xref: /openbmc/linux/tools/perf/builtin-stat.c (revision 8440bb9b)
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('c', "scale", &stat_config.scale, "scale/normalize counters"),
722 	OPT_INCR('v', "verbose", &verbose,
723 		    "be more verbose (show counter open errors, etc)"),
724 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
725 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
726 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
727 		    "display details about each run (only with -r option)"),
728 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
729 		    "null run - dont start any counters"),
730 	OPT_INCR('d', "detailed", &detailed_run,
731 		    "detailed run - start a lot of events"),
732 	OPT_BOOLEAN('S', "sync", &sync_run,
733 		    "call sync() before starting a run"),
734 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
735 			   "print large numbers with thousands\' separators",
736 			   stat__set_big_num),
737 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
738 		    "list of cpus to monitor in system-wide"),
739 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
740 		    "disable CPU count aggregation", AGGR_NONE),
741 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
742 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
743 		   "print counts with custom separator"),
744 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
745 		     "monitor event in cgroup name only", parse_cgroups),
746 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
747 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
748 	OPT_INTEGER(0, "log-fd", &output_fd,
749 		    "log output to fd, instead of stderr"),
750 	OPT_STRING(0, "pre", &pre_cmd, "command",
751 			"command to run prior to the measured command"),
752 	OPT_STRING(0, "post", &post_cmd, "command",
753 			"command to run after to the measured command"),
754 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
755 		    "print counts at regular interval in ms "
756 		    "(overhead is possible for values <= 100ms)"),
757 	OPT_INTEGER(0, "interval-count", &stat_config.times,
758 		    "print counts for fixed number of times"),
759 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
760 		    "clear screen in between new interval"),
761 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
762 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
763 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
764 		     "aggregate counts per processor socket", AGGR_SOCKET),
765 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
766 		     "aggregate counts per physical processor core", AGGR_CORE),
767 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
768 		     "aggregate counts per thread", AGGR_THREAD),
769 	OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
770 		     "ms to wait before starting measurement after program start"),
771 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
772 			"Only print computed metrics. No raw values", enable_metric_only),
773 	OPT_BOOLEAN(0, "topdown", &topdown_run,
774 			"measure topdown level 1 statistics"),
775 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
776 			"measure SMI cost"),
777 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
778 		     "monitor specified metrics or metric groups (separated by ,)",
779 		     parse_metric_groups),
780 	OPT_END()
781 };
782 
783 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
784 				 struct cpu_map *map, int cpu)
785 {
786 	return cpu_map__get_socket(map, cpu, NULL);
787 }
788 
789 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
790 			       struct cpu_map *map, int cpu)
791 {
792 	return cpu_map__get_core(map, cpu, NULL);
793 }
794 
795 static int cpu_map__get_max(struct cpu_map *map)
796 {
797 	int i, max = -1;
798 
799 	for (i = 0; i < map->nr; i++) {
800 		if (map->map[i] > max)
801 			max = map->map[i];
802 	}
803 
804 	return max;
805 }
806 
807 static int perf_stat__get_aggr(struct perf_stat_config *config,
808 			       aggr_get_id_t get_id, struct cpu_map *map, int idx)
809 {
810 	int cpu;
811 
812 	if (idx >= map->nr)
813 		return -1;
814 
815 	cpu = map->map[idx];
816 
817 	if (config->cpus_aggr_map->map[cpu] == -1)
818 		config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
819 
820 	return config->cpus_aggr_map->map[cpu];
821 }
822 
823 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
824 					struct cpu_map *map, int idx)
825 {
826 	return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
827 }
828 
829 static int perf_stat__get_core_cached(struct perf_stat_config *config,
830 				      struct cpu_map *map, int idx)
831 {
832 	return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
833 }
834 
835 static int perf_stat_init_aggr_mode(void)
836 {
837 	int nr;
838 
839 	switch (stat_config.aggr_mode) {
840 	case AGGR_SOCKET:
841 		if (cpu_map__build_socket_map(evsel_list->cpus, &stat_config.aggr_map)) {
842 			perror("cannot build socket map");
843 			return -1;
844 		}
845 		stat_config.aggr_get_id = perf_stat__get_socket_cached;
846 		break;
847 	case AGGR_CORE:
848 		if (cpu_map__build_core_map(evsel_list->cpus, &stat_config.aggr_map)) {
849 			perror("cannot build core map");
850 			return -1;
851 		}
852 		stat_config.aggr_get_id = perf_stat__get_core_cached;
853 		break;
854 	case AGGR_NONE:
855 	case AGGR_GLOBAL:
856 	case AGGR_THREAD:
857 	case AGGR_UNSET:
858 	default:
859 		break;
860 	}
861 
862 	/*
863 	 * The evsel_list->cpus is the base we operate on,
864 	 * taking the highest cpu number to be the size of
865 	 * the aggregation translate cpumap.
866 	 */
867 	nr = cpu_map__get_max(evsel_list->cpus);
868 	stat_config.cpus_aggr_map = cpu_map__empty_new(nr + 1);
869 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
870 }
871 
872 static void perf_stat__exit_aggr_mode(void)
873 {
874 	cpu_map__put(stat_config.aggr_map);
875 	cpu_map__put(stat_config.cpus_aggr_map);
876 	stat_config.aggr_map = NULL;
877 	stat_config.cpus_aggr_map = NULL;
878 }
879 
880 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
881 {
882 	int cpu;
883 
884 	if (idx > map->nr)
885 		return -1;
886 
887 	cpu = map->map[idx];
888 
889 	if (cpu >= env->nr_cpus_avail)
890 		return -1;
891 
892 	return cpu;
893 }
894 
895 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
896 {
897 	struct perf_env *env = data;
898 	int cpu = perf_env__get_cpu(env, map, idx);
899 
900 	return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
901 }
902 
903 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
904 {
905 	struct perf_env *env = data;
906 	int core = -1, cpu = perf_env__get_cpu(env, map, idx);
907 
908 	if (cpu != -1) {
909 		int socket_id = env->cpu[cpu].socket_id;
910 
911 		/*
912 		 * Encode socket in upper 16 bits
913 		 * core_id is relative to socket, and
914 		 * we need a global id. So we combine
915 		 * socket + core id.
916 		 */
917 		core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
918 	}
919 
920 	return core;
921 }
922 
923 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
924 				      struct cpu_map **sockp)
925 {
926 	return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
927 }
928 
929 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
930 				    struct cpu_map **corep)
931 {
932 	return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
933 }
934 
935 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
936 				      struct cpu_map *map, int idx)
937 {
938 	return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
939 }
940 
941 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
942 				    struct cpu_map *map, int idx)
943 {
944 	return perf_env__get_core(map, idx, &perf_stat.session->header.env);
945 }
946 
947 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
948 {
949 	struct perf_env *env = &st->session->header.env;
950 
951 	switch (stat_config.aggr_mode) {
952 	case AGGR_SOCKET:
953 		if (perf_env__build_socket_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
954 			perror("cannot build socket map");
955 			return -1;
956 		}
957 		stat_config.aggr_get_id = perf_stat__get_socket_file;
958 		break;
959 	case AGGR_CORE:
960 		if (perf_env__build_core_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
961 			perror("cannot build core map");
962 			return -1;
963 		}
964 		stat_config.aggr_get_id = perf_stat__get_core_file;
965 		break;
966 	case AGGR_NONE:
967 	case AGGR_GLOBAL:
968 	case AGGR_THREAD:
969 	case AGGR_UNSET:
970 	default:
971 		break;
972 	}
973 
974 	return 0;
975 }
976 
977 static int topdown_filter_events(const char **attr, char **str, bool use_group)
978 {
979 	int off = 0;
980 	int i;
981 	int len = 0;
982 	char *s;
983 
984 	for (i = 0; attr[i]; i++) {
985 		if (pmu_have_event("cpu", attr[i])) {
986 			len += strlen(attr[i]) + 1;
987 			attr[i - off] = attr[i];
988 		} else
989 			off++;
990 	}
991 	attr[i - off] = NULL;
992 
993 	*str = malloc(len + 1 + 2);
994 	if (!*str)
995 		return -1;
996 	s = *str;
997 	if (i - off == 0) {
998 		*s = 0;
999 		return 0;
1000 	}
1001 	if (use_group)
1002 		*s++ = '{';
1003 	for (i = 0; attr[i]; i++) {
1004 		strcpy(s, attr[i]);
1005 		s += strlen(s);
1006 		*s++ = ',';
1007 	}
1008 	if (use_group) {
1009 		s[-1] = '}';
1010 		*s = 0;
1011 	} else
1012 		s[-1] = 0;
1013 	return 0;
1014 }
1015 
1016 __weak bool arch_topdown_check_group(bool *warn)
1017 {
1018 	*warn = false;
1019 	return false;
1020 }
1021 
1022 __weak void arch_topdown_group_warn(void)
1023 {
1024 }
1025 
1026 /*
1027  * Add default attributes, if there were no attributes specified or
1028  * if -d/--detailed, -d -d or -d -d -d is used:
1029  */
1030 static int add_default_attributes(void)
1031 {
1032 	int err;
1033 	struct perf_event_attr default_attrs0[] = {
1034 
1035   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1036   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1037   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1038   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1039 
1040   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1041 };
1042 	struct perf_event_attr frontend_attrs[] = {
1043   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1044 };
1045 	struct perf_event_attr backend_attrs[] = {
1046   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1047 };
1048 	struct perf_event_attr default_attrs1[] = {
1049   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1050   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1051   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1052 
1053 };
1054 
1055 /*
1056  * Detailed stats (-d), covering the L1 and last level data caches:
1057  */
1058 	struct perf_event_attr detailed_attrs[] = {
1059 
1060   { .type = PERF_TYPE_HW_CACHE,
1061     .config =
1062 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1063 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1064 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1065 
1066   { .type = PERF_TYPE_HW_CACHE,
1067     .config =
1068 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1069 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1070 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1071 
1072   { .type = PERF_TYPE_HW_CACHE,
1073     .config =
1074 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1075 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1076 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1077 
1078   { .type = PERF_TYPE_HW_CACHE,
1079     .config =
1080 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1081 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1082 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1083 };
1084 
1085 /*
1086  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1087  */
1088 	struct perf_event_attr very_detailed_attrs[] = {
1089 
1090   { .type = PERF_TYPE_HW_CACHE,
1091     .config =
1092 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1093 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1094 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1095 
1096   { .type = PERF_TYPE_HW_CACHE,
1097     .config =
1098 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1099 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1100 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1101 
1102   { .type = PERF_TYPE_HW_CACHE,
1103     .config =
1104 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1105 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1106 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1107 
1108   { .type = PERF_TYPE_HW_CACHE,
1109     .config =
1110 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1111 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1112 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1113 
1114   { .type = PERF_TYPE_HW_CACHE,
1115     .config =
1116 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1117 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1118 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1119 
1120   { .type = PERF_TYPE_HW_CACHE,
1121     .config =
1122 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1123 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1124 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1125 
1126 };
1127 
1128 /*
1129  * Very, very detailed stats (-d -d -d), adding prefetch events:
1130  */
1131 	struct perf_event_attr very_very_detailed_attrs[] = {
1132 
1133   { .type = PERF_TYPE_HW_CACHE,
1134     .config =
1135 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1136 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1137 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1138 
1139   { .type = PERF_TYPE_HW_CACHE,
1140     .config =
1141 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1142 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1143 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1144 };
1145 	struct parse_events_error errinfo;
1146 
1147 	/* Set attrs if no event is selected and !null_run: */
1148 	if (stat_config.null_run)
1149 		return 0;
1150 
1151 	if (transaction_run) {
1152 		/* Handle -T as -M transaction. Once platform specific metrics
1153 		 * support has been added to the json files, all archictures
1154 		 * will use this approach. To determine transaction support
1155 		 * on an architecture test for such a metric name.
1156 		 */
1157 		if (metricgroup__has_metric("transaction")) {
1158 			struct option opt = { .value = &evsel_list };
1159 
1160 			return metricgroup__parse_groups(&opt, "transaction",
1161 							 &stat_config.metric_events);
1162 		}
1163 
1164 		if (pmu_have_event("cpu", "cycles-ct") &&
1165 		    pmu_have_event("cpu", "el-start"))
1166 			err = parse_events(evsel_list, transaction_attrs,
1167 					   &errinfo);
1168 		else
1169 			err = parse_events(evsel_list,
1170 					   transaction_limited_attrs,
1171 					   &errinfo);
1172 		if (err) {
1173 			fprintf(stderr, "Cannot set up transaction events\n");
1174 			parse_events_print_error(&errinfo, transaction_attrs);
1175 			return -1;
1176 		}
1177 		return 0;
1178 	}
1179 
1180 	if (smi_cost) {
1181 		int smi;
1182 
1183 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1184 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1185 			return -1;
1186 		}
1187 
1188 		if (!smi) {
1189 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1190 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1191 				return -1;
1192 			}
1193 			smi_reset = true;
1194 		}
1195 
1196 		if (pmu_have_event("msr", "aperf") &&
1197 		    pmu_have_event("msr", "smi")) {
1198 			if (!force_metric_only)
1199 				stat_config.metric_only = true;
1200 			err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1201 		} else {
1202 			fprintf(stderr, "To measure SMI cost, it needs "
1203 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1204 			parse_events_print_error(&errinfo, smi_cost_attrs);
1205 			return -1;
1206 		}
1207 		if (err) {
1208 			fprintf(stderr, "Cannot set up SMI cost events\n");
1209 			return -1;
1210 		}
1211 		return 0;
1212 	}
1213 
1214 	if (topdown_run) {
1215 		char *str = NULL;
1216 		bool warn = false;
1217 
1218 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1219 		    stat_config.aggr_mode != AGGR_CORE) {
1220 			pr_err("top down event configuration requires --per-core mode\n");
1221 			return -1;
1222 		}
1223 		stat_config.aggr_mode = AGGR_CORE;
1224 		if (nr_cgroups || !target__has_cpu(&target)) {
1225 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1226 			return -1;
1227 		}
1228 
1229 		if (!force_metric_only)
1230 			stat_config.metric_only = true;
1231 		if (topdown_filter_events(topdown_attrs, &str,
1232 				arch_topdown_check_group(&warn)) < 0) {
1233 			pr_err("Out of memory\n");
1234 			return -1;
1235 		}
1236 		if (topdown_attrs[0] && str) {
1237 			if (warn)
1238 				arch_topdown_group_warn();
1239 			err = parse_events(evsel_list, str, &errinfo);
1240 			if (err) {
1241 				fprintf(stderr,
1242 					"Cannot set up top down events %s: %d\n",
1243 					str, err);
1244 				free(str);
1245 				parse_events_print_error(&errinfo, str);
1246 				return -1;
1247 			}
1248 		} else {
1249 			fprintf(stderr, "System does not support topdown\n");
1250 			return -1;
1251 		}
1252 		free(str);
1253 	}
1254 
1255 	if (!evsel_list->nr_entries) {
1256 		if (target__has_cpu(&target))
1257 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1258 
1259 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1260 			return -1;
1261 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1262 			if (perf_evlist__add_default_attrs(evsel_list,
1263 						frontend_attrs) < 0)
1264 				return -1;
1265 		}
1266 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1267 			if (perf_evlist__add_default_attrs(evsel_list,
1268 						backend_attrs) < 0)
1269 				return -1;
1270 		}
1271 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1272 			return -1;
1273 	}
1274 
1275 	/* Detailed events get appended to the event list: */
1276 
1277 	if (detailed_run <  1)
1278 		return 0;
1279 
1280 	/* Append detailed run extra attributes: */
1281 	if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1282 		return -1;
1283 
1284 	if (detailed_run < 2)
1285 		return 0;
1286 
1287 	/* Append very detailed run extra attributes: */
1288 	if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1289 		return -1;
1290 
1291 	if (detailed_run < 3)
1292 		return 0;
1293 
1294 	/* Append very, very detailed run extra attributes: */
1295 	return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1296 }
1297 
1298 static const char * const stat_record_usage[] = {
1299 	"perf stat record [<options>]",
1300 	NULL,
1301 };
1302 
1303 static void init_features(struct perf_session *session)
1304 {
1305 	int feat;
1306 
1307 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1308 		perf_header__set_feat(&session->header, feat);
1309 
1310 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1311 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1312 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1313 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1314 }
1315 
1316 static int __cmd_record(int argc, const char **argv)
1317 {
1318 	struct perf_session *session;
1319 	struct perf_data *data = &perf_stat.data;
1320 
1321 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
1322 			     PARSE_OPT_STOP_AT_NON_OPTION);
1323 
1324 	if (output_name)
1325 		data->path = output_name;
1326 
1327 	if (stat_config.run_count != 1 || forever) {
1328 		pr_err("Cannot use -r option with perf stat record.\n");
1329 		return -1;
1330 	}
1331 
1332 	session = perf_session__new(data, false, NULL);
1333 	if (session == NULL) {
1334 		pr_err("Perf session creation failed.\n");
1335 		return -1;
1336 	}
1337 
1338 	init_features(session);
1339 
1340 	session->evlist   = evsel_list;
1341 	perf_stat.session = session;
1342 	perf_stat.record  = true;
1343 	return argc;
1344 }
1345 
1346 static int process_stat_round_event(struct perf_session *session,
1347 				    union perf_event *event)
1348 {
1349 	struct stat_round_event *stat_round = &event->stat_round;
1350 	struct perf_evsel *counter;
1351 	struct timespec tsh, *ts = NULL;
1352 	const char **argv = session->header.env.cmdline_argv;
1353 	int argc = session->header.env.nr_cmdline;
1354 
1355 	evlist__for_each_entry(evsel_list, counter)
1356 		perf_stat_process_counter(&stat_config, counter);
1357 
1358 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1359 		update_stats(&walltime_nsecs_stats, stat_round->time);
1360 
1361 	if (stat_config.interval && stat_round->time) {
1362 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
1363 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1364 		ts = &tsh;
1365 	}
1366 
1367 	print_counters(ts, argc, argv);
1368 	return 0;
1369 }
1370 
1371 static
1372 int process_stat_config_event(struct perf_session *session,
1373 			      union perf_event *event)
1374 {
1375 	struct perf_tool *tool = session->tool;
1376 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1377 
1378 	perf_event__read_stat_config(&stat_config, &event->stat_config);
1379 
1380 	if (cpu_map__empty(st->cpus)) {
1381 		if (st->aggr_mode != AGGR_UNSET)
1382 			pr_warning("warning: processing task data, aggregation mode not set\n");
1383 		return 0;
1384 	}
1385 
1386 	if (st->aggr_mode != AGGR_UNSET)
1387 		stat_config.aggr_mode = st->aggr_mode;
1388 
1389 	if (perf_stat.data.is_pipe)
1390 		perf_stat_init_aggr_mode();
1391 	else
1392 		perf_stat_init_aggr_mode_file(st);
1393 
1394 	return 0;
1395 }
1396 
1397 static int set_maps(struct perf_stat *st)
1398 {
1399 	if (!st->cpus || !st->threads)
1400 		return 0;
1401 
1402 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1403 		return -EINVAL;
1404 
1405 	perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
1406 
1407 	if (perf_evlist__alloc_stats(evsel_list, true))
1408 		return -ENOMEM;
1409 
1410 	st->maps_allocated = true;
1411 	return 0;
1412 }
1413 
1414 static
1415 int process_thread_map_event(struct perf_session *session,
1416 			     union perf_event *event)
1417 {
1418 	struct perf_tool *tool = session->tool;
1419 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1420 
1421 	if (st->threads) {
1422 		pr_warning("Extra thread map event, ignoring.\n");
1423 		return 0;
1424 	}
1425 
1426 	st->threads = thread_map__new_event(&event->thread_map);
1427 	if (!st->threads)
1428 		return -ENOMEM;
1429 
1430 	return set_maps(st);
1431 }
1432 
1433 static
1434 int process_cpu_map_event(struct perf_session *session,
1435 			  union perf_event *event)
1436 {
1437 	struct perf_tool *tool = session->tool;
1438 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1439 	struct cpu_map *cpus;
1440 
1441 	if (st->cpus) {
1442 		pr_warning("Extra cpu map event, ignoring.\n");
1443 		return 0;
1444 	}
1445 
1446 	cpus = cpu_map__new_data(&event->cpu_map.data);
1447 	if (!cpus)
1448 		return -ENOMEM;
1449 
1450 	st->cpus = cpus;
1451 	return set_maps(st);
1452 }
1453 
1454 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
1455 {
1456 	int i;
1457 
1458 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
1459 	if (!config->stats)
1460 		return -1;
1461 
1462 	config->stats_num = nthreads;
1463 
1464 	for (i = 0; i < nthreads; i++)
1465 		runtime_stat__init(&config->stats[i]);
1466 
1467 	return 0;
1468 }
1469 
1470 static void runtime_stat_delete(struct perf_stat_config *config)
1471 {
1472 	int i;
1473 
1474 	if (!config->stats)
1475 		return;
1476 
1477 	for (i = 0; i < config->stats_num; i++)
1478 		runtime_stat__exit(&config->stats[i]);
1479 
1480 	free(config->stats);
1481 }
1482 
1483 static const char * const stat_report_usage[] = {
1484 	"perf stat report [<options>]",
1485 	NULL,
1486 };
1487 
1488 static struct perf_stat perf_stat = {
1489 	.tool = {
1490 		.attr		= perf_event__process_attr,
1491 		.event_update	= perf_event__process_event_update,
1492 		.thread_map	= process_thread_map_event,
1493 		.cpu_map	= process_cpu_map_event,
1494 		.stat_config	= process_stat_config_event,
1495 		.stat		= perf_event__process_stat_event,
1496 		.stat_round	= process_stat_round_event,
1497 	},
1498 	.aggr_mode = AGGR_UNSET,
1499 };
1500 
1501 static int __cmd_report(int argc, const char **argv)
1502 {
1503 	struct perf_session *session;
1504 	const struct option options[] = {
1505 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1506 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1507 		     "aggregate counts per processor socket", AGGR_SOCKET),
1508 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1509 		     "aggregate counts per physical processor core", AGGR_CORE),
1510 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1511 		     "disable CPU count aggregation", AGGR_NONE),
1512 	OPT_END()
1513 	};
1514 	struct stat st;
1515 	int ret;
1516 
1517 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
1518 
1519 	if (!input_name || !strlen(input_name)) {
1520 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1521 			input_name = "-";
1522 		else
1523 			input_name = "perf.data";
1524 	}
1525 
1526 	perf_stat.data.path = input_name;
1527 	perf_stat.data.mode = PERF_DATA_MODE_READ;
1528 
1529 	session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
1530 	if (session == NULL)
1531 		return -1;
1532 
1533 	perf_stat.session  = session;
1534 	stat_config.output = stderr;
1535 	evsel_list         = session->evlist;
1536 
1537 	ret = perf_session__process_events(session);
1538 	if (ret)
1539 		return ret;
1540 
1541 	perf_session__delete(session);
1542 	return 0;
1543 }
1544 
1545 static void setup_system_wide(int forks)
1546 {
1547 	/*
1548 	 * Make system wide (-a) the default target if
1549 	 * no target was specified and one of following
1550 	 * conditions is met:
1551 	 *
1552 	 *   - there's no workload specified
1553 	 *   - there is workload specified but all requested
1554 	 *     events are system wide events
1555 	 */
1556 	if (!target__none(&target))
1557 		return;
1558 
1559 	if (!forks)
1560 		target.system_wide = true;
1561 	else {
1562 		struct perf_evsel *counter;
1563 
1564 		evlist__for_each_entry(evsel_list, counter) {
1565 			if (!counter->system_wide)
1566 				return;
1567 		}
1568 
1569 		if (evsel_list->nr_entries)
1570 			target.system_wide = true;
1571 	}
1572 }
1573 
1574 int cmd_stat(int argc, const char **argv)
1575 {
1576 	const char * const stat_usage[] = {
1577 		"perf stat [<options>] [<command>]",
1578 		NULL
1579 	};
1580 	int status = -EINVAL, run_idx;
1581 	const char *mode;
1582 	FILE *output = stderr;
1583 	unsigned int interval, timeout;
1584 	const char * const stat_subcommands[] = { "record", "report" };
1585 
1586 	setlocale(LC_ALL, "");
1587 
1588 	evsel_list = perf_evlist__new();
1589 	if (evsel_list == NULL)
1590 		return -ENOMEM;
1591 
1592 	parse_events__shrink_config_terms();
1593 
1594 	/* String-parsing callback-based options would segfault when negated */
1595 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
1596 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
1597 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
1598 
1599 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1600 					(const char **) stat_usage,
1601 					PARSE_OPT_STOP_AT_NON_OPTION);
1602 	perf_stat__collect_metric_expr(evsel_list);
1603 	perf_stat__init_shadow_stats();
1604 
1605 	if (stat_config.csv_sep) {
1606 		stat_config.csv_output = true;
1607 		if (!strcmp(stat_config.csv_sep, "\\t"))
1608 			stat_config.csv_sep = "\t";
1609 	} else
1610 		stat_config.csv_sep = DEFAULT_SEPARATOR;
1611 
1612 	if (argc && !strncmp(argv[0], "rec", 3)) {
1613 		argc = __cmd_record(argc, argv);
1614 		if (argc < 0)
1615 			return -1;
1616 	} else if (argc && !strncmp(argv[0], "rep", 3))
1617 		return __cmd_report(argc, argv);
1618 
1619 	interval = stat_config.interval;
1620 	timeout = stat_config.timeout;
1621 
1622 	/*
1623 	 * For record command the -o is already taken care of.
1624 	 */
1625 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1626 		output = NULL;
1627 
1628 	if (output_name && output_fd) {
1629 		fprintf(stderr, "cannot use both --output and --log-fd\n");
1630 		parse_options_usage(stat_usage, stat_options, "o", 1);
1631 		parse_options_usage(NULL, stat_options, "log-fd", 0);
1632 		goto out;
1633 	}
1634 
1635 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
1636 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
1637 		goto out;
1638 	}
1639 
1640 	if (stat_config.metric_only && stat_config.run_count > 1) {
1641 		fprintf(stderr, "--metric-only is not supported with -r\n");
1642 		goto out;
1643 	}
1644 
1645 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
1646 		fprintf(stderr, "--table is only supported with -r\n");
1647 		parse_options_usage(stat_usage, stat_options, "r", 1);
1648 		parse_options_usage(NULL, stat_options, "table", 0);
1649 		goto out;
1650 	}
1651 
1652 	if (output_fd < 0) {
1653 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
1654 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1655 		goto out;
1656 	}
1657 
1658 	if (!output) {
1659 		struct timespec tm;
1660 		mode = append_file ? "a" : "w";
1661 
1662 		output = fopen(output_name, mode);
1663 		if (!output) {
1664 			perror("failed to create output file");
1665 			return -1;
1666 		}
1667 		clock_gettime(CLOCK_REALTIME, &tm);
1668 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1669 	} else if (output_fd > 0) {
1670 		mode = append_file ? "a" : "w";
1671 		output = fdopen(output_fd, mode);
1672 		if (!output) {
1673 			perror("Failed opening logfd");
1674 			return -errno;
1675 		}
1676 	}
1677 
1678 	stat_config.output = output;
1679 
1680 	/*
1681 	 * let the spreadsheet do the pretty-printing
1682 	 */
1683 	if (stat_config.csv_output) {
1684 		/* User explicitly passed -B? */
1685 		if (big_num_opt == 1) {
1686 			fprintf(stderr, "-B option not supported with -x\n");
1687 			parse_options_usage(stat_usage, stat_options, "B", 1);
1688 			parse_options_usage(NULL, stat_options, "x", 1);
1689 			goto out;
1690 		} else /* Nope, so disable big number formatting */
1691 			stat_config.big_num = false;
1692 	} else if (big_num_opt == 0) /* User passed --no-big-num */
1693 		stat_config.big_num = false;
1694 
1695 	setup_system_wide(argc);
1696 
1697 	/*
1698 	 * Display user/system times only for single
1699 	 * run and when there's specified tracee.
1700 	 */
1701 	if ((stat_config.run_count == 1) && target__none(&target))
1702 		stat_config.ru_display = true;
1703 
1704 	if (stat_config.run_count < 0) {
1705 		pr_err("Run count must be a positive number\n");
1706 		parse_options_usage(stat_usage, stat_options, "r", 1);
1707 		goto out;
1708 	} else if (stat_config.run_count == 0) {
1709 		forever = true;
1710 		stat_config.run_count = 1;
1711 	}
1712 
1713 	if (stat_config.walltime_run_table) {
1714 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
1715 		if (!stat_config.walltime_run) {
1716 			pr_err("failed to setup -r option");
1717 			goto out;
1718 		}
1719 	}
1720 
1721 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
1722 		!target__has_task(&target)) {
1723 		if (!target.system_wide || target.cpu_list) {
1724 			fprintf(stderr, "The --per-thread option is only "
1725 				"available when monitoring via -p -t -a "
1726 				"options or only --per-thread.\n");
1727 			parse_options_usage(NULL, stat_options, "p", 1);
1728 			parse_options_usage(NULL, stat_options, "t", 1);
1729 			goto out;
1730 		}
1731 	}
1732 
1733 	/*
1734 	 * no_aggr, cgroup are for system-wide only
1735 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1736 	 */
1737 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1738 	      stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1739 	    !target__has_cpu(&target)) {
1740 		fprintf(stderr, "both cgroup and no-aggregation "
1741 			"modes only available in system-wide mode\n");
1742 
1743 		parse_options_usage(stat_usage, stat_options, "G", 1);
1744 		parse_options_usage(NULL, stat_options, "A", 1);
1745 		parse_options_usage(NULL, stat_options, "a", 1);
1746 		goto out;
1747 	}
1748 
1749 	if (add_default_attributes())
1750 		goto out;
1751 
1752 	target__validate(&target);
1753 
1754 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
1755 		target.per_thread = true;
1756 
1757 	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1758 		if (target__has_task(&target)) {
1759 			pr_err("Problems finding threads of monitor\n");
1760 			parse_options_usage(stat_usage, stat_options, "p", 1);
1761 			parse_options_usage(NULL, stat_options, "t", 1);
1762 		} else if (target__has_cpu(&target)) {
1763 			perror("failed to parse CPUs map");
1764 			parse_options_usage(stat_usage, stat_options, "C", 1);
1765 			parse_options_usage(NULL, stat_options, "a", 1);
1766 		}
1767 		goto out;
1768 	}
1769 
1770 	/*
1771 	 * Initialize thread_map with comm names,
1772 	 * so we could print it out on output.
1773 	 */
1774 	if (stat_config.aggr_mode == AGGR_THREAD) {
1775 		thread_map__read_comms(evsel_list->threads);
1776 		if (target.system_wide) {
1777 			if (runtime_stat_new(&stat_config,
1778 				thread_map__nr(evsel_list->threads))) {
1779 				goto out;
1780 			}
1781 		}
1782 	}
1783 
1784 	if (stat_config.times && interval)
1785 		interval_count = true;
1786 	else if (stat_config.times && !interval) {
1787 		pr_err("interval-count option should be used together with "
1788 				"interval-print.\n");
1789 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
1790 		parse_options_usage(stat_usage, stat_options, "I", 1);
1791 		goto out;
1792 	}
1793 
1794 	if (timeout && timeout < 100) {
1795 		if (timeout < 10) {
1796 			pr_err("timeout must be >= 10ms.\n");
1797 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
1798 			goto out;
1799 		} else
1800 			pr_warning("timeout < 100ms. "
1801 				   "The overhead percentage could be high in some cases. "
1802 				   "Please proceed with caution.\n");
1803 	}
1804 	if (timeout && interval) {
1805 		pr_err("timeout option is not supported with interval-print.\n");
1806 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
1807 		parse_options_usage(stat_usage, stat_options, "I", 1);
1808 		goto out;
1809 	}
1810 
1811 	if (perf_evlist__alloc_stats(evsel_list, interval))
1812 		goto out;
1813 
1814 	if (perf_stat_init_aggr_mode())
1815 		goto out;
1816 
1817 	/*
1818 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
1819 	 * while avoiding that older tools show confusing messages.
1820 	 *
1821 	 * However for pipe sessions we need to keep it zero,
1822 	 * because script's perf_evsel__check_attr is triggered
1823 	 * by attr->sample_type != 0, and we can't run it on
1824 	 * stat sessions.
1825 	 */
1826 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
1827 
1828 	/*
1829 	 * We dont want to block the signals - that would cause
1830 	 * child tasks to inherit that and Ctrl-C would not work.
1831 	 * What we want is for Ctrl-C to work in the exec()-ed
1832 	 * task, but being ignored by perf stat itself:
1833 	 */
1834 	atexit(sig_atexit);
1835 	if (!forever)
1836 		signal(SIGINT,  skip_signal);
1837 	signal(SIGCHLD, skip_signal);
1838 	signal(SIGALRM, skip_signal);
1839 	signal(SIGABRT, skip_signal);
1840 
1841 	status = 0;
1842 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
1843 		if (stat_config.run_count != 1 && verbose > 0)
1844 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1845 				run_idx + 1);
1846 
1847 		status = run_perf_stat(argc, argv, run_idx);
1848 		if (forever && status != -1) {
1849 			print_counters(NULL, argc, argv);
1850 			perf_stat__reset_stats();
1851 		}
1852 	}
1853 
1854 	if (!forever && status != -1 && !interval)
1855 		print_counters(NULL, argc, argv);
1856 
1857 	if (STAT_RECORD) {
1858 		/*
1859 		 * We synthesize the kernel mmap record just so that older tools
1860 		 * don't emit warnings about not being able to resolve symbols
1861 		 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
1862 		 * a saner message about no samples being in the perf.data file.
1863 		 *
1864 		 * This also serves to suppress a warning about f_header.data.size == 0
1865 		 * in header.c at the moment 'perf stat record' gets introduced, which
1866 		 * is not really needed once we start adding the stat specific PERF_RECORD_
1867 		 * records, but the need to suppress the kptr_restrict messages in older
1868 		 * tools remain  -acme
1869 		 */
1870 		int fd = perf_data__fd(&perf_stat.data);
1871 		int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
1872 							     process_synthesized_event,
1873 							     &perf_stat.session->machines.host);
1874 		if (err) {
1875 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
1876 				   "older tools may produce warnings about this file\n.");
1877 		}
1878 
1879 		if (!interval) {
1880 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
1881 				pr_err("failed to write stat round event\n");
1882 		}
1883 
1884 		if (!perf_stat.data.is_pipe) {
1885 			perf_stat.session->header.data_size += perf_stat.bytes_written;
1886 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
1887 		}
1888 
1889 		perf_session__delete(perf_stat.session);
1890 	}
1891 
1892 	perf_stat__exit_aggr_mode();
1893 	perf_evlist__free_stats(evsel_list);
1894 out:
1895 	free(stat_config.walltime_run);
1896 
1897 	if (smi_cost && smi_reset)
1898 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
1899 
1900 	perf_evlist__delete(evsel_list);
1901 
1902 	runtime_stat_delete(&stat_config);
1903 
1904 	return status;
1905 }
1906