xref: /openbmc/linux/tools/perf/builtin-stat.c (revision 911b8eac)
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.stop_read_counter && (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 && stat_config.summary) {
901 		stat_config.interval = 0;
902 		stat_config.stop_read_counter = 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 	OPT_BOOLEAN(0, "summary", &stat_config.summary,
1168 		       "print summary for interval mode"),
1169 #ifdef HAVE_LIBPFM
1170 	OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1171 		"libpfm4 event selector. use 'perf list' to list available events",
1172 		parse_libpfm_events_option),
1173 #endif
1174 	OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd]",
1175 		     "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1176 		     "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.",
1177 		      parse_control_option),
1178 	OPT_END()
1179 };
1180 
1181 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1182 				 struct perf_cpu_map *map, int cpu)
1183 {
1184 	return cpu_map__get_socket(map, cpu, NULL);
1185 }
1186 
1187 static int perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1188 			      struct perf_cpu_map *map, int cpu)
1189 {
1190 	return cpu_map__get_die(map, cpu, NULL);
1191 }
1192 
1193 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1194 			       struct perf_cpu_map *map, int cpu)
1195 {
1196 	return cpu_map__get_core(map, cpu, NULL);
1197 }
1198 
1199 static int perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1200 			       struct perf_cpu_map *map, int cpu)
1201 {
1202 	return cpu_map__get_node(map, cpu, NULL);
1203 }
1204 
1205 static int perf_stat__get_aggr(struct perf_stat_config *config,
1206 			       aggr_get_id_t get_id, struct perf_cpu_map *map, int idx)
1207 {
1208 	int cpu;
1209 
1210 	if (idx >= map->nr)
1211 		return -1;
1212 
1213 	cpu = map->map[idx];
1214 
1215 	if (config->cpus_aggr_map->map[cpu] == -1)
1216 		config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
1217 
1218 	return config->cpus_aggr_map->map[cpu];
1219 }
1220 
1221 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
1222 					struct perf_cpu_map *map, int idx)
1223 {
1224 	return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
1225 }
1226 
1227 static int perf_stat__get_die_cached(struct perf_stat_config *config,
1228 					struct perf_cpu_map *map, int idx)
1229 {
1230 	return perf_stat__get_aggr(config, perf_stat__get_die, map, idx);
1231 }
1232 
1233 static int perf_stat__get_core_cached(struct perf_stat_config *config,
1234 				      struct perf_cpu_map *map, int idx)
1235 {
1236 	return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
1237 }
1238 
1239 static int perf_stat__get_node_cached(struct perf_stat_config *config,
1240 				      struct perf_cpu_map *map, int idx)
1241 {
1242 	return perf_stat__get_aggr(config, perf_stat__get_node, map, idx);
1243 }
1244 
1245 static bool term_percore_set(void)
1246 {
1247 	struct evsel *counter;
1248 
1249 	evlist__for_each_entry(evsel_list, counter) {
1250 		if (counter->percore)
1251 			return true;
1252 	}
1253 
1254 	return false;
1255 }
1256 
1257 static int perf_stat_init_aggr_mode(void)
1258 {
1259 	int nr;
1260 
1261 	switch (stat_config.aggr_mode) {
1262 	case AGGR_SOCKET:
1263 		if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1264 			perror("cannot build socket map");
1265 			return -1;
1266 		}
1267 		stat_config.aggr_get_id = perf_stat__get_socket_cached;
1268 		break;
1269 	case AGGR_DIE:
1270 		if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1271 			perror("cannot build die map");
1272 			return -1;
1273 		}
1274 		stat_config.aggr_get_id = perf_stat__get_die_cached;
1275 		break;
1276 	case AGGR_CORE:
1277 		if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1278 			perror("cannot build core map");
1279 			return -1;
1280 		}
1281 		stat_config.aggr_get_id = perf_stat__get_core_cached;
1282 		break;
1283 	case AGGR_NODE:
1284 		if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1285 			perror("cannot build core map");
1286 			return -1;
1287 		}
1288 		stat_config.aggr_get_id = perf_stat__get_node_cached;
1289 		break;
1290 	case AGGR_NONE:
1291 		if (term_percore_set()) {
1292 			if (cpu_map__build_core_map(evsel_list->core.cpus,
1293 						    &stat_config.aggr_map)) {
1294 				perror("cannot build core map");
1295 				return -1;
1296 			}
1297 			stat_config.aggr_get_id = perf_stat__get_core_cached;
1298 		}
1299 		break;
1300 	case AGGR_GLOBAL:
1301 	case AGGR_THREAD:
1302 	case AGGR_UNSET:
1303 	default:
1304 		break;
1305 	}
1306 
1307 	/*
1308 	 * The evsel_list->cpus is the base we operate on,
1309 	 * taking the highest cpu number to be the size of
1310 	 * the aggregation translate cpumap.
1311 	 */
1312 	nr = perf_cpu_map__max(evsel_list->core.cpus);
1313 	stat_config.cpus_aggr_map = perf_cpu_map__empty_new(nr + 1);
1314 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1315 }
1316 
1317 static void perf_stat__exit_aggr_mode(void)
1318 {
1319 	perf_cpu_map__put(stat_config.aggr_map);
1320 	perf_cpu_map__put(stat_config.cpus_aggr_map);
1321 	stat_config.aggr_map = NULL;
1322 	stat_config.cpus_aggr_map = NULL;
1323 }
1324 
1325 static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx)
1326 {
1327 	int cpu;
1328 
1329 	if (idx > map->nr)
1330 		return -1;
1331 
1332 	cpu = map->map[idx];
1333 
1334 	if (cpu >= env->nr_cpus_avail)
1335 		return -1;
1336 
1337 	return cpu;
1338 }
1339 
1340 static int perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data)
1341 {
1342 	struct perf_env *env = data;
1343 	int cpu = perf_env__get_cpu(env, map, idx);
1344 
1345 	return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1346 }
1347 
1348 static int perf_env__get_die(struct perf_cpu_map *map, int idx, void *data)
1349 {
1350 	struct perf_env *env = data;
1351 	int die_id = -1, cpu = perf_env__get_cpu(env, map, idx);
1352 
1353 	if (cpu != -1) {
1354 		/*
1355 		 * Encode socket in bit range 15:8
1356 		 * die_id is relative to socket,
1357 		 * we need a global id. So we combine
1358 		 * socket + die id
1359 		 */
1360 		if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
1361 			return -1;
1362 
1363 		if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
1364 			return -1;
1365 
1366 		die_id = (env->cpu[cpu].socket_id << 8) | (env->cpu[cpu].die_id & 0xff);
1367 	}
1368 
1369 	return die_id;
1370 }
1371 
1372 static int perf_env__get_core(struct perf_cpu_map *map, int idx, void *data)
1373 {
1374 	struct perf_env *env = data;
1375 	int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1376 
1377 	if (cpu != -1) {
1378 		/*
1379 		 * Encode socket in bit range 31:24
1380 		 * encode die id in bit range 23:16
1381 		 * core_id is relative to socket and die,
1382 		 * we need a global id. So we combine
1383 		 * socket + die id + core id
1384 		 */
1385 		if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
1386 			return -1;
1387 
1388 		if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
1389 			return -1;
1390 
1391 		if (WARN_ONCE(env->cpu[cpu].core_id >> 16, "The core id number is too big.\n"))
1392 			return -1;
1393 
1394 		core = (env->cpu[cpu].socket_id << 24) |
1395 		       (env->cpu[cpu].die_id << 16) |
1396 		       (env->cpu[cpu].core_id & 0xffff);
1397 	}
1398 
1399 	return core;
1400 }
1401 
1402 static int perf_env__get_node(struct perf_cpu_map *map, int idx, void *data)
1403 {
1404 	int cpu = perf_env__get_cpu(data, map, idx);
1405 
1406 	return perf_env__numa_node(data, cpu);
1407 }
1408 
1409 static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus,
1410 				      struct perf_cpu_map **sockp)
1411 {
1412 	return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1413 }
1414 
1415 static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus,
1416 				   struct perf_cpu_map **diep)
1417 {
1418 	return cpu_map__build_map(cpus, diep, perf_env__get_die, env);
1419 }
1420 
1421 static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus,
1422 				    struct perf_cpu_map **corep)
1423 {
1424 	return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1425 }
1426 
1427 static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus,
1428 				    struct perf_cpu_map **nodep)
1429 {
1430 	return cpu_map__build_map(cpus, nodep, perf_env__get_node, env);
1431 }
1432 
1433 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1434 				      struct perf_cpu_map *map, int idx)
1435 {
1436 	return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1437 }
1438 static int perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1439 				   struct perf_cpu_map *map, int idx)
1440 {
1441 	return perf_env__get_die(map, idx, &perf_stat.session->header.env);
1442 }
1443 
1444 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1445 				    struct perf_cpu_map *map, int idx)
1446 {
1447 	return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1448 }
1449 
1450 static int perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1451 				    struct perf_cpu_map *map, int idx)
1452 {
1453 	return perf_env__get_node(map, idx, &perf_stat.session->header.env);
1454 }
1455 
1456 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1457 {
1458 	struct perf_env *env = &st->session->header.env;
1459 
1460 	switch (stat_config.aggr_mode) {
1461 	case AGGR_SOCKET:
1462 		if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1463 			perror("cannot build socket map");
1464 			return -1;
1465 		}
1466 		stat_config.aggr_get_id = perf_stat__get_socket_file;
1467 		break;
1468 	case AGGR_DIE:
1469 		if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1470 			perror("cannot build die map");
1471 			return -1;
1472 		}
1473 		stat_config.aggr_get_id = perf_stat__get_die_file;
1474 		break;
1475 	case AGGR_CORE:
1476 		if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1477 			perror("cannot build core map");
1478 			return -1;
1479 		}
1480 		stat_config.aggr_get_id = perf_stat__get_core_file;
1481 		break;
1482 	case AGGR_NODE:
1483 		if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1484 			perror("cannot build core map");
1485 			return -1;
1486 		}
1487 		stat_config.aggr_get_id = perf_stat__get_node_file;
1488 		break;
1489 	case AGGR_NONE:
1490 	case AGGR_GLOBAL:
1491 	case AGGR_THREAD:
1492 	case AGGR_UNSET:
1493 	default:
1494 		break;
1495 	}
1496 
1497 	return 0;
1498 }
1499 
1500 static int topdown_filter_events(const char **attr, char **str, bool use_group)
1501 {
1502 	int off = 0;
1503 	int i;
1504 	int len = 0;
1505 	char *s;
1506 
1507 	for (i = 0; attr[i]; i++) {
1508 		if (pmu_have_event("cpu", attr[i])) {
1509 			len += strlen(attr[i]) + 1;
1510 			attr[i - off] = attr[i];
1511 		} else
1512 			off++;
1513 	}
1514 	attr[i - off] = NULL;
1515 
1516 	*str = malloc(len + 1 + 2);
1517 	if (!*str)
1518 		return -1;
1519 	s = *str;
1520 	if (i - off == 0) {
1521 		*s = 0;
1522 		return 0;
1523 	}
1524 	if (use_group)
1525 		*s++ = '{';
1526 	for (i = 0; attr[i]; i++) {
1527 		strcpy(s, attr[i]);
1528 		s += strlen(s);
1529 		*s++ = ',';
1530 	}
1531 	if (use_group) {
1532 		s[-1] = '}';
1533 		*s = 0;
1534 	} else
1535 		s[-1] = 0;
1536 	return 0;
1537 }
1538 
1539 __weak bool arch_topdown_check_group(bool *warn)
1540 {
1541 	*warn = false;
1542 	return false;
1543 }
1544 
1545 __weak void arch_topdown_group_warn(void)
1546 {
1547 }
1548 
1549 /*
1550  * Add default attributes, if there were no attributes specified or
1551  * if -d/--detailed, -d -d or -d -d -d is used:
1552  */
1553 static int add_default_attributes(void)
1554 {
1555 	int err;
1556 	struct perf_event_attr default_attrs0[] = {
1557 
1558   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1559   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1560   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1561   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1562 
1563   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1564 };
1565 	struct perf_event_attr frontend_attrs[] = {
1566   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1567 };
1568 	struct perf_event_attr backend_attrs[] = {
1569   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1570 };
1571 	struct perf_event_attr default_attrs1[] = {
1572   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1573   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1574   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1575 
1576 };
1577 
1578 /*
1579  * Detailed stats (-d), covering the L1 and last level data caches:
1580  */
1581 	struct perf_event_attr detailed_attrs[] = {
1582 
1583   { .type = PERF_TYPE_HW_CACHE,
1584     .config =
1585 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1586 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1587 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1588 
1589   { .type = PERF_TYPE_HW_CACHE,
1590     .config =
1591 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1592 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1593 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1594 
1595   { .type = PERF_TYPE_HW_CACHE,
1596     .config =
1597 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1598 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1599 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1600 
1601   { .type = PERF_TYPE_HW_CACHE,
1602     .config =
1603 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1604 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1605 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1606 };
1607 
1608 /*
1609  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1610  */
1611 	struct perf_event_attr very_detailed_attrs[] = {
1612 
1613   { .type = PERF_TYPE_HW_CACHE,
1614     .config =
1615 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1616 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1617 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1618 
1619   { .type = PERF_TYPE_HW_CACHE,
1620     .config =
1621 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1622 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1623 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1624 
1625   { .type = PERF_TYPE_HW_CACHE,
1626     .config =
1627 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1628 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1629 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1630 
1631   { .type = PERF_TYPE_HW_CACHE,
1632     .config =
1633 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1634 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1635 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1636 
1637   { .type = PERF_TYPE_HW_CACHE,
1638     .config =
1639 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1640 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1641 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1642 
1643   { .type = PERF_TYPE_HW_CACHE,
1644     .config =
1645 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1646 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1647 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1648 
1649 };
1650 
1651 /*
1652  * Very, very detailed stats (-d -d -d), adding prefetch events:
1653  */
1654 	struct perf_event_attr very_very_detailed_attrs[] = {
1655 
1656   { .type = PERF_TYPE_HW_CACHE,
1657     .config =
1658 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1659 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1660 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1661 
1662   { .type = PERF_TYPE_HW_CACHE,
1663     .config =
1664 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1665 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1666 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1667 };
1668 	struct parse_events_error errinfo;
1669 
1670 	/* Set attrs if no event is selected and !null_run: */
1671 	if (stat_config.null_run)
1672 		return 0;
1673 
1674 	bzero(&errinfo, sizeof(errinfo));
1675 	if (transaction_run) {
1676 		/* Handle -T as -M transaction. Once platform specific metrics
1677 		 * support has been added to the json files, all archictures
1678 		 * will use this approach. To determine transaction support
1679 		 * on an architecture test for such a metric name.
1680 		 */
1681 		if (metricgroup__has_metric("transaction")) {
1682 			struct option opt = { .value = &evsel_list };
1683 
1684 			return metricgroup__parse_groups(&opt, "transaction",
1685 							 stat_config.metric_no_group,
1686 							stat_config.metric_no_merge,
1687 							 &stat_config.metric_events);
1688 		}
1689 
1690 		if (pmu_have_event("cpu", "cycles-ct") &&
1691 		    pmu_have_event("cpu", "el-start"))
1692 			err = parse_events(evsel_list, transaction_attrs,
1693 					   &errinfo);
1694 		else
1695 			err = parse_events(evsel_list,
1696 					   transaction_limited_attrs,
1697 					   &errinfo);
1698 		if (err) {
1699 			fprintf(stderr, "Cannot set up transaction events\n");
1700 			parse_events_print_error(&errinfo, transaction_attrs);
1701 			return -1;
1702 		}
1703 		return 0;
1704 	}
1705 
1706 	if (smi_cost) {
1707 		int smi;
1708 
1709 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1710 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1711 			return -1;
1712 		}
1713 
1714 		if (!smi) {
1715 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1716 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1717 				return -1;
1718 			}
1719 			smi_reset = true;
1720 		}
1721 
1722 		if (pmu_have_event("msr", "aperf") &&
1723 		    pmu_have_event("msr", "smi")) {
1724 			if (!force_metric_only)
1725 				stat_config.metric_only = true;
1726 			err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1727 		} else {
1728 			fprintf(stderr, "To measure SMI cost, it needs "
1729 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1730 			parse_events_print_error(&errinfo, smi_cost_attrs);
1731 			return -1;
1732 		}
1733 		if (err) {
1734 			parse_events_print_error(&errinfo, smi_cost_attrs);
1735 			fprintf(stderr, "Cannot set up SMI cost events\n");
1736 			return -1;
1737 		}
1738 		return 0;
1739 	}
1740 
1741 	if (topdown_run) {
1742 		char *str = NULL;
1743 		bool warn = false;
1744 
1745 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1746 		    stat_config.aggr_mode != AGGR_CORE) {
1747 			pr_err("top down event configuration requires --per-core mode\n");
1748 			return -1;
1749 		}
1750 		stat_config.aggr_mode = AGGR_CORE;
1751 		if (nr_cgroups || !target__has_cpu(&target)) {
1752 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1753 			return -1;
1754 		}
1755 
1756 		if (!force_metric_only)
1757 			stat_config.metric_only = true;
1758 		if (topdown_filter_events(topdown_attrs, &str,
1759 				arch_topdown_check_group(&warn)) < 0) {
1760 			pr_err("Out of memory\n");
1761 			return -1;
1762 		}
1763 		if (topdown_attrs[0] && str) {
1764 			if (warn)
1765 				arch_topdown_group_warn();
1766 			err = parse_events(evsel_list, str, &errinfo);
1767 			if (err) {
1768 				fprintf(stderr,
1769 					"Cannot set up top down events %s: %d\n",
1770 					str, err);
1771 				parse_events_print_error(&errinfo, str);
1772 				free(str);
1773 				return -1;
1774 			}
1775 		} else {
1776 			fprintf(stderr, "System does not support topdown\n");
1777 			return -1;
1778 		}
1779 		free(str);
1780 	}
1781 
1782 	if (!evsel_list->core.nr_entries) {
1783 		if (target__has_cpu(&target))
1784 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1785 
1786 		if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1787 			return -1;
1788 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1789 			if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1790 				return -1;
1791 		}
1792 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1793 			if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1794 				return -1;
1795 		}
1796 		if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1797 			return -1;
1798 	}
1799 
1800 	/* Detailed events get appended to the event list: */
1801 
1802 	if (detailed_run <  1)
1803 		return 0;
1804 
1805 	/* Append detailed run extra attributes: */
1806 	if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1807 		return -1;
1808 
1809 	if (detailed_run < 2)
1810 		return 0;
1811 
1812 	/* Append very detailed run extra attributes: */
1813 	if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1814 		return -1;
1815 
1816 	if (detailed_run < 3)
1817 		return 0;
1818 
1819 	/* Append very, very detailed run extra attributes: */
1820 	return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1821 }
1822 
1823 static const char * const stat_record_usage[] = {
1824 	"perf stat record [<options>]",
1825 	NULL,
1826 };
1827 
1828 static void init_features(struct perf_session *session)
1829 {
1830 	int feat;
1831 
1832 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1833 		perf_header__set_feat(&session->header, feat);
1834 
1835 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1836 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1837 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1838 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1839 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1840 }
1841 
1842 static int __cmd_record(int argc, const char **argv)
1843 {
1844 	struct perf_session *session;
1845 	struct perf_data *data = &perf_stat.data;
1846 
1847 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
1848 			     PARSE_OPT_STOP_AT_NON_OPTION);
1849 
1850 	if (output_name)
1851 		data->path = output_name;
1852 
1853 	if (stat_config.run_count != 1 || forever) {
1854 		pr_err("Cannot use -r option with perf stat record.\n");
1855 		return -1;
1856 	}
1857 
1858 	session = perf_session__new(data, false, NULL);
1859 	if (IS_ERR(session)) {
1860 		pr_err("Perf session creation failed\n");
1861 		return PTR_ERR(session);
1862 	}
1863 
1864 	init_features(session);
1865 
1866 	session->evlist   = evsel_list;
1867 	perf_stat.session = session;
1868 	perf_stat.record  = true;
1869 	return argc;
1870 }
1871 
1872 static int process_stat_round_event(struct perf_session *session,
1873 				    union perf_event *event)
1874 {
1875 	struct perf_record_stat_round *stat_round = &event->stat_round;
1876 	struct evsel *counter;
1877 	struct timespec tsh, *ts = NULL;
1878 	const char **argv = session->header.env.cmdline_argv;
1879 	int argc = session->header.env.nr_cmdline;
1880 
1881 	evlist__for_each_entry(evsel_list, counter)
1882 		perf_stat_process_counter(&stat_config, counter);
1883 
1884 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1885 		update_stats(&walltime_nsecs_stats, stat_round->time);
1886 
1887 	if (stat_config.interval && stat_round->time) {
1888 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
1889 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1890 		ts = &tsh;
1891 	}
1892 
1893 	print_counters(ts, argc, argv);
1894 	return 0;
1895 }
1896 
1897 static
1898 int process_stat_config_event(struct perf_session *session,
1899 			      union perf_event *event)
1900 {
1901 	struct perf_tool *tool = session->tool;
1902 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1903 
1904 	perf_event__read_stat_config(&stat_config, &event->stat_config);
1905 
1906 	if (perf_cpu_map__empty(st->cpus)) {
1907 		if (st->aggr_mode != AGGR_UNSET)
1908 			pr_warning("warning: processing task data, aggregation mode not set\n");
1909 		return 0;
1910 	}
1911 
1912 	if (st->aggr_mode != AGGR_UNSET)
1913 		stat_config.aggr_mode = st->aggr_mode;
1914 
1915 	if (perf_stat.data.is_pipe)
1916 		perf_stat_init_aggr_mode();
1917 	else
1918 		perf_stat_init_aggr_mode_file(st);
1919 
1920 	return 0;
1921 }
1922 
1923 static int set_maps(struct perf_stat *st)
1924 {
1925 	if (!st->cpus || !st->threads)
1926 		return 0;
1927 
1928 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1929 		return -EINVAL;
1930 
1931 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
1932 
1933 	if (perf_evlist__alloc_stats(evsel_list, true))
1934 		return -ENOMEM;
1935 
1936 	st->maps_allocated = true;
1937 	return 0;
1938 }
1939 
1940 static
1941 int process_thread_map_event(struct perf_session *session,
1942 			     union perf_event *event)
1943 {
1944 	struct perf_tool *tool = session->tool;
1945 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1946 
1947 	if (st->threads) {
1948 		pr_warning("Extra thread map event, ignoring.\n");
1949 		return 0;
1950 	}
1951 
1952 	st->threads = thread_map__new_event(&event->thread_map);
1953 	if (!st->threads)
1954 		return -ENOMEM;
1955 
1956 	return set_maps(st);
1957 }
1958 
1959 static
1960 int process_cpu_map_event(struct perf_session *session,
1961 			  union perf_event *event)
1962 {
1963 	struct perf_tool *tool = session->tool;
1964 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1965 	struct perf_cpu_map *cpus;
1966 
1967 	if (st->cpus) {
1968 		pr_warning("Extra cpu map event, ignoring.\n");
1969 		return 0;
1970 	}
1971 
1972 	cpus = cpu_map__new_data(&event->cpu_map.data);
1973 	if (!cpus)
1974 		return -ENOMEM;
1975 
1976 	st->cpus = cpus;
1977 	return set_maps(st);
1978 }
1979 
1980 static const char * const stat_report_usage[] = {
1981 	"perf stat report [<options>]",
1982 	NULL,
1983 };
1984 
1985 static struct perf_stat perf_stat = {
1986 	.tool = {
1987 		.attr		= perf_event__process_attr,
1988 		.event_update	= perf_event__process_event_update,
1989 		.thread_map	= process_thread_map_event,
1990 		.cpu_map	= process_cpu_map_event,
1991 		.stat_config	= process_stat_config_event,
1992 		.stat		= perf_event__process_stat_event,
1993 		.stat_round	= process_stat_round_event,
1994 	},
1995 	.aggr_mode = AGGR_UNSET,
1996 };
1997 
1998 static int __cmd_report(int argc, const char **argv)
1999 {
2000 	struct perf_session *session;
2001 	const struct option options[] = {
2002 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
2003 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2004 		     "aggregate counts per processor socket", AGGR_SOCKET),
2005 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2006 		     "aggregate counts per processor die", AGGR_DIE),
2007 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2008 		     "aggregate counts per physical processor core", AGGR_CORE),
2009 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2010 		     "aggregate counts per numa node", AGGR_NODE),
2011 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2012 		     "disable CPU count aggregation", AGGR_NONE),
2013 	OPT_END()
2014 	};
2015 	struct stat st;
2016 	int ret;
2017 
2018 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
2019 
2020 	if (!input_name || !strlen(input_name)) {
2021 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2022 			input_name = "-";
2023 		else
2024 			input_name = "perf.data";
2025 	}
2026 
2027 	perf_stat.data.path = input_name;
2028 	perf_stat.data.mode = PERF_DATA_MODE_READ;
2029 
2030 	session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2031 	if (IS_ERR(session))
2032 		return PTR_ERR(session);
2033 
2034 	perf_stat.session  = session;
2035 	stat_config.output = stderr;
2036 	evsel_list         = session->evlist;
2037 
2038 	ret = perf_session__process_events(session);
2039 	if (ret)
2040 		return ret;
2041 
2042 	perf_session__delete(session);
2043 	return 0;
2044 }
2045 
2046 static void setup_system_wide(int forks)
2047 {
2048 	/*
2049 	 * Make system wide (-a) the default target if
2050 	 * no target was specified and one of following
2051 	 * conditions is met:
2052 	 *
2053 	 *   - there's no workload specified
2054 	 *   - there is workload specified but all requested
2055 	 *     events are system wide events
2056 	 */
2057 	if (!target__none(&target))
2058 		return;
2059 
2060 	if (!forks)
2061 		target.system_wide = true;
2062 	else {
2063 		struct evsel *counter;
2064 
2065 		evlist__for_each_entry(evsel_list, counter) {
2066 			if (!counter->core.system_wide)
2067 				return;
2068 		}
2069 
2070 		if (evsel_list->core.nr_entries)
2071 			target.system_wide = true;
2072 	}
2073 }
2074 
2075 int cmd_stat(int argc, const char **argv)
2076 {
2077 	const char * const stat_usage[] = {
2078 		"perf stat [<options>] [<command>]",
2079 		NULL
2080 	};
2081 	int status = -EINVAL, run_idx;
2082 	const char *mode;
2083 	FILE *output = stderr;
2084 	unsigned int interval, timeout;
2085 	const char * const stat_subcommands[] = { "record", "report" };
2086 
2087 	setlocale(LC_ALL, "");
2088 
2089 	evsel_list = evlist__new();
2090 	if (evsel_list == NULL)
2091 		return -ENOMEM;
2092 
2093 	parse_events__shrink_config_terms();
2094 
2095 	/* String-parsing callback-based options would segfault when negated */
2096 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2097 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2098 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2099 
2100 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2101 					(const char **) stat_usage,
2102 					PARSE_OPT_STOP_AT_NON_OPTION);
2103 	perf_stat__collect_metric_expr(evsel_list);
2104 	perf_stat__init_shadow_stats();
2105 
2106 	if (stat_config.csv_sep) {
2107 		stat_config.csv_output = true;
2108 		if (!strcmp(stat_config.csv_sep, "\\t"))
2109 			stat_config.csv_sep = "\t";
2110 	} else
2111 		stat_config.csv_sep = DEFAULT_SEPARATOR;
2112 
2113 	if (argc && !strncmp(argv[0], "rec", 3)) {
2114 		argc = __cmd_record(argc, argv);
2115 		if (argc < 0)
2116 			return -1;
2117 	} else if (argc && !strncmp(argv[0], "rep", 3))
2118 		return __cmd_report(argc, argv);
2119 
2120 	interval = stat_config.interval;
2121 	timeout = stat_config.timeout;
2122 
2123 	/*
2124 	 * For record command the -o is already taken care of.
2125 	 */
2126 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2127 		output = NULL;
2128 
2129 	if (output_name && output_fd) {
2130 		fprintf(stderr, "cannot use both --output and --log-fd\n");
2131 		parse_options_usage(stat_usage, stat_options, "o", 1);
2132 		parse_options_usage(NULL, stat_options, "log-fd", 0);
2133 		goto out;
2134 	}
2135 
2136 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2137 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2138 		goto out;
2139 	}
2140 
2141 	if (stat_config.metric_only && stat_config.run_count > 1) {
2142 		fprintf(stderr, "--metric-only is not supported with -r\n");
2143 		goto out;
2144 	}
2145 
2146 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2147 		fprintf(stderr, "--table is only supported with -r\n");
2148 		parse_options_usage(stat_usage, stat_options, "r", 1);
2149 		parse_options_usage(NULL, stat_options, "table", 0);
2150 		goto out;
2151 	}
2152 
2153 	if (output_fd < 0) {
2154 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
2155 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2156 		goto out;
2157 	}
2158 
2159 	if (!output) {
2160 		struct timespec tm;
2161 		mode = append_file ? "a" : "w";
2162 
2163 		output = fopen(output_name, mode);
2164 		if (!output) {
2165 			perror("failed to create output file");
2166 			return -1;
2167 		}
2168 		clock_gettime(CLOCK_REALTIME, &tm);
2169 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2170 	} else if (output_fd > 0) {
2171 		mode = append_file ? "a" : "w";
2172 		output = fdopen(output_fd, mode);
2173 		if (!output) {
2174 			perror("Failed opening logfd");
2175 			return -errno;
2176 		}
2177 	}
2178 
2179 	stat_config.output = output;
2180 
2181 	/*
2182 	 * let the spreadsheet do the pretty-printing
2183 	 */
2184 	if (stat_config.csv_output) {
2185 		/* User explicitly passed -B? */
2186 		if (big_num_opt == 1) {
2187 			fprintf(stderr, "-B option not supported with -x\n");
2188 			parse_options_usage(stat_usage, stat_options, "B", 1);
2189 			parse_options_usage(NULL, stat_options, "x", 1);
2190 			goto out;
2191 		} else /* Nope, so disable big number formatting */
2192 			stat_config.big_num = false;
2193 	} else if (big_num_opt == 0) /* User passed --no-big-num */
2194 		stat_config.big_num = false;
2195 
2196 	setup_system_wide(argc);
2197 
2198 	/*
2199 	 * Display user/system times only for single
2200 	 * run and when there's specified tracee.
2201 	 */
2202 	if ((stat_config.run_count == 1) && target__none(&target))
2203 		stat_config.ru_display = true;
2204 
2205 	if (stat_config.run_count < 0) {
2206 		pr_err("Run count must be a positive number\n");
2207 		parse_options_usage(stat_usage, stat_options, "r", 1);
2208 		goto out;
2209 	} else if (stat_config.run_count == 0) {
2210 		forever = true;
2211 		stat_config.run_count = 1;
2212 	}
2213 
2214 	if (stat_config.walltime_run_table) {
2215 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2216 		if (!stat_config.walltime_run) {
2217 			pr_err("failed to setup -r option");
2218 			goto out;
2219 		}
2220 	}
2221 
2222 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2223 		!target__has_task(&target)) {
2224 		if (!target.system_wide || target.cpu_list) {
2225 			fprintf(stderr, "The --per-thread option is only "
2226 				"available when monitoring via -p -t -a "
2227 				"options or only --per-thread.\n");
2228 			parse_options_usage(NULL, stat_options, "p", 1);
2229 			parse_options_usage(NULL, stat_options, "t", 1);
2230 			goto out;
2231 		}
2232 	}
2233 
2234 	/*
2235 	 * no_aggr, cgroup are for system-wide only
2236 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2237 	 */
2238 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2239 	      stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2240 	    !target__has_cpu(&target)) {
2241 		fprintf(stderr, "both cgroup and no-aggregation "
2242 			"modes only available in system-wide mode\n");
2243 
2244 		parse_options_usage(stat_usage, stat_options, "G", 1);
2245 		parse_options_usage(NULL, stat_options, "A", 1);
2246 		parse_options_usage(NULL, stat_options, "a", 1);
2247 		goto out;
2248 	}
2249 
2250 	if (add_default_attributes())
2251 		goto out;
2252 
2253 	target__validate(&target);
2254 
2255 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2256 		target.per_thread = true;
2257 
2258 	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2259 		if (target__has_task(&target)) {
2260 			pr_err("Problems finding threads of monitor\n");
2261 			parse_options_usage(stat_usage, stat_options, "p", 1);
2262 			parse_options_usage(NULL, stat_options, "t", 1);
2263 		} else if (target__has_cpu(&target)) {
2264 			perror("failed to parse CPUs map");
2265 			parse_options_usage(stat_usage, stat_options, "C", 1);
2266 			parse_options_usage(NULL, stat_options, "a", 1);
2267 		}
2268 		goto out;
2269 	}
2270 
2271 	evlist__check_cpu_maps(evsel_list);
2272 
2273 	/*
2274 	 * Initialize thread_map with comm names,
2275 	 * so we could print it out on output.
2276 	 */
2277 	if (stat_config.aggr_mode == AGGR_THREAD) {
2278 		thread_map__read_comms(evsel_list->core.threads);
2279 		if (target.system_wide) {
2280 			if (runtime_stat_new(&stat_config,
2281 				perf_thread_map__nr(evsel_list->core.threads))) {
2282 				goto out;
2283 			}
2284 		}
2285 	}
2286 
2287 	if (stat_config.aggr_mode == AGGR_NODE)
2288 		cpu__setup_cpunode_map();
2289 
2290 	if (stat_config.times && interval)
2291 		interval_count = true;
2292 	else if (stat_config.times && !interval) {
2293 		pr_err("interval-count option should be used together with "
2294 				"interval-print.\n");
2295 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2296 		parse_options_usage(stat_usage, stat_options, "I", 1);
2297 		goto out;
2298 	}
2299 
2300 	if (timeout && timeout < 100) {
2301 		if (timeout < 10) {
2302 			pr_err("timeout must be >= 10ms.\n");
2303 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2304 			goto out;
2305 		} else
2306 			pr_warning("timeout < 100ms. "
2307 				   "The overhead percentage could be high in some cases. "
2308 				   "Please proceed with caution.\n");
2309 	}
2310 	if (timeout && interval) {
2311 		pr_err("timeout option is not supported with interval-print.\n");
2312 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2313 		parse_options_usage(stat_usage, stat_options, "I", 1);
2314 		goto out;
2315 	}
2316 
2317 	if (perf_evlist__alloc_stats(evsel_list, interval))
2318 		goto out;
2319 
2320 	if (perf_stat_init_aggr_mode())
2321 		goto out;
2322 
2323 	/*
2324 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2325 	 * while avoiding that older tools show confusing messages.
2326 	 *
2327 	 * However for pipe sessions we need to keep it zero,
2328 	 * because script's perf_evsel__check_attr is triggered
2329 	 * by attr->sample_type != 0, and we can't run it on
2330 	 * stat sessions.
2331 	 */
2332 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2333 
2334 	/*
2335 	 * We dont want to block the signals - that would cause
2336 	 * child tasks to inherit that and Ctrl-C would not work.
2337 	 * What we want is for Ctrl-C to work in the exec()-ed
2338 	 * task, but being ignored by perf stat itself:
2339 	 */
2340 	atexit(sig_atexit);
2341 	if (!forever)
2342 		signal(SIGINT,  skip_signal);
2343 	signal(SIGCHLD, skip_signal);
2344 	signal(SIGALRM, skip_signal);
2345 	signal(SIGABRT, skip_signal);
2346 
2347 	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2348 		goto out;
2349 
2350 	status = 0;
2351 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2352 		if (stat_config.run_count != 1 && verbose > 0)
2353 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2354 				run_idx + 1);
2355 
2356 		if (run_idx != 0)
2357 			perf_evlist__reset_prev_raw_counts(evsel_list);
2358 
2359 		status = run_perf_stat(argc, argv, run_idx);
2360 		if (forever && status != -1 && !interval) {
2361 			print_counters(NULL, argc, argv);
2362 			perf_stat__reset_stats();
2363 		}
2364 	}
2365 
2366 	if (!forever && status != -1 && (!interval || stat_config.summary))
2367 		print_counters(NULL, argc, argv);
2368 
2369 	evlist__finalize_ctlfd(evsel_list);
2370 
2371 	if (STAT_RECORD) {
2372 		/*
2373 		 * We synthesize the kernel mmap record just so that older tools
2374 		 * don't emit warnings about not being able to resolve symbols
2375 		 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2376 		 * a saner message about no samples being in the perf.data file.
2377 		 *
2378 		 * This also serves to suppress a warning about f_header.data.size == 0
2379 		 * in header.c at the moment 'perf stat record' gets introduced, which
2380 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2381 		 * records, but the need to suppress the kptr_restrict messages in older
2382 		 * tools remain  -acme
2383 		 */
2384 		int fd = perf_data__fd(&perf_stat.data);
2385 		int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2386 							     process_synthesized_event,
2387 							     &perf_stat.session->machines.host);
2388 		if (err) {
2389 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2390 				   "older tools may produce warnings about this file\n.");
2391 		}
2392 
2393 		if (!interval) {
2394 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2395 				pr_err("failed to write stat round event\n");
2396 		}
2397 
2398 		if (!perf_stat.data.is_pipe) {
2399 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2400 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2401 		}
2402 
2403 		evlist__close(evsel_list);
2404 		perf_session__delete(perf_stat.session);
2405 	}
2406 
2407 	perf_stat__exit_aggr_mode();
2408 	perf_evlist__free_stats(evsel_list);
2409 out:
2410 	zfree(&stat_config.walltime_run);
2411 
2412 	if (smi_cost && smi_reset)
2413 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2414 
2415 	evlist__delete(evsel_list);
2416 
2417 	metricgroup__rblist_exit(&stat_config.metric_events);
2418 	runtime_stat_delete(&stat_config);
2419 
2420 	return status;
2421 }
2422