xref: /linux/tools/perf/builtin-stat.c (revision b0f84a84fff180718995b1269da2988e5b28be42)
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 "perf.h"
44 #include "builtin.h"
45 #include "util/cgroup.h"
46 #include "util/util.h"
47 #include <subcmd/parse-options.h>
48 #include "util/parse-events.h"
49 #include "util/pmu.h"
50 #include "util/event.h"
51 #include "util/evlist.h"
52 #include "util/evsel.h"
53 #include "util/debug.h"
54 #include "util/color.h"
55 #include "util/stat.h"
56 #include "util/header.h"
57 #include "util/cpumap.h"
58 #include "util/thread.h"
59 #include "util/thread_map.h"
60 #include "util/counts.h"
61 #include "util/group.h"
62 #include "util/session.h"
63 #include "util/tool.h"
64 #include "util/string2.h"
65 #include "util/metricgroup.h"
66 #include "util/top.h"
67 #include "asm/bug.h"
68 
69 #include <linux/time64.h>
70 #include <api/fs/fs.h>
71 #include <errno.h>
72 #include <signal.h>
73 #include <stdlib.h>
74 #include <sys/prctl.h>
75 #include <inttypes.h>
76 #include <locale.h>
77 #include <math.h>
78 #include <sys/types.h>
79 #include <sys/stat.h>
80 #include <sys/wait.h>
81 #include <unistd.h>
82 #include <sys/time.h>
83 #include <sys/resource.h>
84 
85 #include "sane_ctype.h"
86 
87 #define DEFAULT_SEPARATOR	" "
88 #define FREEZE_ON_SMI_PATH	"devices/cpu/freeze_on_smi"
89 
90 static void print_counters(struct timespec *ts, int argc, const char **argv);
91 
92 /* Default events used for perf stat -T */
93 static const char *transaction_attrs = {
94 	"task-clock,"
95 	"{"
96 	"instructions,"
97 	"cycles,"
98 	"cpu/cycles-t/,"
99 	"cpu/tx-start/,"
100 	"cpu/el-start/,"
101 	"cpu/cycles-ct/"
102 	"}"
103 };
104 
105 /* More limited version when the CPU does not have all events. */
106 static const char * transaction_limited_attrs = {
107 	"task-clock,"
108 	"{"
109 	"instructions,"
110 	"cycles,"
111 	"cpu/cycles-t/,"
112 	"cpu/tx-start/"
113 	"}"
114 };
115 
116 static const char * topdown_attrs[] = {
117 	"topdown-total-slots",
118 	"topdown-slots-retired",
119 	"topdown-recovery-bubbles",
120 	"topdown-fetch-bubbles",
121 	"topdown-slots-issued",
122 	NULL,
123 };
124 
125 static const char *smi_cost_attrs = {
126 	"{"
127 	"msr/aperf/,"
128 	"msr/smi/,"
129 	"cycles"
130 	"}"
131 };
132 
133 static struct perf_evlist	*evsel_list;
134 
135 static struct target target = {
136 	.uid	= UINT_MAX,
137 };
138 
139 #define METRIC_ONLY_LEN 20
140 
141 static volatile pid_t		child_pid			= -1;
142 static int			detailed_run			=  0;
143 static bool			transaction_run;
144 static bool			topdown_run			= false;
145 static bool			smi_cost			= false;
146 static bool			smi_reset			= false;
147 static int			big_num_opt			=  -1;
148 static bool			group				= false;
149 static const char		*pre_cmd			= NULL;
150 static const char		*post_cmd			= NULL;
151 static bool			sync_run			= false;
152 static bool			forever				= false;
153 static bool			force_metric_only		= false;
154 static struct timespec		ref_time;
155 static bool			append_file;
156 static bool			interval_count;
157 static const char		*output_name;
158 static int			output_fd;
159 
160 struct perf_stat {
161 	bool			 record;
162 	struct perf_data	 data;
163 	struct perf_session	*session;
164 	u64			 bytes_written;
165 	struct perf_tool	 tool;
166 	bool			 maps_allocated;
167 	struct cpu_map		*cpus;
168 	struct thread_map	*threads;
169 	enum aggr_mode		 aggr_mode;
170 };
171 
172 static struct perf_stat		perf_stat;
173 #define STAT_RECORD		perf_stat.record
174 
175 static volatile int done = 0;
176 
177 static struct perf_stat_config stat_config = {
178 	.aggr_mode		= AGGR_GLOBAL,
179 	.scale			= true,
180 	.unit_width		= 4, /* strlen("unit") */
181 	.run_count		= 1,
182 	.metric_only_len	= METRIC_ONLY_LEN,
183 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
184 	.big_num		= true,
185 };
186 
187 static inline void diff_timespec(struct timespec *r, struct timespec *a,
188 				 struct timespec *b)
189 {
190 	r->tv_sec = a->tv_sec - b->tv_sec;
191 	if (a->tv_nsec < b->tv_nsec) {
192 		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
193 		r->tv_sec--;
194 	} else {
195 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
196 	}
197 }
198 
199 static void perf_stat__reset_stats(void)
200 {
201 	int i;
202 
203 	perf_evlist__reset_stats(evsel_list);
204 	perf_stat__reset_shadow_stats();
205 
206 	for (i = 0; i < stat_config.stats_num; i++)
207 		perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
208 }
209 
210 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
211 				     union perf_event *event,
212 				     struct perf_sample *sample __maybe_unused,
213 				     struct machine *machine __maybe_unused)
214 {
215 	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
216 		pr_err("failed to write perf data, error: %m\n");
217 		return -1;
218 	}
219 
220 	perf_stat.bytes_written += event->header.size;
221 	return 0;
222 }
223 
224 static int write_stat_round_event(u64 tm, u64 type)
225 {
226 	return perf_event__synthesize_stat_round(NULL, tm, type,
227 						 process_synthesized_event,
228 						 NULL);
229 }
230 
231 #define WRITE_STAT_ROUND_EVENT(time, interval) \
232 	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
233 
234 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
235 
236 static int
237 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
238 			     struct perf_counts_values *count)
239 {
240 	struct perf_sample_id *sid = SID(counter, cpu, thread);
241 
242 	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
243 					   process_synthesized_event, NULL);
244 }
245 
246 static int read_single_counter(struct perf_evsel *counter, int cpu,
247 			       int thread, struct timespec *rs)
248 {
249 	if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
250 		u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
251 		struct perf_counts_values *count =
252 			perf_counts(counter->counts, cpu, thread);
253 		count->ena = count->run = val;
254 		count->val = val;
255 		return 0;
256 	}
257 	return perf_evsel__read_counter(counter, cpu, thread);
258 }
259 
260 /*
261  * Read out the results of a single counter:
262  * do not aggregate counts across CPUs in system-wide mode
263  */
264 static int read_counter(struct perf_evsel *counter, struct timespec *rs)
265 {
266 	int nthreads = thread_map__nr(evsel_list->threads);
267 	int ncpus, cpu, thread;
268 
269 	if (target__has_cpu(&target) && !target__has_per_thread(&target))
270 		ncpus = perf_evsel__nr_cpus(counter);
271 	else
272 		ncpus = 1;
273 
274 	if (!counter->supported)
275 		return -ENOENT;
276 
277 	if (counter->system_wide)
278 		nthreads = 1;
279 
280 	for (thread = 0; thread < nthreads; thread++) {
281 		for (cpu = 0; cpu < ncpus; cpu++) {
282 			struct perf_counts_values *count;
283 
284 			count = perf_counts(counter->counts, cpu, thread);
285 
286 			/*
287 			 * The leader's group read loads data into its group members
288 			 * (via perf_evsel__read_counter) and sets threir count->loaded.
289 			 */
290 			if (!count->loaded &&
291 			    read_single_counter(counter, cpu, thread, rs)) {
292 				counter->counts->scaled = -1;
293 				perf_counts(counter->counts, cpu, thread)->ena = 0;
294 				perf_counts(counter->counts, cpu, thread)->run = 0;
295 				return -1;
296 			}
297 
298 			count->loaded = false;
299 
300 			if (STAT_RECORD) {
301 				if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
302 					pr_err("failed to write stat event\n");
303 					return -1;
304 				}
305 			}
306 
307 			if (verbose > 1) {
308 				fprintf(stat_config.output,
309 					"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
310 						perf_evsel__name(counter),
311 						cpu,
312 						count->val, count->ena, count->run);
313 			}
314 		}
315 	}
316 
317 	return 0;
318 }
319 
320 static void read_counters(struct timespec *rs)
321 {
322 	struct perf_evsel *counter;
323 	int ret;
324 
325 	evlist__for_each_entry(evsel_list, counter) {
326 		ret = read_counter(counter, rs);
327 		if (ret)
328 			pr_debug("failed to read counter %s\n", counter->name);
329 
330 		if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
331 			pr_warning("failed to process counter %s\n", counter->name);
332 	}
333 }
334 
335 static void process_interval(void)
336 {
337 	struct timespec ts, rs;
338 
339 	clock_gettime(CLOCK_MONOTONIC, &ts);
340 	diff_timespec(&rs, &ts, &ref_time);
341 
342 	read_counters(&rs);
343 
344 	if (STAT_RECORD) {
345 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
346 			pr_err("failed to write stat round event\n");
347 	}
348 
349 	init_stats(&walltime_nsecs_stats);
350 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
351 	print_counters(&rs, 0, NULL);
352 }
353 
354 static void enable_counters(void)
355 {
356 	if (stat_config.initial_delay)
357 		usleep(stat_config.initial_delay * USEC_PER_MSEC);
358 
359 	/*
360 	 * We need to enable counters only if:
361 	 * - we don't have tracee (attaching to task or cpu)
362 	 * - we have initial delay configured
363 	 */
364 	if (!target__none(&target) || stat_config.initial_delay)
365 		perf_evlist__enable(evsel_list);
366 }
367 
368 static void disable_counters(void)
369 {
370 	/*
371 	 * If we don't have tracee (attaching to task or cpu), counters may
372 	 * still be running. To get accurate group ratios, we must stop groups
373 	 * from counting before reading their constituent counters.
374 	 */
375 	if (!target__none(&target))
376 		perf_evlist__disable(evsel_list);
377 }
378 
379 static volatile int workload_exec_errno;
380 
381 /*
382  * perf_evlist__prepare_workload will send a SIGUSR1
383  * if the fork fails, since we asked by setting its
384  * want_signal to true.
385  */
386 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
387 					void *ucontext __maybe_unused)
388 {
389 	workload_exec_errno = info->si_value.sival_int;
390 }
391 
392 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
393 {
394 	return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
395 }
396 
397 static bool is_target_alive(struct target *_target,
398 			    struct thread_map *threads)
399 {
400 	struct stat st;
401 	int i;
402 
403 	if (!target__has_task(_target))
404 		return true;
405 
406 	for (i = 0; i < threads->nr; i++) {
407 		char path[PATH_MAX];
408 
409 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
410 			  threads->map[i].pid);
411 
412 		if (!stat(path, &st))
413 			return true;
414 	}
415 
416 	return false;
417 }
418 
419 static int __run_perf_stat(int argc, const char **argv, int run_idx)
420 {
421 	int interval = stat_config.interval;
422 	int times = stat_config.times;
423 	int timeout = stat_config.timeout;
424 	char msg[BUFSIZ];
425 	unsigned long long t0, t1;
426 	struct perf_evsel *counter;
427 	struct timespec ts;
428 	size_t l;
429 	int status = 0;
430 	const bool forks = (argc > 0);
431 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
432 
433 	if (interval) {
434 		ts.tv_sec  = interval / USEC_PER_MSEC;
435 		ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
436 	} else if (timeout) {
437 		ts.tv_sec  = timeout / USEC_PER_MSEC;
438 		ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
439 	} else {
440 		ts.tv_sec  = 1;
441 		ts.tv_nsec = 0;
442 	}
443 
444 	if (forks) {
445 		if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
446 						  workload_exec_failed_signal) < 0) {
447 			perror("failed to prepare workload");
448 			return -1;
449 		}
450 		child_pid = evsel_list->workload.pid;
451 	}
452 
453 	if (group)
454 		perf_evlist__set_leader(evsel_list);
455 
456 	evlist__for_each_entry(evsel_list, counter) {
457 try_again:
458 		if (create_perf_stat_counter(counter, &stat_config, &target) < 0) {
459 
460 			/* Weak group failed. Reset the group. */
461 			if ((errno == EINVAL || errno == EBADF) &&
462 			    counter->leader != counter &&
463 			    counter->weak_group) {
464 				counter = perf_evlist__reset_weak_group(evsel_list, counter);
465 				goto try_again;
466 			}
467 
468 			/*
469 			 * PPC returns ENXIO for HW counters until 2.6.37
470 			 * (behavior changed with commit b0a873e).
471 			 */
472 			if (errno == EINVAL || errno == ENOSYS ||
473 			    errno == ENOENT || errno == EOPNOTSUPP ||
474 			    errno == ENXIO) {
475 				if (verbose > 0)
476 					ui__warning("%s event is not supported by the kernel.\n",
477 						    perf_evsel__name(counter));
478 				counter->supported = false;
479 
480 				if ((counter->leader != counter) ||
481 				    !(counter->leader->nr_members > 1))
482 					continue;
483 			} else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
484                                 if (verbose > 0)
485                                         ui__warning("%s\n", msg);
486                                 goto try_again;
487 			} else if (target__has_per_thread(&target) &&
488 				   evsel_list->threads &&
489 				   evsel_list->threads->err_thread != -1) {
490 				/*
491 				 * For global --per-thread case, skip current
492 				 * error thread.
493 				 */
494 				if (!thread_map__remove(evsel_list->threads,
495 							evsel_list->threads->err_thread)) {
496 					evsel_list->threads->err_thread = -1;
497 					goto try_again;
498 				}
499 			}
500 
501 			perf_evsel__open_strerror(counter, &target,
502 						  errno, msg, sizeof(msg));
503 			ui__error("%s\n", msg);
504 
505 			if (child_pid != -1)
506 				kill(child_pid, SIGTERM);
507 
508 			return -1;
509 		}
510 		counter->supported = true;
511 
512 		l = strlen(counter->unit);
513 		if (l > stat_config.unit_width)
514 			stat_config.unit_width = l;
515 
516 		if (perf_evsel__should_store_id(counter) &&
517 		    perf_evsel__store_ids(counter, evsel_list))
518 			return -1;
519 	}
520 
521 	if (perf_evlist__apply_filters(evsel_list, &counter)) {
522 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
523 			counter->filter, perf_evsel__name(counter), errno,
524 			str_error_r(errno, msg, sizeof(msg)));
525 		return -1;
526 	}
527 
528 	if (STAT_RECORD) {
529 		int err, fd = perf_data__fd(&perf_stat.data);
530 
531 		if (is_pipe) {
532 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
533 		} else {
534 			err = perf_session__write_header(perf_stat.session, evsel_list,
535 							 fd, false);
536 		}
537 
538 		if (err < 0)
539 			return err;
540 
541 		err = perf_stat_synthesize_config(&stat_config, NULL, evsel_list,
542 						  process_synthesized_event, is_pipe);
543 		if (err < 0)
544 			return err;
545 	}
546 
547 	/*
548 	 * Enable counters and exec the command:
549 	 */
550 	t0 = rdclock();
551 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
552 
553 	if (forks) {
554 		perf_evlist__start_workload(evsel_list);
555 		enable_counters();
556 
557 		if (interval || timeout) {
558 			while (!waitpid(child_pid, &status, WNOHANG)) {
559 				nanosleep(&ts, NULL);
560 				if (timeout)
561 					break;
562 				process_interval();
563 				if (interval_count && !(--times))
564 					break;
565 			}
566 		}
567 		if (child_pid != -1)
568 			wait4(child_pid, &status, 0, &stat_config.ru_data);
569 
570 		if (workload_exec_errno) {
571 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
572 			pr_err("Workload failed: %s\n", emsg);
573 			return -1;
574 		}
575 
576 		if (WIFSIGNALED(status))
577 			psignal(WTERMSIG(status), argv[0]);
578 	} else {
579 		enable_counters();
580 		while (!done) {
581 			nanosleep(&ts, NULL);
582 			if (!is_target_alive(&target, evsel_list->threads))
583 				break;
584 			if (timeout)
585 				break;
586 			if (interval) {
587 				process_interval();
588 				if (interval_count && !(--times))
589 					break;
590 			}
591 		}
592 	}
593 
594 	disable_counters();
595 
596 	t1 = rdclock();
597 
598 	if (stat_config.walltime_run_table)
599 		stat_config.walltime_run[run_idx] = t1 - t0;
600 
601 	update_stats(&walltime_nsecs_stats, t1 - t0);
602 
603 	/*
604 	 * Closing a group leader splits the group, and as we only disable
605 	 * group leaders, results in remaining events becoming enabled. To
606 	 * avoid arbitrary skew, we must read all counters before closing any
607 	 * group leaders.
608 	 */
609 	read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
610 	perf_evlist__close(evsel_list);
611 
612 	return WEXITSTATUS(status);
613 }
614 
615 static int run_perf_stat(int argc, const char **argv, int run_idx)
616 {
617 	int ret;
618 
619 	if (pre_cmd) {
620 		ret = system(pre_cmd);
621 		if (ret)
622 			return ret;
623 	}
624 
625 	if (sync_run)
626 		sync();
627 
628 	ret = __run_perf_stat(argc, argv, run_idx);
629 	if (ret)
630 		return ret;
631 
632 	if (post_cmd) {
633 		ret = system(post_cmd);
634 		if (ret)
635 			return ret;
636 	}
637 
638 	return ret;
639 }
640 
641 static void print_counters(struct timespec *ts, int argc, const char **argv)
642 {
643 	/* Do not print anything if we record to the pipe. */
644 	if (STAT_RECORD && perf_stat.data.is_pipe)
645 		return;
646 
647 	perf_evlist__print_counters(evsel_list, &stat_config, &target,
648 				    ts, argc, argv);
649 }
650 
651 static volatile int signr = -1;
652 
653 static void skip_signal(int signo)
654 {
655 	if ((child_pid == -1) || stat_config.interval)
656 		done = 1;
657 
658 	signr = signo;
659 	/*
660 	 * render child_pid harmless
661 	 * won't send SIGTERM to a random
662 	 * process in case of race condition
663 	 * and fast PID recycling
664 	 */
665 	child_pid = -1;
666 }
667 
668 static void sig_atexit(void)
669 {
670 	sigset_t set, oset;
671 
672 	/*
673 	 * avoid race condition with SIGCHLD handler
674 	 * in skip_signal() which is modifying child_pid
675 	 * goal is to avoid send SIGTERM to a random
676 	 * process
677 	 */
678 	sigemptyset(&set);
679 	sigaddset(&set, SIGCHLD);
680 	sigprocmask(SIG_BLOCK, &set, &oset);
681 
682 	if (child_pid != -1)
683 		kill(child_pid, SIGTERM);
684 
685 	sigprocmask(SIG_SETMASK, &oset, NULL);
686 
687 	if (signr == -1)
688 		return;
689 
690 	signal(signr, SIG_DFL);
691 	kill(getpid(), signr);
692 }
693 
694 static int stat__set_big_num(const struct option *opt __maybe_unused,
695 			     const char *s __maybe_unused, int unset)
696 {
697 	big_num_opt = unset ? 0 : 1;
698 	return 0;
699 }
700 
701 static int enable_metric_only(const struct option *opt __maybe_unused,
702 			      const char *s __maybe_unused, int unset)
703 {
704 	force_metric_only = true;
705 	stat_config.metric_only = !unset;
706 	return 0;
707 }
708 
709 static int parse_metric_groups(const struct option *opt,
710 			       const char *str,
711 			       int unset __maybe_unused)
712 {
713 	return metricgroup__parse_groups(opt, str, &stat_config.metric_events);
714 }
715 
716 static struct option stat_options[] = {
717 	OPT_BOOLEAN('T', "transaction", &transaction_run,
718 		    "hardware transaction statistics"),
719 	OPT_CALLBACK('e', "event", &evsel_list, "event",
720 		     "event selector. use 'perf list' to list available events",
721 		     parse_events_option),
722 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
723 		     "event filter", parse_filter),
724 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
725 		    "child tasks do not inherit counters"),
726 	OPT_STRING('p', "pid", &target.pid, "pid",
727 		   "stat events on existing process id"),
728 	OPT_STRING('t', "tid", &target.tid, "tid",
729 		   "stat events on existing thread id"),
730 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
731 		    "system-wide collection from all CPUs"),
732 	OPT_BOOLEAN('g', "group", &group,
733 		    "put the counters into a counter group"),
734 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
735 		    "Use --no-scale to disable counter scaling for multiplexing"),
736 	OPT_INCR('v', "verbose", &verbose,
737 		    "be more verbose (show counter open errors, etc)"),
738 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
739 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
740 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
741 		    "display details about each run (only with -r option)"),
742 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
743 		    "null run - dont start any counters"),
744 	OPT_INCR('d', "detailed", &detailed_run,
745 		    "detailed run - start a lot of events"),
746 	OPT_BOOLEAN('S', "sync", &sync_run,
747 		    "call sync() before starting a run"),
748 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
749 			   "print large numbers with thousands\' separators",
750 			   stat__set_big_num),
751 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
752 		    "list of cpus to monitor in system-wide"),
753 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
754 		    "disable CPU count aggregation", AGGR_NONE),
755 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
756 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
757 		   "print counts with custom separator"),
758 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
759 		     "monitor event in cgroup name only", parse_cgroups),
760 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
761 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
762 	OPT_INTEGER(0, "log-fd", &output_fd,
763 		    "log output to fd, instead of stderr"),
764 	OPT_STRING(0, "pre", &pre_cmd, "command",
765 			"command to run prior to the measured command"),
766 	OPT_STRING(0, "post", &post_cmd, "command",
767 			"command to run after to the measured command"),
768 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
769 		    "print counts at regular interval in ms "
770 		    "(overhead is possible for values <= 100ms)"),
771 	OPT_INTEGER(0, "interval-count", &stat_config.times,
772 		    "print counts for fixed number of times"),
773 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
774 		    "clear screen in between new interval"),
775 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
776 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
777 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
778 		     "aggregate counts per processor socket", AGGR_SOCKET),
779 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
780 		     "aggregate counts per physical processor core", AGGR_CORE),
781 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
782 		     "aggregate counts per thread", AGGR_THREAD),
783 	OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
784 		     "ms to wait before starting measurement after program start"),
785 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
786 			"Only print computed metrics. No raw values", enable_metric_only),
787 	OPT_BOOLEAN(0, "topdown", &topdown_run,
788 			"measure topdown level 1 statistics"),
789 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
790 			"measure SMI cost"),
791 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
792 		     "monitor specified metrics or metric groups (separated by ,)",
793 		     parse_metric_groups),
794 	OPT_END()
795 };
796 
797 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
798 				 struct cpu_map *map, int cpu)
799 {
800 	return cpu_map__get_socket(map, cpu, NULL);
801 }
802 
803 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
804 			       struct cpu_map *map, int cpu)
805 {
806 	return cpu_map__get_core(map, cpu, NULL);
807 }
808 
809 static int cpu_map__get_max(struct cpu_map *map)
810 {
811 	int i, max = -1;
812 
813 	for (i = 0; i < map->nr; i++) {
814 		if (map->map[i] > max)
815 			max = map->map[i];
816 	}
817 
818 	return max;
819 }
820 
821 static int perf_stat__get_aggr(struct perf_stat_config *config,
822 			       aggr_get_id_t get_id, struct cpu_map *map, int idx)
823 {
824 	int cpu;
825 
826 	if (idx >= map->nr)
827 		return -1;
828 
829 	cpu = map->map[idx];
830 
831 	if (config->cpus_aggr_map->map[cpu] == -1)
832 		config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
833 
834 	return config->cpus_aggr_map->map[cpu];
835 }
836 
837 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
838 					struct cpu_map *map, int idx)
839 {
840 	return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
841 }
842 
843 static int perf_stat__get_core_cached(struct perf_stat_config *config,
844 				      struct cpu_map *map, int idx)
845 {
846 	return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
847 }
848 
849 static bool term_percore_set(void)
850 {
851 	struct perf_evsel *counter;
852 
853 	evlist__for_each_entry(evsel_list, counter) {
854 		if (counter->percore)
855 			return true;
856 	}
857 
858 	return false;
859 }
860 
861 static int perf_stat_init_aggr_mode(void)
862 {
863 	int nr;
864 
865 	switch (stat_config.aggr_mode) {
866 	case AGGR_SOCKET:
867 		if (cpu_map__build_socket_map(evsel_list->cpus, &stat_config.aggr_map)) {
868 			perror("cannot build socket map");
869 			return -1;
870 		}
871 		stat_config.aggr_get_id = perf_stat__get_socket_cached;
872 		break;
873 	case AGGR_CORE:
874 		if (cpu_map__build_core_map(evsel_list->cpus, &stat_config.aggr_map)) {
875 			perror("cannot build core map");
876 			return -1;
877 		}
878 		stat_config.aggr_get_id = perf_stat__get_core_cached;
879 		break;
880 	case AGGR_NONE:
881 		if (term_percore_set()) {
882 			if (cpu_map__build_core_map(evsel_list->cpus,
883 						    &stat_config.aggr_map)) {
884 				perror("cannot build core map");
885 				return -1;
886 			}
887 			stat_config.aggr_get_id = perf_stat__get_core_cached;
888 		}
889 		break;
890 	case AGGR_GLOBAL:
891 	case AGGR_THREAD:
892 	case AGGR_UNSET:
893 	default:
894 		break;
895 	}
896 
897 	/*
898 	 * The evsel_list->cpus is the base we operate on,
899 	 * taking the highest cpu number to be the size of
900 	 * the aggregation translate cpumap.
901 	 */
902 	nr = cpu_map__get_max(evsel_list->cpus);
903 	stat_config.cpus_aggr_map = cpu_map__empty_new(nr + 1);
904 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
905 }
906 
907 static void perf_stat__exit_aggr_mode(void)
908 {
909 	cpu_map__put(stat_config.aggr_map);
910 	cpu_map__put(stat_config.cpus_aggr_map);
911 	stat_config.aggr_map = NULL;
912 	stat_config.cpus_aggr_map = NULL;
913 }
914 
915 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
916 {
917 	int cpu;
918 
919 	if (idx > map->nr)
920 		return -1;
921 
922 	cpu = map->map[idx];
923 
924 	if (cpu >= env->nr_cpus_avail)
925 		return -1;
926 
927 	return cpu;
928 }
929 
930 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
931 {
932 	struct perf_env *env = data;
933 	int cpu = perf_env__get_cpu(env, map, idx);
934 
935 	return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
936 }
937 
938 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
939 {
940 	struct perf_env *env = data;
941 	int core = -1, cpu = perf_env__get_cpu(env, map, idx);
942 
943 	if (cpu != -1) {
944 		int socket_id = env->cpu[cpu].socket_id;
945 
946 		/*
947 		 * Encode socket in upper 16 bits
948 		 * core_id is relative to socket, and
949 		 * we need a global id. So we combine
950 		 * socket + core id.
951 		 */
952 		core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
953 	}
954 
955 	return core;
956 }
957 
958 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
959 				      struct cpu_map **sockp)
960 {
961 	return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
962 }
963 
964 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
965 				    struct cpu_map **corep)
966 {
967 	return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
968 }
969 
970 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
971 				      struct cpu_map *map, int idx)
972 {
973 	return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
974 }
975 
976 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
977 				    struct cpu_map *map, int idx)
978 {
979 	return perf_env__get_core(map, idx, &perf_stat.session->header.env);
980 }
981 
982 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
983 {
984 	struct perf_env *env = &st->session->header.env;
985 
986 	switch (stat_config.aggr_mode) {
987 	case AGGR_SOCKET:
988 		if (perf_env__build_socket_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
989 			perror("cannot build socket map");
990 			return -1;
991 		}
992 		stat_config.aggr_get_id = perf_stat__get_socket_file;
993 		break;
994 	case AGGR_CORE:
995 		if (perf_env__build_core_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
996 			perror("cannot build core map");
997 			return -1;
998 		}
999 		stat_config.aggr_get_id = perf_stat__get_core_file;
1000 		break;
1001 	case AGGR_NONE:
1002 	case AGGR_GLOBAL:
1003 	case AGGR_THREAD:
1004 	case AGGR_UNSET:
1005 	default:
1006 		break;
1007 	}
1008 
1009 	return 0;
1010 }
1011 
1012 static int topdown_filter_events(const char **attr, char **str, bool use_group)
1013 {
1014 	int off = 0;
1015 	int i;
1016 	int len = 0;
1017 	char *s;
1018 
1019 	for (i = 0; attr[i]; i++) {
1020 		if (pmu_have_event("cpu", attr[i])) {
1021 			len += strlen(attr[i]) + 1;
1022 			attr[i - off] = attr[i];
1023 		} else
1024 			off++;
1025 	}
1026 	attr[i - off] = NULL;
1027 
1028 	*str = malloc(len + 1 + 2);
1029 	if (!*str)
1030 		return -1;
1031 	s = *str;
1032 	if (i - off == 0) {
1033 		*s = 0;
1034 		return 0;
1035 	}
1036 	if (use_group)
1037 		*s++ = '{';
1038 	for (i = 0; attr[i]; i++) {
1039 		strcpy(s, attr[i]);
1040 		s += strlen(s);
1041 		*s++ = ',';
1042 	}
1043 	if (use_group) {
1044 		s[-1] = '}';
1045 		*s = 0;
1046 	} else
1047 		s[-1] = 0;
1048 	return 0;
1049 }
1050 
1051 __weak bool arch_topdown_check_group(bool *warn)
1052 {
1053 	*warn = false;
1054 	return false;
1055 }
1056 
1057 __weak void arch_topdown_group_warn(void)
1058 {
1059 }
1060 
1061 /*
1062  * Add default attributes, if there were no attributes specified or
1063  * if -d/--detailed, -d -d or -d -d -d is used:
1064  */
1065 static int add_default_attributes(void)
1066 {
1067 	int err;
1068 	struct perf_event_attr default_attrs0[] = {
1069 
1070   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1071   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1072   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1073   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1074 
1075   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1076 };
1077 	struct perf_event_attr frontend_attrs[] = {
1078   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1079 };
1080 	struct perf_event_attr backend_attrs[] = {
1081   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1082 };
1083 	struct perf_event_attr default_attrs1[] = {
1084   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1085   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1086   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1087 
1088 };
1089 
1090 /*
1091  * Detailed stats (-d), covering the L1 and last level data caches:
1092  */
1093 	struct perf_event_attr detailed_attrs[] = {
1094 
1095   { .type = PERF_TYPE_HW_CACHE,
1096     .config =
1097 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1098 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1099 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1100 
1101   { .type = PERF_TYPE_HW_CACHE,
1102     .config =
1103 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1104 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1105 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1106 
1107   { .type = PERF_TYPE_HW_CACHE,
1108     .config =
1109 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1110 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1111 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1112 
1113   { .type = PERF_TYPE_HW_CACHE,
1114     .config =
1115 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1116 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1117 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1118 };
1119 
1120 /*
1121  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1122  */
1123 	struct perf_event_attr very_detailed_attrs[] = {
1124 
1125   { .type = PERF_TYPE_HW_CACHE,
1126     .config =
1127 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1128 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1129 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1130 
1131   { .type = PERF_TYPE_HW_CACHE,
1132     .config =
1133 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1134 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1135 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1136 
1137   { .type = PERF_TYPE_HW_CACHE,
1138     .config =
1139 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1140 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1141 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1142 
1143   { .type = PERF_TYPE_HW_CACHE,
1144     .config =
1145 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1146 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1147 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1148 
1149   { .type = PERF_TYPE_HW_CACHE,
1150     .config =
1151 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1152 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1153 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1154 
1155   { .type = PERF_TYPE_HW_CACHE,
1156     .config =
1157 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1158 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1159 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1160 
1161 };
1162 
1163 /*
1164  * Very, very detailed stats (-d -d -d), adding prefetch events:
1165  */
1166 	struct perf_event_attr very_very_detailed_attrs[] = {
1167 
1168   { .type = PERF_TYPE_HW_CACHE,
1169     .config =
1170 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1171 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1172 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1173 
1174   { .type = PERF_TYPE_HW_CACHE,
1175     .config =
1176 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1177 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1178 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1179 };
1180 	struct parse_events_error errinfo;
1181 
1182 	/* Set attrs if no event is selected and !null_run: */
1183 	if (stat_config.null_run)
1184 		return 0;
1185 
1186 	if (transaction_run) {
1187 		/* Handle -T as -M transaction. Once platform specific metrics
1188 		 * support has been added to the json files, all archictures
1189 		 * will use this approach. To determine transaction support
1190 		 * on an architecture test for such a metric name.
1191 		 */
1192 		if (metricgroup__has_metric("transaction")) {
1193 			struct option opt = { .value = &evsel_list };
1194 
1195 			return metricgroup__parse_groups(&opt, "transaction",
1196 							 &stat_config.metric_events);
1197 		}
1198 
1199 		if (pmu_have_event("cpu", "cycles-ct") &&
1200 		    pmu_have_event("cpu", "el-start"))
1201 			err = parse_events(evsel_list, transaction_attrs,
1202 					   &errinfo);
1203 		else
1204 			err = parse_events(evsel_list,
1205 					   transaction_limited_attrs,
1206 					   &errinfo);
1207 		if (err) {
1208 			fprintf(stderr, "Cannot set up transaction events\n");
1209 			parse_events_print_error(&errinfo, transaction_attrs);
1210 			return -1;
1211 		}
1212 		return 0;
1213 	}
1214 
1215 	if (smi_cost) {
1216 		int smi;
1217 
1218 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1219 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1220 			return -1;
1221 		}
1222 
1223 		if (!smi) {
1224 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1225 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1226 				return -1;
1227 			}
1228 			smi_reset = true;
1229 		}
1230 
1231 		if (pmu_have_event("msr", "aperf") &&
1232 		    pmu_have_event("msr", "smi")) {
1233 			if (!force_metric_only)
1234 				stat_config.metric_only = true;
1235 			err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1236 		} else {
1237 			fprintf(stderr, "To measure SMI cost, it needs "
1238 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1239 			parse_events_print_error(&errinfo, smi_cost_attrs);
1240 			return -1;
1241 		}
1242 		if (err) {
1243 			fprintf(stderr, "Cannot set up SMI cost events\n");
1244 			return -1;
1245 		}
1246 		return 0;
1247 	}
1248 
1249 	if (topdown_run) {
1250 		char *str = NULL;
1251 		bool warn = false;
1252 
1253 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1254 		    stat_config.aggr_mode != AGGR_CORE) {
1255 			pr_err("top down event configuration requires --per-core mode\n");
1256 			return -1;
1257 		}
1258 		stat_config.aggr_mode = AGGR_CORE;
1259 		if (nr_cgroups || !target__has_cpu(&target)) {
1260 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1261 			return -1;
1262 		}
1263 
1264 		if (!force_metric_only)
1265 			stat_config.metric_only = true;
1266 		if (topdown_filter_events(topdown_attrs, &str,
1267 				arch_topdown_check_group(&warn)) < 0) {
1268 			pr_err("Out of memory\n");
1269 			return -1;
1270 		}
1271 		if (topdown_attrs[0] && str) {
1272 			if (warn)
1273 				arch_topdown_group_warn();
1274 			err = parse_events(evsel_list, str, &errinfo);
1275 			if (err) {
1276 				fprintf(stderr,
1277 					"Cannot set up top down events %s: %d\n",
1278 					str, err);
1279 				free(str);
1280 				parse_events_print_error(&errinfo, str);
1281 				return -1;
1282 			}
1283 		} else {
1284 			fprintf(stderr, "System does not support topdown\n");
1285 			return -1;
1286 		}
1287 		free(str);
1288 	}
1289 
1290 	if (!evsel_list->nr_entries) {
1291 		if (target__has_cpu(&target))
1292 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1293 
1294 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1295 			return -1;
1296 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1297 			if (perf_evlist__add_default_attrs(evsel_list,
1298 						frontend_attrs) < 0)
1299 				return -1;
1300 		}
1301 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1302 			if (perf_evlist__add_default_attrs(evsel_list,
1303 						backend_attrs) < 0)
1304 				return -1;
1305 		}
1306 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1307 			return -1;
1308 	}
1309 
1310 	/* Detailed events get appended to the event list: */
1311 
1312 	if (detailed_run <  1)
1313 		return 0;
1314 
1315 	/* Append detailed run extra attributes: */
1316 	if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1317 		return -1;
1318 
1319 	if (detailed_run < 2)
1320 		return 0;
1321 
1322 	/* Append very detailed run extra attributes: */
1323 	if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1324 		return -1;
1325 
1326 	if (detailed_run < 3)
1327 		return 0;
1328 
1329 	/* Append very, very detailed run extra attributes: */
1330 	return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1331 }
1332 
1333 static const char * const stat_record_usage[] = {
1334 	"perf stat record [<options>]",
1335 	NULL,
1336 };
1337 
1338 static void init_features(struct perf_session *session)
1339 {
1340 	int feat;
1341 
1342 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1343 		perf_header__set_feat(&session->header, feat);
1344 
1345 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1346 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1347 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1348 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1349 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1350 }
1351 
1352 static int __cmd_record(int argc, const char **argv)
1353 {
1354 	struct perf_session *session;
1355 	struct perf_data *data = &perf_stat.data;
1356 
1357 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
1358 			     PARSE_OPT_STOP_AT_NON_OPTION);
1359 
1360 	if (output_name)
1361 		data->path = output_name;
1362 
1363 	if (stat_config.run_count != 1 || forever) {
1364 		pr_err("Cannot use -r option with perf stat record.\n");
1365 		return -1;
1366 	}
1367 
1368 	session = perf_session__new(data, false, NULL);
1369 	if (session == NULL) {
1370 		pr_err("Perf session creation failed.\n");
1371 		return -1;
1372 	}
1373 
1374 	init_features(session);
1375 
1376 	session->evlist   = evsel_list;
1377 	perf_stat.session = session;
1378 	perf_stat.record  = true;
1379 	return argc;
1380 }
1381 
1382 static int process_stat_round_event(struct perf_session *session,
1383 				    union perf_event *event)
1384 {
1385 	struct stat_round_event *stat_round = &event->stat_round;
1386 	struct perf_evsel *counter;
1387 	struct timespec tsh, *ts = NULL;
1388 	const char **argv = session->header.env.cmdline_argv;
1389 	int argc = session->header.env.nr_cmdline;
1390 
1391 	evlist__for_each_entry(evsel_list, counter)
1392 		perf_stat_process_counter(&stat_config, counter);
1393 
1394 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1395 		update_stats(&walltime_nsecs_stats, stat_round->time);
1396 
1397 	if (stat_config.interval && stat_round->time) {
1398 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
1399 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1400 		ts = &tsh;
1401 	}
1402 
1403 	print_counters(ts, argc, argv);
1404 	return 0;
1405 }
1406 
1407 static
1408 int process_stat_config_event(struct perf_session *session,
1409 			      union perf_event *event)
1410 {
1411 	struct perf_tool *tool = session->tool;
1412 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1413 
1414 	perf_event__read_stat_config(&stat_config, &event->stat_config);
1415 
1416 	if (cpu_map__empty(st->cpus)) {
1417 		if (st->aggr_mode != AGGR_UNSET)
1418 			pr_warning("warning: processing task data, aggregation mode not set\n");
1419 		return 0;
1420 	}
1421 
1422 	if (st->aggr_mode != AGGR_UNSET)
1423 		stat_config.aggr_mode = st->aggr_mode;
1424 
1425 	if (perf_stat.data.is_pipe)
1426 		perf_stat_init_aggr_mode();
1427 	else
1428 		perf_stat_init_aggr_mode_file(st);
1429 
1430 	return 0;
1431 }
1432 
1433 static int set_maps(struct perf_stat *st)
1434 {
1435 	if (!st->cpus || !st->threads)
1436 		return 0;
1437 
1438 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1439 		return -EINVAL;
1440 
1441 	perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
1442 
1443 	if (perf_evlist__alloc_stats(evsel_list, true))
1444 		return -ENOMEM;
1445 
1446 	st->maps_allocated = true;
1447 	return 0;
1448 }
1449 
1450 static
1451 int process_thread_map_event(struct perf_session *session,
1452 			     union perf_event *event)
1453 {
1454 	struct perf_tool *tool = session->tool;
1455 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1456 
1457 	if (st->threads) {
1458 		pr_warning("Extra thread map event, ignoring.\n");
1459 		return 0;
1460 	}
1461 
1462 	st->threads = thread_map__new_event(&event->thread_map);
1463 	if (!st->threads)
1464 		return -ENOMEM;
1465 
1466 	return set_maps(st);
1467 }
1468 
1469 static
1470 int process_cpu_map_event(struct perf_session *session,
1471 			  union perf_event *event)
1472 {
1473 	struct perf_tool *tool = session->tool;
1474 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1475 	struct cpu_map *cpus;
1476 
1477 	if (st->cpus) {
1478 		pr_warning("Extra cpu map event, ignoring.\n");
1479 		return 0;
1480 	}
1481 
1482 	cpus = cpu_map__new_data(&event->cpu_map.data);
1483 	if (!cpus)
1484 		return -ENOMEM;
1485 
1486 	st->cpus = cpus;
1487 	return set_maps(st);
1488 }
1489 
1490 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
1491 {
1492 	int i;
1493 
1494 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
1495 	if (!config->stats)
1496 		return -1;
1497 
1498 	config->stats_num = nthreads;
1499 
1500 	for (i = 0; i < nthreads; i++)
1501 		runtime_stat__init(&config->stats[i]);
1502 
1503 	return 0;
1504 }
1505 
1506 static void runtime_stat_delete(struct perf_stat_config *config)
1507 {
1508 	int i;
1509 
1510 	if (!config->stats)
1511 		return;
1512 
1513 	for (i = 0; i < config->stats_num; i++)
1514 		runtime_stat__exit(&config->stats[i]);
1515 
1516 	free(config->stats);
1517 }
1518 
1519 static const char * const stat_report_usage[] = {
1520 	"perf stat report [<options>]",
1521 	NULL,
1522 };
1523 
1524 static struct perf_stat perf_stat = {
1525 	.tool = {
1526 		.attr		= perf_event__process_attr,
1527 		.event_update	= perf_event__process_event_update,
1528 		.thread_map	= process_thread_map_event,
1529 		.cpu_map	= process_cpu_map_event,
1530 		.stat_config	= process_stat_config_event,
1531 		.stat		= perf_event__process_stat_event,
1532 		.stat_round	= process_stat_round_event,
1533 	},
1534 	.aggr_mode = AGGR_UNSET,
1535 };
1536 
1537 static int __cmd_report(int argc, const char **argv)
1538 {
1539 	struct perf_session *session;
1540 	const struct option options[] = {
1541 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1542 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1543 		     "aggregate counts per processor socket", AGGR_SOCKET),
1544 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1545 		     "aggregate counts per physical processor core", AGGR_CORE),
1546 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1547 		     "disable CPU count aggregation", AGGR_NONE),
1548 	OPT_END()
1549 	};
1550 	struct stat st;
1551 	int ret;
1552 
1553 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
1554 
1555 	if (!input_name || !strlen(input_name)) {
1556 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1557 			input_name = "-";
1558 		else
1559 			input_name = "perf.data";
1560 	}
1561 
1562 	perf_stat.data.path = input_name;
1563 	perf_stat.data.mode = PERF_DATA_MODE_READ;
1564 
1565 	session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
1566 	if (session == NULL)
1567 		return -1;
1568 
1569 	perf_stat.session  = session;
1570 	stat_config.output = stderr;
1571 	evsel_list         = session->evlist;
1572 
1573 	ret = perf_session__process_events(session);
1574 	if (ret)
1575 		return ret;
1576 
1577 	perf_session__delete(session);
1578 	return 0;
1579 }
1580 
1581 static void setup_system_wide(int forks)
1582 {
1583 	/*
1584 	 * Make system wide (-a) the default target if
1585 	 * no target was specified and one of following
1586 	 * conditions is met:
1587 	 *
1588 	 *   - there's no workload specified
1589 	 *   - there is workload specified but all requested
1590 	 *     events are system wide events
1591 	 */
1592 	if (!target__none(&target))
1593 		return;
1594 
1595 	if (!forks)
1596 		target.system_wide = true;
1597 	else {
1598 		struct perf_evsel *counter;
1599 
1600 		evlist__for_each_entry(evsel_list, counter) {
1601 			if (!counter->system_wide)
1602 				return;
1603 		}
1604 
1605 		if (evsel_list->nr_entries)
1606 			target.system_wide = true;
1607 	}
1608 }
1609 
1610 int cmd_stat(int argc, const char **argv)
1611 {
1612 	const char * const stat_usage[] = {
1613 		"perf stat [<options>] [<command>]",
1614 		NULL
1615 	};
1616 	int status = -EINVAL, run_idx;
1617 	const char *mode;
1618 	FILE *output = stderr;
1619 	unsigned int interval, timeout;
1620 	const char * const stat_subcommands[] = { "record", "report" };
1621 
1622 	setlocale(LC_ALL, "");
1623 
1624 	evsel_list = perf_evlist__new();
1625 	if (evsel_list == NULL)
1626 		return -ENOMEM;
1627 
1628 	parse_events__shrink_config_terms();
1629 
1630 	/* String-parsing callback-based options would segfault when negated */
1631 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
1632 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
1633 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
1634 
1635 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1636 					(const char **) stat_usage,
1637 					PARSE_OPT_STOP_AT_NON_OPTION);
1638 	perf_stat__collect_metric_expr(evsel_list);
1639 	perf_stat__init_shadow_stats();
1640 
1641 	if (stat_config.csv_sep) {
1642 		stat_config.csv_output = true;
1643 		if (!strcmp(stat_config.csv_sep, "\\t"))
1644 			stat_config.csv_sep = "\t";
1645 	} else
1646 		stat_config.csv_sep = DEFAULT_SEPARATOR;
1647 
1648 	if (argc && !strncmp(argv[0], "rec", 3)) {
1649 		argc = __cmd_record(argc, argv);
1650 		if (argc < 0)
1651 			return -1;
1652 	} else if (argc && !strncmp(argv[0], "rep", 3))
1653 		return __cmd_report(argc, argv);
1654 
1655 	interval = stat_config.interval;
1656 	timeout = stat_config.timeout;
1657 
1658 	/*
1659 	 * For record command the -o is already taken care of.
1660 	 */
1661 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1662 		output = NULL;
1663 
1664 	if (output_name && output_fd) {
1665 		fprintf(stderr, "cannot use both --output and --log-fd\n");
1666 		parse_options_usage(stat_usage, stat_options, "o", 1);
1667 		parse_options_usage(NULL, stat_options, "log-fd", 0);
1668 		goto out;
1669 	}
1670 
1671 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
1672 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
1673 		goto out;
1674 	}
1675 
1676 	if (stat_config.metric_only && stat_config.run_count > 1) {
1677 		fprintf(stderr, "--metric-only is not supported with -r\n");
1678 		goto out;
1679 	}
1680 
1681 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
1682 		fprintf(stderr, "--table is only supported with -r\n");
1683 		parse_options_usage(stat_usage, stat_options, "r", 1);
1684 		parse_options_usage(NULL, stat_options, "table", 0);
1685 		goto out;
1686 	}
1687 
1688 	if (output_fd < 0) {
1689 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
1690 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1691 		goto out;
1692 	}
1693 
1694 	if (!output) {
1695 		struct timespec tm;
1696 		mode = append_file ? "a" : "w";
1697 
1698 		output = fopen(output_name, mode);
1699 		if (!output) {
1700 			perror("failed to create output file");
1701 			return -1;
1702 		}
1703 		clock_gettime(CLOCK_REALTIME, &tm);
1704 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1705 	} else if (output_fd > 0) {
1706 		mode = append_file ? "a" : "w";
1707 		output = fdopen(output_fd, mode);
1708 		if (!output) {
1709 			perror("Failed opening logfd");
1710 			return -errno;
1711 		}
1712 	}
1713 
1714 	stat_config.output = output;
1715 
1716 	/*
1717 	 * let the spreadsheet do the pretty-printing
1718 	 */
1719 	if (stat_config.csv_output) {
1720 		/* User explicitly passed -B? */
1721 		if (big_num_opt == 1) {
1722 			fprintf(stderr, "-B option not supported with -x\n");
1723 			parse_options_usage(stat_usage, stat_options, "B", 1);
1724 			parse_options_usage(NULL, stat_options, "x", 1);
1725 			goto out;
1726 		} else /* Nope, so disable big number formatting */
1727 			stat_config.big_num = false;
1728 	} else if (big_num_opt == 0) /* User passed --no-big-num */
1729 		stat_config.big_num = false;
1730 
1731 	setup_system_wide(argc);
1732 
1733 	/*
1734 	 * Display user/system times only for single
1735 	 * run and when there's specified tracee.
1736 	 */
1737 	if ((stat_config.run_count == 1) && target__none(&target))
1738 		stat_config.ru_display = true;
1739 
1740 	if (stat_config.run_count < 0) {
1741 		pr_err("Run count must be a positive number\n");
1742 		parse_options_usage(stat_usage, stat_options, "r", 1);
1743 		goto out;
1744 	} else if (stat_config.run_count == 0) {
1745 		forever = true;
1746 		stat_config.run_count = 1;
1747 	}
1748 
1749 	if (stat_config.walltime_run_table) {
1750 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
1751 		if (!stat_config.walltime_run) {
1752 			pr_err("failed to setup -r option");
1753 			goto out;
1754 		}
1755 	}
1756 
1757 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
1758 		!target__has_task(&target)) {
1759 		if (!target.system_wide || target.cpu_list) {
1760 			fprintf(stderr, "The --per-thread option is only "
1761 				"available when monitoring via -p -t -a "
1762 				"options or only --per-thread.\n");
1763 			parse_options_usage(NULL, stat_options, "p", 1);
1764 			parse_options_usage(NULL, stat_options, "t", 1);
1765 			goto out;
1766 		}
1767 	}
1768 
1769 	/*
1770 	 * no_aggr, cgroup are for system-wide only
1771 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1772 	 */
1773 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1774 	      stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1775 	    !target__has_cpu(&target)) {
1776 		fprintf(stderr, "both cgroup and no-aggregation "
1777 			"modes only available in system-wide mode\n");
1778 
1779 		parse_options_usage(stat_usage, stat_options, "G", 1);
1780 		parse_options_usage(NULL, stat_options, "A", 1);
1781 		parse_options_usage(NULL, stat_options, "a", 1);
1782 		goto out;
1783 	}
1784 
1785 	if (add_default_attributes())
1786 		goto out;
1787 
1788 	target__validate(&target);
1789 
1790 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
1791 		target.per_thread = true;
1792 
1793 	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1794 		if (target__has_task(&target)) {
1795 			pr_err("Problems finding threads of monitor\n");
1796 			parse_options_usage(stat_usage, stat_options, "p", 1);
1797 			parse_options_usage(NULL, stat_options, "t", 1);
1798 		} else if (target__has_cpu(&target)) {
1799 			perror("failed to parse CPUs map");
1800 			parse_options_usage(stat_usage, stat_options, "C", 1);
1801 			parse_options_usage(NULL, stat_options, "a", 1);
1802 		}
1803 		goto out;
1804 	}
1805 
1806 	/*
1807 	 * Initialize thread_map with comm names,
1808 	 * so we could print it out on output.
1809 	 */
1810 	if (stat_config.aggr_mode == AGGR_THREAD) {
1811 		thread_map__read_comms(evsel_list->threads);
1812 		if (target.system_wide) {
1813 			if (runtime_stat_new(&stat_config,
1814 				thread_map__nr(evsel_list->threads))) {
1815 				goto out;
1816 			}
1817 		}
1818 	}
1819 
1820 	if (stat_config.times && interval)
1821 		interval_count = true;
1822 	else if (stat_config.times && !interval) {
1823 		pr_err("interval-count option should be used together with "
1824 				"interval-print.\n");
1825 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
1826 		parse_options_usage(stat_usage, stat_options, "I", 1);
1827 		goto out;
1828 	}
1829 
1830 	if (timeout && timeout < 100) {
1831 		if (timeout < 10) {
1832 			pr_err("timeout must be >= 10ms.\n");
1833 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
1834 			goto out;
1835 		} else
1836 			pr_warning("timeout < 100ms. "
1837 				   "The overhead percentage could be high in some cases. "
1838 				   "Please proceed with caution.\n");
1839 	}
1840 	if (timeout && interval) {
1841 		pr_err("timeout option is not supported with interval-print.\n");
1842 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
1843 		parse_options_usage(stat_usage, stat_options, "I", 1);
1844 		goto out;
1845 	}
1846 
1847 	if (perf_evlist__alloc_stats(evsel_list, interval))
1848 		goto out;
1849 
1850 	if (perf_stat_init_aggr_mode())
1851 		goto out;
1852 
1853 	/*
1854 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
1855 	 * while avoiding that older tools show confusing messages.
1856 	 *
1857 	 * However for pipe sessions we need to keep it zero,
1858 	 * because script's perf_evsel__check_attr is triggered
1859 	 * by attr->sample_type != 0, and we can't run it on
1860 	 * stat sessions.
1861 	 */
1862 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
1863 
1864 	/*
1865 	 * We dont want to block the signals - that would cause
1866 	 * child tasks to inherit that and Ctrl-C would not work.
1867 	 * What we want is for Ctrl-C to work in the exec()-ed
1868 	 * task, but being ignored by perf stat itself:
1869 	 */
1870 	atexit(sig_atexit);
1871 	if (!forever)
1872 		signal(SIGINT,  skip_signal);
1873 	signal(SIGCHLD, skip_signal);
1874 	signal(SIGALRM, skip_signal);
1875 	signal(SIGABRT, skip_signal);
1876 
1877 	status = 0;
1878 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
1879 		if (stat_config.run_count != 1 && verbose > 0)
1880 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1881 				run_idx + 1);
1882 
1883 		status = run_perf_stat(argc, argv, run_idx);
1884 		if (forever && status != -1) {
1885 			print_counters(NULL, argc, argv);
1886 			perf_stat__reset_stats();
1887 		}
1888 	}
1889 
1890 	if (!forever && status != -1 && !interval)
1891 		print_counters(NULL, argc, argv);
1892 
1893 	if (STAT_RECORD) {
1894 		/*
1895 		 * We synthesize the kernel mmap record just so that older tools
1896 		 * don't emit warnings about not being able to resolve symbols
1897 		 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
1898 		 * a saner message about no samples being in the perf.data file.
1899 		 *
1900 		 * This also serves to suppress a warning about f_header.data.size == 0
1901 		 * in header.c at the moment 'perf stat record' gets introduced, which
1902 		 * is not really needed once we start adding the stat specific PERF_RECORD_
1903 		 * records, but the need to suppress the kptr_restrict messages in older
1904 		 * tools remain  -acme
1905 		 */
1906 		int fd = perf_data__fd(&perf_stat.data);
1907 		int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
1908 							     process_synthesized_event,
1909 							     &perf_stat.session->machines.host);
1910 		if (err) {
1911 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
1912 				   "older tools may produce warnings about this file\n.");
1913 		}
1914 
1915 		if (!interval) {
1916 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
1917 				pr_err("failed to write stat round event\n");
1918 		}
1919 
1920 		if (!perf_stat.data.is_pipe) {
1921 			perf_stat.session->header.data_size += perf_stat.bytes_written;
1922 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
1923 		}
1924 
1925 		perf_session__delete(perf_stat.session);
1926 	}
1927 
1928 	perf_stat__exit_aggr_mode();
1929 	perf_evlist__free_stats(evsel_list);
1930 out:
1931 	free(stat_config.walltime_run);
1932 
1933 	if (smi_cost && smi_reset)
1934 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
1935 
1936 	perf_evlist__delete(evsel_list);
1937 
1938 	runtime_stat_delete(&stat_config);
1939 
1940 	return status;
1941 }
1942