xref: /openbmc/linux/tools/perf/builtin-stat.c (revision 2fa49589)
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/drv_configs.h"
56 #include "util/color.h"
57 #include "util/stat.h"
58 #include "util/header.h"
59 #include "util/cpumap.h"
60 #include "util/thread.h"
61 #include "util/thread_map.h"
62 #include "util/counts.h"
63 #include "util/group.h"
64 #include "util/session.h"
65 #include "util/tool.h"
66 #include "util/string2.h"
67 #include "util/metricgroup.h"
68 #include "util/top.h"
69 #include "asm/bug.h"
70 
71 #include <linux/time64.h>
72 #include <api/fs/fs.h>
73 #include <errno.h>
74 #include <signal.h>
75 #include <stdlib.h>
76 #include <sys/prctl.h>
77 #include <inttypes.h>
78 #include <locale.h>
79 #include <math.h>
80 #include <sys/types.h>
81 #include <sys/stat.h>
82 #include <sys/wait.h>
83 #include <unistd.h>
84 #include <sys/time.h>
85 #include <sys/resource.h>
86 #include <sys/wait.h>
87 
88 #include "sane_ctype.h"
89 
90 #define DEFAULT_SEPARATOR	" "
91 #define FREEZE_ON_SMI_PATH	"devices/cpu/freeze_on_smi"
92 
93 static void print_counters(struct timespec *ts, int argc, const char **argv);
94 
95 /* Default events used for perf stat -T */
96 static const char *transaction_attrs = {
97 	"task-clock,"
98 	"{"
99 	"instructions,"
100 	"cycles,"
101 	"cpu/cycles-t/,"
102 	"cpu/tx-start/,"
103 	"cpu/el-start/,"
104 	"cpu/cycles-ct/"
105 	"}"
106 };
107 
108 /* More limited version when the CPU does not have all events. */
109 static const char * transaction_limited_attrs = {
110 	"task-clock,"
111 	"{"
112 	"instructions,"
113 	"cycles,"
114 	"cpu/cycles-t/,"
115 	"cpu/tx-start/"
116 	"}"
117 };
118 
119 static const char * topdown_attrs[] = {
120 	"topdown-total-slots",
121 	"topdown-slots-retired",
122 	"topdown-recovery-bubbles",
123 	"topdown-fetch-bubbles",
124 	"topdown-slots-issued",
125 	NULL,
126 };
127 
128 static const char *smi_cost_attrs = {
129 	"{"
130 	"msr/aperf/,"
131 	"msr/smi/,"
132 	"cycles"
133 	"}"
134 };
135 
136 static struct perf_evlist	*evsel_list;
137 
138 static struct target target = {
139 	.uid	= UINT_MAX,
140 };
141 
142 #define METRIC_ONLY_LEN 20
143 
144 static volatile pid_t		child_pid			= -1;
145 static int			detailed_run			=  0;
146 static bool			transaction_run;
147 static bool			topdown_run			= false;
148 static bool			smi_cost			= false;
149 static bool			smi_reset			= false;
150 static int			big_num_opt			=  -1;
151 static bool			group				= false;
152 static const char		*pre_cmd			= NULL;
153 static const char		*post_cmd			= NULL;
154 static bool			sync_run			= false;
155 static bool			forever				= false;
156 static bool			force_metric_only		= false;
157 static struct timespec		ref_time;
158 static bool			append_file;
159 static bool			interval_count;
160 static const char		*output_name;
161 static int			output_fd;
162 
163 struct perf_stat {
164 	bool			 record;
165 	struct perf_data	 data;
166 	struct perf_session	*session;
167 	u64			 bytes_written;
168 	struct perf_tool	 tool;
169 	bool			 maps_allocated;
170 	struct cpu_map		*cpus;
171 	struct thread_map	*threads;
172 	enum aggr_mode		 aggr_mode;
173 };
174 
175 static struct perf_stat		perf_stat;
176 #define STAT_RECORD		perf_stat.record
177 
178 static volatile int done = 0;
179 
180 static struct perf_stat_config stat_config = {
181 	.aggr_mode		= AGGR_GLOBAL,
182 	.scale			= true,
183 	.unit_width		= 4, /* strlen("unit") */
184 	.run_count		= 1,
185 	.metric_only_len	= METRIC_ONLY_LEN,
186 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
187 	.big_num		= true,
188 };
189 
190 static inline void diff_timespec(struct timespec *r, struct timespec *a,
191 				 struct timespec *b)
192 {
193 	r->tv_sec = a->tv_sec - b->tv_sec;
194 	if (a->tv_nsec < b->tv_nsec) {
195 		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
196 		r->tv_sec--;
197 	} else {
198 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
199 	}
200 }
201 
202 static void perf_stat__reset_stats(void)
203 {
204 	int i;
205 
206 	perf_evlist__reset_stats(evsel_list);
207 	perf_stat__reset_shadow_stats();
208 
209 	for (i = 0; i < stat_config.stats_num; i++)
210 		perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
211 }
212 
213 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
214 				     union perf_event *event,
215 				     struct perf_sample *sample __maybe_unused,
216 				     struct machine *machine __maybe_unused)
217 {
218 	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
219 		pr_err("failed to write perf data, error: %m\n");
220 		return -1;
221 	}
222 
223 	perf_stat.bytes_written += event->header.size;
224 	return 0;
225 }
226 
227 static int write_stat_round_event(u64 tm, u64 type)
228 {
229 	return perf_event__synthesize_stat_round(NULL, tm, type,
230 						 process_synthesized_event,
231 						 NULL);
232 }
233 
234 #define WRITE_STAT_ROUND_EVENT(time, interval) \
235 	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
236 
237 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
238 
239 static int
240 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
241 			     struct perf_counts_values *count)
242 {
243 	struct perf_sample_id *sid = SID(counter, cpu, thread);
244 
245 	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
246 					   process_synthesized_event, NULL);
247 }
248 
249 /*
250  * Read out the results of a single counter:
251  * do not aggregate counts across CPUs in system-wide mode
252  */
253 static int read_counter(struct perf_evsel *counter)
254 {
255 	int nthreads = thread_map__nr(evsel_list->threads);
256 	int ncpus, cpu, thread;
257 
258 	if (target__has_cpu(&target) && !target__has_per_thread(&target))
259 		ncpus = perf_evsel__nr_cpus(counter);
260 	else
261 		ncpus = 1;
262 
263 	if (!counter->supported)
264 		return -ENOENT;
265 
266 	if (counter->system_wide)
267 		nthreads = 1;
268 
269 	for (thread = 0; thread < nthreads; thread++) {
270 		for (cpu = 0; cpu < ncpus; cpu++) {
271 			struct perf_counts_values *count;
272 
273 			count = perf_counts(counter->counts, cpu, thread);
274 
275 			/*
276 			 * The leader's group read loads data into its group members
277 			 * (via perf_evsel__read_counter) and sets threir count->loaded.
278 			 */
279 			if (!count->loaded &&
280 			    perf_evsel__read_counter(counter, cpu, thread)) {
281 				counter->counts->scaled = -1;
282 				perf_counts(counter->counts, cpu, thread)->ena = 0;
283 				perf_counts(counter->counts, cpu, thread)->run = 0;
284 				return -1;
285 			}
286 
287 			count->loaded = false;
288 
289 			if (STAT_RECORD) {
290 				if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
291 					pr_err("failed to write stat event\n");
292 					return -1;
293 				}
294 			}
295 
296 			if (verbose > 1) {
297 				fprintf(stat_config.output,
298 					"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
299 						perf_evsel__name(counter),
300 						cpu,
301 						count->val, count->ena, count->run);
302 			}
303 		}
304 	}
305 
306 	return 0;
307 }
308 
309 static void read_counters(void)
310 {
311 	struct perf_evsel *counter;
312 	int ret;
313 
314 	evlist__for_each_entry(evsel_list, counter) {
315 		ret = read_counter(counter);
316 		if (ret)
317 			pr_debug("failed to read counter %s\n", counter->name);
318 
319 		if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
320 			pr_warning("failed to process counter %s\n", counter->name);
321 	}
322 }
323 
324 static void process_interval(void)
325 {
326 	struct timespec ts, rs;
327 
328 	read_counters();
329 
330 	clock_gettime(CLOCK_MONOTONIC, &ts);
331 	diff_timespec(&rs, &ts, &ref_time);
332 
333 	if (STAT_RECORD) {
334 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
335 			pr_err("failed to write stat round event\n");
336 	}
337 
338 	init_stats(&walltime_nsecs_stats);
339 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
340 	print_counters(&rs, 0, NULL);
341 }
342 
343 static void enable_counters(void)
344 {
345 	if (stat_config.initial_delay)
346 		usleep(stat_config.initial_delay * USEC_PER_MSEC);
347 
348 	/*
349 	 * We need to enable counters only if:
350 	 * - we don't have tracee (attaching to task or cpu)
351 	 * - we have initial delay configured
352 	 */
353 	if (!target__none(&target) || stat_config.initial_delay)
354 		perf_evlist__enable(evsel_list);
355 }
356 
357 static void disable_counters(void)
358 {
359 	/*
360 	 * If we don't have tracee (attaching to task or cpu), counters may
361 	 * still be running. To get accurate group ratios, we must stop groups
362 	 * from counting before reading their constituent counters.
363 	 */
364 	if (!target__none(&target))
365 		perf_evlist__disable(evsel_list);
366 }
367 
368 static volatile int workload_exec_errno;
369 
370 /*
371  * perf_evlist__prepare_workload will send a SIGUSR1
372  * if the fork fails, since we asked by setting its
373  * want_signal to true.
374  */
375 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
376 					void *ucontext __maybe_unused)
377 {
378 	workload_exec_errno = info->si_value.sival_int;
379 }
380 
381 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
382 {
383 	return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
384 }
385 
386 static bool is_target_alive(struct target *_target,
387 			    struct thread_map *threads)
388 {
389 	struct stat st;
390 	int i;
391 
392 	if (!target__has_task(_target))
393 		return true;
394 
395 	for (i = 0; i < threads->nr; i++) {
396 		char path[PATH_MAX];
397 
398 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
399 			  threads->map[i].pid);
400 
401 		if (!stat(path, &st))
402 			return true;
403 	}
404 
405 	return false;
406 }
407 
408 static int __run_perf_stat(int argc, const char **argv, int run_idx)
409 {
410 	int interval = stat_config.interval;
411 	int times = stat_config.times;
412 	int timeout = stat_config.timeout;
413 	char msg[BUFSIZ];
414 	unsigned long long t0, t1;
415 	struct perf_evsel *counter;
416 	struct timespec ts;
417 	size_t l;
418 	int status = 0;
419 	const bool forks = (argc > 0);
420 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
421 	struct perf_evsel_config_term *err_term;
422 
423 	if (interval) {
424 		ts.tv_sec  = interval / USEC_PER_MSEC;
425 		ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
426 	} else if (timeout) {
427 		ts.tv_sec  = timeout / USEC_PER_MSEC;
428 		ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
429 	} else {
430 		ts.tv_sec  = 1;
431 		ts.tv_nsec = 0;
432 	}
433 
434 	if (forks) {
435 		if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
436 						  workload_exec_failed_signal) < 0) {
437 			perror("failed to prepare workload");
438 			return -1;
439 		}
440 		child_pid = evsel_list->workload.pid;
441 	}
442 
443 	if (group)
444 		perf_evlist__set_leader(evsel_list);
445 
446 	evlist__for_each_entry(evsel_list, counter) {
447 try_again:
448 		if (create_perf_stat_counter(counter, &stat_config, &target) < 0) {
449 
450 			/* Weak group failed. Reset the group. */
451 			if ((errno == EINVAL || errno == EBADF) &&
452 			    counter->leader != counter &&
453 			    counter->weak_group) {
454 				counter = perf_evlist__reset_weak_group(evsel_list, counter);
455 				goto try_again;
456 			}
457 
458 			/*
459 			 * PPC returns ENXIO for HW counters until 2.6.37
460 			 * (behavior changed with commit b0a873e).
461 			 */
462 			if (errno == EINVAL || errno == ENOSYS ||
463 			    errno == ENOENT || errno == EOPNOTSUPP ||
464 			    errno == ENXIO) {
465 				if (verbose > 0)
466 					ui__warning("%s event is not supported by the kernel.\n",
467 						    perf_evsel__name(counter));
468 				counter->supported = false;
469 
470 				if ((counter->leader != counter) ||
471 				    !(counter->leader->nr_members > 1))
472 					continue;
473 			} else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
474                                 if (verbose > 0)
475                                         ui__warning("%s\n", msg);
476                                 goto try_again;
477 			} else if (target__has_per_thread(&target) &&
478 				   evsel_list->threads &&
479 				   evsel_list->threads->err_thread != -1) {
480 				/*
481 				 * For global --per-thread case, skip current
482 				 * error thread.
483 				 */
484 				if (!thread_map__remove(evsel_list->threads,
485 							evsel_list->threads->err_thread)) {
486 					evsel_list->threads->err_thread = -1;
487 					goto try_again;
488 				}
489 			}
490 
491 			perf_evsel__open_strerror(counter, &target,
492 						  errno, msg, sizeof(msg));
493 			ui__error("%s\n", msg);
494 
495 			if (child_pid != -1)
496 				kill(child_pid, SIGTERM);
497 
498 			return -1;
499 		}
500 		counter->supported = true;
501 
502 		l = strlen(counter->unit);
503 		if (l > stat_config.unit_width)
504 			stat_config.unit_width = l;
505 
506 		if (perf_evsel__should_store_id(counter) &&
507 		    perf_evsel__store_ids(counter, evsel_list))
508 			return -1;
509 	}
510 
511 	if (perf_evlist__apply_filters(evsel_list, &counter)) {
512 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
513 			counter->filter, perf_evsel__name(counter), errno,
514 			str_error_r(errno, msg, sizeof(msg)));
515 		return -1;
516 	}
517 
518 	if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
519 		pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
520 		      err_term->val.drv_cfg, perf_evsel__name(counter), errno,
521 		      str_error_r(errno, msg, sizeof(msg)));
522 		return -1;
523 	}
524 
525 	if (STAT_RECORD) {
526 		int err, fd = perf_data__fd(&perf_stat.data);
527 
528 		if (is_pipe) {
529 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
530 		} else {
531 			err = perf_session__write_header(perf_stat.session, evsel_list,
532 							 fd, false);
533 		}
534 
535 		if (err < 0)
536 			return err;
537 
538 		err = perf_stat_synthesize_config(&stat_config, NULL, evsel_list,
539 						  process_synthesized_event, is_pipe);
540 		if (err < 0)
541 			return err;
542 	}
543 
544 	/*
545 	 * Enable counters and exec the command:
546 	 */
547 	t0 = rdclock();
548 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
549 
550 	if (forks) {
551 		perf_evlist__start_workload(evsel_list);
552 		enable_counters();
553 
554 		if (interval || timeout) {
555 			while (!waitpid(child_pid, &status, WNOHANG)) {
556 				nanosleep(&ts, NULL);
557 				if (timeout)
558 					break;
559 				process_interval();
560 				if (interval_count && !(--times))
561 					break;
562 			}
563 		}
564 		if (child_pid != -1)
565 			wait4(child_pid, &status, 0, &stat_config.ru_data);
566 
567 		if (workload_exec_errno) {
568 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
569 			pr_err("Workload failed: %s\n", emsg);
570 			return -1;
571 		}
572 
573 		if (WIFSIGNALED(status))
574 			psignal(WTERMSIG(status), argv[0]);
575 	} else {
576 		enable_counters();
577 		while (!done) {
578 			nanosleep(&ts, NULL);
579 			if (!is_target_alive(&target, evsel_list->threads))
580 				break;
581 			if (timeout)
582 				break;
583 			if (interval) {
584 				process_interval();
585 				if (interval_count && !(--times))
586 					break;
587 			}
588 		}
589 	}
590 
591 	disable_counters();
592 
593 	t1 = rdclock();
594 
595 	if (stat_config.walltime_run_table)
596 		stat_config.walltime_run[run_idx] = t1 - t0;
597 
598 	update_stats(&walltime_nsecs_stats, t1 - t0);
599 
600 	/*
601 	 * Closing a group leader splits the group, and as we only disable
602 	 * group leaders, results in remaining events becoming enabled. To
603 	 * avoid arbitrary skew, we must read all counters before closing any
604 	 * group leaders.
605 	 */
606 	read_counters();
607 	perf_evlist__close(evsel_list);
608 
609 	return WEXITSTATUS(status);
610 }
611 
612 static int run_perf_stat(int argc, const char **argv, int run_idx)
613 {
614 	int ret;
615 
616 	if (pre_cmd) {
617 		ret = system(pre_cmd);
618 		if (ret)
619 			return ret;
620 	}
621 
622 	if (sync_run)
623 		sync();
624 
625 	ret = __run_perf_stat(argc, argv, run_idx);
626 	if (ret)
627 		return ret;
628 
629 	if (post_cmd) {
630 		ret = system(post_cmd);
631 		if (ret)
632 			return ret;
633 	}
634 
635 	return ret;
636 }
637 
638 static void print_counters(struct timespec *ts, int argc, const char **argv)
639 {
640 	/* Do not print anything if we record to the pipe. */
641 	if (STAT_RECORD && perf_stat.data.is_pipe)
642 		return;
643 
644 	perf_evlist__print_counters(evsel_list, &stat_config, &target,
645 				    ts, argc, argv);
646 }
647 
648 static volatile int signr = -1;
649 
650 static void skip_signal(int signo)
651 {
652 	if ((child_pid == -1) || stat_config.interval)
653 		done = 1;
654 
655 	signr = signo;
656 	/*
657 	 * render child_pid harmless
658 	 * won't send SIGTERM to a random
659 	 * process in case of race condition
660 	 * and fast PID recycling
661 	 */
662 	child_pid = -1;
663 }
664 
665 static void sig_atexit(void)
666 {
667 	sigset_t set, oset;
668 
669 	/*
670 	 * avoid race condition with SIGCHLD handler
671 	 * in skip_signal() which is modifying child_pid
672 	 * goal is to avoid send SIGTERM to a random
673 	 * process
674 	 */
675 	sigemptyset(&set);
676 	sigaddset(&set, SIGCHLD);
677 	sigprocmask(SIG_BLOCK, &set, &oset);
678 
679 	if (child_pid != -1)
680 		kill(child_pid, SIGTERM);
681 
682 	sigprocmask(SIG_SETMASK, &oset, NULL);
683 
684 	if (signr == -1)
685 		return;
686 
687 	signal(signr, SIG_DFL);
688 	kill(getpid(), signr);
689 }
690 
691 static int stat__set_big_num(const struct option *opt __maybe_unused,
692 			     const char *s __maybe_unused, int unset)
693 {
694 	big_num_opt = unset ? 0 : 1;
695 	return 0;
696 }
697 
698 static int enable_metric_only(const struct option *opt __maybe_unused,
699 			      const char *s __maybe_unused, int unset)
700 {
701 	force_metric_only = true;
702 	stat_config.metric_only = !unset;
703 	return 0;
704 }
705 
706 static int parse_metric_groups(const struct option *opt,
707 			       const char *str,
708 			       int unset __maybe_unused)
709 {
710 	return metricgroup__parse_groups(opt, str, &stat_config.metric_events);
711 }
712 
713 static struct option stat_options[] = {
714 	OPT_BOOLEAN('T', "transaction", &transaction_run,
715 		    "hardware transaction statistics"),
716 	OPT_CALLBACK('e', "event", &evsel_list, "event",
717 		     "event selector. use 'perf list' to list available events",
718 		     parse_events_option),
719 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
720 		     "event filter", parse_filter),
721 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
722 		    "child tasks do not inherit counters"),
723 	OPT_STRING('p', "pid", &target.pid, "pid",
724 		   "stat events on existing process id"),
725 	OPT_STRING('t', "tid", &target.tid, "tid",
726 		   "stat events on existing thread id"),
727 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
728 		    "system-wide collection from all CPUs"),
729 	OPT_BOOLEAN('g', "group", &group,
730 		    "put the counters into a counter group"),
731 	OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
732 	OPT_INCR('v', "verbose", &verbose,
733 		    "be more verbose (show counter open errors, etc)"),
734 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
735 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
736 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
737 		    "display details about each run (only with -r option)"),
738 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
739 		    "null run - dont start any counters"),
740 	OPT_INCR('d', "detailed", &detailed_run,
741 		    "detailed run - start a lot of events"),
742 	OPT_BOOLEAN('S', "sync", &sync_run,
743 		    "call sync() before starting a run"),
744 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
745 			   "print large numbers with thousands\' separators",
746 			   stat__set_big_num),
747 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
748 		    "list of cpus to monitor in system-wide"),
749 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
750 		    "disable CPU count aggregation", AGGR_NONE),
751 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
752 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
753 		   "print counts with custom separator"),
754 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
755 		     "monitor event in cgroup name only", parse_cgroups),
756 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
757 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
758 	OPT_INTEGER(0, "log-fd", &output_fd,
759 		    "log output to fd, instead of stderr"),
760 	OPT_STRING(0, "pre", &pre_cmd, "command",
761 			"command to run prior to the measured command"),
762 	OPT_STRING(0, "post", &post_cmd, "command",
763 			"command to run after to the measured command"),
764 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
765 		    "print counts at regular interval in ms "
766 		    "(overhead is possible for values <= 100ms)"),
767 	OPT_INTEGER(0, "interval-count", &stat_config.times,
768 		    "print counts for fixed number of times"),
769 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
770 		    "clear screen in between new interval"),
771 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
772 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
773 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
774 		     "aggregate counts per processor socket", AGGR_SOCKET),
775 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
776 		     "aggregate counts per physical processor core", AGGR_CORE),
777 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
778 		     "aggregate counts per thread", AGGR_THREAD),
779 	OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
780 		     "ms to wait before starting measurement after program start"),
781 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
782 			"Only print computed metrics. No raw values", enable_metric_only),
783 	OPT_BOOLEAN(0, "topdown", &topdown_run,
784 			"measure topdown level 1 statistics"),
785 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
786 			"measure SMI cost"),
787 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
788 		     "monitor specified metrics or metric groups (separated by ,)",
789 		     parse_metric_groups),
790 	OPT_END()
791 };
792 
793 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
794 				 struct cpu_map *map, int cpu)
795 {
796 	return cpu_map__get_socket(map, cpu, NULL);
797 }
798 
799 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
800 			       struct cpu_map *map, int cpu)
801 {
802 	return cpu_map__get_core(map, cpu, NULL);
803 }
804 
805 static int cpu_map__get_max(struct cpu_map *map)
806 {
807 	int i, max = -1;
808 
809 	for (i = 0; i < map->nr; i++) {
810 		if (map->map[i] > max)
811 			max = map->map[i];
812 	}
813 
814 	return max;
815 }
816 
817 static int perf_stat__get_aggr(struct perf_stat_config *config,
818 			       aggr_get_id_t get_id, struct cpu_map *map, int idx)
819 {
820 	int cpu;
821 
822 	if (idx >= map->nr)
823 		return -1;
824 
825 	cpu = map->map[idx];
826 
827 	if (config->cpus_aggr_map->map[cpu] == -1)
828 		config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
829 
830 	return config->cpus_aggr_map->map[cpu];
831 }
832 
833 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
834 					struct cpu_map *map, int idx)
835 {
836 	return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
837 }
838 
839 static int perf_stat__get_core_cached(struct perf_stat_config *config,
840 				      struct cpu_map *map, int idx)
841 {
842 	return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
843 }
844 
845 static int perf_stat_init_aggr_mode(void)
846 {
847 	int nr;
848 
849 	switch (stat_config.aggr_mode) {
850 	case AGGR_SOCKET:
851 		if (cpu_map__build_socket_map(evsel_list->cpus, &stat_config.aggr_map)) {
852 			perror("cannot build socket map");
853 			return -1;
854 		}
855 		stat_config.aggr_get_id = perf_stat__get_socket_cached;
856 		break;
857 	case AGGR_CORE:
858 		if (cpu_map__build_core_map(evsel_list->cpus, &stat_config.aggr_map)) {
859 			perror("cannot build core map");
860 			return -1;
861 		}
862 		stat_config.aggr_get_id = perf_stat__get_core_cached;
863 		break;
864 	case AGGR_NONE:
865 	case AGGR_GLOBAL:
866 	case AGGR_THREAD:
867 	case AGGR_UNSET:
868 	default:
869 		break;
870 	}
871 
872 	/*
873 	 * The evsel_list->cpus is the base we operate on,
874 	 * taking the highest cpu number to be the size of
875 	 * the aggregation translate cpumap.
876 	 */
877 	nr = cpu_map__get_max(evsel_list->cpus);
878 	stat_config.cpus_aggr_map = cpu_map__empty_new(nr + 1);
879 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
880 }
881 
882 static void perf_stat__exit_aggr_mode(void)
883 {
884 	cpu_map__put(stat_config.aggr_map);
885 	cpu_map__put(stat_config.cpus_aggr_map);
886 	stat_config.aggr_map = NULL;
887 	stat_config.cpus_aggr_map = NULL;
888 }
889 
890 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
891 {
892 	int cpu;
893 
894 	if (idx > map->nr)
895 		return -1;
896 
897 	cpu = map->map[idx];
898 
899 	if (cpu >= env->nr_cpus_avail)
900 		return -1;
901 
902 	return cpu;
903 }
904 
905 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
906 {
907 	struct perf_env *env = data;
908 	int cpu = perf_env__get_cpu(env, map, idx);
909 
910 	return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
911 }
912 
913 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
914 {
915 	struct perf_env *env = data;
916 	int core = -1, cpu = perf_env__get_cpu(env, map, idx);
917 
918 	if (cpu != -1) {
919 		int socket_id = env->cpu[cpu].socket_id;
920 
921 		/*
922 		 * Encode socket in upper 16 bits
923 		 * core_id is relative to socket, and
924 		 * we need a global id. So we combine
925 		 * socket + core id.
926 		 */
927 		core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
928 	}
929 
930 	return core;
931 }
932 
933 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
934 				      struct cpu_map **sockp)
935 {
936 	return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
937 }
938 
939 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
940 				    struct cpu_map **corep)
941 {
942 	return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
943 }
944 
945 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
946 				      struct cpu_map *map, int idx)
947 {
948 	return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
949 }
950 
951 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
952 				    struct cpu_map *map, int idx)
953 {
954 	return perf_env__get_core(map, idx, &perf_stat.session->header.env);
955 }
956 
957 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
958 {
959 	struct perf_env *env = &st->session->header.env;
960 
961 	switch (stat_config.aggr_mode) {
962 	case AGGR_SOCKET:
963 		if (perf_env__build_socket_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
964 			perror("cannot build socket map");
965 			return -1;
966 		}
967 		stat_config.aggr_get_id = perf_stat__get_socket_file;
968 		break;
969 	case AGGR_CORE:
970 		if (perf_env__build_core_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
971 			perror("cannot build core map");
972 			return -1;
973 		}
974 		stat_config.aggr_get_id = perf_stat__get_core_file;
975 		break;
976 	case AGGR_NONE:
977 	case AGGR_GLOBAL:
978 	case AGGR_THREAD:
979 	case AGGR_UNSET:
980 	default:
981 		break;
982 	}
983 
984 	return 0;
985 }
986 
987 static int topdown_filter_events(const char **attr, char **str, bool use_group)
988 {
989 	int off = 0;
990 	int i;
991 	int len = 0;
992 	char *s;
993 
994 	for (i = 0; attr[i]; i++) {
995 		if (pmu_have_event("cpu", attr[i])) {
996 			len += strlen(attr[i]) + 1;
997 			attr[i - off] = attr[i];
998 		} else
999 			off++;
1000 	}
1001 	attr[i - off] = NULL;
1002 
1003 	*str = malloc(len + 1 + 2);
1004 	if (!*str)
1005 		return -1;
1006 	s = *str;
1007 	if (i - off == 0) {
1008 		*s = 0;
1009 		return 0;
1010 	}
1011 	if (use_group)
1012 		*s++ = '{';
1013 	for (i = 0; attr[i]; i++) {
1014 		strcpy(s, attr[i]);
1015 		s += strlen(s);
1016 		*s++ = ',';
1017 	}
1018 	if (use_group) {
1019 		s[-1] = '}';
1020 		*s = 0;
1021 	} else
1022 		s[-1] = 0;
1023 	return 0;
1024 }
1025 
1026 __weak bool arch_topdown_check_group(bool *warn)
1027 {
1028 	*warn = false;
1029 	return false;
1030 }
1031 
1032 __weak void arch_topdown_group_warn(void)
1033 {
1034 }
1035 
1036 /*
1037  * Add default attributes, if there were no attributes specified or
1038  * if -d/--detailed, -d -d or -d -d -d is used:
1039  */
1040 static int add_default_attributes(void)
1041 {
1042 	int err;
1043 	struct perf_event_attr default_attrs0[] = {
1044 
1045   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1046   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1047   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1048   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1049 
1050   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1051 };
1052 	struct perf_event_attr frontend_attrs[] = {
1053   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1054 };
1055 	struct perf_event_attr backend_attrs[] = {
1056   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1057 };
1058 	struct perf_event_attr default_attrs1[] = {
1059   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1060   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1061   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1062 
1063 };
1064 
1065 /*
1066  * Detailed stats (-d), covering the L1 and last level data caches:
1067  */
1068 	struct perf_event_attr detailed_attrs[] = {
1069 
1070   { .type = PERF_TYPE_HW_CACHE,
1071     .config =
1072 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1073 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1074 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1075 
1076   { .type = PERF_TYPE_HW_CACHE,
1077     .config =
1078 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1079 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1080 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1081 
1082   { .type = PERF_TYPE_HW_CACHE,
1083     .config =
1084 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1085 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1086 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1087 
1088   { .type = PERF_TYPE_HW_CACHE,
1089     .config =
1090 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1091 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1092 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1093 };
1094 
1095 /*
1096  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1097  */
1098 	struct perf_event_attr very_detailed_attrs[] = {
1099 
1100   { .type = PERF_TYPE_HW_CACHE,
1101     .config =
1102 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1103 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1104 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1105 
1106   { .type = PERF_TYPE_HW_CACHE,
1107     .config =
1108 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1109 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1110 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1111 
1112   { .type = PERF_TYPE_HW_CACHE,
1113     .config =
1114 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1115 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1116 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1117 
1118   { .type = PERF_TYPE_HW_CACHE,
1119     .config =
1120 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1121 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1122 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1123 
1124   { .type = PERF_TYPE_HW_CACHE,
1125     .config =
1126 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1127 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1128 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1129 
1130   { .type = PERF_TYPE_HW_CACHE,
1131     .config =
1132 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1133 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1134 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1135 
1136 };
1137 
1138 /*
1139  * Very, very detailed stats (-d -d -d), adding prefetch events:
1140  */
1141 	struct perf_event_attr very_very_detailed_attrs[] = {
1142 
1143   { .type = PERF_TYPE_HW_CACHE,
1144     .config =
1145 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1146 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1147 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1148 
1149   { .type = PERF_TYPE_HW_CACHE,
1150     .config =
1151 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1152 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1153 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1154 };
1155 	struct parse_events_error errinfo;
1156 
1157 	/* Set attrs if no event is selected and !null_run: */
1158 	if (stat_config.null_run)
1159 		return 0;
1160 
1161 	if (transaction_run) {
1162 		/* Handle -T as -M transaction. Once platform specific metrics
1163 		 * support has been added to the json files, all archictures
1164 		 * will use this approach. To determine transaction support
1165 		 * on an architecture test for such a metric name.
1166 		 */
1167 		if (metricgroup__has_metric("transaction")) {
1168 			struct option opt = { .value = &evsel_list };
1169 
1170 			return metricgroup__parse_groups(&opt, "transaction",
1171 							 &stat_config.metric_events);
1172 		}
1173 
1174 		if (pmu_have_event("cpu", "cycles-ct") &&
1175 		    pmu_have_event("cpu", "el-start"))
1176 			err = parse_events(evsel_list, transaction_attrs,
1177 					   &errinfo);
1178 		else
1179 			err = parse_events(evsel_list,
1180 					   transaction_limited_attrs,
1181 					   &errinfo);
1182 		if (err) {
1183 			fprintf(stderr, "Cannot set up transaction events\n");
1184 			parse_events_print_error(&errinfo, transaction_attrs);
1185 			return -1;
1186 		}
1187 		return 0;
1188 	}
1189 
1190 	if (smi_cost) {
1191 		int smi;
1192 
1193 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1194 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1195 			return -1;
1196 		}
1197 
1198 		if (!smi) {
1199 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1200 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1201 				return -1;
1202 			}
1203 			smi_reset = true;
1204 		}
1205 
1206 		if (pmu_have_event("msr", "aperf") &&
1207 		    pmu_have_event("msr", "smi")) {
1208 			if (!force_metric_only)
1209 				stat_config.metric_only = true;
1210 			err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1211 		} else {
1212 			fprintf(stderr, "To measure SMI cost, it needs "
1213 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1214 			parse_events_print_error(&errinfo, smi_cost_attrs);
1215 			return -1;
1216 		}
1217 		if (err) {
1218 			fprintf(stderr, "Cannot set up SMI cost events\n");
1219 			return -1;
1220 		}
1221 		return 0;
1222 	}
1223 
1224 	if (topdown_run) {
1225 		char *str = NULL;
1226 		bool warn = false;
1227 
1228 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1229 		    stat_config.aggr_mode != AGGR_CORE) {
1230 			pr_err("top down event configuration requires --per-core mode\n");
1231 			return -1;
1232 		}
1233 		stat_config.aggr_mode = AGGR_CORE;
1234 		if (nr_cgroups || !target__has_cpu(&target)) {
1235 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1236 			return -1;
1237 		}
1238 
1239 		if (!force_metric_only)
1240 			stat_config.metric_only = true;
1241 		if (topdown_filter_events(topdown_attrs, &str,
1242 				arch_topdown_check_group(&warn)) < 0) {
1243 			pr_err("Out of memory\n");
1244 			return -1;
1245 		}
1246 		if (topdown_attrs[0] && str) {
1247 			if (warn)
1248 				arch_topdown_group_warn();
1249 			err = parse_events(evsel_list, str, &errinfo);
1250 			if (err) {
1251 				fprintf(stderr,
1252 					"Cannot set up top down events %s: %d\n",
1253 					str, err);
1254 				free(str);
1255 				parse_events_print_error(&errinfo, str);
1256 				return -1;
1257 			}
1258 		} else {
1259 			fprintf(stderr, "System does not support topdown\n");
1260 			return -1;
1261 		}
1262 		free(str);
1263 	}
1264 
1265 	if (!evsel_list->nr_entries) {
1266 		if (target__has_cpu(&target))
1267 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1268 
1269 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1270 			return -1;
1271 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1272 			if (perf_evlist__add_default_attrs(evsel_list,
1273 						frontend_attrs) < 0)
1274 				return -1;
1275 		}
1276 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1277 			if (perf_evlist__add_default_attrs(evsel_list,
1278 						backend_attrs) < 0)
1279 				return -1;
1280 		}
1281 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1282 			return -1;
1283 	}
1284 
1285 	/* Detailed events get appended to the event list: */
1286 
1287 	if (detailed_run <  1)
1288 		return 0;
1289 
1290 	/* Append detailed run extra attributes: */
1291 	if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1292 		return -1;
1293 
1294 	if (detailed_run < 2)
1295 		return 0;
1296 
1297 	/* Append very detailed run extra attributes: */
1298 	if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1299 		return -1;
1300 
1301 	if (detailed_run < 3)
1302 		return 0;
1303 
1304 	/* Append very, very detailed run extra attributes: */
1305 	return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1306 }
1307 
1308 static const char * const stat_record_usage[] = {
1309 	"perf stat record [<options>]",
1310 	NULL,
1311 };
1312 
1313 static void init_features(struct perf_session *session)
1314 {
1315 	int feat;
1316 
1317 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1318 		perf_header__set_feat(&session->header, feat);
1319 
1320 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1321 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1322 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1323 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1324 }
1325 
1326 static int __cmd_record(int argc, const char **argv)
1327 {
1328 	struct perf_session *session;
1329 	struct perf_data *data = &perf_stat.data;
1330 
1331 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
1332 			     PARSE_OPT_STOP_AT_NON_OPTION);
1333 
1334 	if (output_name)
1335 		data->file.path = output_name;
1336 
1337 	if (stat_config.run_count != 1 || forever) {
1338 		pr_err("Cannot use -r option with perf stat record.\n");
1339 		return -1;
1340 	}
1341 
1342 	session = perf_session__new(data, false, NULL);
1343 	if (session == NULL) {
1344 		pr_err("Perf session creation failed.\n");
1345 		return -1;
1346 	}
1347 
1348 	init_features(session);
1349 
1350 	session->evlist   = evsel_list;
1351 	perf_stat.session = session;
1352 	perf_stat.record  = true;
1353 	return argc;
1354 }
1355 
1356 static int process_stat_round_event(struct perf_session *session,
1357 				    union perf_event *event)
1358 {
1359 	struct stat_round_event *stat_round = &event->stat_round;
1360 	struct perf_evsel *counter;
1361 	struct timespec tsh, *ts = NULL;
1362 	const char **argv = session->header.env.cmdline_argv;
1363 	int argc = session->header.env.nr_cmdline;
1364 
1365 	evlist__for_each_entry(evsel_list, counter)
1366 		perf_stat_process_counter(&stat_config, counter);
1367 
1368 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1369 		update_stats(&walltime_nsecs_stats, stat_round->time);
1370 
1371 	if (stat_config.interval && stat_round->time) {
1372 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
1373 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1374 		ts = &tsh;
1375 	}
1376 
1377 	print_counters(ts, argc, argv);
1378 	return 0;
1379 }
1380 
1381 static
1382 int process_stat_config_event(struct perf_session *session,
1383 			      union perf_event *event)
1384 {
1385 	struct perf_tool *tool = session->tool;
1386 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1387 
1388 	perf_event__read_stat_config(&stat_config, &event->stat_config);
1389 
1390 	if (cpu_map__empty(st->cpus)) {
1391 		if (st->aggr_mode != AGGR_UNSET)
1392 			pr_warning("warning: processing task data, aggregation mode not set\n");
1393 		return 0;
1394 	}
1395 
1396 	if (st->aggr_mode != AGGR_UNSET)
1397 		stat_config.aggr_mode = st->aggr_mode;
1398 
1399 	if (perf_stat.data.is_pipe)
1400 		perf_stat_init_aggr_mode();
1401 	else
1402 		perf_stat_init_aggr_mode_file(st);
1403 
1404 	return 0;
1405 }
1406 
1407 static int set_maps(struct perf_stat *st)
1408 {
1409 	if (!st->cpus || !st->threads)
1410 		return 0;
1411 
1412 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1413 		return -EINVAL;
1414 
1415 	perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
1416 
1417 	if (perf_evlist__alloc_stats(evsel_list, true))
1418 		return -ENOMEM;
1419 
1420 	st->maps_allocated = true;
1421 	return 0;
1422 }
1423 
1424 static
1425 int process_thread_map_event(struct perf_session *session,
1426 			     union perf_event *event)
1427 {
1428 	struct perf_tool *tool = session->tool;
1429 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1430 
1431 	if (st->threads) {
1432 		pr_warning("Extra thread map event, ignoring.\n");
1433 		return 0;
1434 	}
1435 
1436 	st->threads = thread_map__new_event(&event->thread_map);
1437 	if (!st->threads)
1438 		return -ENOMEM;
1439 
1440 	return set_maps(st);
1441 }
1442 
1443 static
1444 int process_cpu_map_event(struct perf_session *session,
1445 			  union perf_event *event)
1446 {
1447 	struct perf_tool *tool = session->tool;
1448 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1449 	struct cpu_map *cpus;
1450 
1451 	if (st->cpus) {
1452 		pr_warning("Extra cpu map event, ignoring.\n");
1453 		return 0;
1454 	}
1455 
1456 	cpus = cpu_map__new_data(&event->cpu_map.data);
1457 	if (!cpus)
1458 		return -ENOMEM;
1459 
1460 	st->cpus = cpus;
1461 	return set_maps(st);
1462 }
1463 
1464 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
1465 {
1466 	int i;
1467 
1468 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
1469 	if (!config->stats)
1470 		return -1;
1471 
1472 	config->stats_num = nthreads;
1473 
1474 	for (i = 0; i < nthreads; i++)
1475 		runtime_stat__init(&config->stats[i]);
1476 
1477 	return 0;
1478 }
1479 
1480 static void runtime_stat_delete(struct perf_stat_config *config)
1481 {
1482 	int i;
1483 
1484 	if (!config->stats)
1485 		return;
1486 
1487 	for (i = 0; i < config->stats_num; i++)
1488 		runtime_stat__exit(&config->stats[i]);
1489 
1490 	free(config->stats);
1491 }
1492 
1493 static const char * const stat_report_usage[] = {
1494 	"perf stat report [<options>]",
1495 	NULL,
1496 };
1497 
1498 static struct perf_stat perf_stat = {
1499 	.tool = {
1500 		.attr		= perf_event__process_attr,
1501 		.event_update	= perf_event__process_event_update,
1502 		.thread_map	= process_thread_map_event,
1503 		.cpu_map	= process_cpu_map_event,
1504 		.stat_config	= process_stat_config_event,
1505 		.stat		= perf_event__process_stat_event,
1506 		.stat_round	= process_stat_round_event,
1507 	},
1508 	.aggr_mode = AGGR_UNSET,
1509 };
1510 
1511 static int __cmd_report(int argc, const char **argv)
1512 {
1513 	struct perf_session *session;
1514 	const struct option options[] = {
1515 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1516 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1517 		     "aggregate counts per processor socket", AGGR_SOCKET),
1518 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1519 		     "aggregate counts per physical processor core", AGGR_CORE),
1520 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1521 		     "disable CPU count aggregation", AGGR_NONE),
1522 	OPT_END()
1523 	};
1524 	struct stat st;
1525 	int ret;
1526 
1527 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
1528 
1529 	if (!input_name || !strlen(input_name)) {
1530 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1531 			input_name = "-";
1532 		else
1533 			input_name = "perf.data";
1534 	}
1535 
1536 	perf_stat.data.file.path = input_name;
1537 	perf_stat.data.mode      = PERF_DATA_MODE_READ;
1538 
1539 	session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
1540 	if (session == NULL)
1541 		return -1;
1542 
1543 	perf_stat.session  = session;
1544 	stat_config.output = stderr;
1545 	evsel_list         = session->evlist;
1546 
1547 	ret = perf_session__process_events(session);
1548 	if (ret)
1549 		return ret;
1550 
1551 	perf_session__delete(session);
1552 	return 0;
1553 }
1554 
1555 static void setup_system_wide(int forks)
1556 {
1557 	/*
1558 	 * Make system wide (-a) the default target if
1559 	 * no target was specified and one of following
1560 	 * conditions is met:
1561 	 *
1562 	 *   - there's no workload specified
1563 	 *   - there is workload specified but all requested
1564 	 *     events are system wide events
1565 	 */
1566 	if (!target__none(&target))
1567 		return;
1568 
1569 	if (!forks)
1570 		target.system_wide = true;
1571 	else {
1572 		struct perf_evsel *counter;
1573 
1574 		evlist__for_each_entry(evsel_list, counter) {
1575 			if (!counter->system_wide)
1576 				return;
1577 		}
1578 
1579 		if (evsel_list->nr_entries)
1580 			target.system_wide = true;
1581 	}
1582 }
1583 
1584 int cmd_stat(int argc, const char **argv)
1585 {
1586 	const char * const stat_usage[] = {
1587 		"perf stat [<options>] [<command>]",
1588 		NULL
1589 	};
1590 	int status = -EINVAL, run_idx;
1591 	const char *mode;
1592 	FILE *output = stderr;
1593 	unsigned int interval, timeout;
1594 	const char * const stat_subcommands[] = { "record", "report" };
1595 
1596 	setlocale(LC_ALL, "");
1597 
1598 	evsel_list = perf_evlist__new();
1599 	if (evsel_list == NULL)
1600 		return -ENOMEM;
1601 
1602 	parse_events__shrink_config_terms();
1603 
1604 	/* String-parsing callback-based options would segfault when negated */
1605 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
1606 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
1607 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
1608 
1609 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1610 					(const char **) stat_usage,
1611 					PARSE_OPT_STOP_AT_NON_OPTION);
1612 	perf_stat__collect_metric_expr(evsel_list);
1613 	perf_stat__init_shadow_stats();
1614 
1615 	if (stat_config.csv_sep) {
1616 		stat_config.csv_output = true;
1617 		if (!strcmp(stat_config.csv_sep, "\\t"))
1618 			stat_config.csv_sep = "\t";
1619 	} else
1620 		stat_config.csv_sep = DEFAULT_SEPARATOR;
1621 
1622 	if (argc && !strncmp(argv[0], "rec", 3)) {
1623 		argc = __cmd_record(argc, argv);
1624 		if (argc < 0)
1625 			return -1;
1626 	} else if (argc && !strncmp(argv[0], "rep", 3))
1627 		return __cmd_report(argc, argv);
1628 
1629 	interval = stat_config.interval;
1630 	timeout = stat_config.timeout;
1631 
1632 	/*
1633 	 * For record command the -o is already taken care of.
1634 	 */
1635 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1636 		output = NULL;
1637 
1638 	if (output_name && output_fd) {
1639 		fprintf(stderr, "cannot use both --output and --log-fd\n");
1640 		parse_options_usage(stat_usage, stat_options, "o", 1);
1641 		parse_options_usage(NULL, stat_options, "log-fd", 0);
1642 		goto out;
1643 	}
1644 
1645 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
1646 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
1647 		goto out;
1648 	}
1649 
1650 	if (stat_config.metric_only && stat_config.run_count > 1) {
1651 		fprintf(stderr, "--metric-only is not supported with -r\n");
1652 		goto out;
1653 	}
1654 
1655 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
1656 		fprintf(stderr, "--table is only supported with -r\n");
1657 		parse_options_usage(stat_usage, stat_options, "r", 1);
1658 		parse_options_usage(NULL, stat_options, "table", 0);
1659 		goto out;
1660 	}
1661 
1662 	if (output_fd < 0) {
1663 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
1664 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1665 		goto out;
1666 	}
1667 
1668 	if (!output) {
1669 		struct timespec tm;
1670 		mode = append_file ? "a" : "w";
1671 
1672 		output = fopen(output_name, mode);
1673 		if (!output) {
1674 			perror("failed to create output file");
1675 			return -1;
1676 		}
1677 		clock_gettime(CLOCK_REALTIME, &tm);
1678 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1679 	} else if (output_fd > 0) {
1680 		mode = append_file ? "a" : "w";
1681 		output = fdopen(output_fd, mode);
1682 		if (!output) {
1683 			perror("Failed opening logfd");
1684 			return -errno;
1685 		}
1686 	}
1687 
1688 	stat_config.output = output;
1689 
1690 	/*
1691 	 * let the spreadsheet do the pretty-printing
1692 	 */
1693 	if (stat_config.csv_output) {
1694 		/* User explicitly passed -B? */
1695 		if (big_num_opt == 1) {
1696 			fprintf(stderr, "-B option not supported with -x\n");
1697 			parse_options_usage(stat_usage, stat_options, "B", 1);
1698 			parse_options_usage(NULL, stat_options, "x", 1);
1699 			goto out;
1700 		} else /* Nope, so disable big number formatting */
1701 			stat_config.big_num = false;
1702 	} else if (big_num_opt == 0) /* User passed --no-big-num */
1703 		stat_config.big_num = false;
1704 
1705 	setup_system_wide(argc);
1706 
1707 	/*
1708 	 * Display user/system times only for single
1709 	 * run and when there's specified tracee.
1710 	 */
1711 	if ((stat_config.run_count == 1) && target__none(&target))
1712 		stat_config.ru_display = true;
1713 
1714 	if (stat_config.run_count < 0) {
1715 		pr_err("Run count must be a positive number\n");
1716 		parse_options_usage(stat_usage, stat_options, "r", 1);
1717 		goto out;
1718 	} else if (stat_config.run_count == 0) {
1719 		forever = true;
1720 		stat_config.run_count = 1;
1721 	}
1722 
1723 	if (stat_config.walltime_run_table) {
1724 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
1725 		if (!stat_config.walltime_run) {
1726 			pr_err("failed to setup -r option");
1727 			goto out;
1728 		}
1729 	}
1730 
1731 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
1732 		!target__has_task(&target)) {
1733 		if (!target.system_wide || target.cpu_list) {
1734 			fprintf(stderr, "The --per-thread option is only "
1735 				"available when monitoring via -p -t -a "
1736 				"options or only --per-thread.\n");
1737 			parse_options_usage(NULL, stat_options, "p", 1);
1738 			parse_options_usage(NULL, stat_options, "t", 1);
1739 			goto out;
1740 		}
1741 	}
1742 
1743 	/*
1744 	 * no_aggr, cgroup are for system-wide only
1745 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1746 	 */
1747 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1748 	      stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1749 	    !target__has_cpu(&target)) {
1750 		fprintf(stderr, "both cgroup and no-aggregation "
1751 			"modes only available in system-wide mode\n");
1752 
1753 		parse_options_usage(stat_usage, stat_options, "G", 1);
1754 		parse_options_usage(NULL, stat_options, "A", 1);
1755 		parse_options_usage(NULL, stat_options, "a", 1);
1756 		goto out;
1757 	}
1758 
1759 	if (add_default_attributes())
1760 		goto out;
1761 
1762 	target__validate(&target);
1763 
1764 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
1765 		target.per_thread = true;
1766 
1767 	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1768 		if (target__has_task(&target)) {
1769 			pr_err("Problems finding threads of monitor\n");
1770 			parse_options_usage(stat_usage, stat_options, "p", 1);
1771 			parse_options_usage(NULL, stat_options, "t", 1);
1772 		} else if (target__has_cpu(&target)) {
1773 			perror("failed to parse CPUs map");
1774 			parse_options_usage(stat_usage, stat_options, "C", 1);
1775 			parse_options_usage(NULL, stat_options, "a", 1);
1776 		}
1777 		goto out;
1778 	}
1779 
1780 	/*
1781 	 * Initialize thread_map with comm names,
1782 	 * so we could print it out on output.
1783 	 */
1784 	if (stat_config.aggr_mode == AGGR_THREAD) {
1785 		thread_map__read_comms(evsel_list->threads);
1786 		if (target.system_wide) {
1787 			if (runtime_stat_new(&stat_config,
1788 				thread_map__nr(evsel_list->threads))) {
1789 				goto out;
1790 			}
1791 		}
1792 	}
1793 
1794 	if (stat_config.times && interval)
1795 		interval_count = true;
1796 	else if (stat_config.times && !interval) {
1797 		pr_err("interval-count option should be used together with "
1798 				"interval-print.\n");
1799 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
1800 		parse_options_usage(stat_usage, stat_options, "I", 1);
1801 		goto out;
1802 	}
1803 
1804 	if (timeout && timeout < 100) {
1805 		if (timeout < 10) {
1806 			pr_err("timeout must be >= 10ms.\n");
1807 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
1808 			goto out;
1809 		} else
1810 			pr_warning("timeout < 100ms. "
1811 				   "The overhead percentage could be high in some cases. "
1812 				   "Please proceed with caution.\n");
1813 	}
1814 	if (timeout && interval) {
1815 		pr_err("timeout option is not supported with interval-print.\n");
1816 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
1817 		parse_options_usage(stat_usage, stat_options, "I", 1);
1818 		goto out;
1819 	}
1820 
1821 	if (perf_evlist__alloc_stats(evsel_list, interval))
1822 		goto out;
1823 
1824 	if (perf_stat_init_aggr_mode())
1825 		goto out;
1826 
1827 	/*
1828 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
1829 	 * while avoiding that older tools show confusing messages.
1830 	 *
1831 	 * However for pipe sessions we need to keep it zero,
1832 	 * because script's perf_evsel__check_attr is triggered
1833 	 * by attr->sample_type != 0, and we can't run it on
1834 	 * stat sessions.
1835 	 */
1836 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
1837 
1838 	/*
1839 	 * We dont want to block the signals - that would cause
1840 	 * child tasks to inherit that and Ctrl-C would not work.
1841 	 * What we want is for Ctrl-C to work in the exec()-ed
1842 	 * task, but being ignored by perf stat itself:
1843 	 */
1844 	atexit(sig_atexit);
1845 	if (!forever)
1846 		signal(SIGINT,  skip_signal);
1847 	signal(SIGCHLD, skip_signal);
1848 	signal(SIGALRM, skip_signal);
1849 	signal(SIGABRT, skip_signal);
1850 
1851 	status = 0;
1852 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
1853 		if (stat_config.run_count != 1 && verbose > 0)
1854 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1855 				run_idx + 1);
1856 
1857 		status = run_perf_stat(argc, argv, run_idx);
1858 		if (forever && status != -1) {
1859 			print_counters(NULL, argc, argv);
1860 			perf_stat__reset_stats();
1861 		}
1862 	}
1863 
1864 	if (!forever && status != -1 && !interval)
1865 		print_counters(NULL, argc, argv);
1866 
1867 	if (STAT_RECORD) {
1868 		/*
1869 		 * We synthesize the kernel mmap record just so that older tools
1870 		 * don't emit warnings about not being able to resolve symbols
1871 		 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
1872 		 * a saner message about no samples being in the perf.data file.
1873 		 *
1874 		 * This also serves to suppress a warning about f_header.data.size == 0
1875 		 * in header.c at the moment 'perf stat record' gets introduced, which
1876 		 * is not really needed once we start adding the stat specific PERF_RECORD_
1877 		 * records, but the need to suppress the kptr_restrict messages in older
1878 		 * tools remain  -acme
1879 		 */
1880 		int fd = perf_data__fd(&perf_stat.data);
1881 		int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
1882 							     process_synthesized_event,
1883 							     &perf_stat.session->machines.host);
1884 		if (err) {
1885 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
1886 				   "older tools may produce warnings about this file\n.");
1887 		}
1888 
1889 		if (!interval) {
1890 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
1891 				pr_err("failed to write stat round event\n");
1892 		}
1893 
1894 		if (!perf_stat.data.is_pipe) {
1895 			perf_stat.session->header.data_size += perf_stat.bytes_written;
1896 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
1897 		}
1898 
1899 		perf_session__delete(perf_stat.session);
1900 	}
1901 
1902 	perf_stat__exit_aggr_mode();
1903 	perf_evlist__free_stats(evsel_list);
1904 out:
1905 	free(stat_config.walltime_run);
1906 
1907 	if (smi_cost && smi_reset)
1908 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
1909 
1910 	perf_evlist__delete(evsel_list);
1911 
1912 	runtime_stat_delete(&stat_config);
1913 
1914 	return status;
1915 }
1916