xref: /linux/tools/perf/builtin-stat.c (revision 18f90d372cf35b387663f1567de701e5393f6eb5)
1 /*
2  * builtin-stat.c
3  *
4  * Builtin stat command: Give a precise performance counters summary
5  * overview about any workload, CPU or specific PID.
6  *
7  * Sample output:
8 
9    $ perf stat ./hackbench 10
10 
11   Time: 0.118
12 
13   Performance counter stats for './hackbench 10':
14 
15        1708.761321 task-clock                #   11.037 CPUs utilized
16             41,190 context-switches          #    0.024 M/sec
17              6,735 CPU-migrations            #    0.004 M/sec
18             17,318 page-faults               #    0.010 M/sec
19      5,205,202,243 cycles                    #    3.046 GHz
20      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
21      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
22      2,603,501,247 instructions              #    0.50  insns per cycle
23                                              #    1.48  stalled cycles per insn
24        484,357,498 branches                  #  283.455 M/sec
25          6,388,934 branch-misses             #    1.32% of all branches
26 
27         0.154822978  seconds time elapsed
28 
29  *
30  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
31  *
32  * Improvements and fixes by:
33  *
34  *   Arjan van de Ven <arjan@linux.intel.com>
35  *   Yanmin Zhang <yanmin.zhang@intel.com>
36  *   Wu Fengguang <fengguang.wu@intel.com>
37  *   Mike Galbraith <efault@gmx.de>
38  *   Paul Mackerras <paulus@samba.org>
39  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
40  *
41  * Released under the GPL v2. (and only v2, not any later version)
42  */
43 
44 #include "perf.h"
45 #include "builtin.h"
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include <subcmd/parse-options.h>
49 #include "util/parse-events.h"
50 #include "util/pmu.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/color.h"
56 #include "util/stat.h"
57 #include "util/header.h"
58 #include "util/cpumap.h"
59 #include "util/thread.h"
60 #include "util/thread_map.h"
61 #include "util/counts.h"
62 #include "util/group.h"
63 #include "util/session.h"
64 #include "util/tool.h"
65 #include "util/string2.h"
66 #include "util/metricgroup.h"
67 #include "util/top.h"
68 #include "asm/bug.h"
69 
70 #include <linux/time64.h>
71 #include <api/fs/fs.h>
72 #include <errno.h>
73 #include <signal.h>
74 #include <stdlib.h>
75 #include <sys/prctl.h>
76 #include <inttypes.h>
77 #include <locale.h>
78 #include <math.h>
79 #include <sys/types.h>
80 #include <sys/stat.h>
81 #include <sys/wait.h>
82 #include <unistd.h>
83 #include <sys/time.h>
84 #include <sys/resource.h>
85 
86 #include "sane_ctype.h"
87 
88 #define DEFAULT_SEPARATOR	" "
89 #define FREEZE_ON_SMI_PATH	"devices/cpu/freeze_on_smi"
90 
91 static void print_counters(struct timespec *ts, int argc, const char **argv);
92 
93 /* Default events used for perf stat -T */
94 static const char *transaction_attrs = {
95 	"task-clock,"
96 	"{"
97 	"instructions,"
98 	"cycles,"
99 	"cpu/cycles-t/,"
100 	"cpu/tx-start/,"
101 	"cpu/el-start/,"
102 	"cpu/cycles-ct/"
103 	"}"
104 };
105 
106 /* More limited version when the CPU does not have all events. */
107 static const char * transaction_limited_attrs = {
108 	"task-clock,"
109 	"{"
110 	"instructions,"
111 	"cycles,"
112 	"cpu/cycles-t/,"
113 	"cpu/tx-start/"
114 	"}"
115 };
116 
117 static const char * topdown_attrs[] = {
118 	"topdown-total-slots",
119 	"topdown-slots-retired",
120 	"topdown-recovery-bubbles",
121 	"topdown-fetch-bubbles",
122 	"topdown-slots-issued",
123 	NULL,
124 };
125 
126 static const char *smi_cost_attrs = {
127 	"{"
128 	"msr/aperf/,"
129 	"msr/smi/,"
130 	"cycles"
131 	"}"
132 };
133 
134 static struct perf_evlist	*evsel_list;
135 
136 static struct target target = {
137 	.uid	= UINT_MAX,
138 };
139 
140 #define METRIC_ONLY_LEN 20
141 
142 static volatile pid_t		child_pid			= -1;
143 static int			detailed_run			=  0;
144 static bool			transaction_run;
145 static bool			topdown_run			= false;
146 static bool			smi_cost			= false;
147 static bool			smi_reset			= false;
148 static int			big_num_opt			=  -1;
149 static bool			group				= false;
150 static const char		*pre_cmd			= NULL;
151 static const char		*post_cmd			= NULL;
152 static bool			sync_run			= false;
153 static bool			forever				= false;
154 static bool			force_metric_only		= false;
155 static struct timespec		ref_time;
156 static bool			append_file;
157 static bool			interval_count;
158 static const char		*output_name;
159 static int			output_fd;
160 
161 struct perf_stat {
162 	bool			 record;
163 	struct perf_data	 data;
164 	struct perf_session	*session;
165 	u64			 bytes_written;
166 	struct perf_tool	 tool;
167 	bool			 maps_allocated;
168 	struct cpu_map		*cpus;
169 	struct thread_map	*threads;
170 	enum aggr_mode		 aggr_mode;
171 };
172 
173 static struct perf_stat		perf_stat;
174 #define STAT_RECORD		perf_stat.record
175 
176 static volatile int done = 0;
177 
178 static struct perf_stat_config stat_config = {
179 	.aggr_mode		= AGGR_GLOBAL,
180 	.scale			= true,
181 	.unit_width		= 4, /* strlen("unit") */
182 	.run_count		= 1,
183 	.metric_only_len	= METRIC_ONLY_LEN,
184 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
185 	.big_num		= true,
186 };
187 
188 static inline void diff_timespec(struct timespec *r, struct timespec *a,
189 				 struct timespec *b)
190 {
191 	r->tv_sec = a->tv_sec - b->tv_sec;
192 	if (a->tv_nsec < b->tv_nsec) {
193 		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
194 		r->tv_sec--;
195 	} else {
196 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
197 	}
198 }
199 
200 static void perf_stat__reset_stats(void)
201 {
202 	int i;
203 
204 	perf_evlist__reset_stats(evsel_list);
205 	perf_stat__reset_shadow_stats();
206 
207 	for (i = 0; i < stat_config.stats_num; i++)
208 		perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
209 }
210 
211 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
212 				     union perf_event *event,
213 				     struct perf_sample *sample __maybe_unused,
214 				     struct machine *machine __maybe_unused)
215 {
216 	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
217 		pr_err("failed to write perf data, error: %m\n");
218 		return -1;
219 	}
220 
221 	perf_stat.bytes_written += event->header.size;
222 	return 0;
223 }
224 
225 static int write_stat_round_event(u64 tm, u64 type)
226 {
227 	return perf_event__synthesize_stat_round(NULL, tm, type,
228 						 process_synthesized_event,
229 						 NULL);
230 }
231 
232 #define WRITE_STAT_ROUND_EVENT(time, interval) \
233 	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
234 
235 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
236 
237 static int
238 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
239 			     struct perf_counts_values *count)
240 {
241 	struct perf_sample_id *sid = SID(counter, cpu, thread);
242 
243 	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
244 					   process_synthesized_event, NULL);
245 }
246 
247 /*
248  * Read out the results of a single counter:
249  * do not aggregate counts across CPUs in system-wide mode
250  */
251 static int read_counter(struct perf_evsel *counter)
252 {
253 	int nthreads = thread_map__nr(evsel_list->threads);
254 	int ncpus, cpu, thread;
255 
256 	if (target__has_cpu(&target) && !target__has_per_thread(&target))
257 		ncpus = perf_evsel__nr_cpus(counter);
258 	else
259 		ncpus = 1;
260 
261 	if (!counter->supported)
262 		return -ENOENT;
263 
264 	if (counter->system_wide)
265 		nthreads = 1;
266 
267 	for (thread = 0; thread < nthreads; thread++) {
268 		for (cpu = 0; cpu < ncpus; cpu++) {
269 			struct perf_counts_values *count;
270 
271 			count = perf_counts(counter->counts, cpu, thread);
272 
273 			/*
274 			 * The leader's group read loads data into its group members
275 			 * (via perf_evsel__read_counter) and sets threir count->loaded.
276 			 */
277 			if (!count->loaded &&
278 			    perf_evsel__read_counter(counter, cpu, thread)) {
279 				counter->counts->scaled = -1;
280 				perf_counts(counter->counts, cpu, thread)->ena = 0;
281 				perf_counts(counter->counts, cpu, thread)->run = 0;
282 				return -1;
283 			}
284 
285 			count->loaded = false;
286 
287 			if (STAT_RECORD) {
288 				if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
289 					pr_err("failed to write stat event\n");
290 					return -1;
291 				}
292 			}
293 
294 			if (verbose > 1) {
295 				fprintf(stat_config.output,
296 					"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
297 						perf_evsel__name(counter),
298 						cpu,
299 						count->val, count->ena, count->run);
300 			}
301 		}
302 	}
303 
304 	return 0;
305 }
306 
307 static void read_counters(void)
308 {
309 	struct perf_evsel *counter;
310 	int ret;
311 
312 	evlist__for_each_entry(evsel_list, counter) {
313 		ret = read_counter(counter);
314 		if (ret)
315 			pr_debug("failed to read counter %s\n", counter->name);
316 
317 		if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
318 			pr_warning("failed to process counter %s\n", counter->name);
319 	}
320 }
321 
322 static void process_interval(void)
323 {
324 	struct timespec ts, rs;
325 
326 	read_counters();
327 
328 	clock_gettime(CLOCK_MONOTONIC, &ts);
329 	diff_timespec(&rs, &ts, &ref_time);
330 
331 	if (STAT_RECORD) {
332 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
333 			pr_err("failed to write stat round event\n");
334 	}
335 
336 	init_stats(&walltime_nsecs_stats);
337 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
338 	print_counters(&rs, 0, NULL);
339 }
340 
341 static void enable_counters(void)
342 {
343 	if (stat_config.initial_delay)
344 		usleep(stat_config.initial_delay * USEC_PER_MSEC);
345 
346 	/*
347 	 * We need to enable counters only if:
348 	 * - we don't have tracee (attaching to task or cpu)
349 	 * - we have initial delay configured
350 	 */
351 	if (!target__none(&target) || stat_config.initial_delay)
352 		perf_evlist__enable(evsel_list);
353 }
354 
355 static void disable_counters(void)
356 {
357 	/*
358 	 * If we don't have tracee (attaching to task or cpu), counters may
359 	 * still be running. To get accurate group ratios, we must stop groups
360 	 * from counting before reading their constituent counters.
361 	 */
362 	if (!target__none(&target))
363 		perf_evlist__disable(evsel_list);
364 }
365 
366 static volatile int workload_exec_errno;
367 
368 /*
369  * perf_evlist__prepare_workload will send a SIGUSR1
370  * if the fork fails, since we asked by setting its
371  * want_signal to true.
372  */
373 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
374 					void *ucontext __maybe_unused)
375 {
376 	workload_exec_errno = info->si_value.sival_int;
377 }
378 
379 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
380 {
381 	return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
382 }
383 
384 static bool is_target_alive(struct target *_target,
385 			    struct thread_map *threads)
386 {
387 	struct stat st;
388 	int i;
389 
390 	if (!target__has_task(_target))
391 		return true;
392 
393 	for (i = 0; i < threads->nr; i++) {
394 		char path[PATH_MAX];
395 
396 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
397 			  threads->map[i].pid);
398 
399 		if (!stat(path, &st))
400 			return true;
401 	}
402 
403 	return false;
404 }
405 
406 static int __run_perf_stat(int argc, const char **argv, int run_idx)
407 {
408 	int interval = stat_config.interval;
409 	int times = stat_config.times;
410 	int timeout = stat_config.timeout;
411 	char msg[BUFSIZ];
412 	unsigned long long t0, t1;
413 	struct perf_evsel *counter;
414 	struct timespec ts;
415 	size_t l;
416 	int status = 0;
417 	const bool forks = (argc > 0);
418 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
419 
420 	if (interval) {
421 		ts.tv_sec  = interval / USEC_PER_MSEC;
422 		ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
423 	} else if (timeout) {
424 		ts.tv_sec  = timeout / USEC_PER_MSEC;
425 		ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
426 	} else {
427 		ts.tv_sec  = 1;
428 		ts.tv_nsec = 0;
429 	}
430 
431 	if (forks) {
432 		if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
433 						  workload_exec_failed_signal) < 0) {
434 			perror("failed to prepare workload");
435 			return -1;
436 		}
437 		child_pid = evsel_list->workload.pid;
438 	}
439 
440 	if (group)
441 		perf_evlist__set_leader(evsel_list);
442 
443 	evlist__for_each_entry(evsel_list, counter) {
444 try_again:
445 		if (create_perf_stat_counter(counter, &stat_config, &target) < 0) {
446 
447 			/* Weak group failed. Reset the group. */
448 			if ((errno == EINVAL || errno == EBADF) &&
449 			    counter->leader != counter &&
450 			    counter->weak_group) {
451 				counter = perf_evlist__reset_weak_group(evsel_list, counter);
452 				goto try_again;
453 			}
454 
455 			/*
456 			 * PPC returns ENXIO for HW counters until 2.6.37
457 			 * (behavior changed with commit b0a873e).
458 			 */
459 			if (errno == EINVAL || errno == ENOSYS ||
460 			    errno == ENOENT || errno == EOPNOTSUPP ||
461 			    errno == ENXIO) {
462 				if (verbose > 0)
463 					ui__warning("%s event is not supported by the kernel.\n",
464 						    perf_evsel__name(counter));
465 				counter->supported = false;
466 
467 				if ((counter->leader != counter) ||
468 				    !(counter->leader->nr_members > 1))
469 					continue;
470 			} else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
471                                 if (verbose > 0)
472                                         ui__warning("%s\n", msg);
473                                 goto try_again;
474 			} else if (target__has_per_thread(&target) &&
475 				   evsel_list->threads &&
476 				   evsel_list->threads->err_thread != -1) {
477 				/*
478 				 * For global --per-thread case, skip current
479 				 * error thread.
480 				 */
481 				if (!thread_map__remove(evsel_list->threads,
482 							evsel_list->threads->err_thread)) {
483 					evsel_list->threads->err_thread = -1;
484 					goto try_again;
485 				}
486 			}
487 
488 			perf_evsel__open_strerror(counter, &target,
489 						  errno, msg, sizeof(msg));
490 			ui__error("%s\n", msg);
491 
492 			if (child_pid != -1)
493 				kill(child_pid, SIGTERM);
494 
495 			return -1;
496 		}
497 		counter->supported = true;
498 
499 		l = strlen(counter->unit);
500 		if (l > stat_config.unit_width)
501 			stat_config.unit_width = l;
502 
503 		if (perf_evsel__should_store_id(counter) &&
504 		    perf_evsel__store_ids(counter, evsel_list))
505 			return -1;
506 	}
507 
508 	if (perf_evlist__apply_filters(evsel_list, &counter)) {
509 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
510 			counter->filter, perf_evsel__name(counter), errno,
511 			str_error_r(errno, msg, sizeof(msg)));
512 		return -1;
513 	}
514 
515 	if (STAT_RECORD) {
516 		int err, fd = perf_data__fd(&perf_stat.data);
517 
518 		if (is_pipe) {
519 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
520 		} else {
521 			err = perf_session__write_header(perf_stat.session, evsel_list,
522 							 fd, false);
523 		}
524 
525 		if (err < 0)
526 			return err;
527 
528 		err = perf_stat_synthesize_config(&stat_config, NULL, evsel_list,
529 						  process_synthesized_event, is_pipe);
530 		if (err < 0)
531 			return err;
532 	}
533 
534 	/*
535 	 * Enable counters and exec the command:
536 	 */
537 	t0 = rdclock();
538 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
539 
540 	if (forks) {
541 		perf_evlist__start_workload(evsel_list);
542 		enable_counters();
543 
544 		if (interval || timeout) {
545 			while (!waitpid(child_pid, &status, WNOHANG)) {
546 				nanosleep(&ts, NULL);
547 				if (timeout)
548 					break;
549 				process_interval();
550 				if (interval_count && !(--times))
551 					break;
552 			}
553 		}
554 		if (child_pid != -1)
555 			wait4(child_pid, &status, 0, &stat_config.ru_data);
556 
557 		if (workload_exec_errno) {
558 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
559 			pr_err("Workload failed: %s\n", emsg);
560 			return -1;
561 		}
562 
563 		if (WIFSIGNALED(status))
564 			psignal(WTERMSIG(status), argv[0]);
565 	} else {
566 		enable_counters();
567 		while (!done) {
568 			nanosleep(&ts, NULL);
569 			if (!is_target_alive(&target, evsel_list->threads))
570 				break;
571 			if (timeout)
572 				break;
573 			if (interval) {
574 				process_interval();
575 				if (interval_count && !(--times))
576 					break;
577 			}
578 		}
579 	}
580 
581 	disable_counters();
582 
583 	t1 = rdclock();
584 
585 	if (stat_config.walltime_run_table)
586 		stat_config.walltime_run[run_idx] = t1 - t0;
587 
588 	update_stats(&walltime_nsecs_stats, t1 - t0);
589 
590 	/*
591 	 * Closing a group leader splits the group, and as we only disable
592 	 * group leaders, results in remaining events becoming enabled. To
593 	 * avoid arbitrary skew, we must read all counters before closing any
594 	 * group leaders.
595 	 */
596 	read_counters();
597 	perf_evlist__close(evsel_list);
598 
599 	return WEXITSTATUS(status);
600 }
601 
602 static int run_perf_stat(int argc, const char **argv, int run_idx)
603 {
604 	int ret;
605 
606 	if (pre_cmd) {
607 		ret = system(pre_cmd);
608 		if (ret)
609 			return ret;
610 	}
611 
612 	if (sync_run)
613 		sync();
614 
615 	ret = __run_perf_stat(argc, argv, run_idx);
616 	if (ret)
617 		return ret;
618 
619 	if (post_cmd) {
620 		ret = system(post_cmd);
621 		if (ret)
622 			return ret;
623 	}
624 
625 	return ret;
626 }
627 
628 static void print_counters(struct timespec *ts, int argc, const char **argv)
629 {
630 	/* Do not print anything if we record to the pipe. */
631 	if (STAT_RECORD && perf_stat.data.is_pipe)
632 		return;
633 
634 	perf_evlist__print_counters(evsel_list, &stat_config, &target,
635 				    ts, argc, argv);
636 }
637 
638 static volatile int signr = -1;
639 
640 static void skip_signal(int signo)
641 {
642 	if ((child_pid == -1) || stat_config.interval)
643 		done = 1;
644 
645 	signr = signo;
646 	/*
647 	 * render child_pid harmless
648 	 * won't send SIGTERM to a random
649 	 * process in case of race condition
650 	 * and fast PID recycling
651 	 */
652 	child_pid = -1;
653 }
654 
655 static void sig_atexit(void)
656 {
657 	sigset_t set, oset;
658 
659 	/*
660 	 * avoid race condition with SIGCHLD handler
661 	 * in skip_signal() which is modifying child_pid
662 	 * goal is to avoid send SIGTERM to a random
663 	 * process
664 	 */
665 	sigemptyset(&set);
666 	sigaddset(&set, SIGCHLD);
667 	sigprocmask(SIG_BLOCK, &set, &oset);
668 
669 	if (child_pid != -1)
670 		kill(child_pid, SIGTERM);
671 
672 	sigprocmask(SIG_SETMASK, &oset, NULL);
673 
674 	if (signr == -1)
675 		return;
676 
677 	signal(signr, SIG_DFL);
678 	kill(getpid(), signr);
679 }
680 
681 static int stat__set_big_num(const struct option *opt __maybe_unused,
682 			     const char *s __maybe_unused, int unset)
683 {
684 	big_num_opt = unset ? 0 : 1;
685 	return 0;
686 }
687 
688 static int enable_metric_only(const struct option *opt __maybe_unused,
689 			      const char *s __maybe_unused, int unset)
690 {
691 	force_metric_only = true;
692 	stat_config.metric_only = !unset;
693 	return 0;
694 }
695 
696 static int parse_metric_groups(const struct option *opt,
697 			       const char *str,
698 			       int unset __maybe_unused)
699 {
700 	return metricgroup__parse_groups(opt, str, &stat_config.metric_events);
701 }
702 
703 static struct option stat_options[] = {
704 	OPT_BOOLEAN('T', "transaction", &transaction_run,
705 		    "hardware transaction statistics"),
706 	OPT_CALLBACK('e', "event", &evsel_list, "event",
707 		     "event selector. use 'perf list' to list available events",
708 		     parse_events_option),
709 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
710 		     "event filter", parse_filter),
711 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
712 		    "child tasks do not inherit counters"),
713 	OPT_STRING('p', "pid", &target.pid, "pid",
714 		   "stat events on existing process id"),
715 	OPT_STRING('t', "tid", &target.tid, "tid",
716 		   "stat events on existing thread id"),
717 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
718 		    "system-wide collection from all CPUs"),
719 	OPT_BOOLEAN('g', "group", &group,
720 		    "put the counters into a counter group"),
721 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
722 		    "Use --no-scale to disable counter scaling for multiplexing"),
723 	OPT_INCR('v', "verbose", &verbose,
724 		    "be more verbose (show counter open errors, etc)"),
725 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
726 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
727 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
728 		    "display details about each run (only with -r option)"),
729 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
730 		    "null run - dont start any counters"),
731 	OPT_INCR('d', "detailed", &detailed_run,
732 		    "detailed run - start a lot of events"),
733 	OPT_BOOLEAN('S', "sync", &sync_run,
734 		    "call sync() before starting a run"),
735 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
736 			   "print large numbers with thousands\' separators",
737 			   stat__set_big_num),
738 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
739 		    "list of cpus to monitor in system-wide"),
740 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
741 		    "disable CPU count aggregation", AGGR_NONE),
742 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
743 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
744 		   "print counts with custom separator"),
745 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
746 		     "monitor event in cgroup name only", parse_cgroups),
747 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
748 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
749 	OPT_INTEGER(0, "log-fd", &output_fd,
750 		    "log output to fd, instead of stderr"),
751 	OPT_STRING(0, "pre", &pre_cmd, "command",
752 			"command to run prior to the measured command"),
753 	OPT_STRING(0, "post", &post_cmd, "command",
754 			"command to run after to the measured command"),
755 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
756 		    "print counts at regular interval in ms "
757 		    "(overhead is possible for values <= 100ms)"),
758 	OPT_INTEGER(0, "interval-count", &stat_config.times,
759 		    "print counts for fixed number of times"),
760 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
761 		    "clear screen in between new interval"),
762 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
763 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
764 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
765 		     "aggregate counts per processor socket", AGGR_SOCKET),
766 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
767 		     "aggregate counts per physical processor core", AGGR_CORE),
768 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
769 		     "aggregate counts per thread", AGGR_THREAD),
770 	OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
771 		     "ms to wait before starting measurement after program start"),
772 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
773 			"Only print computed metrics. No raw values", enable_metric_only),
774 	OPT_BOOLEAN(0, "topdown", &topdown_run,
775 			"measure topdown level 1 statistics"),
776 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
777 			"measure SMI cost"),
778 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
779 		     "monitor specified metrics or metric groups (separated by ,)",
780 		     parse_metric_groups),
781 	OPT_END()
782 };
783 
784 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
785 				 struct cpu_map *map, int cpu)
786 {
787 	return cpu_map__get_socket(map, cpu, NULL);
788 }
789 
790 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
791 			       struct cpu_map *map, int cpu)
792 {
793 	return cpu_map__get_core(map, cpu, NULL);
794 }
795 
796 static int cpu_map__get_max(struct cpu_map *map)
797 {
798 	int i, max = -1;
799 
800 	for (i = 0; i < map->nr; i++) {
801 		if (map->map[i] > max)
802 			max = map->map[i];
803 	}
804 
805 	return max;
806 }
807 
808 static int perf_stat__get_aggr(struct perf_stat_config *config,
809 			       aggr_get_id_t get_id, struct cpu_map *map, int idx)
810 {
811 	int cpu;
812 
813 	if (idx >= map->nr)
814 		return -1;
815 
816 	cpu = map->map[idx];
817 
818 	if (config->cpus_aggr_map->map[cpu] == -1)
819 		config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
820 
821 	return config->cpus_aggr_map->map[cpu];
822 }
823 
824 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
825 					struct cpu_map *map, int idx)
826 {
827 	return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
828 }
829 
830 static int perf_stat__get_core_cached(struct perf_stat_config *config,
831 				      struct cpu_map *map, int idx)
832 {
833 	return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
834 }
835 
836 static int perf_stat_init_aggr_mode(void)
837 {
838 	int nr;
839 
840 	switch (stat_config.aggr_mode) {
841 	case AGGR_SOCKET:
842 		if (cpu_map__build_socket_map(evsel_list->cpus, &stat_config.aggr_map)) {
843 			perror("cannot build socket map");
844 			return -1;
845 		}
846 		stat_config.aggr_get_id = perf_stat__get_socket_cached;
847 		break;
848 	case AGGR_CORE:
849 		if (cpu_map__build_core_map(evsel_list->cpus, &stat_config.aggr_map)) {
850 			perror("cannot build core map");
851 			return -1;
852 		}
853 		stat_config.aggr_get_id = perf_stat__get_core_cached;
854 		break;
855 	case AGGR_NONE:
856 	case AGGR_GLOBAL:
857 	case AGGR_THREAD:
858 	case AGGR_UNSET:
859 	default:
860 		break;
861 	}
862 
863 	/*
864 	 * The evsel_list->cpus is the base we operate on,
865 	 * taking the highest cpu number to be the size of
866 	 * the aggregation translate cpumap.
867 	 */
868 	nr = cpu_map__get_max(evsel_list->cpus);
869 	stat_config.cpus_aggr_map = cpu_map__empty_new(nr + 1);
870 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
871 }
872 
873 static void perf_stat__exit_aggr_mode(void)
874 {
875 	cpu_map__put(stat_config.aggr_map);
876 	cpu_map__put(stat_config.cpus_aggr_map);
877 	stat_config.aggr_map = NULL;
878 	stat_config.cpus_aggr_map = NULL;
879 }
880 
881 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
882 {
883 	int cpu;
884 
885 	if (idx > map->nr)
886 		return -1;
887 
888 	cpu = map->map[idx];
889 
890 	if (cpu >= env->nr_cpus_avail)
891 		return -1;
892 
893 	return cpu;
894 }
895 
896 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
897 {
898 	struct perf_env *env = data;
899 	int cpu = perf_env__get_cpu(env, map, idx);
900 
901 	return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
902 }
903 
904 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
905 {
906 	struct perf_env *env = data;
907 	int core = -1, cpu = perf_env__get_cpu(env, map, idx);
908 
909 	if (cpu != -1) {
910 		int socket_id = env->cpu[cpu].socket_id;
911 
912 		/*
913 		 * Encode socket in upper 16 bits
914 		 * core_id is relative to socket, and
915 		 * we need a global id. So we combine
916 		 * socket + core id.
917 		 */
918 		core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
919 	}
920 
921 	return core;
922 }
923 
924 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
925 				      struct cpu_map **sockp)
926 {
927 	return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
928 }
929 
930 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
931 				    struct cpu_map **corep)
932 {
933 	return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
934 }
935 
936 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
937 				      struct cpu_map *map, int idx)
938 {
939 	return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
940 }
941 
942 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
943 				    struct cpu_map *map, int idx)
944 {
945 	return perf_env__get_core(map, idx, &perf_stat.session->header.env);
946 }
947 
948 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
949 {
950 	struct perf_env *env = &st->session->header.env;
951 
952 	switch (stat_config.aggr_mode) {
953 	case AGGR_SOCKET:
954 		if (perf_env__build_socket_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
955 			perror("cannot build socket map");
956 			return -1;
957 		}
958 		stat_config.aggr_get_id = perf_stat__get_socket_file;
959 		break;
960 	case AGGR_CORE:
961 		if (perf_env__build_core_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
962 			perror("cannot build core map");
963 			return -1;
964 		}
965 		stat_config.aggr_get_id = perf_stat__get_core_file;
966 		break;
967 	case AGGR_NONE:
968 	case AGGR_GLOBAL:
969 	case AGGR_THREAD:
970 	case AGGR_UNSET:
971 	default:
972 		break;
973 	}
974 
975 	return 0;
976 }
977 
978 static int topdown_filter_events(const char **attr, char **str, bool use_group)
979 {
980 	int off = 0;
981 	int i;
982 	int len = 0;
983 	char *s;
984 
985 	for (i = 0; attr[i]; i++) {
986 		if (pmu_have_event("cpu", attr[i])) {
987 			len += strlen(attr[i]) + 1;
988 			attr[i - off] = attr[i];
989 		} else
990 			off++;
991 	}
992 	attr[i - off] = NULL;
993 
994 	*str = malloc(len + 1 + 2);
995 	if (!*str)
996 		return -1;
997 	s = *str;
998 	if (i - off == 0) {
999 		*s = 0;
1000 		return 0;
1001 	}
1002 	if (use_group)
1003 		*s++ = '{';
1004 	for (i = 0; attr[i]; i++) {
1005 		strcpy(s, attr[i]);
1006 		s += strlen(s);
1007 		*s++ = ',';
1008 	}
1009 	if (use_group) {
1010 		s[-1] = '}';
1011 		*s = 0;
1012 	} else
1013 		s[-1] = 0;
1014 	return 0;
1015 }
1016 
1017 __weak bool arch_topdown_check_group(bool *warn)
1018 {
1019 	*warn = false;
1020 	return false;
1021 }
1022 
1023 __weak void arch_topdown_group_warn(void)
1024 {
1025 }
1026 
1027 /*
1028  * Add default attributes, if there were no attributes specified or
1029  * if -d/--detailed, -d -d or -d -d -d is used:
1030  */
1031 static int add_default_attributes(void)
1032 {
1033 	int err;
1034 	struct perf_event_attr default_attrs0[] = {
1035 
1036   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1037   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1038   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1039   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1040 
1041   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1042 };
1043 	struct perf_event_attr frontend_attrs[] = {
1044   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1045 };
1046 	struct perf_event_attr backend_attrs[] = {
1047   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1048 };
1049 	struct perf_event_attr default_attrs1[] = {
1050   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1051   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1052   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1053 
1054 };
1055 
1056 /*
1057  * Detailed stats (-d), covering the L1 and last level data caches:
1058  */
1059 	struct perf_event_attr detailed_attrs[] = {
1060 
1061   { .type = PERF_TYPE_HW_CACHE,
1062     .config =
1063 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1064 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1065 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1066 
1067   { .type = PERF_TYPE_HW_CACHE,
1068     .config =
1069 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1070 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1071 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1072 
1073   { .type = PERF_TYPE_HW_CACHE,
1074     .config =
1075 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1076 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1077 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1078 
1079   { .type = PERF_TYPE_HW_CACHE,
1080     .config =
1081 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1082 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1083 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1084 };
1085 
1086 /*
1087  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1088  */
1089 	struct perf_event_attr very_detailed_attrs[] = {
1090 
1091   { .type = PERF_TYPE_HW_CACHE,
1092     .config =
1093 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1094 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1095 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1096 
1097   { .type = PERF_TYPE_HW_CACHE,
1098     .config =
1099 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1100 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1101 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1102 
1103   { .type = PERF_TYPE_HW_CACHE,
1104     .config =
1105 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1106 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1107 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1108 
1109   { .type = PERF_TYPE_HW_CACHE,
1110     .config =
1111 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1112 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1113 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1114 
1115   { .type = PERF_TYPE_HW_CACHE,
1116     .config =
1117 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1118 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1119 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1120 
1121   { .type = PERF_TYPE_HW_CACHE,
1122     .config =
1123 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1124 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1125 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1126 
1127 };
1128 
1129 /*
1130  * Very, very detailed stats (-d -d -d), adding prefetch events:
1131  */
1132 	struct perf_event_attr very_very_detailed_attrs[] = {
1133 
1134   { .type = PERF_TYPE_HW_CACHE,
1135     .config =
1136 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1137 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1138 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1139 
1140   { .type = PERF_TYPE_HW_CACHE,
1141     .config =
1142 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1143 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1144 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1145 };
1146 	struct parse_events_error errinfo;
1147 
1148 	/* Set attrs if no event is selected and !null_run: */
1149 	if (stat_config.null_run)
1150 		return 0;
1151 
1152 	if (transaction_run) {
1153 		/* Handle -T as -M transaction. Once platform specific metrics
1154 		 * support has been added to the json files, all archictures
1155 		 * will use this approach. To determine transaction support
1156 		 * on an architecture test for such a metric name.
1157 		 */
1158 		if (metricgroup__has_metric("transaction")) {
1159 			struct option opt = { .value = &evsel_list };
1160 
1161 			return metricgroup__parse_groups(&opt, "transaction",
1162 							 &stat_config.metric_events);
1163 		}
1164 
1165 		if (pmu_have_event("cpu", "cycles-ct") &&
1166 		    pmu_have_event("cpu", "el-start"))
1167 			err = parse_events(evsel_list, transaction_attrs,
1168 					   &errinfo);
1169 		else
1170 			err = parse_events(evsel_list,
1171 					   transaction_limited_attrs,
1172 					   &errinfo);
1173 		if (err) {
1174 			fprintf(stderr, "Cannot set up transaction events\n");
1175 			parse_events_print_error(&errinfo, transaction_attrs);
1176 			return -1;
1177 		}
1178 		return 0;
1179 	}
1180 
1181 	if (smi_cost) {
1182 		int smi;
1183 
1184 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1185 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1186 			return -1;
1187 		}
1188 
1189 		if (!smi) {
1190 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1191 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1192 				return -1;
1193 			}
1194 			smi_reset = true;
1195 		}
1196 
1197 		if (pmu_have_event("msr", "aperf") &&
1198 		    pmu_have_event("msr", "smi")) {
1199 			if (!force_metric_only)
1200 				stat_config.metric_only = true;
1201 			err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1202 		} else {
1203 			fprintf(stderr, "To measure SMI cost, it needs "
1204 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1205 			parse_events_print_error(&errinfo, smi_cost_attrs);
1206 			return -1;
1207 		}
1208 		if (err) {
1209 			fprintf(stderr, "Cannot set up SMI cost events\n");
1210 			return -1;
1211 		}
1212 		return 0;
1213 	}
1214 
1215 	if (topdown_run) {
1216 		char *str = NULL;
1217 		bool warn = false;
1218 
1219 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1220 		    stat_config.aggr_mode != AGGR_CORE) {
1221 			pr_err("top down event configuration requires --per-core mode\n");
1222 			return -1;
1223 		}
1224 		stat_config.aggr_mode = AGGR_CORE;
1225 		if (nr_cgroups || !target__has_cpu(&target)) {
1226 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1227 			return -1;
1228 		}
1229 
1230 		if (!force_metric_only)
1231 			stat_config.metric_only = true;
1232 		if (topdown_filter_events(topdown_attrs, &str,
1233 				arch_topdown_check_group(&warn)) < 0) {
1234 			pr_err("Out of memory\n");
1235 			return -1;
1236 		}
1237 		if (topdown_attrs[0] && str) {
1238 			if (warn)
1239 				arch_topdown_group_warn();
1240 			err = parse_events(evsel_list, str, &errinfo);
1241 			if (err) {
1242 				fprintf(stderr,
1243 					"Cannot set up top down events %s: %d\n",
1244 					str, err);
1245 				free(str);
1246 				parse_events_print_error(&errinfo, str);
1247 				return -1;
1248 			}
1249 		} else {
1250 			fprintf(stderr, "System does not support topdown\n");
1251 			return -1;
1252 		}
1253 		free(str);
1254 	}
1255 
1256 	if (!evsel_list->nr_entries) {
1257 		if (target__has_cpu(&target))
1258 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1259 
1260 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1261 			return -1;
1262 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1263 			if (perf_evlist__add_default_attrs(evsel_list,
1264 						frontend_attrs) < 0)
1265 				return -1;
1266 		}
1267 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1268 			if (perf_evlist__add_default_attrs(evsel_list,
1269 						backend_attrs) < 0)
1270 				return -1;
1271 		}
1272 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1273 			return -1;
1274 	}
1275 
1276 	/* Detailed events get appended to the event list: */
1277 
1278 	if (detailed_run <  1)
1279 		return 0;
1280 
1281 	/* Append detailed run extra attributes: */
1282 	if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1283 		return -1;
1284 
1285 	if (detailed_run < 2)
1286 		return 0;
1287 
1288 	/* Append very detailed run extra attributes: */
1289 	if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1290 		return -1;
1291 
1292 	if (detailed_run < 3)
1293 		return 0;
1294 
1295 	/* Append very, very detailed run extra attributes: */
1296 	return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1297 }
1298 
1299 static const char * const stat_record_usage[] = {
1300 	"perf stat record [<options>]",
1301 	NULL,
1302 };
1303 
1304 static void init_features(struct perf_session *session)
1305 {
1306 	int feat;
1307 
1308 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1309 		perf_header__set_feat(&session->header, feat);
1310 
1311 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1312 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1313 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1314 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1315 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1316 }
1317 
1318 static int __cmd_record(int argc, const char **argv)
1319 {
1320 	struct perf_session *session;
1321 	struct perf_data *data = &perf_stat.data;
1322 
1323 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
1324 			     PARSE_OPT_STOP_AT_NON_OPTION);
1325 
1326 	if (output_name)
1327 		data->path = output_name;
1328 
1329 	if (stat_config.run_count != 1 || forever) {
1330 		pr_err("Cannot use -r option with perf stat record.\n");
1331 		return -1;
1332 	}
1333 
1334 	session = perf_session__new(data, false, NULL);
1335 	if (session == NULL) {
1336 		pr_err("Perf session creation failed.\n");
1337 		return -1;
1338 	}
1339 
1340 	init_features(session);
1341 
1342 	session->evlist   = evsel_list;
1343 	perf_stat.session = session;
1344 	perf_stat.record  = true;
1345 	return argc;
1346 }
1347 
1348 static int process_stat_round_event(struct perf_session *session,
1349 				    union perf_event *event)
1350 {
1351 	struct stat_round_event *stat_round = &event->stat_round;
1352 	struct perf_evsel *counter;
1353 	struct timespec tsh, *ts = NULL;
1354 	const char **argv = session->header.env.cmdline_argv;
1355 	int argc = session->header.env.nr_cmdline;
1356 
1357 	evlist__for_each_entry(evsel_list, counter)
1358 		perf_stat_process_counter(&stat_config, counter);
1359 
1360 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1361 		update_stats(&walltime_nsecs_stats, stat_round->time);
1362 
1363 	if (stat_config.interval && stat_round->time) {
1364 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
1365 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1366 		ts = &tsh;
1367 	}
1368 
1369 	print_counters(ts, argc, argv);
1370 	return 0;
1371 }
1372 
1373 static
1374 int process_stat_config_event(struct perf_session *session,
1375 			      union perf_event *event)
1376 {
1377 	struct perf_tool *tool = session->tool;
1378 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1379 
1380 	perf_event__read_stat_config(&stat_config, &event->stat_config);
1381 
1382 	if (cpu_map__empty(st->cpus)) {
1383 		if (st->aggr_mode != AGGR_UNSET)
1384 			pr_warning("warning: processing task data, aggregation mode not set\n");
1385 		return 0;
1386 	}
1387 
1388 	if (st->aggr_mode != AGGR_UNSET)
1389 		stat_config.aggr_mode = st->aggr_mode;
1390 
1391 	if (perf_stat.data.is_pipe)
1392 		perf_stat_init_aggr_mode();
1393 	else
1394 		perf_stat_init_aggr_mode_file(st);
1395 
1396 	return 0;
1397 }
1398 
1399 static int set_maps(struct perf_stat *st)
1400 {
1401 	if (!st->cpus || !st->threads)
1402 		return 0;
1403 
1404 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1405 		return -EINVAL;
1406 
1407 	perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
1408 
1409 	if (perf_evlist__alloc_stats(evsel_list, true))
1410 		return -ENOMEM;
1411 
1412 	st->maps_allocated = true;
1413 	return 0;
1414 }
1415 
1416 static
1417 int process_thread_map_event(struct perf_session *session,
1418 			     union perf_event *event)
1419 {
1420 	struct perf_tool *tool = session->tool;
1421 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1422 
1423 	if (st->threads) {
1424 		pr_warning("Extra thread map event, ignoring.\n");
1425 		return 0;
1426 	}
1427 
1428 	st->threads = thread_map__new_event(&event->thread_map);
1429 	if (!st->threads)
1430 		return -ENOMEM;
1431 
1432 	return set_maps(st);
1433 }
1434 
1435 static
1436 int process_cpu_map_event(struct perf_session *session,
1437 			  union perf_event *event)
1438 {
1439 	struct perf_tool *tool = session->tool;
1440 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1441 	struct cpu_map *cpus;
1442 
1443 	if (st->cpus) {
1444 		pr_warning("Extra cpu map event, ignoring.\n");
1445 		return 0;
1446 	}
1447 
1448 	cpus = cpu_map__new_data(&event->cpu_map.data);
1449 	if (!cpus)
1450 		return -ENOMEM;
1451 
1452 	st->cpus = cpus;
1453 	return set_maps(st);
1454 }
1455 
1456 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
1457 {
1458 	int i;
1459 
1460 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
1461 	if (!config->stats)
1462 		return -1;
1463 
1464 	config->stats_num = nthreads;
1465 
1466 	for (i = 0; i < nthreads; i++)
1467 		runtime_stat__init(&config->stats[i]);
1468 
1469 	return 0;
1470 }
1471 
1472 static void runtime_stat_delete(struct perf_stat_config *config)
1473 {
1474 	int i;
1475 
1476 	if (!config->stats)
1477 		return;
1478 
1479 	for (i = 0; i < config->stats_num; i++)
1480 		runtime_stat__exit(&config->stats[i]);
1481 
1482 	free(config->stats);
1483 }
1484 
1485 static const char * const stat_report_usage[] = {
1486 	"perf stat report [<options>]",
1487 	NULL,
1488 };
1489 
1490 static struct perf_stat perf_stat = {
1491 	.tool = {
1492 		.attr		= perf_event__process_attr,
1493 		.event_update	= perf_event__process_event_update,
1494 		.thread_map	= process_thread_map_event,
1495 		.cpu_map	= process_cpu_map_event,
1496 		.stat_config	= process_stat_config_event,
1497 		.stat		= perf_event__process_stat_event,
1498 		.stat_round	= process_stat_round_event,
1499 	},
1500 	.aggr_mode = AGGR_UNSET,
1501 };
1502 
1503 static int __cmd_report(int argc, const char **argv)
1504 {
1505 	struct perf_session *session;
1506 	const struct option options[] = {
1507 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1508 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1509 		     "aggregate counts per processor socket", AGGR_SOCKET),
1510 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1511 		     "aggregate counts per physical processor core", AGGR_CORE),
1512 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1513 		     "disable CPU count aggregation", AGGR_NONE),
1514 	OPT_END()
1515 	};
1516 	struct stat st;
1517 	int ret;
1518 
1519 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
1520 
1521 	if (!input_name || !strlen(input_name)) {
1522 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1523 			input_name = "-";
1524 		else
1525 			input_name = "perf.data";
1526 	}
1527 
1528 	perf_stat.data.path = input_name;
1529 	perf_stat.data.mode = PERF_DATA_MODE_READ;
1530 
1531 	session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
1532 	if (session == NULL)
1533 		return -1;
1534 
1535 	perf_stat.session  = session;
1536 	stat_config.output = stderr;
1537 	evsel_list         = session->evlist;
1538 
1539 	ret = perf_session__process_events(session);
1540 	if (ret)
1541 		return ret;
1542 
1543 	perf_session__delete(session);
1544 	return 0;
1545 }
1546 
1547 static void setup_system_wide(int forks)
1548 {
1549 	/*
1550 	 * Make system wide (-a) the default target if
1551 	 * no target was specified and one of following
1552 	 * conditions is met:
1553 	 *
1554 	 *   - there's no workload specified
1555 	 *   - there is workload specified but all requested
1556 	 *     events are system wide events
1557 	 */
1558 	if (!target__none(&target))
1559 		return;
1560 
1561 	if (!forks)
1562 		target.system_wide = true;
1563 	else {
1564 		struct perf_evsel *counter;
1565 
1566 		evlist__for_each_entry(evsel_list, counter) {
1567 			if (!counter->system_wide)
1568 				return;
1569 		}
1570 
1571 		if (evsel_list->nr_entries)
1572 			target.system_wide = true;
1573 	}
1574 }
1575 
1576 int cmd_stat(int argc, const char **argv)
1577 {
1578 	const char * const stat_usage[] = {
1579 		"perf stat [<options>] [<command>]",
1580 		NULL
1581 	};
1582 	int status = -EINVAL, run_idx;
1583 	const char *mode;
1584 	FILE *output = stderr;
1585 	unsigned int interval, timeout;
1586 	const char * const stat_subcommands[] = { "record", "report" };
1587 
1588 	setlocale(LC_ALL, "");
1589 
1590 	evsel_list = perf_evlist__new();
1591 	if (evsel_list == NULL)
1592 		return -ENOMEM;
1593 
1594 	parse_events__shrink_config_terms();
1595 
1596 	/* String-parsing callback-based options would segfault when negated */
1597 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
1598 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
1599 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
1600 
1601 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1602 					(const char **) stat_usage,
1603 					PARSE_OPT_STOP_AT_NON_OPTION);
1604 	perf_stat__collect_metric_expr(evsel_list);
1605 	perf_stat__init_shadow_stats();
1606 
1607 	if (stat_config.csv_sep) {
1608 		stat_config.csv_output = true;
1609 		if (!strcmp(stat_config.csv_sep, "\\t"))
1610 			stat_config.csv_sep = "\t";
1611 	} else
1612 		stat_config.csv_sep = DEFAULT_SEPARATOR;
1613 
1614 	if (argc && !strncmp(argv[0], "rec", 3)) {
1615 		argc = __cmd_record(argc, argv);
1616 		if (argc < 0)
1617 			return -1;
1618 	} else if (argc && !strncmp(argv[0], "rep", 3))
1619 		return __cmd_report(argc, argv);
1620 
1621 	interval = stat_config.interval;
1622 	timeout = stat_config.timeout;
1623 
1624 	/*
1625 	 * For record command the -o is already taken care of.
1626 	 */
1627 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1628 		output = NULL;
1629 
1630 	if (output_name && output_fd) {
1631 		fprintf(stderr, "cannot use both --output and --log-fd\n");
1632 		parse_options_usage(stat_usage, stat_options, "o", 1);
1633 		parse_options_usage(NULL, stat_options, "log-fd", 0);
1634 		goto out;
1635 	}
1636 
1637 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
1638 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
1639 		goto out;
1640 	}
1641 
1642 	if (stat_config.metric_only && stat_config.run_count > 1) {
1643 		fprintf(stderr, "--metric-only is not supported with -r\n");
1644 		goto out;
1645 	}
1646 
1647 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
1648 		fprintf(stderr, "--table is only supported with -r\n");
1649 		parse_options_usage(stat_usage, stat_options, "r", 1);
1650 		parse_options_usage(NULL, stat_options, "table", 0);
1651 		goto out;
1652 	}
1653 
1654 	if (output_fd < 0) {
1655 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
1656 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1657 		goto out;
1658 	}
1659 
1660 	if (!output) {
1661 		struct timespec tm;
1662 		mode = append_file ? "a" : "w";
1663 
1664 		output = fopen(output_name, mode);
1665 		if (!output) {
1666 			perror("failed to create output file");
1667 			return -1;
1668 		}
1669 		clock_gettime(CLOCK_REALTIME, &tm);
1670 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1671 	} else if (output_fd > 0) {
1672 		mode = append_file ? "a" : "w";
1673 		output = fdopen(output_fd, mode);
1674 		if (!output) {
1675 			perror("Failed opening logfd");
1676 			return -errno;
1677 		}
1678 	}
1679 
1680 	stat_config.output = output;
1681 
1682 	/*
1683 	 * let the spreadsheet do the pretty-printing
1684 	 */
1685 	if (stat_config.csv_output) {
1686 		/* User explicitly passed -B? */
1687 		if (big_num_opt == 1) {
1688 			fprintf(stderr, "-B option not supported with -x\n");
1689 			parse_options_usage(stat_usage, stat_options, "B", 1);
1690 			parse_options_usage(NULL, stat_options, "x", 1);
1691 			goto out;
1692 		} else /* Nope, so disable big number formatting */
1693 			stat_config.big_num = false;
1694 	} else if (big_num_opt == 0) /* User passed --no-big-num */
1695 		stat_config.big_num = false;
1696 
1697 	setup_system_wide(argc);
1698 
1699 	/*
1700 	 * Display user/system times only for single
1701 	 * run and when there's specified tracee.
1702 	 */
1703 	if ((stat_config.run_count == 1) && target__none(&target))
1704 		stat_config.ru_display = true;
1705 
1706 	if (stat_config.run_count < 0) {
1707 		pr_err("Run count must be a positive number\n");
1708 		parse_options_usage(stat_usage, stat_options, "r", 1);
1709 		goto out;
1710 	} else if (stat_config.run_count == 0) {
1711 		forever = true;
1712 		stat_config.run_count = 1;
1713 	}
1714 
1715 	if (stat_config.walltime_run_table) {
1716 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
1717 		if (!stat_config.walltime_run) {
1718 			pr_err("failed to setup -r option");
1719 			goto out;
1720 		}
1721 	}
1722 
1723 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
1724 		!target__has_task(&target)) {
1725 		if (!target.system_wide || target.cpu_list) {
1726 			fprintf(stderr, "The --per-thread option is only "
1727 				"available when monitoring via -p -t -a "
1728 				"options or only --per-thread.\n");
1729 			parse_options_usage(NULL, stat_options, "p", 1);
1730 			parse_options_usage(NULL, stat_options, "t", 1);
1731 			goto out;
1732 		}
1733 	}
1734 
1735 	/*
1736 	 * no_aggr, cgroup are for system-wide only
1737 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1738 	 */
1739 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1740 	      stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1741 	    !target__has_cpu(&target)) {
1742 		fprintf(stderr, "both cgroup and no-aggregation "
1743 			"modes only available in system-wide mode\n");
1744 
1745 		parse_options_usage(stat_usage, stat_options, "G", 1);
1746 		parse_options_usage(NULL, stat_options, "A", 1);
1747 		parse_options_usage(NULL, stat_options, "a", 1);
1748 		goto out;
1749 	}
1750 
1751 	if (add_default_attributes())
1752 		goto out;
1753 
1754 	target__validate(&target);
1755 
1756 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
1757 		target.per_thread = true;
1758 
1759 	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1760 		if (target__has_task(&target)) {
1761 			pr_err("Problems finding threads of monitor\n");
1762 			parse_options_usage(stat_usage, stat_options, "p", 1);
1763 			parse_options_usage(NULL, stat_options, "t", 1);
1764 		} else if (target__has_cpu(&target)) {
1765 			perror("failed to parse CPUs map");
1766 			parse_options_usage(stat_usage, stat_options, "C", 1);
1767 			parse_options_usage(NULL, stat_options, "a", 1);
1768 		}
1769 		goto out;
1770 	}
1771 
1772 	/*
1773 	 * Initialize thread_map with comm names,
1774 	 * so we could print it out on output.
1775 	 */
1776 	if (stat_config.aggr_mode == AGGR_THREAD) {
1777 		thread_map__read_comms(evsel_list->threads);
1778 		if (target.system_wide) {
1779 			if (runtime_stat_new(&stat_config,
1780 				thread_map__nr(evsel_list->threads))) {
1781 				goto out;
1782 			}
1783 		}
1784 	}
1785 
1786 	if (stat_config.times && interval)
1787 		interval_count = true;
1788 	else if (stat_config.times && !interval) {
1789 		pr_err("interval-count option should be used together with "
1790 				"interval-print.\n");
1791 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
1792 		parse_options_usage(stat_usage, stat_options, "I", 1);
1793 		goto out;
1794 	}
1795 
1796 	if (timeout && timeout < 100) {
1797 		if (timeout < 10) {
1798 			pr_err("timeout must be >= 10ms.\n");
1799 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
1800 			goto out;
1801 		} else
1802 			pr_warning("timeout < 100ms. "
1803 				   "The overhead percentage could be high in some cases. "
1804 				   "Please proceed with caution.\n");
1805 	}
1806 	if (timeout && interval) {
1807 		pr_err("timeout option is not supported with interval-print.\n");
1808 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
1809 		parse_options_usage(stat_usage, stat_options, "I", 1);
1810 		goto out;
1811 	}
1812 
1813 	if (perf_evlist__alloc_stats(evsel_list, interval))
1814 		goto out;
1815 
1816 	if (perf_stat_init_aggr_mode())
1817 		goto out;
1818 
1819 	/*
1820 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
1821 	 * while avoiding that older tools show confusing messages.
1822 	 *
1823 	 * However for pipe sessions we need to keep it zero,
1824 	 * because script's perf_evsel__check_attr is triggered
1825 	 * by attr->sample_type != 0, and we can't run it on
1826 	 * stat sessions.
1827 	 */
1828 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
1829 
1830 	/*
1831 	 * We dont want to block the signals - that would cause
1832 	 * child tasks to inherit that and Ctrl-C would not work.
1833 	 * What we want is for Ctrl-C to work in the exec()-ed
1834 	 * task, but being ignored by perf stat itself:
1835 	 */
1836 	atexit(sig_atexit);
1837 	if (!forever)
1838 		signal(SIGINT,  skip_signal);
1839 	signal(SIGCHLD, skip_signal);
1840 	signal(SIGALRM, skip_signal);
1841 	signal(SIGABRT, skip_signal);
1842 
1843 	status = 0;
1844 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
1845 		if (stat_config.run_count != 1 && verbose > 0)
1846 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1847 				run_idx + 1);
1848 
1849 		status = run_perf_stat(argc, argv, run_idx);
1850 		if (forever && status != -1) {
1851 			print_counters(NULL, argc, argv);
1852 			perf_stat__reset_stats();
1853 		}
1854 	}
1855 
1856 	if (!forever && status != -1 && !interval)
1857 		print_counters(NULL, argc, argv);
1858 
1859 	if (STAT_RECORD) {
1860 		/*
1861 		 * We synthesize the kernel mmap record just so that older tools
1862 		 * don't emit warnings about not being able to resolve symbols
1863 		 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
1864 		 * a saner message about no samples being in the perf.data file.
1865 		 *
1866 		 * This also serves to suppress a warning about f_header.data.size == 0
1867 		 * in header.c at the moment 'perf stat record' gets introduced, which
1868 		 * is not really needed once we start adding the stat specific PERF_RECORD_
1869 		 * records, but the need to suppress the kptr_restrict messages in older
1870 		 * tools remain  -acme
1871 		 */
1872 		int fd = perf_data__fd(&perf_stat.data);
1873 		int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
1874 							     process_synthesized_event,
1875 							     &perf_stat.session->machines.host);
1876 		if (err) {
1877 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
1878 				   "older tools may produce warnings about this file\n.");
1879 		}
1880 
1881 		if (!interval) {
1882 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
1883 				pr_err("failed to write stat round event\n");
1884 		}
1885 
1886 		if (!perf_stat.data.is_pipe) {
1887 			perf_stat.session->header.data_size += perf_stat.bytes_written;
1888 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
1889 		}
1890 
1891 		perf_session__delete(perf_stat.session);
1892 	}
1893 
1894 	perf_stat__exit_aggr_mode();
1895 	perf_evlist__free_stats(evsel_list);
1896 out:
1897 	free(stat_config.walltime_run);
1898 
1899 	if (smi_cost && smi_reset)
1900 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
1901 
1902 	perf_evlist__delete(evsel_list);
1903 
1904 	runtime_stat_delete(&stat_config);
1905 
1906 	return status;
1907 }
1908