
1697 			fprintf(stderr, "Cannot set up transaction events\n");
1698 			parse_events_print_error(&errinfo, transaction_attrs);
1699 			return -1;
1700 		}
1701 		return 0;
1702 	}
1703 
1704 	if (smi_cost) {
1705 		int smi;
1706 
1707 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1708 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1709 			return -1;
1710 		}
1711 
1712 		if (!smi) {
1713 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1714 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1715 				return -1;
1716 			}
1717 			smi_reset = true;
1718 		}
1719 
1720 		if (pmu_have_event("msr", "aperf") &&
1721 		    pmu_have_event("msr", "smi")) {
1722 			if (!force_metric_only)
1723 				stat_config.metric_only = true;
1724 			err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1725 		} else {
1726 			fprintf(stderr, "To measure SMI cost, it needs "
1727 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1728 			parse_events_print_error(&errinfo, smi_cost_attrs);
1729 			return -1;
1730 		}
1731 		if (err) {
1732 			parse_events_print_error(&errinfo, smi_cost_attrs);
1733 			fprintf(stderr, "Cannot set up SMI cost events\n");
1734 			return -1;
1735 		}
1736 		return 0;
1737 	}
1738 
1739 	if (topdown_run) {
1740 		char *str = NULL;
1741 		bool warn = false;
1742 
1743 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1744 		    stat_config.aggr_mode != AGGR_CORE) {
1745 			pr_err("top down event configuration requires --per-core mode\n");
1746 			return -1;
1747 		}
1748 		stat_config.aggr_mode = AGGR_CORE;
1749 		if (nr_cgroups || !target__has_cpu(&target)) {
1750 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1751 			return -1;
1752 		}
1753 
1754 		if (!force_metric_only)
1755 			stat_config.metric_only = true;
1756 		if (topdown_filter_events(topdown_attrs, &str,
1757 				arch_topdown_check_group(&warn)) < 0) {
1758 			pr_err("Out of memory\n");
1759 			return -1;
1760 		}
1761 		if (topdown_attrs[0] && str) {
1762 			if (warn)
1763 				arch_topdown_group_warn();
1764 			err = parse_events(evsel_list, str, &errinfo);
1765 			if (err) {
1766 				fprintf(stderr,
1767 					"Cannot set up top down events %s: %d\n",
1768 					str, err);
1769 				parse_events_print_error(&errinfo, str);
1770 				free(str);
1771 				return -1;
1772 			}
1773 		} else {
1774 			fprintf(stderr, "System does not support topdown\n");
1775 			return -1;
1776 		}
1777 		free(str);
1778 	}
1779 
1780 	if (!evsel_list->core.nr_entries) {
1781 		if (target__has_cpu(&target))
1782 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1783 
1784 		if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1785 			return -1;
1786 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1787 			if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1788 				return -1;
1789 		}
1790 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1791 			if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1792 				return -1;
1793 		}
1794 		if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1795 			return -1;
1796 	}
1797 
1798 	/* Detailed events get appended to the event list: */
1799 
1800 	if (detailed_run <  1)
1801 		return 0;
1802 
1803 	/* Append detailed run extra attributes: */
1804 	if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1805 		return -1;
1806 
1807 	if (detailed_run < 2)
1808 		return 0;
1809 
1810 	/* Append very detailed run extra attributes: */
1811 	if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1812 		return -1;
1813 
1814 	if (detailed_run < 3)
1815 		return 0;
1816 
1817 	/* Append very, very detailed run extra attributes: */
1818 	return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1819 }
1820 
1821 static const char * const stat_record_usage[] = {
1822 	"perf stat record [<options>]",
1823 	NULL,
1824 };
1825 
1826 static void init_features(struct perf_session *session)
1827 {
1828 	int feat;
1829 
1830 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1831 		perf_header__set_feat(&session->header, feat);
1832 
1833 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1834 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1835 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1836 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1837 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1838 }
1839 
1840 static int __cmd_record(int argc, const char **argv)
1841 {
1842 	struct perf_session *session;
1843 	struct perf_data *data = &perf_stat.data;
1844 
1845 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
1846 			     PARSE_OPT_STOP_AT_NON_OPTION);
1847 
1848 	if (output_name)
1849 		data->path = output_name;
1850 
1851 	if (stat_config.run_count != 1 || forever) {
1852 		pr_err("Cannot use -r option with perf stat record.\n");
1853 		return -1;
1854 	}
1855 
1856 	session = perf_session__new(data, false, NULL);
1857 	if (IS_ERR(session)) {
1858 		pr_err("Perf session creation failed\n");
1859 		return PTR_ERR(session);
1860 	}
1861 
1862 	init_features(session);
1863 
1864 	session->evlist   = evsel_list;
1865 	perf_stat.session = session;
1866 	perf_stat.record  = true;
1867 	return argc;
1868 }
1869 
1870 static int process_stat_round_event(struct perf_session *session,
1871 				    union perf_event *event)
1872 {
1873 	struct perf_record_stat_round *stat_round = &event->stat_round;
1874 	struct evsel *counter;
1875 	struct timespec tsh, *ts = NULL;
1876 	const char **argv = session->header.env.cmdline_argv;
1877 	int argc = session->header.env.nr_cmdline;
1878 
1879 	evlist__for_each_entry(evsel_list, counter)
1880 		perf_stat_process_counter(&stat_config, counter);
1881 
1882 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1883 		update_stats(&walltime_nsecs_stats, stat_round->time);
1884 
1885 	if (stat_config.interval && stat_round->time) {
1886 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
1887 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1888 		ts = &tsh;
1889 	}
1890 
1891 	print_counters(ts, argc, argv);
1892 	return 0;
1893 }
1894 
1895 static
1896 int process_stat_config_event(struct perf_session *session,
1897 			      union perf_event *event)
1898 {
1899 	struct perf_tool *tool = session->tool;
1900 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1901 
1902 	perf_event__read_stat_config(&stat_config, &event->stat_config);
1903 
1904 	if (perf_cpu_map__empty(st->cpus)) {
1905 		if (st->aggr_mode != AGGR_UNSET)
1906 			pr_warning("warning: processing task data, aggregation mode not set\n");
1907 		return 0;
1908 	}
1909 
1910 	if (st->aggr_mode != AGGR_UNSET)
1911 		stat_config.aggr_mode = st->aggr_mode;
1912 
1913 	if (perf_stat.data.is_pipe)
1914 		perf_stat_init_aggr_mode();
1915 	else
1916 		perf_stat_init_aggr_mode_file(st);
1917 
1918 	return 0;
1919 }
1920 
1921 static int set_maps(struct perf_stat *st)
1922 {
1923 	if (!st->cpus || !st->threads)
1924 		return 0;
1925 
1926 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1927 		return -EINVAL;
1928 
1929 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
1930 
1931 	if (perf_evlist__alloc_stats(evsel_list, true))
1932 		return -ENOMEM;
1933 
1934 	st->maps_allocated = true;
1935 	return 0;
1936 }
1937 
1938 static
1939 int process_thread_map_event(struct perf_session *session,
1940 			     union perf_event *event)
1941 {
1942 	struct perf_tool *tool = session->tool;
1943 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1944 
1945 	if (st->threads) {
1946 		pr_warning("Extra thread map event, ignoring.\n");
1947 		return 0;
1948 	}
1949 
1950 	st->threads = thread_map__new_event(&event->thread_map);
1951 	if (!st->threads)
1952 		return -ENOMEM;
1953 
1954 	return set_maps(st);
1955 }
1956 
1957 static
1958 int process_cpu_map_event(struct perf_session *session,
1959 			  union perf_event *event)
1960 {
1961 	struct perf_tool *tool = session->tool;
1962 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1963 	struct perf_cpu_map *cpus;
1964 
1965 	if (st->cpus) {
1966 		pr_warning("Extra cpu map event, ignoring.\n");
1967 		return 0;
1968 	}
1969 
1970 	cpus = cpu_map__new_data(&event->cpu_map.data);
1971 	if (!cpus)
1972 		return -ENOMEM;
1973 
1974 	st->cpus = cpus;
1975 	return set_maps(st);
1976 }
1977 
1978 static const char * const stat_report_usage[] = {
1979 	"perf stat report [<options>]",
1980 	NULL,
1981 };
1982 
1983 static struct perf_stat perf_stat = {
1984 	.tool = {
1985 		.attr		= perf_event__process_attr,
1986 		.event_update	= perf_event__process_event_update,
1987 		.thread_map	= process_thread_map_event,
1988 		.cpu_map	= process_cpu_map_event,
1989 		.stat_config	= process_stat_config_event,
1990 		.stat		= perf_event__process_stat_event,
1991 		.stat_round	= process_stat_round_event,
1992 	},
1993 	.aggr_mode = AGGR_UNSET,
1994 };
1995 
1996 static int __cmd_report(int argc, const char **argv)
1997 {
1998 	struct perf_session *session;
1999 	const struct option options[] = {
2000 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2001 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2002 		     "aggregate counts per processor socket", AGGR_SOCKET),
2003 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2004 		     "aggregate counts per processor die", AGGR_DIE),
2005 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2006 		     "aggregate counts per physical processor core", AGGR_CORE),
2007 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2008 		     "aggregate counts per numa node", AGGR_NODE),
2009 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2010 		     "disable CPU count aggregation", AGGR_NONE),
2011 	OPT_END()
2012 	};
2013 	struct stat st;
2014 	int ret;
2015 
2016 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2017 
2018 	if (!input_name || !strlen(input_name)) {
2019 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2020 			input_name = "-";
2021 		else
2022 			input_name = "perf.data";
2023 	}
2024 
2025 	perf_stat.data.path = input_name;
2026 	perf_stat.data.mode = PERF_DATA_MODE_READ;
2027 
2028 	session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2029 	if (IS_ERR(session))
2030 		return PTR_ERR(session);
2031 
2032 	perf_stat.session  = session;
2033 	stat_config.output = stderr;
2034 	evsel_list         = session->evlist;
2035 
2036 	ret = perf_session__process_events(session);
2037 	if (ret)
2038 		return ret;
2039 
2040 	perf_session__delete(session);
2041 	return 0;
2042 }
2043 
2044 static void setup_system_wide(int forks)
2045 {
2046 	/*
2047 	 * Make system wide (-a) the default target if
2048 	 * no target was specified and one of following
2049 	 * conditions is met:
2050 	 *
2051 	 *   - there's no workload specified
2052 	 *   - there is workload specified but all requested
2053 	 *     events are system wide events
2054 	 */
2055 	if (!target__none(&target))
2056 		return;
2057 
2058 	if (!forks)
2059 		target.system_wide = true;
2060 	else {
2061 		struct evsel *counter;
2062 
2063 		evlist__for_each_entry(evsel_list, counter) {
2064 			if (!counter->core.system_wide)
2065 				return;
2066 		}
2067 
2068 		if (evsel_list->core.nr_entries)
2069 			target.system_wide = true;
2070 	}
2071 }
2072 
2073 int cmd_stat(int argc, const char **argv)
2074 {
2075 	const char * const stat_usage[] = {
2076 		"perf stat [<options>] [<command>]",
2077 		NULL
2078 	};
2079 	int status = -EINVAL, run_idx;
2080 	const char *mode;
2081 	FILE *output = stderr;
2082 	unsigned int interval, timeout;
2083 	const char * const stat_subcommands[] = { "record", "report" };
2084 
2085 	setlocale(LC_ALL, "");
2086 
2087 	evsel_list = evlist__new();
2088 	if (evsel_list == NULL)
2089 		return -ENOMEM;
2090 
2091 	parse_events__shrink_config_terms();
2092 
2093 	/* String-parsing callback-based options would segfault when negated */
2094 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2095 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2096 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2097 
2098 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2099 					(const char **) stat_usage,
2100 					PARSE_OPT_STOP_AT_NON_OPTION);
2101 	perf_stat__collect_metric_expr(evsel_list);
2102 	perf_stat__init_shadow_stats();
2103 
2104 	if (stat_config.csv_sep) {
2105 		stat_config.csv_output = true;
2106 		if (!strcmp(stat_config.csv_sep, "\\t"))
2107 			stat_config.csv_sep = "\t";
2108 	} else
2109 		stat_config.csv_sep = DEFAULT_SEPARATOR;
2110 
2111 	if (argc && !strncmp(argv[0], "rec", 3)) {
2112 		argc = __cmd_record(argc, argv);
2113 		if (argc < 0)
2114 			return -1;
2115 	} else if (argc && !strncmp(argv[0], "rep", 3))
2116 		return __cmd_report(argc, argv);
2117 
2118 	interval = stat_config.interval;
2119 	timeout = stat_config.timeout;
2120 
2121 	/*
2122 	 * For record command the -o is already taken care of.
2123 	 */
2124 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2125 		output = NULL;
2126 
2127 	if (output_name && output_fd) {
2128 		fprintf(stderr, "cannot use both --output and --log-fd\n");
2129 		parse_options_usage(stat_usage, stat_options, "o", 1);
2130 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2131 		goto out;
2132 	}
2133 
2134 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2135 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2136 		goto out;
2137 	}
2138 
2139 	if (stat_config.metric_only && stat_config.run_count > 1) {
2140 		fprintf(stderr, "--metric-only is not supported with -r\n");
2141 		goto out;
2142 	}
2143 
2144 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2145 		fprintf(stderr, "--table is only supported with -r\n");
2146 		parse_options_usage(stat_usage, stat_options, "r", 1);
2147 		parse_options_usage(NULL, stat_options, "table", 0);
2148 		goto out;
2149 	}
2150 
2151 	if (output_fd < 0) {
2152 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2153 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2154 		goto out;
2155 	}
2156 
2157 	if (!output) {
2158 		struct timespec tm;
2159 		mode = append_file ? "a" : "w";
2160 
2161 		output = fopen(output_name, mode);
2162 		if (!output) {
2163 			perror("failed to create output file");
2164 			return -1;
2165 		}
2166 		clock_gettime(CLOCK_REALTIME, &tm);
2167 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2168 	} else if (output_fd > 0) {
2169 		mode = append_file ? "a" : "w";
2170 		output = fdopen(output_fd, mode);
2171 		if (!output) {
2172 			perror("Failed opening logfd");
2173 			return -errno;
2174 		}
2175 	}
2176 
2177 	stat_config.output = output;
2178 
2179 	/*
2180 	 * let the spreadsheet do the pretty-printing
2181 	 */
2182 	if (stat_config.csv_output) {
2183 		/* User explicitly passed -B? */
2184 		if (big_num_opt == 1) {
2185 			fprintf(stderr, "-B option not supported with -x\n");
2186 			parse_options_usage(stat_usage, stat_options, "B", 1);
2187 			parse_options_usage(NULL, stat_options, "x", 1);
2188 			goto out;
2189 		} else /* Nope, so disable big number formatting */
2190 			stat_config.big_num = false;
2191 	} else if (big_num_opt == 0) /* User passed --no-big-num */
2192 		stat_config.big_num = false;
2193 
2194 	setup_system_wide(argc);
2195 
2196 	/*
2197 	 * Display user/system times only for single
2198 	 * run and when there's specified tracee.
2199 	 */
2200 	if ((stat_config.run_count == 1) && target__none(&target))
2201 		stat_config.ru_display = true;
2202 
2203 	if (stat_config.run_count < 0) {
2204 		pr_err("Run count must be a positive number\n");
2205 		parse_options_usage(stat_usage, stat_options, "r", 1);
2206 		goto out;
2207 	} else if (stat_config.run_count == 0) {
2208 		forever = true;
2209 		stat_config.run_count = 1;
2210 	}
2211 
2212 	if (stat_config.walltime_run_table) {
2213 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2214 		if (!stat_config.walltime_run) {
2215 			pr_err("failed to setup -r option");
2216 			goto out;
2217 		}
2218 	}
2219 
2220 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2221 		!target__has_task(&target)) {
2222 		if (!target.system_wide || target.cpu_list) {
2223 			fprintf(stderr, "The --per-thread option is only "
2224 				"available when monitoring via -p -t -a "
2225 				"options or only --per-thread.\n");
2226 			parse_options_usage(NULL, stat_options, "p", 1);
2227 			parse_options_usage(NULL, stat_options, "t", 1);
2228 			goto out;
2229 		}
2230 	}
2231 
2232 	/*
2233 	 * no_aggr, cgroup are for system-wide only
2234 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2235 	 */
2236 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2237 	      stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2238 	    !target__has_cpu(&target)) {
2239 		fprintf(stderr, "both cgroup and no-aggregation "
2240 			"modes only available in system-wide mode\n");
2241 
2242 		parse_options_usage(stat_usage, stat_options, "G", 1);
2243 		parse_options_usage(NULL, stat_options, "A", 1);
2244 		parse_options_usage(NULL, stat_options, "a", 1);
2245 		goto out;
2246 	}
2247 
2248 	if (add_default_attributes())
2249 		goto out;
2250 
2251 	target__validate(&target);
2252 
2253 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2254 		target.per_thread = true;
2255 
2256 	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2257 		if (target__has_task(&target)) {
2258 			pr_err("Problems finding threads of monitor\n");
2259 			parse_options_usage(stat_usage, stat_options, "p", 1);
2260 			parse_options_usage(NULL, stat_options, "t", 1);
2261 		} else if (target__has_cpu(&target)) {
2262 			perror("failed to parse CPUs map");
2263 			parse_options_usage(stat_usage, stat_options, "C", 1);
2264 			parse_options_usage(NULL, stat_options, "a", 1);
2265 		}
2266 		goto out;
2267 	}
2268 
2269 	evlist__check_cpu_maps(evsel_list);
2270 
2271 	/*
2272 	 * Initialize thread_map with comm names,
2273 	 * so we could print it out on output.
2274 	 */
2275 	if (stat_config.aggr_mode == AGGR_THREAD) {
2276 		thread_map__read_comms(evsel_list->core.threads);
2277 		if (target.system_wide) {
2278 			if (runtime_stat_new(&stat_config,
2279 				perf_thread_map__nr(evsel_list->core.threads))) {
2280 				goto out;
2281 			}
2282 		}
2283 	}
2284 
2285 	if (stat_config.aggr_mode == AGGR_NODE)
2286 		cpu__setup_cpunode_map();
2287 
2288 	if (stat_config.times && interval)
2289 		interval_count = true;
2290 	else if (stat_config.times && !interval) {
2291 		pr_err("interval-count option should be used together with "
2292 				"interval-print.\n");
2293 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2294 		parse_options_usage(stat_usage, stat_options, "I", 1);
2295 		goto out;
2296 	}
2297 
2298 	if (timeout && timeout < 100) {
2299 		if (timeout < 10) {
2300 			pr_err("timeout must be >= 10ms.\n");
2301 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2302 			goto out;
2303 		} else
2304 			pr_warning("timeout < 100ms. "
2305 				   "The overhead percentage could be high in some cases. "
2306 				   "Please proceed with caution.\n");
2307 	}
2308 	if (timeout && interval) {
2309 		pr_err("timeout option is not supported with interval-print.\n");
2310 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2311 		parse_options_usage(stat_usage, stat_options, "I", 1);
2312 		goto out;
2313 	}
2314 
2315 	if (perf_evlist__alloc_stats(evsel_list, interval))
2316 		goto out;
2317 
2318 	if (perf_stat_init_aggr_mode())
2319 		goto out;
2320 
2321 	/*
2322 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2323 	 * while avoiding that older tools show confusing messages.
2324 	 *
2325 	 * However for pipe sessions we need to keep it zero,
2326 	 * because script's perf_evsel__check_attr is triggered
2327 	 * by attr->sample_type != 0, and we can't run it on
2328 	 * stat sessions.
2329 	 */
2330 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2331 
2332 	/*
2333 	 * We dont want to block the signals - that would cause
2334 	 * child tasks to inherit that and Ctrl-C would not work.
2335 	 * What we want is for Ctrl-C to work in the exec()-ed
2336 	 * task, but being ignored by perf stat itself:
2337 	 */
2338 	atexit(sig_atexit);
2339 	if (!forever)
2340 		signal(SIGINT,  skip_signal);
2341 	signal(SIGCHLD, skip_signal);
2342 	signal(SIGALRM, skip_signal);
2343 	signal(SIGABRT, skip_signal);
2344 
2345 	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2346 		goto out;
2347 
2348 	status = 0;
2349 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2350 		if (stat_config.run_count != 1 && verbose > 0)
2351 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2352 				run_idx + 1);
2353 
2354 		if (run_idx != 0)
2355 			perf_evlist__reset_prev_raw_counts(evsel_list);
2356 
2357 		status = run_perf_stat(argc, argv, run_idx);
2358 		if (forever && status != -1 && !interval) {
2359 			print_counters(NULL, argc, argv);
2360 			perf_stat__reset_stats();
2361 		}
2362 	}
2363 
2364 	if (!forever && status != -1 && (!interval || stat_config.summary))
2365 		print_counters(NULL, argc, argv);
2366 
2367 	evlist__finalize_ctlfd(evsel_list);
2368 
2369 	if (STAT_RECORD) {
2370 		/*
2371 		 * We synthesize the kernel mmap record just so that older tools
2372 		 * don't emit warnings about not being able to resolve symbols
2373 		 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2374 		 * a saner message about no samples being in the perf.data file.
2375 		 *
2376 		 * This also serves to suppress a warning about f_header.data.size == 0
2377 		 * in header.c at the moment 'perf stat record' gets introduced, which
2378 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2379 		 * records, but the need to suppress the kptr_restrict messages in older
2380 		 * tools remain  -acme
2381 		 */
2382 		int fd = perf_data__fd(&perf_stat.data);
2383 		int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2384 							     process_synthesized_event,
2385 							     &perf_stat.session->machines.host);
2386 		if (err) {
2387 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2388 				   "older tools may produce warnings about this file\n.");
2389 		}
2390 
2391 		if (!interval) {
2392 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2393 				pr_err("failed to write stat round event\n");
2394 		}
2395 
2396 		if (!perf_stat.data.is_pipe) {
2397 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2398 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2399 		}
2400 
2401 		evlist__close(evsel_list);
2402 		perf_session__delete(perf_stat.session);
2403 	}
2404 
2405 	perf_stat__exit_aggr_mode();
2406 	perf_evlist__free_stats(evsel_list);
2407 out:
2408 	zfree(&stat_config.walltime_run);
2409 
2410 	if (smi_cost && smi_reset)
2411 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2412 
2413 	evlist__delete(evsel_list);
2414 
2415 	metricgroup__rblist_exit(&stat_config.metric_events);
2416 	runtime_stat_delete(&stat_config);
2417 
2418 	return status;
2419 }
2420