xref: /openbmc/linux/tools/perf/builtin-stat.c (revision c6fddb28)
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 "asm/bug.h"
70 
71 #include <linux/time64.h>
72 #include <linux/zalloc.h>
73 #include <api/fs/fs.h>
74 #include <errno.h>
75 #include <signal.h>
76 #include <stdlib.h>
77 #include <sys/prctl.h>
78 #include <inttypes.h>
79 #include <locale.h>
80 #include <math.h>
81 #include <sys/types.h>
82 #include <sys/stat.h>
83 #include <sys/wait.h>
84 #include <unistd.h>
85 #include <sys/time.h>
86 #include <sys/resource.h>
87 #include <linux/err.h>
88 
89 #include <linux/ctype.h>
90 #include <perf/evlist.h>
91 
92 #define DEFAULT_SEPARATOR	" "
93 #define FREEZE_ON_SMI_PATH	"devices/cpu/freeze_on_smi"
94 
95 static void print_counters(struct timespec *ts, int argc, const char **argv);
96 
97 /* Default events used for perf stat -T */
98 static const char *transaction_attrs = {
99 	"task-clock,"
100 	"{"
101 	"instructions,"
102 	"cycles,"
103 	"cpu/cycles-t/,"
104 	"cpu/tx-start/,"
105 	"cpu/el-start/,"
106 	"cpu/cycles-ct/"
107 	"}"
108 };
109 
110 /* More limited version when the CPU does not have all events. */
111 static const char * transaction_limited_attrs = {
112 	"task-clock,"
113 	"{"
114 	"instructions,"
115 	"cycles,"
116 	"cpu/cycles-t/,"
117 	"cpu/tx-start/"
118 	"}"
119 };
120 
121 static const char * topdown_attrs[] = {
122 	"topdown-total-slots",
123 	"topdown-slots-retired",
124 	"topdown-recovery-bubbles",
125 	"topdown-fetch-bubbles",
126 	"topdown-slots-issued",
127 	NULL,
128 };
129 
130 static const char *smi_cost_attrs = {
131 	"{"
132 	"msr/aperf/,"
133 	"msr/smi/,"
134 	"cycles"
135 	"}"
136 };
137 
138 static struct evlist	*evsel_list;
139 
140 static struct target target = {
141 	.uid	= UINT_MAX,
142 };
143 
144 #define METRIC_ONLY_LEN 20
145 
146 static volatile pid_t		child_pid			= -1;
147 static int			detailed_run			=  0;
148 static bool			transaction_run;
149 static bool			topdown_run			= false;
150 static bool			smi_cost			= false;
151 static bool			smi_reset			= false;
152 static int			big_num_opt			=  -1;
153 static bool			group				= false;
154 static const char		*pre_cmd			= NULL;
155 static const char		*post_cmd			= NULL;
156 static bool			sync_run			= false;
157 static bool			forever				= false;
158 static bool			force_metric_only		= false;
159 static struct timespec		ref_time;
160 static bool			append_file;
161 static bool			interval_count;
162 static const char		*output_name;
163 static int			output_fd;
164 
165 struct perf_stat {
166 	bool			 record;
167 	struct perf_data	 data;
168 	struct perf_session	*session;
169 	u64			 bytes_written;
170 	struct perf_tool	 tool;
171 	bool			 maps_allocated;
172 	struct perf_cpu_map	*cpus;
173 	struct perf_thread_map *threads;
174 	enum aggr_mode		 aggr_mode;
175 };
176 
177 static struct perf_stat		perf_stat;
178 #define STAT_RECORD		perf_stat.record
179 
180 static volatile int done = 0;
181 
182 static struct perf_stat_config stat_config = {
183 	.aggr_mode		= AGGR_GLOBAL,
184 	.scale			= true,
185 	.unit_width		= 4, /* strlen("unit") */
186 	.run_count		= 1,
187 	.metric_only_len	= METRIC_ONLY_LEN,
188 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
189 	.big_num		= true,
190 };
191 
192 static inline void diff_timespec(struct timespec *r, struct timespec *a,
193 				 struct timespec *b)
194 {
195 	r->tv_sec = a->tv_sec - b->tv_sec;
196 	if (a->tv_nsec < b->tv_nsec) {
197 		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
198 		r->tv_sec--;
199 	} else {
200 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
201 	}
202 }
203 
204 static void perf_stat__reset_stats(void)
205 {
206 	int i;
207 
208 	perf_evlist__reset_stats(evsel_list);
209 	perf_stat__reset_shadow_stats();
210 
211 	for (i = 0; i < stat_config.stats_num; i++)
212 		perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
213 }
214 
215 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
216 				     union perf_event *event,
217 				     struct perf_sample *sample __maybe_unused,
218 				     struct machine *machine __maybe_unused)
219 {
220 	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
221 		pr_err("failed to write perf data, error: %m\n");
222 		return -1;
223 	}
224 
225 	perf_stat.bytes_written += event->header.size;
226 	return 0;
227 }
228 
229 static int write_stat_round_event(u64 tm, u64 type)
230 {
231 	return perf_event__synthesize_stat_round(NULL, tm, type,
232 						 process_synthesized_event,
233 						 NULL);
234 }
235 
236 #define WRITE_STAT_ROUND_EVENT(time, interval) \
237 	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
238 
239 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
240 
241 static int
242 perf_evsel__write_stat_event(struct evsel *counter, u32 cpu, u32 thread,
243 			     struct perf_counts_values *count)
244 {
245 	struct perf_sample_id *sid = SID(counter, cpu, thread);
246 
247 	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
248 					   process_synthesized_event, NULL);
249 }
250 
251 static int read_single_counter(struct evsel *counter, int cpu,
252 			       int thread, struct timespec *rs)
253 {
254 	if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
255 		u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
256 		struct perf_counts_values *count =
257 			perf_counts(counter->counts, cpu, thread);
258 		count->ena = count->run = val;
259 		count->val = val;
260 		return 0;
261 	}
262 	return perf_evsel__read_counter(counter, cpu, thread);
263 }
264 
265 /*
266  * Read out the results of a single counter:
267  * do not aggregate counts across CPUs in system-wide mode
268  */
269 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu)
270 {
271 	int nthreads = perf_thread_map__nr(evsel_list->core.threads);
272 	int thread;
273 
274 	if (!counter->supported)
275 		return -ENOENT;
276 
277 	if (counter->core.system_wide)
278 		nthreads = 1;
279 
280 	for (thread = 0; thread < nthreads; thread++) {
281 		struct perf_counts_values *count;
282 
283 		count = perf_counts(counter->counts, cpu, thread);
284 
285 		/*
286 		 * The leader's group read loads data into its group members
287 		 * (via perf_evsel__read_counter()) and sets their count->loaded.
288 		 */
289 		if (!perf_counts__is_loaded(counter->counts, cpu, thread) &&
290 		    read_single_counter(counter, cpu, thread, rs)) {
291 			counter->counts->scaled = -1;
292 			perf_counts(counter->counts, cpu, thread)->ena = 0;
293 			perf_counts(counter->counts, cpu, thread)->run = 0;
294 			return -1;
295 		}
296 
297 		perf_counts__set_loaded(counter->counts, cpu, thread, false);
298 
299 		if (STAT_RECORD) {
300 			if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
301 				pr_err("failed to write stat event\n");
302 				return -1;
303 			}
304 		}
305 
306 		if (verbose > 1) {
307 			fprintf(stat_config.output,
308 				"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
309 					perf_evsel__name(counter),
310 					cpu,
311 					count->val, count->ena, count->run);
312 		}
313 	}
314 
315 	return 0;
316 }
317 
318 static void read_counters(struct timespec *rs)
319 {
320 	struct evsel *counter;
321 	struct affinity affinity;
322 	int i, ncpus, cpu;
323 
324 	if (affinity__setup(&affinity) < 0)
325 		return;
326 
327 	ncpus = perf_cpu_map__nr(evsel_list->core.all_cpus);
328 	if (!target__has_cpu(&target) || target__has_per_thread(&target))
329 		ncpus = 1;
330 	evlist__for_each_cpu(evsel_list, i, cpu) {
331 		if (i >= ncpus)
332 			break;
333 		affinity__set(&affinity, cpu);
334 
335 		evlist__for_each_entry(evsel_list, counter) {
336 			if (evsel__cpu_iter_skip(counter, cpu))
337 				continue;
338 			if (!counter->err) {
339 				counter->err = read_counter_cpu(counter, rs,
340 								counter->cpu_iter - 1);
341 			}
342 		}
343 	}
344 	affinity__cleanup(&affinity);
345 
346 	evlist__for_each_entry(evsel_list, counter) {
347 		if (counter->err)
348 			pr_debug("failed to read counter %s\n", counter->name);
349 		if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
350 			pr_warning("failed to process counter %s\n", counter->name);
351 		counter->err = 0;
352 	}
353 }
354 
355 static void process_interval(void)
356 {
357 	struct timespec ts, rs;
358 
359 	clock_gettime(CLOCK_MONOTONIC, &ts);
360 	diff_timespec(&rs, &ts, &ref_time);
361 
362 	read_counters(&rs);
363 
364 	if (STAT_RECORD) {
365 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
366 			pr_err("failed to write stat round event\n");
367 	}
368 
369 	init_stats(&walltime_nsecs_stats);
370 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
371 	print_counters(&rs, 0, NULL);
372 }
373 
374 static void enable_counters(void)
375 {
376 	if (stat_config.initial_delay)
377 		usleep(stat_config.initial_delay * USEC_PER_MSEC);
378 
379 	/*
380 	 * We need to enable counters only if:
381 	 * - we don't have tracee (attaching to task or cpu)
382 	 * - we have initial delay configured
383 	 */
384 	if (!target__none(&target) || stat_config.initial_delay)
385 		evlist__enable(evsel_list);
386 }
387 
388 static void disable_counters(void)
389 {
390 	/*
391 	 * If we don't have tracee (attaching to task or cpu), counters may
392 	 * still be running. To get accurate group ratios, we must stop groups
393 	 * from counting before reading their constituent counters.
394 	 */
395 	if (!target__none(&target))
396 		evlist__disable(evsel_list);
397 }
398 
399 static volatile int workload_exec_errno;
400 
401 /*
402  * perf_evlist__prepare_workload will send a SIGUSR1
403  * if the fork fails, since we asked by setting its
404  * want_signal to true.
405  */
406 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
407 					void *ucontext __maybe_unused)
408 {
409 	workload_exec_errno = info->si_value.sival_int;
410 }
411 
412 static bool perf_evsel__should_store_id(struct evsel *counter)
413 {
414 	return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
415 }
416 
417 static bool is_target_alive(struct target *_target,
418 			    struct perf_thread_map *threads)
419 {
420 	struct stat st;
421 	int i;
422 
423 	if (!target__has_task(_target))
424 		return true;
425 
426 	for (i = 0; i < threads->nr; i++) {
427 		char path[PATH_MAX];
428 
429 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
430 			  threads->map[i].pid);
431 
432 		if (!stat(path, &st))
433 			return true;
434 	}
435 
436 	return false;
437 }
438 
439 enum counter_recovery {
440 	COUNTER_SKIP,
441 	COUNTER_RETRY,
442 	COUNTER_FATAL,
443 };
444 
445 static enum counter_recovery stat_handle_error(struct evsel *counter)
446 {
447 	char msg[BUFSIZ];
448 	/*
449 	 * PPC returns ENXIO for HW counters until 2.6.37
450 	 * (behavior changed with commit b0a873e).
451 	 */
452 	if (errno == EINVAL || errno == ENOSYS ||
453 	    errno == ENOENT || errno == EOPNOTSUPP ||
454 	    errno == ENXIO) {
455 		if (verbose > 0)
456 			ui__warning("%s event is not supported by the kernel.\n",
457 				    perf_evsel__name(counter));
458 		counter->supported = false;
459 		/*
460 		 * errored is a sticky flag that means one of the counter's
461 		 * cpu event had a problem and needs to be reexamined.
462 		 */
463 		counter->errored = true;
464 
465 		if ((counter->leader != counter) ||
466 		    !(counter->leader->core.nr_members > 1))
467 			return COUNTER_SKIP;
468 	} else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
469 		if (verbose > 0)
470 			ui__warning("%s\n", msg);
471 		return COUNTER_RETRY;
472 	} else if (target__has_per_thread(&target) &&
473 		   evsel_list->core.threads &&
474 		   evsel_list->core.threads->err_thread != -1) {
475 		/*
476 		 * For global --per-thread case, skip current
477 		 * error thread.
478 		 */
479 		if (!thread_map__remove(evsel_list->core.threads,
480 					evsel_list->core.threads->err_thread)) {
481 			evsel_list->core.threads->err_thread = -1;
482 			return COUNTER_RETRY;
483 		}
484 	}
485 
486 	perf_evsel__open_strerror(counter, &target,
487 				  errno, msg, sizeof(msg));
488 	ui__error("%s\n", msg);
489 
490 	if (child_pid != -1)
491 		kill(child_pid, SIGTERM);
492 	return COUNTER_FATAL;
493 }
494 
495 static int __run_perf_stat(int argc, const char **argv, int run_idx)
496 {
497 	int interval = stat_config.interval;
498 	int times = stat_config.times;
499 	int timeout = stat_config.timeout;
500 	char msg[BUFSIZ];
501 	unsigned long long t0, t1;
502 	struct evsel *counter;
503 	struct timespec ts;
504 	size_t l;
505 	int status = 0;
506 	const bool forks = (argc > 0);
507 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
508 	struct affinity affinity;
509 	int i, cpu;
510 	bool second_pass = false;
511 
512 	if (interval) {
513 		ts.tv_sec  = interval / USEC_PER_MSEC;
514 		ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
515 	} else if (timeout) {
516 		ts.tv_sec  = timeout / USEC_PER_MSEC;
517 		ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
518 	} else {
519 		ts.tv_sec  = 1;
520 		ts.tv_nsec = 0;
521 	}
522 
523 	if (forks) {
524 		if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
525 						  workload_exec_failed_signal) < 0) {
526 			perror("failed to prepare workload");
527 			return -1;
528 		}
529 		child_pid = evsel_list->workload.pid;
530 	}
531 
532 	if (group)
533 		perf_evlist__set_leader(evsel_list);
534 
535 	if (affinity__setup(&affinity) < 0)
536 		return -1;
537 
538 	evlist__for_each_cpu (evsel_list, i, cpu) {
539 		affinity__set(&affinity, cpu);
540 
541 		evlist__for_each_entry(evsel_list, counter) {
542 			if (evsel__cpu_iter_skip(counter, cpu))
543 				continue;
544 			if (counter->reset_group || counter->errored)
545 				continue;
546 try_again:
547 			if (create_perf_stat_counter(counter, &stat_config, &target,
548 						     counter->cpu_iter - 1) < 0) {
549 
550 				/*
551 				 * Weak group failed. We cannot just undo this here
552 				 * because earlier CPUs might be in group mode, and the kernel
553 				 * doesn't support mixing group and non group reads. Defer
554 				 * it to later.
555 				 * Don't close here because we're in the wrong affinity.
556 				 */
557 				if ((errno == EINVAL || errno == EBADF) &&
558 				    counter->leader != counter &&
559 				    counter->weak_group) {
560 					perf_evlist__reset_weak_group(evsel_list, counter, false);
561 					assert(counter->reset_group);
562 					second_pass = true;
563 					continue;
564 				}
565 
566 				switch (stat_handle_error(counter)) {
567 				case COUNTER_FATAL:
568 					return -1;
569 				case COUNTER_RETRY:
570 					goto try_again;
571 				case COUNTER_SKIP:
572 					continue;
573 				default:
574 					break;
575 				}
576 
577 			}
578 			counter->supported = true;
579 		}
580 	}
581 
582 	if (second_pass) {
583 		/*
584 		 * Now redo all the weak group after closing them,
585 		 * and also close errored counters.
586 		 */
587 
588 		evlist__for_each_cpu(evsel_list, i, cpu) {
589 			affinity__set(&affinity, cpu);
590 			/* First close errored or weak retry */
591 			evlist__for_each_entry(evsel_list, counter) {
592 				if (!counter->reset_group && !counter->errored)
593 					continue;
594 				if (evsel__cpu_iter_skip_no_inc(counter, cpu))
595 					continue;
596 				perf_evsel__close_cpu(&counter->core, counter->cpu_iter);
597 			}
598 			/* Now reopen weak */
599 			evlist__for_each_entry(evsel_list, counter) {
600 				if (!counter->reset_group && !counter->errored)
601 					continue;
602 				if (evsel__cpu_iter_skip(counter, cpu))
603 					continue;
604 				if (!counter->reset_group)
605 					continue;
606 try_again_reset:
607 				pr_debug2("reopening weak %s\n", perf_evsel__name(counter));
608 				if (create_perf_stat_counter(counter, &stat_config, &target,
609 							     counter->cpu_iter - 1) < 0) {
610 
611 					switch (stat_handle_error(counter)) {
612 					case COUNTER_FATAL:
613 						return -1;
614 					case COUNTER_RETRY:
615 						goto try_again_reset;
616 					case COUNTER_SKIP:
617 						continue;
618 					default:
619 						break;
620 					}
621 				}
622 				counter->supported = true;
623 			}
624 		}
625 	}
626 	affinity__cleanup(&affinity);
627 
628 	evlist__for_each_entry(evsel_list, counter) {
629 		if (!counter->supported) {
630 			perf_evsel__free_fd(&counter->core);
631 			continue;
632 		}
633 
634 		l = strlen(counter->unit);
635 		if (l > stat_config.unit_width)
636 			stat_config.unit_width = l;
637 
638 		if (perf_evsel__should_store_id(counter) &&
639 		    perf_evsel__store_ids(counter, evsel_list))
640 			return -1;
641 	}
642 
643 	if (perf_evlist__apply_filters(evsel_list, &counter)) {
644 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
645 			counter->filter, perf_evsel__name(counter), errno,
646 			str_error_r(errno, msg, sizeof(msg)));
647 		return -1;
648 	}
649 
650 	if (STAT_RECORD) {
651 		int err, fd = perf_data__fd(&perf_stat.data);
652 
653 		if (is_pipe) {
654 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
655 		} else {
656 			err = perf_session__write_header(perf_stat.session, evsel_list,
657 							 fd, false);
658 		}
659 
660 		if (err < 0)
661 			return err;
662 
663 		err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
664 							 process_synthesized_event, is_pipe);
665 		if (err < 0)
666 			return err;
667 	}
668 
669 	/*
670 	 * Enable counters and exec the command:
671 	 */
672 	t0 = rdclock();
673 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
674 
675 	if (forks) {
676 		perf_evlist__start_workload(evsel_list);
677 		enable_counters();
678 
679 		if (interval || timeout) {
680 			while (!waitpid(child_pid, &status, WNOHANG)) {
681 				nanosleep(&ts, NULL);
682 				if (timeout)
683 					break;
684 				process_interval();
685 				if (interval_count && !(--times))
686 					break;
687 			}
688 		}
689 		if (child_pid != -1) {
690 			if (timeout)
691 				kill(child_pid, SIGTERM);
692 			wait4(child_pid, &status, 0, &stat_config.ru_data);
693 		}
694 
695 		if (workload_exec_errno) {
696 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
697 			pr_err("Workload failed: %s\n", emsg);
698 			return -1;
699 		}
700 
701 		if (WIFSIGNALED(status))
702 			psignal(WTERMSIG(status), argv[0]);
703 	} else {
704 		enable_counters();
705 		while (!done) {
706 			nanosleep(&ts, NULL);
707 			if (!is_target_alive(&target, evsel_list->core.threads))
708 				break;
709 			if (timeout)
710 				break;
711 			if (interval) {
712 				process_interval();
713 				if (interval_count && !(--times))
714 					break;
715 			}
716 		}
717 	}
718 
719 	disable_counters();
720 
721 	t1 = rdclock();
722 
723 	if (stat_config.walltime_run_table)
724 		stat_config.walltime_run[run_idx] = t1 - t0;
725 
726 	update_stats(&walltime_nsecs_stats, t1 - t0);
727 
728 	/*
729 	 * Closing a group leader splits the group, and as we only disable
730 	 * group leaders, results in remaining events becoming enabled. To
731 	 * avoid arbitrary skew, we must read all counters before closing any
732 	 * group leaders.
733 	 */
734 	read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
735 
736 	/*
737 	 * We need to keep evsel_list alive, because it's processed
738 	 * later the evsel_list will be closed after.
739 	 */
740 	if (!STAT_RECORD)
741 		evlist__close(evsel_list);
742 
743 	return WEXITSTATUS(status);
744 }
745 
746 static int run_perf_stat(int argc, const char **argv, int run_idx)
747 {
748 	int ret;
749 
750 	if (pre_cmd) {
751 		ret = system(pre_cmd);
752 		if (ret)
753 			return ret;
754 	}
755 
756 	if (sync_run)
757 		sync();
758 
759 	ret = __run_perf_stat(argc, argv, run_idx);
760 	if (ret)
761 		return ret;
762 
763 	if (post_cmd) {
764 		ret = system(post_cmd);
765 		if (ret)
766 			return ret;
767 	}
768 
769 	return ret;
770 }
771 
772 static void print_counters(struct timespec *ts, int argc, const char **argv)
773 {
774 	/* Do not print anything if we record to the pipe. */
775 	if (STAT_RECORD && perf_stat.data.is_pipe)
776 		return;
777 
778 	perf_evlist__print_counters(evsel_list, &stat_config, &target,
779 				    ts, argc, argv);
780 }
781 
782 static volatile int signr = -1;
783 
784 static void skip_signal(int signo)
785 {
786 	if ((child_pid == -1) || stat_config.interval)
787 		done = 1;
788 
789 	signr = signo;
790 	/*
791 	 * render child_pid harmless
792 	 * won't send SIGTERM to a random
793 	 * process in case of race condition
794 	 * and fast PID recycling
795 	 */
796 	child_pid = -1;
797 }
798 
799 static void sig_atexit(void)
800 {
801 	sigset_t set, oset;
802 
803 	/*
804 	 * avoid race condition with SIGCHLD handler
805 	 * in skip_signal() which is modifying child_pid
806 	 * goal is to avoid send SIGTERM to a random
807 	 * process
808 	 */
809 	sigemptyset(&set);
810 	sigaddset(&set, SIGCHLD);
811 	sigprocmask(SIG_BLOCK, &set, &oset);
812 
813 	if (child_pid != -1)
814 		kill(child_pid, SIGTERM);
815 
816 	sigprocmask(SIG_SETMASK, &oset, NULL);
817 
818 	if (signr == -1)
819 		return;
820 
821 	signal(signr, SIG_DFL);
822 	kill(getpid(), signr);
823 }
824 
825 static int stat__set_big_num(const struct option *opt __maybe_unused,
826 			     const char *s __maybe_unused, int unset)
827 {
828 	big_num_opt = unset ? 0 : 1;
829 	return 0;
830 }
831 
832 static int enable_metric_only(const struct option *opt __maybe_unused,
833 			      const char *s __maybe_unused, int unset)
834 {
835 	force_metric_only = true;
836 	stat_config.metric_only = !unset;
837 	return 0;
838 }
839 
840 static int parse_metric_groups(const struct option *opt,
841 			       const char *str,
842 			       int unset __maybe_unused)
843 {
844 	return metricgroup__parse_groups(opt, str, &stat_config.metric_events);
845 }
846 
847 static struct option stat_options[] = {
848 	OPT_BOOLEAN('T', "transaction", &transaction_run,
849 		    "hardware transaction statistics"),
850 	OPT_CALLBACK('e', "event", &evsel_list, "event",
851 		     "event selector. use 'perf list' to list available events",
852 		     parse_events_option),
853 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
854 		     "event filter", parse_filter),
855 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
856 		    "child tasks do not inherit counters"),
857 	OPT_STRING('p', "pid", &target.pid, "pid",
858 		   "stat events on existing process id"),
859 	OPT_STRING('t', "tid", &target.tid, "tid",
860 		   "stat events on existing thread id"),
861 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
862 		    "system-wide collection from all CPUs"),
863 	OPT_BOOLEAN('g', "group", &group,
864 		    "put the counters into a counter group"),
865 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
866 		    "Use --no-scale to disable counter scaling for multiplexing"),
867 	OPT_INCR('v', "verbose", &verbose,
868 		    "be more verbose (show counter open errors, etc)"),
869 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
870 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
871 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
872 		    "display details about each run (only with -r option)"),
873 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
874 		    "null run - dont start any counters"),
875 	OPT_INCR('d', "detailed", &detailed_run,
876 		    "detailed run - start a lot of events"),
877 	OPT_BOOLEAN('S', "sync", &sync_run,
878 		    "call sync() before starting a run"),
879 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
880 			   "print large numbers with thousands\' separators",
881 			   stat__set_big_num),
882 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
883 		    "list of cpus to monitor in system-wide"),
884 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
885 		    "disable CPU count aggregation", AGGR_NONE),
886 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
887 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
888 		   "print counts with custom separator"),
889 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
890 		     "monitor event in cgroup name only", parse_cgroups),
891 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
892 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
893 	OPT_INTEGER(0, "log-fd", &output_fd,
894 		    "log output to fd, instead of stderr"),
895 	OPT_STRING(0, "pre", &pre_cmd, "command",
896 			"command to run prior to the measured command"),
897 	OPT_STRING(0, "post", &post_cmd, "command",
898 			"command to run after to the measured command"),
899 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
900 		    "print counts at regular interval in ms "
901 		    "(overhead is possible for values <= 100ms)"),
902 	OPT_INTEGER(0, "interval-count", &stat_config.times,
903 		    "print counts for fixed number of times"),
904 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
905 		    "clear screen in between new interval"),
906 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
907 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
908 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
909 		     "aggregate counts per processor socket", AGGR_SOCKET),
910 	OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
911 		     "aggregate counts per processor die", AGGR_DIE),
912 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
913 		     "aggregate counts per physical processor core", AGGR_CORE),
914 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
915 		     "aggregate counts per thread", AGGR_THREAD),
916 	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
917 		     "aggregate counts per numa node", AGGR_NODE),
918 	OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
919 		     "ms to wait before starting measurement after program start"),
920 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
921 			"Only print computed metrics. No raw values", enable_metric_only),
922 	OPT_BOOLEAN(0, "topdown", &topdown_run,
923 			"measure topdown level 1 statistics"),
924 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
925 			"measure SMI cost"),
926 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
927 		     "monitor specified metrics or metric groups (separated by ,)",
928 		     parse_metric_groups),
929 	OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
930 			 "Configure all used events to run in kernel space.",
931 			 PARSE_OPT_EXCLUSIVE),
932 	OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
933 			 "Configure all used events to run in user space.",
934 			 PARSE_OPT_EXCLUSIVE),
935 	OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
936 		    "Use with 'percore' event qualifier to show the event "
937 		    "counts of one hardware thread by sum up total hardware "
938 		    "threads of same physical core"),
939 	OPT_END()
940 };
941 
942 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
943 				 struct perf_cpu_map *map, int cpu)
944 {
945 	return cpu_map__get_socket(map, cpu, NULL);
946 }
947 
948 static int perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
949 			      struct perf_cpu_map *map, int cpu)
950 {
951 	return cpu_map__get_die(map, cpu, NULL);
952 }
953 
954 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
955 			       struct perf_cpu_map *map, int cpu)
956 {
957 	return cpu_map__get_core(map, cpu, NULL);
958 }
959 
960 static int perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
961 			       struct perf_cpu_map *map, int cpu)
962 {
963 	return cpu_map__get_node(map, cpu, NULL);
964 }
965 
966 static int perf_stat__get_aggr(struct perf_stat_config *config,
967 			       aggr_get_id_t get_id, struct perf_cpu_map *map, int idx)
968 {
969 	int cpu;
970 
971 	if (idx >= map->nr)
972 		return -1;
973 
974 	cpu = map->map[idx];
975 
976 	if (config->cpus_aggr_map->map[cpu] == -1)
977 		config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
978 
979 	return config->cpus_aggr_map->map[cpu];
980 }
981 
982 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
983 					struct perf_cpu_map *map, int idx)
984 {
985 	return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
986 }
987 
988 static int perf_stat__get_die_cached(struct perf_stat_config *config,
989 					struct perf_cpu_map *map, int idx)
990 {
991 	return perf_stat__get_aggr(config, perf_stat__get_die, map, idx);
992 }
993 
994 static int perf_stat__get_core_cached(struct perf_stat_config *config,
995 				      struct perf_cpu_map *map, int idx)
996 {
997 	return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
998 }
999 
1000 static int perf_stat__get_node_cached(struct perf_stat_config *config,
1001 				      struct perf_cpu_map *map, int idx)
1002 {
1003 	return perf_stat__get_aggr(config, perf_stat__get_node, map, idx);
1004 }
1005 
1006 static bool term_percore_set(void)
1007 {
1008 	struct evsel *counter;
1009 
1010 	evlist__for_each_entry(evsel_list, counter) {
1011 		if (counter->percore)
1012 			return true;
1013 	}
1014 
1015 	return false;
1016 }
1017 
1018 static int perf_stat_init_aggr_mode(void)
1019 {
1020 	int nr;
1021 
1022 	switch (stat_config.aggr_mode) {
1023 	case AGGR_SOCKET:
1024 		if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1025 			perror("cannot build socket map");
1026 			return -1;
1027 		}
1028 		stat_config.aggr_get_id = perf_stat__get_socket_cached;
1029 		break;
1030 	case AGGR_DIE:
1031 		if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1032 			perror("cannot build die map");
1033 			return -1;
1034 		}
1035 		stat_config.aggr_get_id = perf_stat__get_die_cached;
1036 		break;
1037 	case AGGR_CORE:
1038 		if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1039 			perror("cannot build core map");
1040 			return -1;
1041 		}
1042 		stat_config.aggr_get_id = perf_stat__get_core_cached;
1043 		break;
1044 	case AGGR_NODE:
1045 		if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1046 			perror("cannot build core map");
1047 			return -1;
1048 		}
1049 		stat_config.aggr_get_id = perf_stat__get_node_cached;
1050 		break;
1051 	case AGGR_NONE:
1052 		if (term_percore_set()) {
1053 			if (cpu_map__build_core_map(evsel_list->core.cpus,
1054 						    &stat_config.aggr_map)) {
1055 				perror("cannot build core map");
1056 				return -1;
1057 			}
1058 			stat_config.aggr_get_id = perf_stat__get_core_cached;
1059 		}
1060 		break;
1061 	case AGGR_GLOBAL:
1062 	case AGGR_THREAD:
1063 	case AGGR_UNSET:
1064 	default:
1065 		break;
1066 	}
1067 
1068 	/*
1069 	 * The evsel_list->cpus is the base we operate on,
1070 	 * taking the highest cpu number to be the size of
1071 	 * the aggregation translate cpumap.
1072 	 */
1073 	nr = perf_cpu_map__max(evsel_list->core.cpus);
1074 	stat_config.cpus_aggr_map = perf_cpu_map__empty_new(nr + 1);
1075 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1076 }
1077 
1078 static void perf_stat__exit_aggr_mode(void)
1079 {
1080 	perf_cpu_map__put(stat_config.aggr_map);
1081 	perf_cpu_map__put(stat_config.cpus_aggr_map);
1082 	stat_config.aggr_map = NULL;
1083 	stat_config.cpus_aggr_map = NULL;
1084 }
1085 
1086 static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx)
1087 {
1088 	int cpu;
1089 
1090 	if (idx > map->nr)
1091 		return -1;
1092 
1093 	cpu = map->map[idx];
1094 
1095 	if (cpu >= env->nr_cpus_avail)
1096 		return -1;
1097 
1098 	return cpu;
1099 }
1100 
1101 static int perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data)
1102 {
1103 	struct perf_env *env = data;
1104 	int cpu = perf_env__get_cpu(env, map, idx);
1105 
1106 	return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1107 }
1108 
1109 static int perf_env__get_die(struct perf_cpu_map *map, int idx, void *data)
1110 {
1111 	struct perf_env *env = data;
1112 	int die_id = -1, cpu = perf_env__get_cpu(env, map, idx);
1113 
1114 	if (cpu != -1) {
1115 		/*
1116 		 * Encode socket in bit range 15:8
1117 		 * die_id is relative to socket,
1118 		 * we need a global id. So we combine
1119 		 * socket + die id
1120 		 */
1121 		if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
1122 			return -1;
1123 
1124 		if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
1125 			return -1;
1126 
1127 		die_id = (env->cpu[cpu].socket_id << 8) | (env->cpu[cpu].die_id & 0xff);
1128 	}
1129 
1130 	return die_id;
1131 }
1132 
1133 static int perf_env__get_core(struct perf_cpu_map *map, int idx, void *data)
1134 {
1135 	struct perf_env *env = data;
1136 	int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1137 
1138 	if (cpu != -1) {
1139 		/*
1140 		 * Encode socket in bit range 31:24
1141 		 * encode die id in bit range 23:16
1142 		 * core_id is relative to socket and die,
1143 		 * we need a global id. So we combine
1144 		 * socket + die id + core id
1145 		 */
1146 		if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
1147 			return -1;
1148 
1149 		if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
1150 			return -1;
1151 
1152 		if (WARN_ONCE(env->cpu[cpu].core_id >> 16, "The core id number is too big.\n"))
1153 			return -1;
1154 
1155 		core = (env->cpu[cpu].socket_id << 24) |
1156 		       (env->cpu[cpu].die_id << 16) |
1157 		       (env->cpu[cpu].core_id & 0xffff);
1158 	}
1159 
1160 	return core;
1161 }
1162 
1163 static int perf_env__get_node(struct perf_cpu_map *map, int idx, void *data)
1164 {
1165 	int cpu = perf_env__get_cpu(data, map, idx);
1166 
1167 	return perf_env__numa_node(data, cpu);
1168 }
1169 
1170 static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus,
1171 				      struct perf_cpu_map **sockp)
1172 {
1173 	return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1174 }
1175 
1176 static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus,
1177 				   struct perf_cpu_map **diep)
1178 {
1179 	return cpu_map__build_map(cpus, diep, perf_env__get_die, env);
1180 }
1181 
1182 static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus,
1183 				    struct perf_cpu_map **corep)
1184 {
1185 	return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1186 }
1187 
1188 static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus,
1189 				    struct perf_cpu_map **nodep)
1190 {
1191 	return cpu_map__build_map(cpus, nodep, perf_env__get_node, env);
1192 }
1193 
1194 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1195 				      struct perf_cpu_map *map, int idx)
1196 {
1197 	return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1198 }
1199 static int perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1200 				   struct perf_cpu_map *map, int idx)
1201 {
1202 	return perf_env__get_die(map, idx, &perf_stat.session->header.env);
1203 }
1204 
1205 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1206 				    struct perf_cpu_map *map, int idx)
1207 {
1208 	return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1209 }
1210 
1211 static int perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1212 				    struct perf_cpu_map *map, int idx)
1213 {
1214 	return perf_env__get_node(map, idx, &perf_stat.session->header.env);
1215 }
1216 
1217 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1218 {
1219 	struct perf_env *env = &st->session->header.env;
1220 
1221 	switch (stat_config.aggr_mode) {
1222 	case AGGR_SOCKET:
1223 		if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1224 			perror("cannot build socket map");
1225 			return -1;
1226 		}
1227 		stat_config.aggr_get_id = perf_stat__get_socket_file;
1228 		break;
1229 	case AGGR_DIE:
1230 		if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1231 			perror("cannot build die map");
1232 			return -1;
1233 		}
1234 		stat_config.aggr_get_id = perf_stat__get_die_file;
1235 		break;
1236 	case AGGR_CORE:
1237 		if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1238 			perror("cannot build core map");
1239 			return -1;
1240 		}
1241 		stat_config.aggr_get_id = perf_stat__get_core_file;
1242 		break;
1243 	case AGGR_NODE:
1244 		if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1245 			perror("cannot build core map");
1246 			return -1;
1247 		}
1248 		stat_config.aggr_get_id = perf_stat__get_node_file;
1249 		break;
1250 	case AGGR_NONE:
1251 	case AGGR_GLOBAL:
1252 	case AGGR_THREAD:
1253 	case AGGR_UNSET:
1254 	default:
1255 		break;
1256 	}
1257 
1258 	return 0;
1259 }
1260 
1261 static int topdown_filter_events(const char **attr, char **str, bool use_group)
1262 {
1263 	int off = 0;
1264 	int i;
1265 	int len = 0;
1266 	char *s;
1267 
1268 	for (i = 0; attr[i]; i++) {
1269 		if (pmu_have_event("cpu", attr[i])) {
1270 			len += strlen(attr[i]) + 1;
1271 			attr[i - off] = attr[i];
1272 		} else
1273 			off++;
1274 	}
1275 	attr[i - off] = NULL;
1276 
1277 	*str = malloc(len + 1 + 2);
1278 	if (!*str)
1279 		return -1;
1280 	s = *str;
1281 	if (i - off == 0) {
1282 		*s = 0;
1283 		return 0;
1284 	}
1285 	if (use_group)
1286 		*s++ = '{';
1287 	for (i = 0; attr[i]; i++) {
1288 		strcpy(s, attr[i]);
1289 		s += strlen(s);
1290 		*s++ = ',';
1291 	}
1292 	if (use_group) {
1293 		s[-1] = '}';
1294 		*s = 0;
1295 	} else
1296 		s[-1] = 0;
1297 	return 0;
1298 }
1299 
1300 __weak bool arch_topdown_check_group(bool *warn)
1301 {
1302 	*warn = false;
1303 	return false;
1304 }
1305 
1306 __weak void arch_topdown_group_warn(void)
1307 {
1308 }
1309 
1310 /*
1311  * Add default attributes, if there were no attributes specified or
1312  * if -d/--detailed, -d -d or -d -d -d is used:
1313  */
1314 static int add_default_attributes(void)
1315 {
1316 	int err;
1317 	struct perf_event_attr default_attrs0[] = {
1318 
1319   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1320   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1321   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1322   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1323 
1324   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1325 };
1326 	struct perf_event_attr frontend_attrs[] = {
1327   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1328 };
1329 	struct perf_event_attr backend_attrs[] = {
1330   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1331 };
1332 	struct perf_event_attr default_attrs1[] = {
1333   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1334   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1335   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1336 
1337 };
1338 
1339 /*
1340  * Detailed stats (-d), covering the L1 and last level data caches:
1341  */
1342 	struct perf_event_attr detailed_attrs[] = {
1343 
1344   { .type = PERF_TYPE_HW_CACHE,
1345     .config =
1346 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1347 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1348 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1349 
1350   { .type = PERF_TYPE_HW_CACHE,
1351     .config =
1352 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1353 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1354 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1355 
1356   { .type = PERF_TYPE_HW_CACHE,
1357     .config =
1358 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1359 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1360 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1361 
1362   { .type = PERF_TYPE_HW_CACHE,
1363     .config =
1364 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1365 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1366 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1367 };
1368 
1369 /*
1370  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1371  */
1372 	struct perf_event_attr very_detailed_attrs[] = {
1373 
1374   { .type = PERF_TYPE_HW_CACHE,
1375     .config =
1376 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1377 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1378 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1379 
1380   { .type = PERF_TYPE_HW_CACHE,
1381     .config =
1382 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1383 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1384 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1385 
1386   { .type = PERF_TYPE_HW_CACHE,
1387     .config =
1388 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1389 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1390 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1391 
1392   { .type = PERF_TYPE_HW_CACHE,
1393     .config =
1394 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1395 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1396 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1397 
1398   { .type = PERF_TYPE_HW_CACHE,
1399     .config =
1400 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1401 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1402 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1403 
1404   { .type = PERF_TYPE_HW_CACHE,
1405     .config =
1406 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1407 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1408 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1409 
1410 };
1411 
1412 /*
1413  * Very, very detailed stats (-d -d -d), adding prefetch events:
1414  */
1415 	struct perf_event_attr very_very_detailed_attrs[] = {
1416 
1417   { .type = PERF_TYPE_HW_CACHE,
1418     .config =
1419 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1420 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1421 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1422 
1423   { .type = PERF_TYPE_HW_CACHE,
1424     .config =
1425 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1426 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1427 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1428 };
1429 	struct parse_events_error errinfo;
1430 
1431 	/* Set attrs if no event is selected and !null_run: */
1432 	if (stat_config.null_run)
1433 		return 0;
1434 
1435 	bzero(&errinfo, sizeof(errinfo));
1436 	if (transaction_run) {
1437 		/* Handle -T as -M transaction. Once platform specific metrics
1438 		 * support has been added to the json files, all archictures
1439 		 * will use this approach. To determine transaction support
1440 		 * on an architecture test for such a metric name.
1441 		 */
1442 		if (metricgroup__has_metric("transaction")) {
1443 			struct option opt = { .value = &evsel_list };
1444 
1445 			return metricgroup__parse_groups(&opt, "transaction",
1446 							 &stat_config.metric_events);
1447 		}
1448 
1449 		if (pmu_have_event("cpu", "cycles-ct") &&
1450 		    pmu_have_event("cpu", "el-start"))
1451 			err = parse_events(evsel_list, transaction_attrs,
1452 					   &errinfo);
1453 		else
1454 			err = parse_events(evsel_list,
1455 					   transaction_limited_attrs,
1456 					   &errinfo);
1457 		if (err) {
1458 			fprintf(stderr, "Cannot set up transaction events\n");
1459 			parse_events_print_error(&errinfo, transaction_attrs);
1460 			return -1;
1461 		}
1462 		return 0;
1463 	}
1464 
1465 	if (smi_cost) {
1466 		int smi;
1467 
1468 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1469 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1470 			return -1;
1471 		}
1472 
1473 		if (!smi) {
1474 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1475 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1476 				return -1;
1477 			}
1478 			smi_reset = true;
1479 		}
1480 
1481 		if (pmu_have_event("msr", "aperf") &&
1482 		    pmu_have_event("msr", "smi")) {
1483 			if (!force_metric_only)
1484 				stat_config.metric_only = true;
1485 			err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1486 		} else {
1487 			fprintf(stderr, "To measure SMI cost, it needs "
1488 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1489 			parse_events_print_error(&errinfo, smi_cost_attrs);
1490 			return -1;
1491 		}
1492 		if (err) {
1493 			parse_events_print_error(&errinfo, smi_cost_attrs);
1494 			fprintf(stderr, "Cannot set up SMI cost events\n");
1495 			return -1;
1496 		}
1497 		return 0;
1498 	}
1499 
1500 	if (topdown_run) {
1501 		char *str = NULL;
1502 		bool warn = false;
1503 
1504 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1505 		    stat_config.aggr_mode != AGGR_CORE) {
1506 			pr_err("top down event configuration requires --per-core mode\n");
1507 			return -1;
1508 		}
1509 		stat_config.aggr_mode = AGGR_CORE;
1510 		if (nr_cgroups || !target__has_cpu(&target)) {
1511 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1512 			return -1;
1513 		}
1514 
1515 		if (!force_metric_only)
1516 			stat_config.metric_only = true;
1517 		if (topdown_filter_events(topdown_attrs, &str,
1518 				arch_topdown_check_group(&warn)) < 0) {
1519 			pr_err("Out of memory\n");
1520 			return -1;
1521 		}
1522 		if (topdown_attrs[0] && str) {
1523 			if (warn)
1524 				arch_topdown_group_warn();
1525 			err = parse_events(evsel_list, str, &errinfo);
1526 			if (err) {
1527 				fprintf(stderr,
1528 					"Cannot set up top down events %s: %d\n",
1529 					str, err);
1530 				parse_events_print_error(&errinfo, str);
1531 				free(str);
1532 				return -1;
1533 			}
1534 		} else {
1535 			fprintf(stderr, "System does not support topdown\n");
1536 			return -1;
1537 		}
1538 		free(str);
1539 	}
1540 
1541 	if (!evsel_list->core.nr_entries) {
1542 		if (target__has_cpu(&target))
1543 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1544 
1545 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1546 			return -1;
1547 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1548 			if (perf_evlist__add_default_attrs(evsel_list,
1549 						frontend_attrs) < 0)
1550 				return -1;
1551 		}
1552 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1553 			if (perf_evlist__add_default_attrs(evsel_list,
1554 						backend_attrs) < 0)
1555 				return -1;
1556 		}
1557 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1558 			return -1;
1559 	}
1560 
1561 	/* Detailed events get appended to the event list: */
1562 
1563 	if (detailed_run <  1)
1564 		return 0;
1565 
1566 	/* Append detailed run extra attributes: */
1567 	if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1568 		return -1;
1569 
1570 	if (detailed_run < 2)
1571 		return 0;
1572 
1573 	/* Append very detailed run extra attributes: */
1574 	if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1575 		return -1;
1576 
1577 	if (detailed_run < 3)
1578 		return 0;
1579 
1580 	/* Append very, very detailed run extra attributes: */
1581 	return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1582 }
1583 
1584 static const char * const stat_record_usage[] = {
1585 	"perf stat record [<options>]",
1586 	NULL,
1587 };
1588 
1589 static void init_features(struct perf_session *session)
1590 {
1591 	int feat;
1592 
1593 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1594 		perf_header__set_feat(&session->header, feat);
1595 
1596 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1597 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1598 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1599 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1600 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1601 }
1602 
1603 static int __cmd_record(int argc, const char **argv)
1604 {
1605 	struct perf_session *session;
1606 	struct perf_data *data = &perf_stat.data;
1607 
1608 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
1609 			     PARSE_OPT_STOP_AT_NON_OPTION);
1610 
1611 	if (output_name)
1612 		data->path = output_name;
1613 
1614 	if (stat_config.run_count != 1 || forever) {
1615 		pr_err("Cannot use -r option with perf stat record.\n");
1616 		return -1;
1617 	}
1618 
1619 	session = perf_session__new(data, false, NULL);
1620 	if (IS_ERR(session)) {
1621 		pr_err("Perf session creation failed\n");
1622 		return PTR_ERR(session);
1623 	}
1624 
1625 	init_features(session);
1626 
1627 	session->evlist   = evsel_list;
1628 	perf_stat.session = session;
1629 	perf_stat.record  = true;
1630 	return argc;
1631 }
1632 
1633 static int process_stat_round_event(struct perf_session *session,
1634 				    union perf_event *event)
1635 {
1636 	struct perf_record_stat_round *stat_round = &event->stat_round;
1637 	struct evsel *counter;
1638 	struct timespec tsh, *ts = NULL;
1639 	const char **argv = session->header.env.cmdline_argv;
1640 	int argc = session->header.env.nr_cmdline;
1641 
1642 	evlist__for_each_entry(evsel_list, counter)
1643 		perf_stat_process_counter(&stat_config, counter);
1644 
1645 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1646 		update_stats(&walltime_nsecs_stats, stat_round->time);
1647 
1648 	if (stat_config.interval && stat_round->time) {
1649 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
1650 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1651 		ts = &tsh;
1652 	}
1653 
1654 	print_counters(ts, argc, argv);
1655 	return 0;
1656 }
1657 
1658 static
1659 int process_stat_config_event(struct perf_session *session,
1660 			      union perf_event *event)
1661 {
1662 	struct perf_tool *tool = session->tool;
1663 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1664 
1665 	perf_event__read_stat_config(&stat_config, &event->stat_config);
1666 
1667 	if (perf_cpu_map__empty(st->cpus)) {
1668 		if (st->aggr_mode != AGGR_UNSET)
1669 			pr_warning("warning: processing task data, aggregation mode not set\n");
1670 		return 0;
1671 	}
1672 
1673 	if (st->aggr_mode != AGGR_UNSET)
1674 		stat_config.aggr_mode = st->aggr_mode;
1675 
1676 	if (perf_stat.data.is_pipe)
1677 		perf_stat_init_aggr_mode();
1678 	else
1679 		perf_stat_init_aggr_mode_file(st);
1680 
1681 	return 0;
1682 }
1683 
1684 static int set_maps(struct perf_stat *st)
1685 {
1686 	if (!st->cpus || !st->threads)
1687 		return 0;
1688 
1689 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1690 		return -EINVAL;
1691 
1692 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
1693 
1694 	if (perf_evlist__alloc_stats(evsel_list, true))
1695 		return -ENOMEM;
1696 
1697 	st->maps_allocated = true;
1698 	return 0;
1699 }
1700 
1701 static
1702 int process_thread_map_event(struct perf_session *session,
1703 			     union perf_event *event)
1704 {
1705 	struct perf_tool *tool = session->tool;
1706 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1707 
1708 	if (st->threads) {
1709 		pr_warning("Extra thread map event, ignoring.\n");
1710 		return 0;
1711 	}
1712 
1713 	st->threads = thread_map__new_event(&event->thread_map);
1714 	if (!st->threads)
1715 		return -ENOMEM;
1716 
1717 	return set_maps(st);
1718 }
1719 
1720 static
1721 int process_cpu_map_event(struct perf_session *session,
1722 			  union perf_event *event)
1723 {
1724 	struct perf_tool *tool = session->tool;
1725 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1726 	struct perf_cpu_map *cpus;
1727 
1728 	if (st->cpus) {
1729 		pr_warning("Extra cpu map event, ignoring.\n");
1730 		return 0;
1731 	}
1732 
1733 	cpus = cpu_map__new_data(&event->cpu_map.data);
1734 	if (!cpus)
1735 		return -ENOMEM;
1736 
1737 	st->cpus = cpus;
1738 	return set_maps(st);
1739 }
1740 
1741 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
1742 {
1743 	int i;
1744 
1745 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
1746 	if (!config->stats)
1747 		return -1;
1748 
1749 	config->stats_num = nthreads;
1750 
1751 	for (i = 0; i < nthreads; i++)
1752 		runtime_stat__init(&config->stats[i]);
1753 
1754 	return 0;
1755 }
1756 
1757 static void runtime_stat_delete(struct perf_stat_config *config)
1758 {
1759 	int i;
1760 
1761 	if (!config->stats)
1762 		return;
1763 
1764 	for (i = 0; i < config->stats_num; i++)
1765 		runtime_stat__exit(&config->stats[i]);
1766 
1767 	zfree(&config->stats);
1768 }
1769 
1770 static const char * const stat_report_usage[] = {
1771 	"perf stat report [<options>]",
1772 	NULL,
1773 };
1774 
1775 static struct perf_stat perf_stat = {
1776 	.tool = {
1777 		.attr		= perf_event__process_attr,
1778 		.event_update	= perf_event__process_event_update,
1779 		.thread_map	= process_thread_map_event,
1780 		.cpu_map	= process_cpu_map_event,
1781 		.stat_config	= process_stat_config_event,
1782 		.stat		= perf_event__process_stat_event,
1783 		.stat_round	= process_stat_round_event,
1784 	},
1785 	.aggr_mode = AGGR_UNSET,
1786 };
1787 
1788 static int __cmd_report(int argc, const char **argv)
1789 {
1790 	struct perf_session *session;
1791 	const struct option options[] = {
1792 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1793 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1794 		     "aggregate counts per processor socket", AGGR_SOCKET),
1795 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
1796 		     "aggregate counts per processor die", AGGR_DIE),
1797 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1798 		     "aggregate counts per physical processor core", AGGR_CORE),
1799 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
1800 		     "aggregate counts per numa node", AGGR_NODE),
1801 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1802 		     "disable CPU count aggregation", AGGR_NONE),
1803 	OPT_END()
1804 	};
1805 	struct stat st;
1806 	int ret;
1807 
1808 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
1809 
1810 	if (!input_name || !strlen(input_name)) {
1811 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1812 			input_name = "-";
1813 		else
1814 			input_name = "perf.data";
1815 	}
1816 
1817 	perf_stat.data.path = input_name;
1818 	perf_stat.data.mode = PERF_DATA_MODE_READ;
1819 
1820 	session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
1821 	if (IS_ERR(session))
1822 		return PTR_ERR(session);
1823 
1824 	perf_stat.session  = session;
1825 	stat_config.output = stderr;
1826 	evsel_list         = session->evlist;
1827 
1828 	ret = perf_session__process_events(session);
1829 	if (ret)
1830 		return ret;
1831 
1832 	perf_session__delete(session);
1833 	return 0;
1834 }
1835 
1836 static void setup_system_wide(int forks)
1837 {
1838 	/*
1839 	 * Make system wide (-a) the default target if
1840 	 * no target was specified and one of following
1841 	 * conditions is met:
1842 	 *
1843 	 *   - there's no workload specified
1844 	 *   - there is workload specified but all requested
1845 	 *     events are system wide events
1846 	 */
1847 	if (!target__none(&target))
1848 		return;
1849 
1850 	if (!forks)
1851 		target.system_wide = true;
1852 	else {
1853 		struct evsel *counter;
1854 
1855 		evlist__for_each_entry(evsel_list, counter) {
1856 			if (!counter->core.system_wide)
1857 				return;
1858 		}
1859 
1860 		if (evsel_list->core.nr_entries)
1861 			target.system_wide = true;
1862 	}
1863 }
1864 
1865 int cmd_stat(int argc, const char **argv)
1866 {
1867 	const char * const stat_usage[] = {
1868 		"perf stat [<options>] [<command>]",
1869 		NULL
1870 	};
1871 	int status = -EINVAL, run_idx;
1872 	const char *mode;
1873 	FILE *output = stderr;
1874 	unsigned int interval, timeout;
1875 	const char * const stat_subcommands[] = { "record", "report" };
1876 
1877 	setlocale(LC_ALL, "");
1878 
1879 	evsel_list = evlist__new();
1880 	if (evsel_list == NULL)
1881 		return -ENOMEM;
1882 
1883 	parse_events__shrink_config_terms();
1884 
1885 	/* String-parsing callback-based options would segfault when negated */
1886 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
1887 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
1888 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
1889 
1890 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1891 					(const char **) stat_usage,
1892 					PARSE_OPT_STOP_AT_NON_OPTION);
1893 	perf_stat__collect_metric_expr(evsel_list);
1894 	perf_stat__init_shadow_stats();
1895 
1896 	if (stat_config.csv_sep) {
1897 		stat_config.csv_output = true;
1898 		if (!strcmp(stat_config.csv_sep, "\\t"))
1899 			stat_config.csv_sep = "\t";
1900 	} else
1901 		stat_config.csv_sep = DEFAULT_SEPARATOR;
1902 
1903 	if (argc && !strncmp(argv[0], "rec", 3)) {
1904 		argc = __cmd_record(argc, argv);
1905 		if (argc < 0)
1906 			return -1;
1907 	} else if (argc && !strncmp(argv[0], "rep", 3))
1908 		return __cmd_report(argc, argv);
1909 
1910 	interval = stat_config.interval;
1911 	timeout = stat_config.timeout;
1912 
1913 	/*
1914 	 * For record command the -o is already taken care of.
1915 	 */
1916 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1917 		output = NULL;
1918 
1919 	if (output_name && output_fd) {
1920 		fprintf(stderr, "cannot use both --output and --log-fd\n");
1921 		parse_options_usage(stat_usage, stat_options, "o", 1);
1922 		parse_options_usage(NULL, stat_options, "log-fd", 0);
1923 		goto out;
1924 	}
1925 
1926 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
1927 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
1928 		goto out;
1929 	}
1930 
1931 	if (stat_config.metric_only && stat_config.run_count > 1) {
1932 		fprintf(stderr, "--metric-only is not supported with -r\n");
1933 		goto out;
1934 	}
1935 
1936 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
1937 		fprintf(stderr, "--table is only supported with -r\n");
1938 		parse_options_usage(stat_usage, stat_options, "r", 1);
1939 		parse_options_usage(NULL, stat_options, "table", 0);
1940 		goto out;
1941 	}
1942 
1943 	if (output_fd < 0) {
1944 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
1945 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1946 		goto out;
1947 	}
1948 
1949 	if (!output) {
1950 		struct timespec tm;
1951 		mode = append_file ? "a" : "w";
1952 
1953 		output = fopen(output_name, mode);
1954 		if (!output) {
1955 			perror("failed to create output file");
1956 			return -1;
1957 		}
1958 		clock_gettime(CLOCK_REALTIME, &tm);
1959 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1960 	} else if (output_fd > 0) {
1961 		mode = append_file ? "a" : "w";
1962 		output = fdopen(output_fd, mode);
1963 		if (!output) {
1964 			perror("Failed opening logfd");
1965 			return -errno;
1966 		}
1967 	}
1968 
1969 	stat_config.output = output;
1970 
1971 	/*
1972 	 * let the spreadsheet do the pretty-printing
1973 	 */
1974 	if (stat_config.csv_output) {
1975 		/* User explicitly passed -B? */
1976 		if (big_num_opt == 1) {
1977 			fprintf(stderr, "-B option not supported with -x\n");
1978 			parse_options_usage(stat_usage, stat_options, "B", 1);
1979 			parse_options_usage(NULL, stat_options, "x", 1);
1980 			goto out;
1981 		} else /* Nope, so disable big number formatting */
1982 			stat_config.big_num = false;
1983 	} else if (big_num_opt == 0) /* User passed --no-big-num */
1984 		stat_config.big_num = false;
1985 
1986 	setup_system_wide(argc);
1987 
1988 	/*
1989 	 * Display user/system times only for single
1990 	 * run and when there's specified tracee.
1991 	 */
1992 	if ((stat_config.run_count == 1) && target__none(&target))
1993 		stat_config.ru_display = true;
1994 
1995 	if (stat_config.run_count < 0) {
1996 		pr_err("Run count must be a positive number\n");
1997 		parse_options_usage(stat_usage, stat_options, "r", 1);
1998 		goto out;
1999 	} else if (stat_config.run_count == 0) {
2000 		forever = true;
2001 		stat_config.run_count = 1;
2002 	}
2003 
2004 	if (stat_config.walltime_run_table) {
2005 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2006 		if (!stat_config.walltime_run) {
2007 			pr_err("failed to setup -r option");
2008 			goto out;
2009 		}
2010 	}
2011 
2012 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2013 		!target__has_task(&target)) {
2014 		if (!target.system_wide || target.cpu_list) {
2015 			fprintf(stderr, "The --per-thread option is only "
2016 				"available when monitoring via -p -t -a "
2017 				"options or only --per-thread.\n");
2018 			parse_options_usage(NULL, stat_options, "p", 1);
2019 			parse_options_usage(NULL, stat_options, "t", 1);
2020 			goto out;
2021 		}
2022 	}
2023 
2024 	/*
2025 	 * no_aggr, cgroup are for system-wide only
2026 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2027 	 */
2028 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2029 	      stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2030 	    !target__has_cpu(&target)) {
2031 		fprintf(stderr, "both cgroup and no-aggregation "
2032 			"modes only available in system-wide mode\n");
2033 
2034 		parse_options_usage(stat_usage, stat_options, "G", 1);
2035 		parse_options_usage(NULL, stat_options, "A", 1);
2036 		parse_options_usage(NULL, stat_options, "a", 1);
2037 		goto out;
2038 	}
2039 
2040 	if (add_default_attributes())
2041 		goto out;
2042 
2043 	target__validate(&target);
2044 
2045 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2046 		target.per_thread = true;
2047 
2048 	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2049 		if (target__has_task(&target)) {
2050 			pr_err("Problems finding threads of monitor\n");
2051 			parse_options_usage(stat_usage, stat_options, "p", 1);
2052 			parse_options_usage(NULL, stat_options, "t", 1);
2053 		} else if (target__has_cpu(&target)) {
2054 			perror("failed to parse CPUs map");
2055 			parse_options_usage(stat_usage, stat_options, "C", 1);
2056 			parse_options_usage(NULL, stat_options, "a", 1);
2057 		}
2058 		goto out;
2059 	}
2060 
2061 	/*
2062 	 * Initialize thread_map with comm names,
2063 	 * so we could print it out on output.
2064 	 */
2065 	if (stat_config.aggr_mode == AGGR_THREAD) {
2066 		thread_map__read_comms(evsel_list->core.threads);
2067 		if (target.system_wide) {
2068 			if (runtime_stat_new(&stat_config,
2069 				perf_thread_map__nr(evsel_list->core.threads))) {
2070 				goto out;
2071 			}
2072 		}
2073 	}
2074 
2075 	if (stat_config.aggr_mode == AGGR_NODE)
2076 		cpu__setup_cpunode_map();
2077 
2078 	if (stat_config.times && interval)
2079 		interval_count = true;
2080 	else if (stat_config.times && !interval) {
2081 		pr_err("interval-count option should be used together with "
2082 				"interval-print.\n");
2083 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2084 		parse_options_usage(stat_usage, stat_options, "I", 1);
2085 		goto out;
2086 	}
2087 
2088 	if (timeout && timeout < 100) {
2089 		if (timeout < 10) {
2090 			pr_err("timeout must be >= 10ms.\n");
2091 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2092 			goto out;
2093 		} else
2094 			pr_warning("timeout < 100ms. "
2095 				   "The overhead percentage could be high in some cases. "
2096 				   "Please proceed with caution.\n");
2097 	}
2098 	if (timeout && interval) {
2099 		pr_err("timeout option is not supported with interval-print.\n");
2100 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2101 		parse_options_usage(stat_usage, stat_options, "I", 1);
2102 		goto out;
2103 	}
2104 
2105 	if (perf_evlist__alloc_stats(evsel_list, interval))
2106 		goto out;
2107 
2108 	if (perf_stat_init_aggr_mode())
2109 		goto out;
2110 
2111 	/*
2112 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2113 	 * while avoiding that older tools show confusing messages.
2114 	 *
2115 	 * However for pipe sessions we need to keep it zero,
2116 	 * because script's perf_evsel__check_attr is triggered
2117 	 * by attr->sample_type != 0, and we can't run it on
2118 	 * stat sessions.
2119 	 */
2120 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2121 
2122 	/*
2123 	 * We dont want to block the signals - that would cause
2124 	 * child tasks to inherit that and Ctrl-C would not work.
2125 	 * What we want is for Ctrl-C to work in the exec()-ed
2126 	 * task, but being ignored by perf stat itself:
2127 	 */
2128 	atexit(sig_atexit);
2129 	if (!forever)
2130 		signal(SIGINT,  skip_signal);
2131 	signal(SIGCHLD, skip_signal);
2132 	signal(SIGALRM, skip_signal);
2133 	signal(SIGABRT, skip_signal);
2134 
2135 	status = 0;
2136 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2137 		if (stat_config.run_count != 1 && verbose > 0)
2138 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2139 				run_idx + 1);
2140 
2141 		if (run_idx != 0)
2142 			perf_evlist__reset_prev_raw_counts(evsel_list);
2143 
2144 		status = run_perf_stat(argc, argv, run_idx);
2145 		if (forever && status != -1 && !interval) {
2146 			print_counters(NULL, argc, argv);
2147 			perf_stat__reset_stats();
2148 		}
2149 	}
2150 
2151 	if (!forever && status != -1 && !interval)
2152 		print_counters(NULL, argc, argv);
2153 
2154 	if (STAT_RECORD) {
2155 		/*
2156 		 * We synthesize the kernel mmap record just so that older tools
2157 		 * don't emit warnings about not being able to resolve symbols
2158 		 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2159 		 * a saner message about no samples being in the perf.data file.
2160 		 *
2161 		 * This also serves to suppress a warning about f_header.data.size == 0
2162 		 * in header.c at the moment 'perf stat record' gets introduced, which
2163 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2164 		 * records, but the need to suppress the kptr_restrict messages in older
2165 		 * tools remain  -acme
2166 		 */
2167 		int fd = perf_data__fd(&perf_stat.data);
2168 		int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2169 							     process_synthesized_event,
2170 							     &perf_stat.session->machines.host);
2171 		if (err) {
2172 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2173 				   "older tools may produce warnings about this file\n.");
2174 		}
2175 
2176 		if (!interval) {
2177 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2178 				pr_err("failed to write stat round event\n");
2179 		}
2180 
2181 		if (!perf_stat.data.is_pipe) {
2182 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2183 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2184 		}
2185 
2186 		evlist__close(evsel_list);
2187 		perf_session__delete(perf_stat.session);
2188 	}
2189 
2190 	perf_stat__exit_aggr_mode();
2191 	perf_evlist__free_stats(evsel_list);
2192 out:
2193 	zfree(&stat_config.walltime_run);
2194 
2195 	if (smi_cost && smi_reset)
2196 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2197 
2198 	evlist__delete(evsel_list);
2199 
2200 	runtime_stat_delete(&stat_config);
2201 
2202 	return status;
2203 }
2204