xref: /openbmc/linux/tools/perf/builtin-stat.c (revision 15e3ae36)
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 			wait4(child_pid, &status, 0, &stat_config.ru_data);
691 
692 		if (workload_exec_errno) {
693 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
694 			pr_err("Workload failed: %s\n", emsg);
695 			return -1;
696 		}
697 
698 		if (WIFSIGNALED(status))
699 			psignal(WTERMSIG(status), argv[0]);
700 	} else {
701 		enable_counters();
702 		while (!done) {
703 			nanosleep(&ts, NULL);
704 			if (!is_target_alive(&target, evsel_list->core.threads))
705 				break;
706 			if (timeout)
707 				break;
708 			if (interval) {
709 				process_interval();
710 				if (interval_count && !(--times))
711 					break;
712 			}
713 		}
714 	}
715 
716 	disable_counters();
717 
718 	t1 = rdclock();
719 
720 	if (stat_config.walltime_run_table)
721 		stat_config.walltime_run[run_idx] = t1 - t0;
722 
723 	update_stats(&walltime_nsecs_stats, t1 - t0);
724 
725 	/*
726 	 * Closing a group leader splits the group, and as we only disable
727 	 * group leaders, results in remaining events becoming enabled. To
728 	 * avoid arbitrary skew, we must read all counters before closing any
729 	 * group leaders.
730 	 */
731 	read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
732 
733 	/*
734 	 * We need to keep evsel_list alive, because it's processed
735 	 * later the evsel_list will be closed after.
736 	 */
737 	if (!STAT_RECORD)
738 		evlist__close(evsel_list);
739 
740 	return WEXITSTATUS(status);
741 }
742 
743 static int run_perf_stat(int argc, const char **argv, int run_idx)
744 {
745 	int ret;
746 
747 	if (pre_cmd) {
748 		ret = system(pre_cmd);
749 		if (ret)
750 			return ret;
751 	}
752 
753 	if (sync_run)
754 		sync();
755 
756 	ret = __run_perf_stat(argc, argv, run_idx);
757 	if (ret)
758 		return ret;
759 
760 	if (post_cmd) {
761 		ret = system(post_cmd);
762 		if (ret)
763 			return ret;
764 	}
765 
766 	return ret;
767 }
768 
769 static void print_counters(struct timespec *ts, int argc, const char **argv)
770 {
771 	/* Do not print anything if we record to the pipe. */
772 	if (STAT_RECORD && perf_stat.data.is_pipe)
773 		return;
774 
775 	perf_evlist__print_counters(evsel_list, &stat_config, &target,
776 				    ts, argc, argv);
777 }
778 
779 static volatile int signr = -1;
780 
781 static void skip_signal(int signo)
782 {
783 	if ((child_pid == -1) || stat_config.interval)
784 		done = 1;
785 
786 	signr = signo;
787 	/*
788 	 * render child_pid harmless
789 	 * won't send SIGTERM to a random
790 	 * process in case of race condition
791 	 * and fast PID recycling
792 	 */
793 	child_pid = -1;
794 }
795 
796 static void sig_atexit(void)
797 {
798 	sigset_t set, oset;
799 
800 	/*
801 	 * avoid race condition with SIGCHLD handler
802 	 * in skip_signal() which is modifying child_pid
803 	 * goal is to avoid send SIGTERM to a random
804 	 * process
805 	 */
806 	sigemptyset(&set);
807 	sigaddset(&set, SIGCHLD);
808 	sigprocmask(SIG_BLOCK, &set, &oset);
809 
810 	if (child_pid != -1)
811 		kill(child_pid, SIGTERM);
812 
813 	sigprocmask(SIG_SETMASK, &oset, NULL);
814 
815 	if (signr == -1)
816 		return;
817 
818 	signal(signr, SIG_DFL);
819 	kill(getpid(), signr);
820 }
821 
822 static int stat__set_big_num(const struct option *opt __maybe_unused,
823 			     const char *s __maybe_unused, int unset)
824 {
825 	big_num_opt = unset ? 0 : 1;
826 	return 0;
827 }
828 
829 static int enable_metric_only(const struct option *opt __maybe_unused,
830 			      const char *s __maybe_unused, int unset)
831 {
832 	force_metric_only = true;
833 	stat_config.metric_only = !unset;
834 	return 0;
835 }
836 
837 static int parse_metric_groups(const struct option *opt,
838 			       const char *str,
839 			       int unset __maybe_unused)
840 {
841 	return metricgroup__parse_groups(opt, str, &stat_config.metric_events);
842 }
843 
844 static struct option stat_options[] = {
845 	OPT_BOOLEAN('T', "transaction", &transaction_run,
846 		    "hardware transaction statistics"),
847 	OPT_CALLBACK('e', "event", &evsel_list, "event",
848 		     "event selector. use 'perf list' to list available events",
849 		     parse_events_option),
850 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
851 		     "event filter", parse_filter),
852 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
853 		    "child tasks do not inherit counters"),
854 	OPT_STRING('p', "pid", &target.pid, "pid",
855 		   "stat events on existing process id"),
856 	OPT_STRING('t', "tid", &target.tid, "tid",
857 		   "stat events on existing thread id"),
858 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
859 		    "system-wide collection from all CPUs"),
860 	OPT_BOOLEAN('g', "group", &group,
861 		    "put the counters into a counter group"),
862 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
863 		    "Use --no-scale to disable counter scaling for multiplexing"),
864 	OPT_INCR('v', "verbose", &verbose,
865 		    "be more verbose (show counter open errors, etc)"),
866 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
867 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
868 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
869 		    "display details about each run (only with -r option)"),
870 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
871 		    "null run - dont start any counters"),
872 	OPT_INCR('d', "detailed", &detailed_run,
873 		    "detailed run - start a lot of events"),
874 	OPT_BOOLEAN('S', "sync", &sync_run,
875 		    "call sync() before starting a run"),
876 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
877 			   "print large numbers with thousands\' separators",
878 			   stat__set_big_num),
879 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
880 		    "list of cpus to monitor in system-wide"),
881 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
882 		    "disable CPU count aggregation", AGGR_NONE),
883 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
884 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
885 		   "print counts with custom separator"),
886 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
887 		     "monitor event in cgroup name only", parse_cgroups),
888 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
889 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
890 	OPT_INTEGER(0, "log-fd", &output_fd,
891 		    "log output to fd, instead of stderr"),
892 	OPT_STRING(0, "pre", &pre_cmd, "command",
893 			"command to run prior to the measured command"),
894 	OPT_STRING(0, "post", &post_cmd, "command",
895 			"command to run after to the measured command"),
896 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
897 		    "print counts at regular interval in ms "
898 		    "(overhead is possible for values <= 100ms)"),
899 	OPT_INTEGER(0, "interval-count", &stat_config.times,
900 		    "print counts for fixed number of times"),
901 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
902 		    "clear screen in between new interval"),
903 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
904 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
905 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
906 		     "aggregate counts per processor socket", AGGR_SOCKET),
907 	OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
908 		     "aggregate counts per processor die", AGGR_DIE),
909 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
910 		     "aggregate counts per physical processor core", AGGR_CORE),
911 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
912 		     "aggregate counts per thread", AGGR_THREAD),
913 	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
914 		     "aggregate counts per numa node", AGGR_NODE),
915 	OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
916 		     "ms to wait before starting measurement after program start"),
917 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
918 			"Only print computed metrics. No raw values", enable_metric_only),
919 	OPT_BOOLEAN(0, "topdown", &topdown_run,
920 			"measure topdown level 1 statistics"),
921 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
922 			"measure SMI cost"),
923 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
924 		     "monitor specified metrics or metric groups (separated by ,)",
925 		     parse_metric_groups),
926 	OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
927 			 "Configure all used events to run in kernel space.",
928 			 PARSE_OPT_EXCLUSIVE),
929 	OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
930 			 "Configure all used events to run in user space.",
931 			 PARSE_OPT_EXCLUSIVE),
932 	OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
933 		    "Use with 'percore' event qualifier to show the event "
934 		    "counts of one hardware thread by sum up total hardware "
935 		    "threads of same physical core"),
936 	OPT_END()
937 };
938 
939 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
940 				 struct perf_cpu_map *map, int cpu)
941 {
942 	return cpu_map__get_socket(map, cpu, NULL);
943 }
944 
945 static int perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
946 			      struct perf_cpu_map *map, int cpu)
947 {
948 	return cpu_map__get_die(map, cpu, NULL);
949 }
950 
951 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
952 			       struct perf_cpu_map *map, int cpu)
953 {
954 	return cpu_map__get_core(map, cpu, NULL);
955 }
956 
957 static int perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
958 			       struct perf_cpu_map *map, int cpu)
959 {
960 	return cpu_map__get_node(map, cpu, NULL);
961 }
962 
963 static int perf_stat__get_aggr(struct perf_stat_config *config,
964 			       aggr_get_id_t get_id, struct perf_cpu_map *map, int idx)
965 {
966 	int cpu;
967 
968 	if (idx >= map->nr)
969 		return -1;
970 
971 	cpu = map->map[idx];
972 
973 	if (config->cpus_aggr_map->map[cpu] == -1)
974 		config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
975 
976 	return config->cpus_aggr_map->map[cpu];
977 }
978 
979 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
980 					struct perf_cpu_map *map, int idx)
981 {
982 	return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
983 }
984 
985 static int perf_stat__get_die_cached(struct perf_stat_config *config,
986 					struct perf_cpu_map *map, int idx)
987 {
988 	return perf_stat__get_aggr(config, perf_stat__get_die, map, idx);
989 }
990 
991 static int perf_stat__get_core_cached(struct perf_stat_config *config,
992 				      struct perf_cpu_map *map, int idx)
993 {
994 	return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
995 }
996 
997 static int perf_stat__get_node_cached(struct perf_stat_config *config,
998 				      struct perf_cpu_map *map, int idx)
999 {
1000 	return perf_stat__get_aggr(config, perf_stat__get_node, map, idx);
1001 }
1002 
1003 static bool term_percore_set(void)
1004 {
1005 	struct evsel *counter;
1006 
1007 	evlist__for_each_entry(evsel_list, counter) {
1008 		if (counter->percore)
1009 			return true;
1010 	}
1011 
1012 	return false;
1013 }
1014 
1015 static int perf_stat_init_aggr_mode(void)
1016 {
1017 	int nr;
1018 
1019 	switch (stat_config.aggr_mode) {
1020 	case AGGR_SOCKET:
1021 		if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1022 			perror("cannot build socket map");
1023 			return -1;
1024 		}
1025 		stat_config.aggr_get_id = perf_stat__get_socket_cached;
1026 		break;
1027 	case AGGR_DIE:
1028 		if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1029 			perror("cannot build die map");
1030 			return -1;
1031 		}
1032 		stat_config.aggr_get_id = perf_stat__get_die_cached;
1033 		break;
1034 	case AGGR_CORE:
1035 		if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1036 			perror("cannot build core map");
1037 			return -1;
1038 		}
1039 		stat_config.aggr_get_id = perf_stat__get_core_cached;
1040 		break;
1041 	case AGGR_NODE:
1042 		if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1043 			perror("cannot build core map");
1044 			return -1;
1045 		}
1046 		stat_config.aggr_get_id = perf_stat__get_node_cached;
1047 		break;
1048 	case AGGR_NONE:
1049 		if (term_percore_set()) {
1050 			if (cpu_map__build_core_map(evsel_list->core.cpus,
1051 						    &stat_config.aggr_map)) {
1052 				perror("cannot build core map");
1053 				return -1;
1054 			}
1055 			stat_config.aggr_get_id = perf_stat__get_core_cached;
1056 		}
1057 		break;
1058 	case AGGR_GLOBAL:
1059 	case AGGR_THREAD:
1060 	case AGGR_UNSET:
1061 	default:
1062 		break;
1063 	}
1064 
1065 	/*
1066 	 * The evsel_list->cpus is the base we operate on,
1067 	 * taking the highest cpu number to be the size of
1068 	 * the aggregation translate cpumap.
1069 	 */
1070 	nr = perf_cpu_map__max(evsel_list->core.cpus);
1071 	stat_config.cpus_aggr_map = perf_cpu_map__empty_new(nr + 1);
1072 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1073 }
1074 
1075 static void perf_stat__exit_aggr_mode(void)
1076 {
1077 	perf_cpu_map__put(stat_config.aggr_map);
1078 	perf_cpu_map__put(stat_config.cpus_aggr_map);
1079 	stat_config.aggr_map = NULL;
1080 	stat_config.cpus_aggr_map = NULL;
1081 }
1082 
1083 static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx)
1084 {
1085 	int cpu;
1086 
1087 	if (idx > map->nr)
1088 		return -1;
1089 
1090 	cpu = map->map[idx];
1091 
1092 	if (cpu >= env->nr_cpus_avail)
1093 		return -1;
1094 
1095 	return cpu;
1096 }
1097 
1098 static int perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data)
1099 {
1100 	struct perf_env *env = data;
1101 	int cpu = perf_env__get_cpu(env, map, idx);
1102 
1103 	return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1104 }
1105 
1106 static int perf_env__get_die(struct perf_cpu_map *map, int idx, void *data)
1107 {
1108 	struct perf_env *env = data;
1109 	int die_id = -1, cpu = perf_env__get_cpu(env, map, idx);
1110 
1111 	if (cpu != -1) {
1112 		/*
1113 		 * Encode socket in bit range 15:8
1114 		 * die_id is relative to socket,
1115 		 * we need a global id. So we combine
1116 		 * socket + die id
1117 		 */
1118 		if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
1119 			return -1;
1120 
1121 		if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
1122 			return -1;
1123 
1124 		die_id = (env->cpu[cpu].socket_id << 8) | (env->cpu[cpu].die_id & 0xff);
1125 	}
1126 
1127 	return die_id;
1128 }
1129 
1130 static int perf_env__get_core(struct perf_cpu_map *map, int idx, void *data)
1131 {
1132 	struct perf_env *env = data;
1133 	int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1134 
1135 	if (cpu != -1) {
1136 		/*
1137 		 * Encode socket in bit range 31:24
1138 		 * encode die id in bit range 23:16
1139 		 * core_id is relative to socket and die,
1140 		 * we need a global id. So we combine
1141 		 * socket + die id + core id
1142 		 */
1143 		if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
1144 			return -1;
1145 
1146 		if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
1147 			return -1;
1148 
1149 		if (WARN_ONCE(env->cpu[cpu].core_id >> 16, "The core id number is too big.\n"))
1150 			return -1;
1151 
1152 		core = (env->cpu[cpu].socket_id << 24) |
1153 		       (env->cpu[cpu].die_id << 16) |
1154 		       (env->cpu[cpu].core_id & 0xffff);
1155 	}
1156 
1157 	return core;
1158 }
1159 
1160 static int perf_env__get_node(struct perf_cpu_map *map, int idx, void *data)
1161 {
1162 	int cpu = perf_env__get_cpu(data, map, idx);
1163 
1164 	return perf_env__numa_node(data, cpu);
1165 }
1166 
1167 static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus,
1168 				      struct perf_cpu_map **sockp)
1169 {
1170 	return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1171 }
1172 
1173 static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus,
1174 				   struct perf_cpu_map **diep)
1175 {
1176 	return cpu_map__build_map(cpus, diep, perf_env__get_die, env);
1177 }
1178 
1179 static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus,
1180 				    struct perf_cpu_map **corep)
1181 {
1182 	return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1183 }
1184 
1185 static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus,
1186 				    struct perf_cpu_map **nodep)
1187 {
1188 	return cpu_map__build_map(cpus, nodep, perf_env__get_node, env);
1189 }
1190 
1191 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1192 				      struct perf_cpu_map *map, int idx)
1193 {
1194 	return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1195 }
1196 static int perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1197 				   struct perf_cpu_map *map, int idx)
1198 {
1199 	return perf_env__get_die(map, idx, &perf_stat.session->header.env);
1200 }
1201 
1202 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1203 				    struct perf_cpu_map *map, int idx)
1204 {
1205 	return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1206 }
1207 
1208 static int perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1209 				    struct perf_cpu_map *map, int idx)
1210 {
1211 	return perf_env__get_node(map, idx, &perf_stat.session->header.env);
1212 }
1213 
1214 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1215 {
1216 	struct perf_env *env = &st->session->header.env;
1217 
1218 	switch (stat_config.aggr_mode) {
1219 	case AGGR_SOCKET:
1220 		if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1221 			perror("cannot build socket map");
1222 			return -1;
1223 		}
1224 		stat_config.aggr_get_id = perf_stat__get_socket_file;
1225 		break;
1226 	case AGGR_DIE:
1227 		if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1228 			perror("cannot build die map");
1229 			return -1;
1230 		}
1231 		stat_config.aggr_get_id = perf_stat__get_die_file;
1232 		break;
1233 	case AGGR_CORE:
1234 		if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1235 			perror("cannot build core map");
1236 			return -1;
1237 		}
1238 		stat_config.aggr_get_id = perf_stat__get_core_file;
1239 		break;
1240 	case AGGR_NODE:
1241 		if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1242 			perror("cannot build core map");
1243 			return -1;
1244 		}
1245 		stat_config.aggr_get_id = perf_stat__get_node_file;
1246 		break;
1247 	case AGGR_NONE:
1248 	case AGGR_GLOBAL:
1249 	case AGGR_THREAD:
1250 	case AGGR_UNSET:
1251 	default:
1252 		break;
1253 	}
1254 
1255 	return 0;
1256 }
1257 
1258 static int topdown_filter_events(const char **attr, char **str, bool use_group)
1259 {
1260 	int off = 0;
1261 	int i;
1262 	int len = 0;
1263 	char *s;
1264 
1265 	for (i = 0; attr[i]; i++) {
1266 		if (pmu_have_event("cpu", attr[i])) {
1267 			len += strlen(attr[i]) + 1;
1268 			attr[i - off] = attr[i];
1269 		} else
1270 			off++;
1271 	}
1272 	attr[i - off] = NULL;
1273 
1274 	*str = malloc(len + 1 + 2);
1275 	if (!*str)
1276 		return -1;
1277 	s = *str;
1278 	if (i - off == 0) {
1279 		*s = 0;
1280 		return 0;
1281 	}
1282 	if (use_group)
1283 		*s++ = '{';
1284 	for (i = 0; attr[i]; i++) {
1285 		strcpy(s, attr[i]);
1286 		s += strlen(s);
1287 		*s++ = ',';
1288 	}
1289 	if (use_group) {
1290 		s[-1] = '}';
1291 		*s = 0;
1292 	} else
1293 		s[-1] = 0;
1294 	return 0;
1295 }
1296 
1297 __weak bool arch_topdown_check_group(bool *warn)
1298 {
1299 	*warn = false;
1300 	return false;
1301 }
1302 
1303 __weak void arch_topdown_group_warn(void)
1304 {
1305 }
1306 
1307 /*
1308  * Add default attributes, if there were no attributes specified or
1309  * if -d/--detailed, -d -d or -d -d -d is used:
1310  */
1311 static int add_default_attributes(void)
1312 {
1313 	int err;
1314 	struct perf_event_attr default_attrs0[] = {
1315 
1316   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1317   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1318   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1319   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1320 
1321   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1322 };
1323 	struct perf_event_attr frontend_attrs[] = {
1324   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1325 };
1326 	struct perf_event_attr backend_attrs[] = {
1327   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1328 };
1329 	struct perf_event_attr default_attrs1[] = {
1330   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1331   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1332   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1333 
1334 };
1335 
1336 /*
1337  * Detailed stats (-d), covering the L1 and last level data caches:
1338  */
1339 	struct perf_event_attr detailed_attrs[] = {
1340 
1341   { .type = PERF_TYPE_HW_CACHE,
1342     .config =
1343 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1344 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1345 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1346 
1347   { .type = PERF_TYPE_HW_CACHE,
1348     .config =
1349 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1350 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1351 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1352 
1353   { .type = PERF_TYPE_HW_CACHE,
1354     .config =
1355 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1356 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1357 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1358 
1359   { .type = PERF_TYPE_HW_CACHE,
1360     .config =
1361 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1362 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1363 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1364 };
1365 
1366 /*
1367  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1368  */
1369 	struct perf_event_attr very_detailed_attrs[] = {
1370 
1371   { .type = PERF_TYPE_HW_CACHE,
1372     .config =
1373 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1374 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1375 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1376 
1377   { .type = PERF_TYPE_HW_CACHE,
1378     .config =
1379 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1380 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1381 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1382 
1383   { .type = PERF_TYPE_HW_CACHE,
1384     .config =
1385 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1386 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1387 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1388 
1389   { .type = PERF_TYPE_HW_CACHE,
1390     .config =
1391 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1392 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1393 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1394 
1395   { .type = PERF_TYPE_HW_CACHE,
1396     .config =
1397 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1398 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1399 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1400 
1401   { .type = PERF_TYPE_HW_CACHE,
1402     .config =
1403 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1404 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1405 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1406 
1407 };
1408 
1409 /*
1410  * Very, very detailed stats (-d -d -d), adding prefetch events:
1411  */
1412 	struct perf_event_attr very_very_detailed_attrs[] = {
1413 
1414   { .type = PERF_TYPE_HW_CACHE,
1415     .config =
1416 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1417 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1418 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1419 
1420   { .type = PERF_TYPE_HW_CACHE,
1421     .config =
1422 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1423 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1424 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1425 };
1426 	struct parse_events_error errinfo;
1427 
1428 	/* Set attrs if no event is selected and !null_run: */
1429 	if (stat_config.null_run)
1430 		return 0;
1431 
1432 	bzero(&errinfo, sizeof(errinfo));
1433 	if (transaction_run) {
1434 		/* Handle -T as -M transaction. Once platform specific metrics
1435 		 * support has been added to the json files, all archictures
1436 		 * will use this approach. To determine transaction support
1437 		 * on an architecture test for such a metric name.
1438 		 */
1439 		if (metricgroup__has_metric("transaction")) {
1440 			struct option opt = { .value = &evsel_list };
1441 
1442 			return metricgroup__parse_groups(&opt, "transaction",
1443 							 &stat_config.metric_events);
1444 		}
1445 
1446 		if (pmu_have_event("cpu", "cycles-ct") &&
1447 		    pmu_have_event("cpu", "el-start"))
1448 			err = parse_events(evsel_list, transaction_attrs,
1449 					   &errinfo);
1450 		else
1451 			err = parse_events(evsel_list,
1452 					   transaction_limited_attrs,
1453 					   &errinfo);
1454 		if (err) {
1455 			fprintf(stderr, "Cannot set up transaction events\n");
1456 			parse_events_print_error(&errinfo, transaction_attrs);
1457 			return -1;
1458 		}
1459 		return 0;
1460 	}
1461 
1462 	if (smi_cost) {
1463 		int smi;
1464 
1465 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1466 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1467 			return -1;
1468 		}
1469 
1470 		if (!smi) {
1471 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1472 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1473 				return -1;
1474 			}
1475 			smi_reset = true;
1476 		}
1477 
1478 		if (pmu_have_event("msr", "aperf") &&
1479 		    pmu_have_event("msr", "smi")) {
1480 			if (!force_metric_only)
1481 				stat_config.metric_only = true;
1482 			err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1483 		} else {
1484 			fprintf(stderr, "To measure SMI cost, it needs "
1485 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1486 			parse_events_print_error(&errinfo, smi_cost_attrs);
1487 			return -1;
1488 		}
1489 		if (err) {
1490 			parse_events_print_error(&errinfo, smi_cost_attrs);
1491 			fprintf(stderr, "Cannot set up SMI cost events\n");
1492 			return -1;
1493 		}
1494 		return 0;
1495 	}
1496 
1497 	if (topdown_run) {
1498 		char *str = NULL;
1499 		bool warn = false;
1500 
1501 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1502 		    stat_config.aggr_mode != AGGR_CORE) {
1503 			pr_err("top down event configuration requires --per-core mode\n");
1504 			return -1;
1505 		}
1506 		stat_config.aggr_mode = AGGR_CORE;
1507 		if (nr_cgroups || !target__has_cpu(&target)) {
1508 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1509 			return -1;
1510 		}
1511 
1512 		if (!force_metric_only)
1513 			stat_config.metric_only = true;
1514 		if (topdown_filter_events(topdown_attrs, &str,
1515 				arch_topdown_check_group(&warn)) < 0) {
1516 			pr_err("Out of memory\n");
1517 			return -1;
1518 		}
1519 		if (topdown_attrs[0] && str) {
1520 			if (warn)
1521 				arch_topdown_group_warn();
1522 			err = parse_events(evsel_list, str, &errinfo);
1523 			if (err) {
1524 				fprintf(stderr,
1525 					"Cannot set up top down events %s: %d\n",
1526 					str, err);
1527 				parse_events_print_error(&errinfo, str);
1528 				free(str);
1529 				return -1;
1530 			}
1531 		} else {
1532 			fprintf(stderr, "System does not support topdown\n");
1533 			return -1;
1534 		}
1535 		free(str);
1536 	}
1537 
1538 	if (!evsel_list->core.nr_entries) {
1539 		if (target__has_cpu(&target))
1540 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1541 
1542 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1543 			return -1;
1544 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1545 			if (perf_evlist__add_default_attrs(evsel_list,
1546 						frontend_attrs) < 0)
1547 				return -1;
1548 		}
1549 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1550 			if (perf_evlist__add_default_attrs(evsel_list,
1551 						backend_attrs) < 0)
1552 				return -1;
1553 		}
1554 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1555 			return -1;
1556 	}
1557 
1558 	/* Detailed events get appended to the event list: */
1559 
1560 	if (detailed_run <  1)
1561 		return 0;
1562 
1563 	/* Append detailed run extra attributes: */
1564 	if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1565 		return -1;
1566 
1567 	if (detailed_run < 2)
1568 		return 0;
1569 
1570 	/* Append very detailed run extra attributes: */
1571 	if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1572 		return -1;
1573 
1574 	if (detailed_run < 3)
1575 		return 0;
1576 
1577 	/* Append very, very detailed run extra attributes: */
1578 	return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1579 }
1580 
1581 static const char * const stat_record_usage[] = {
1582 	"perf stat record [<options>]",
1583 	NULL,
1584 };
1585 
1586 static void init_features(struct perf_session *session)
1587 {
1588 	int feat;
1589 
1590 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1591 		perf_header__set_feat(&session->header, feat);
1592 
1593 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1594 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1595 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1596 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1597 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1598 }
1599 
1600 static int __cmd_record(int argc, const char **argv)
1601 {
1602 	struct perf_session *session;
1603 	struct perf_data *data = &perf_stat.data;
1604 
1605 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
1606 			     PARSE_OPT_STOP_AT_NON_OPTION);
1607 
1608 	if (output_name)
1609 		data->path = output_name;
1610 
1611 	if (stat_config.run_count != 1 || forever) {
1612 		pr_err("Cannot use -r option with perf stat record.\n");
1613 		return -1;
1614 	}
1615 
1616 	session = perf_session__new(data, false, NULL);
1617 	if (IS_ERR(session)) {
1618 		pr_err("Perf session creation failed\n");
1619 		return PTR_ERR(session);
1620 	}
1621 
1622 	init_features(session);
1623 
1624 	session->evlist   = evsel_list;
1625 	perf_stat.session = session;
1626 	perf_stat.record  = true;
1627 	return argc;
1628 }
1629 
1630 static int process_stat_round_event(struct perf_session *session,
1631 				    union perf_event *event)
1632 {
1633 	struct perf_record_stat_round *stat_round = &event->stat_round;
1634 	struct evsel *counter;
1635 	struct timespec tsh, *ts = NULL;
1636 	const char **argv = session->header.env.cmdline_argv;
1637 	int argc = session->header.env.nr_cmdline;
1638 
1639 	evlist__for_each_entry(evsel_list, counter)
1640 		perf_stat_process_counter(&stat_config, counter);
1641 
1642 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1643 		update_stats(&walltime_nsecs_stats, stat_round->time);
1644 
1645 	if (stat_config.interval && stat_round->time) {
1646 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
1647 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1648 		ts = &tsh;
1649 	}
1650 
1651 	print_counters(ts, argc, argv);
1652 	return 0;
1653 }
1654 
1655 static
1656 int process_stat_config_event(struct perf_session *session,
1657 			      union perf_event *event)
1658 {
1659 	struct perf_tool *tool = session->tool;
1660 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1661 
1662 	perf_event__read_stat_config(&stat_config, &event->stat_config);
1663 
1664 	if (perf_cpu_map__empty(st->cpus)) {
1665 		if (st->aggr_mode != AGGR_UNSET)
1666 			pr_warning("warning: processing task data, aggregation mode not set\n");
1667 		return 0;
1668 	}
1669 
1670 	if (st->aggr_mode != AGGR_UNSET)
1671 		stat_config.aggr_mode = st->aggr_mode;
1672 
1673 	if (perf_stat.data.is_pipe)
1674 		perf_stat_init_aggr_mode();
1675 	else
1676 		perf_stat_init_aggr_mode_file(st);
1677 
1678 	return 0;
1679 }
1680 
1681 static int set_maps(struct perf_stat *st)
1682 {
1683 	if (!st->cpus || !st->threads)
1684 		return 0;
1685 
1686 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1687 		return -EINVAL;
1688 
1689 	perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
1690 
1691 	if (perf_evlist__alloc_stats(evsel_list, true))
1692 		return -ENOMEM;
1693 
1694 	st->maps_allocated = true;
1695 	return 0;
1696 }
1697 
1698 static
1699 int process_thread_map_event(struct perf_session *session,
1700 			     union perf_event *event)
1701 {
1702 	struct perf_tool *tool = session->tool;
1703 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1704 
1705 	if (st->threads) {
1706 		pr_warning("Extra thread map event, ignoring.\n");
1707 		return 0;
1708 	}
1709 
1710 	st->threads = thread_map__new_event(&event->thread_map);
1711 	if (!st->threads)
1712 		return -ENOMEM;
1713 
1714 	return set_maps(st);
1715 }
1716 
1717 static
1718 int process_cpu_map_event(struct perf_session *session,
1719 			  union perf_event *event)
1720 {
1721 	struct perf_tool *tool = session->tool;
1722 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1723 	struct perf_cpu_map *cpus;
1724 
1725 	if (st->cpus) {
1726 		pr_warning("Extra cpu map event, ignoring.\n");
1727 		return 0;
1728 	}
1729 
1730 	cpus = cpu_map__new_data(&event->cpu_map.data);
1731 	if (!cpus)
1732 		return -ENOMEM;
1733 
1734 	st->cpus = cpus;
1735 	return set_maps(st);
1736 }
1737 
1738 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
1739 {
1740 	int i;
1741 
1742 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
1743 	if (!config->stats)
1744 		return -1;
1745 
1746 	config->stats_num = nthreads;
1747 
1748 	for (i = 0; i < nthreads; i++)
1749 		runtime_stat__init(&config->stats[i]);
1750 
1751 	return 0;
1752 }
1753 
1754 static void runtime_stat_delete(struct perf_stat_config *config)
1755 {
1756 	int i;
1757 
1758 	if (!config->stats)
1759 		return;
1760 
1761 	for (i = 0; i < config->stats_num; i++)
1762 		runtime_stat__exit(&config->stats[i]);
1763 
1764 	zfree(&config->stats);
1765 }
1766 
1767 static const char * const stat_report_usage[] = {
1768 	"perf stat report [<options>]",
1769 	NULL,
1770 };
1771 
1772 static struct perf_stat perf_stat = {
1773 	.tool = {
1774 		.attr		= perf_event__process_attr,
1775 		.event_update	= perf_event__process_event_update,
1776 		.thread_map	= process_thread_map_event,
1777 		.cpu_map	= process_cpu_map_event,
1778 		.stat_config	= process_stat_config_event,
1779 		.stat		= perf_event__process_stat_event,
1780 		.stat_round	= process_stat_round_event,
1781 	},
1782 	.aggr_mode = AGGR_UNSET,
1783 };
1784 
1785 static int __cmd_report(int argc, const char **argv)
1786 {
1787 	struct perf_session *session;
1788 	const struct option options[] = {
1789 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1790 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1791 		     "aggregate counts per processor socket", AGGR_SOCKET),
1792 	OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
1793 		     "aggregate counts per processor die", AGGR_DIE),
1794 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1795 		     "aggregate counts per physical processor core", AGGR_CORE),
1796 	OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
1797 		     "aggregate counts per numa node", AGGR_NODE),
1798 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1799 		     "disable CPU count aggregation", AGGR_NONE),
1800 	OPT_END()
1801 	};
1802 	struct stat st;
1803 	int ret;
1804 
1805 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
1806 
1807 	if (!input_name || !strlen(input_name)) {
1808 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1809 			input_name = "-";
1810 		else
1811 			input_name = "perf.data";
1812 	}
1813 
1814 	perf_stat.data.path = input_name;
1815 	perf_stat.data.mode = PERF_DATA_MODE_READ;
1816 
1817 	session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
1818 	if (IS_ERR(session))
1819 		return PTR_ERR(session);
1820 
1821 	perf_stat.session  = session;
1822 	stat_config.output = stderr;
1823 	evsel_list         = session->evlist;
1824 
1825 	ret = perf_session__process_events(session);
1826 	if (ret)
1827 		return ret;
1828 
1829 	perf_session__delete(session);
1830 	return 0;
1831 }
1832 
1833 static void setup_system_wide(int forks)
1834 {
1835 	/*
1836 	 * Make system wide (-a) the default target if
1837 	 * no target was specified and one of following
1838 	 * conditions is met:
1839 	 *
1840 	 *   - there's no workload specified
1841 	 *   - there is workload specified but all requested
1842 	 *     events are system wide events
1843 	 */
1844 	if (!target__none(&target))
1845 		return;
1846 
1847 	if (!forks)
1848 		target.system_wide = true;
1849 	else {
1850 		struct evsel *counter;
1851 
1852 		evlist__for_each_entry(evsel_list, counter) {
1853 			if (!counter->core.system_wide)
1854 				return;
1855 		}
1856 
1857 		if (evsel_list->core.nr_entries)
1858 			target.system_wide = true;
1859 	}
1860 }
1861 
1862 int cmd_stat(int argc, const char **argv)
1863 {
1864 	const char * const stat_usage[] = {
1865 		"perf stat [<options>] [<command>]",
1866 		NULL
1867 	};
1868 	int status = -EINVAL, run_idx;
1869 	const char *mode;
1870 	FILE *output = stderr;
1871 	unsigned int interval, timeout;
1872 	const char * const stat_subcommands[] = { "record", "report" };
1873 
1874 	setlocale(LC_ALL, "");
1875 
1876 	evsel_list = evlist__new();
1877 	if (evsel_list == NULL)
1878 		return -ENOMEM;
1879 
1880 	parse_events__shrink_config_terms();
1881 
1882 	/* String-parsing callback-based options would segfault when negated */
1883 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
1884 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
1885 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
1886 
1887 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1888 					(const char **) stat_usage,
1889 					PARSE_OPT_STOP_AT_NON_OPTION);
1890 	perf_stat__collect_metric_expr(evsel_list);
1891 	perf_stat__init_shadow_stats();
1892 
1893 	if (stat_config.csv_sep) {
1894 		stat_config.csv_output = true;
1895 		if (!strcmp(stat_config.csv_sep, "\\t"))
1896 			stat_config.csv_sep = "\t";
1897 	} else
1898 		stat_config.csv_sep = DEFAULT_SEPARATOR;
1899 
1900 	if (argc && !strncmp(argv[0], "rec", 3)) {
1901 		argc = __cmd_record(argc, argv);
1902 		if (argc < 0)
1903 			return -1;
1904 	} else if (argc && !strncmp(argv[0], "rep", 3))
1905 		return __cmd_report(argc, argv);
1906 
1907 	interval = stat_config.interval;
1908 	timeout = stat_config.timeout;
1909 
1910 	/*
1911 	 * For record command the -o is already taken care of.
1912 	 */
1913 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1914 		output = NULL;
1915 
1916 	if (output_name && output_fd) {
1917 		fprintf(stderr, "cannot use both --output and --log-fd\n");
1918 		parse_options_usage(stat_usage, stat_options, "o", 1);
1919 		parse_options_usage(NULL, stat_options, "log-fd", 0);
1920 		goto out;
1921 	}
1922 
1923 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
1924 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
1925 		goto out;
1926 	}
1927 
1928 	if (stat_config.metric_only && stat_config.run_count > 1) {
1929 		fprintf(stderr, "--metric-only is not supported with -r\n");
1930 		goto out;
1931 	}
1932 
1933 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
1934 		fprintf(stderr, "--table is only supported with -r\n");
1935 		parse_options_usage(stat_usage, stat_options, "r", 1);
1936 		parse_options_usage(NULL, stat_options, "table", 0);
1937 		goto out;
1938 	}
1939 
1940 	if (output_fd < 0) {
1941 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
1942 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1943 		goto out;
1944 	}
1945 
1946 	if (!output) {
1947 		struct timespec tm;
1948 		mode = append_file ? "a" : "w";
1949 
1950 		output = fopen(output_name, mode);
1951 		if (!output) {
1952 			perror("failed to create output file");
1953 			return -1;
1954 		}
1955 		clock_gettime(CLOCK_REALTIME, &tm);
1956 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1957 	} else if (output_fd > 0) {
1958 		mode = append_file ? "a" : "w";
1959 		output = fdopen(output_fd, mode);
1960 		if (!output) {
1961 			perror("Failed opening logfd");
1962 			return -errno;
1963 		}
1964 	}
1965 
1966 	stat_config.output = output;
1967 
1968 	/*
1969 	 * let the spreadsheet do the pretty-printing
1970 	 */
1971 	if (stat_config.csv_output) {
1972 		/* User explicitly passed -B? */
1973 		if (big_num_opt == 1) {
1974 			fprintf(stderr, "-B option not supported with -x\n");
1975 			parse_options_usage(stat_usage, stat_options, "B", 1);
1976 			parse_options_usage(NULL, stat_options, "x", 1);
1977 			goto out;
1978 		} else /* Nope, so disable big number formatting */
1979 			stat_config.big_num = false;
1980 	} else if (big_num_opt == 0) /* User passed --no-big-num */
1981 		stat_config.big_num = false;
1982 
1983 	setup_system_wide(argc);
1984 
1985 	/*
1986 	 * Display user/system times only for single
1987 	 * run and when there's specified tracee.
1988 	 */
1989 	if ((stat_config.run_count == 1) && target__none(&target))
1990 		stat_config.ru_display = true;
1991 
1992 	if (stat_config.run_count < 0) {
1993 		pr_err("Run count must be a positive number\n");
1994 		parse_options_usage(stat_usage, stat_options, "r", 1);
1995 		goto out;
1996 	} else if (stat_config.run_count == 0) {
1997 		forever = true;
1998 		stat_config.run_count = 1;
1999 	}
2000 
2001 	if (stat_config.walltime_run_table) {
2002 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2003 		if (!stat_config.walltime_run) {
2004 			pr_err("failed to setup -r option");
2005 			goto out;
2006 		}
2007 	}
2008 
2009 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
2010 		!target__has_task(&target)) {
2011 		if (!target.system_wide || target.cpu_list) {
2012 			fprintf(stderr, "The --per-thread option is only "
2013 				"available when monitoring via -p -t -a "
2014 				"options or only --per-thread.\n");
2015 			parse_options_usage(NULL, stat_options, "p", 1);
2016 			parse_options_usage(NULL, stat_options, "t", 1);
2017 			goto out;
2018 		}
2019 	}
2020 
2021 	/*
2022 	 * no_aggr, cgroup are for system-wide only
2023 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2024 	 */
2025 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2026 	      stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2027 	    !target__has_cpu(&target)) {
2028 		fprintf(stderr, "both cgroup and no-aggregation "
2029 			"modes only available in system-wide mode\n");
2030 
2031 		parse_options_usage(stat_usage, stat_options, "G", 1);
2032 		parse_options_usage(NULL, stat_options, "A", 1);
2033 		parse_options_usage(NULL, stat_options, "a", 1);
2034 		goto out;
2035 	}
2036 
2037 	if (add_default_attributes())
2038 		goto out;
2039 
2040 	target__validate(&target);
2041 
2042 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2043 		target.per_thread = true;
2044 
2045 	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2046 		if (target__has_task(&target)) {
2047 			pr_err("Problems finding threads of monitor\n");
2048 			parse_options_usage(stat_usage, stat_options, "p", 1);
2049 			parse_options_usage(NULL, stat_options, "t", 1);
2050 		} else if (target__has_cpu(&target)) {
2051 			perror("failed to parse CPUs map");
2052 			parse_options_usage(stat_usage, stat_options, "C", 1);
2053 			parse_options_usage(NULL, stat_options, "a", 1);
2054 		}
2055 		goto out;
2056 	}
2057 
2058 	/*
2059 	 * Initialize thread_map with comm names,
2060 	 * so we could print it out on output.
2061 	 */
2062 	if (stat_config.aggr_mode == AGGR_THREAD) {
2063 		thread_map__read_comms(evsel_list->core.threads);
2064 		if (target.system_wide) {
2065 			if (runtime_stat_new(&stat_config,
2066 				perf_thread_map__nr(evsel_list->core.threads))) {
2067 				goto out;
2068 			}
2069 		}
2070 	}
2071 
2072 	if (stat_config.aggr_mode == AGGR_NODE)
2073 		cpu__setup_cpunode_map();
2074 
2075 	if (stat_config.times && interval)
2076 		interval_count = true;
2077 	else if (stat_config.times && !interval) {
2078 		pr_err("interval-count option should be used together with "
2079 				"interval-print.\n");
2080 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2081 		parse_options_usage(stat_usage, stat_options, "I", 1);
2082 		goto out;
2083 	}
2084 
2085 	if (timeout && timeout < 100) {
2086 		if (timeout < 10) {
2087 			pr_err("timeout must be >= 10ms.\n");
2088 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
2089 			goto out;
2090 		} else
2091 			pr_warning("timeout < 100ms. "
2092 				   "The overhead percentage could be high in some cases. "
2093 				   "Please proceed with caution.\n");
2094 	}
2095 	if (timeout && interval) {
2096 		pr_err("timeout option is not supported with interval-print.\n");
2097 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
2098 		parse_options_usage(stat_usage, stat_options, "I", 1);
2099 		goto out;
2100 	}
2101 
2102 	if (perf_evlist__alloc_stats(evsel_list, interval))
2103 		goto out;
2104 
2105 	if (perf_stat_init_aggr_mode())
2106 		goto out;
2107 
2108 	/*
2109 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2110 	 * while avoiding that older tools show confusing messages.
2111 	 *
2112 	 * However for pipe sessions we need to keep it zero,
2113 	 * because script's perf_evsel__check_attr is triggered
2114 	 * by attr->sample_type != 0, and we can't run it on
2115 	 * stat sessions.
2116 	 */
2117 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2118 
2119 	/*
2120 	 * We dont want to block the signals - that would cause
2121 	 * child tasks to inherit that and Ctrl-C would not work.
2122 	 * What we want is for Ctrl-C to work in the exec()-ed
2123 	 * task, but being ignored by perf stat itself:
2124 	 */
2125 	atexit(sig_atexit);
2126 	if (!forever)
2127 		signal(SIGINT,  skip_signal);
2128 	signal(SIGCHLD, skip_signal);
2129 	signal(SIGALRM, skip_signal);
2130 	signal(SIGABRT, skip_signal);
2131 
2132 	status = 0;
2133 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2134 		if (stat_config.run_count != 1 && verbose > 0)
2135 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2136 				run_idx + 1);
2137 
2138 		if (run_idx != 0)
2139 			perf_evlist__reset_prev_raw_counts(evsel_list);
2140 
2141 		status = run_perf_stat(argc, argv, run_idx);
2142 		if (forever && status != -1 && !interval) {
2143 			print_counters(NULL, argc, argv);
2144 			perf_stat__reset_stats();
2145 		}
2146 	}
2147 
2148 	if (!forever && status != -1 && !interval)
2149 		print_counters(NULL, argc, argv);
2150 
2151 	if (STAT_RECORD) {
2152 		/*
2153 		 * We synthesize the kernel mmap record just so that older tools
2154 		 * don't emit warnings about not being able to resolve symbols
2155 		 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2156 		 * a saner message about no samples being in the perf.data file.
2157 		 *
2158 		 * This also serves to suppress a warning about f_header.data.size == 0
2159 		 * in header.c at the moment 'perf stat record' gets introduced, which
2160 		 * is not really needed once we start adding the stat specific PERF_RECORD_
2161 		 * records, but the need to suppress the kptr_restrict messages in older
2162 		 * tools remain  -acme
2163 		 */
2164 		int fd = perf_data__fd(&perf_stat.data);
2165 		int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2166 							     process_synthesized_event,
2167 							     &perf_stat.session->machines.host);
2168 		if (err) {
2169 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2170 				   "older tools may produce warnings about this file\n.");
2171 		}
2172 
2173 		if (!interval) {
2174 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2175 				pr_err("failed to write stat round event\n");
2176 		}
2177 
2178 		if (!perf_stat.data.is_pipe) {
2179 			perf_stat.session->header.data_size += perf_stat.bytes_written;
2180 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2181 		}
2182 
2183 		evlist__close(evsel_list);
2184 		perf_session__delete(perf_stat.session);
2185 	}
2186 
2187 	perf_stat__exit_aggr_mode();
2188 	perf_evlist__free_stats(evsel_list);
2189 out:
2190 	zfree(&stat_config.walltime_run);
2191 
2192 	if (smi_cost && smi_reset)
2193 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2194 
2195 	evlist__delete(evsel_list);
2196 
2197 	runtime_stat_delete(&stat_config);
2198 
2199 	return status;
2200 }
2201