xref: /openbmc/linux/tools/perf/builtin-stat.c (revision e3d786a3)
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 struct perf_evsel *perf_evsel__reset_weak_group(struct perf_evsel *evsel)
387 {
388 	struct perf_evsel *c2, *leader;
389 	bool is_open = true;
390 
391 	leader = evsel->leader;
392 	pr_debug("Weak group for %s/%d failed\n",
393 			leader->name, leader->nr_members);
394 
395 	/*
396 	 * for_each_group_member doesn't work here because it doesn't
397 	 * include the first entry.
398 	 */
399 	evlist__for_each_entry(evsel_list, c2) {
400 		if (c2 == evsel)
401 			is_open = false;
402 		if (c2->leader == leader) {
403 			if (is_open)
404 				perf_evsel__close(c2);
405 			c2->leader = c2;
406 			c2->nr_members = 0;
407 		}
408 	}
409 	return leader;
410 }
411 
412 static bool is_target_alive(struct target *_target,
413 			    struct thread_map *threads)
414 {
415 	struct stat st;
416 	int i;
417 
418 	if (!target__has_task(_target))
419 		return true;
420 
421 	for (i = 0; i < threads->nr; i++) {
422 		char path[PATH_MAX];
423 
424 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
425 			  threads->map[i].pid);
426 
427 		if (!stat(path, &st))
428 			return true;
429 	}
430 
431 	return false;
432 }
433 
434 static int __run_perf_stat(int argc, const char **argv, int run_idx)
435 {
436 	int interval = stat_config.interval;
437 	int times = stat_config.times;
438 	int timeout = stat_config.timeout;
439 	char msg[BUFSIZ];
440 	unsigned long long t0, t1;
441 	struct perf_evsel *counter;
442 	struct timespec ts;
443 	size_t l;
444 	int status = 0;
445 	const bool forks = (argc > 0);
446 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
447 	struct perf_evsel_config_term *err_term;
448 
449 	if (interval) {
450 		ts.tv_sec  = interval / USEC_PER_MSEC;
451 		ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
452 	} else if (timeout) {
453 		ts.tv_sec  = timeout / USEC_PER_MSEC;
454 		ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
455 	} else {
456 		ts.tv_sec  = 1;
457 		ts.tv_nsec = 0;
458 	}
459 
460 	if (forks) {
461 		if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
462 						  workload_exec_failed_signal) < 0) {
463 			perror("failed to prepare workload");
464 			return -1;
465 		}
466 		child_pid = evsel_list->workload.pid;
467 	}
468 
469 	if (group)
470 		perf_evlist__set_leader(evsel_list);
471 
472 	evlist__for_each_entry(evsel_list, counter) {
473 try_again:
474 		if (create_perf_stat_counter(counter, &stat_config, &target) < 0) {
475 
476 			/* Weak group failed. Reset the group. */
477 			if ((errno == EINVAL || errno == EBADF) &&
478 			    counter->leader != counter &&
479 			    counter->weak_group) {
480 				counter = perf_evsel__reset_weak_group(counter);
481 				goto try_again;
482 			}
483 
484 			/*
485 			 * PPC returns ENXIO for HW counters until 2.6.37
486 			 * (behavior changed with commit b0a873e).
487 			 */
488 			if (errno == EINVAL || errno == ENOSYS ||
489 			    errno == ENOENT || errno == EOPNOTSUPP ||
490 			    errno == ENXIO) {
491 				if (verbose > 0)
492 					ui__warning("%s event is not supported by the kernel.\n",
493 						    perf_evsel__name(counter));
494 				counter->supported = false;
495 
496 				if ((counter->leader != counter) ||
497 				    !(counter->leader->nr_members > 1))
498 					continue;
499 			} else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
500                                 if (verbose > 0)
501                                         ui__warning("%s\n", msg);
502                                 goto try_again;
503 			} else if (target__has_per_thread(&target) &&
504 				   evsel_list->threads &&
505 				   evsel_list->threads->err_thread != -1) {
506 				/*
507 				 * For global --per-thread case, skip current
508 				 * error thread.
509 				 */
510 				if (!thread_map__remove(evsel_list->threads,
511 							evsel_list->threads->err_thread)) {
512 					evsel_list->threads->err_thread = -1;
513 					goto try_again;
514 				}
515 			}
516 
517 			perf_evsel__open_strerror(counter, &target,
518 						  errno, msg, sizeof(msg));
519 			ui__error("%s\n", msg);
520 
521 			if (child_pid != -1)
522 				kill(child_pid, SIGTERM);
523 
524 			return -1;
525 		}
526 		counter->supported = true;
527 
528 		l = strlen(counter->unit);
529 		if (l > stat_config.unit_width)
530 			stat_config.unit_width = l;
531 
532 		if (perf_evsel__should_store_id(counter) &&
533 		    perf_evsel__store_ids(counter, evsel_list))
534 			return -1;
535 	}
536 
537 	if (perf_evlist__apply_filters(evsel_list, &counter)) {
538 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
539 			counter->filter, perf_evsel__name(counter), errno,
540 			str_error_r(errno, msg, sizeof(msg)));
541 		return -1;
542 	}
543 
544 	if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
545 		pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
546 		      err_term->val.drv_cfg, perf_evsel__name(counter), errno,
547 		      str_error_r(errno, msg, sizeof(msg)));
548 		return -1;
549 	}
550 
551 	if (STAT_RECORD) {
552 		int err, fd = perf_data__fd(&perf_stat.data);
553 
554 		if (is_pipe) {
555 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
556 		} else {
557 			err = perf_session__write_header(perf_stat.session, evsel_list,
558 							 fd, false);
559 		}
560 
561 		if (err < 0)
562 			return err;
563 
564 		err = perf_stat_synthesize_config(&stat_config, NULL, evsel_list,
565 						  process_synthesized_event, is_pipe);
566 		if (err < 0)
567 			return err;
568 	}
569 
570 	/*
571 	 * Enable counters and exec the command:
572 	 */
573 	t0 = rdclock();
574 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
575 
576 	if (forks) {
577 		perf_evlist__start_workload(evsel_list);
578 		enable_counters();
579 
580 		if (interval || timeout) {
581 			while (!waitpid(child_pid, &status, WNOHANG)) {
582 				nanosleep(&ts, NULL);
583 				if (timeout)
584 					break;
585 				process_interval();
586 				if (interval_count && !(--times))
587 					break;
588 			}
589 		}
590 		wait4(child_pid, &status, 0, &stat_config.ru_data);
591 
592 		if (workload_exec_errno) {
593 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
594 			pr_err("Workload failed: %s\n", emsg);
595 			return -1;
596 		}
597 
598 		if (WIFSIGNALED(status))
599 			psignal(WTERMSIG(status), argv[0]);
600 	} else {
601 		enable_counters();
602 		while (!done) {
603 			nanosleep(&ts, NULL);
604 			if (!is_target_alive(&target, evsel_list->threads))
605 				break;
606 			if (timeout)
607 				break;
608 			if (interval) {
609 				process_interval();
610 				if (interval_count && !(--times))
611 					break;
612 			}
613 		}
614 	}
615 
616 	disable_counters();
617 
618 	t1 = rdclock();
619 
620 	if (stat_config.walltime_run_table)
621 		stat_config.walltime_run[run_idx] = t1 - t0;
622 
623 	update_stats(&walltime_nsecs_stats, t1 - t0);
624 
625 	/*
626 	 * Closing a group leader splits the group, and as we only disable
627 	 * group leaders, results in remaining events becoming enabled. To
628 	 * avoid arbitrary skew, we must read all counters before closing any
629 	 * group leaders.
630 	 */
631 	read_counters();
632 	perf_evlist__close(evsel_list);
633 
634 	return WEXITSTATUS(status);
635 }
636 
637 static int run_perf_stat(int argc, const char **argv, int run_idx)
638 {
639 	int ret;
640 
641 	if (pre_cmd) {
642 		ret = system(pre_cmd);
643 		if (ret)
644 			return ret;
645 	}
646 
647 	if (sync_run)
648 		sync();
649 
650 	ret = __run_perf_stat(argc, argv, run_idx);
651 	if (ret)
652 		return ret;
653 
654 	if (post_cmd) {
655 		ret = system(post_cmd);
656 		if (ret)
657 			return ret;
658 	}
659 
660 	return ret;
661 }
662 
663 static void print_counters(struct timespec *ts, int argc, const char **argv)
664 {
665 	/* Do not print anything if we record to the pipe. */
666 	if (STAT_RECORD && perf_stat.data.is_pipe)
667 		return;
668 
669 	perf_evlist__print_counters(evsel_list, &stat_config, &target,
670 				    ts, argc, argv);
671 }
672 
673 static volatile int signr = -1;
674 
675 static void skip_signal(int signo)
676 {
677 	if ((child_pid == -1) || stat_config.interval)
678 		done = 1;
679 
680 	signr = signo;
681 	/*
682 	 * render child_pid harmless
683 	 * won't send SIGTERM to a random
684 	 * process in case of race condition
685 	 * and fast PID recycling
686 	 */
687 	child_pid = -1;
688 }
689 
690 static void sig_atexit(void)
691 {
692 	sigset_t set, oset;
693 
694 	/*
695 	 * avoid race condition with SIGCHLD handler
696 	 * in skip_signal() which is modifying child_pid
697 	 * goal is to avoid send SIGTERM to a random
698 	 * process
699 	 */
700 	sigemptyset(&set);
701 	sigaddset(&set, SIGCHLD);
702 	sigprocmask(SIG_BLOCK, &set, &oset);
703 
704 	if (child_pid != -1)
705 		kill(child_pid, SIGTERM);
706 
707 	sigprocmask(SIG_SETMASK, &oset, NULL);
708 
709 	if (signr == -1)
710 		return;
711 
712 	signal(signr, SIG_DFL);
713 	kill(getpid(), signr);
714 }
715 
716 static int stat__set_big_num(const struct option *opt __maybe_unused,
717 			     const char *s __maybe_unused, int unset)
718 {
719 	big_num_opt = unset ? 0 : 1;
720 	return 0;
721 }
722 
723 static int enable_metric_only(const struct option *opt __maybe_unused,
724 			      const char *s __maybe_unused, int unset)
725 {
726 	force_metric_only = true;
727 	stat_config.metric_only = !unset;
728 	return 0;
729 }
730 
731 static int parse_metric_groups(const struct option *opt,
732 			       const char *str,
733 			       int unset __maybe_unused)
734 {
735 	return metricgroup__parse_groups(opt, str, &stat_config.metric_events);
736 }
737 
738 static const struct option stat_options[] = {
739 	OPT_BOOLEAN('T', "transaction", &transaction_run,
740 		    "hardware transaction statistics"),
741 	OPT_CALLBACK('e', "event", &evsel_list, "event",
742 		     "event selector. use 'perf list' to list available events",
743 		     parse_events_option),
744 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
745 		     "event filter", parse_filter),
746 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
747 		    "child tasks do not inherit counters"),
748 	OPT_STRING('p', "pid", &target.pid, "pid",
749 		   "stat events on existing process id"),
750 	OPT_STRING('t', "tid", &target.tid, "tid",
751 		   "stat events on existing thread id"),
752 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
753 		    "system-wide collection from all CPUs"),
754 	OPT_BOOLEAN('g', "group", &group,
755 		    "put the counters into a counter group"),
756 	OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
757 	OPT_INCR('v', "verbose", &verbose,
758 		    "be more verbose (show counter open errors, etc)"),
759 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
760 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
761 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
762 		    "display details about each run (only with -r option)"),
763 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
764 		    "null run - dont start any counters"),
765 	OPT_INCR('d', "detailed", &detailed_run,
766 		    "detailed run - start a lot of events"),
767 	OPT_BOOLEAN('S', "sync", &sync_run,
768 		    "call sync() before starting a run"),
769 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
770 			   "print large numbers with thousands\' separators",
771 			   stat__set_big_num),
772 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
773 		    "list of cpus to monitor in system-wide"),
774 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
775 		    "disable CPU count aggregation", AGGR_NONE),
776 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
777 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
778 		   "print counts with custom separator"),
779 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
780 		     "monitor event in cgroup name only", parse_cgroups),
781 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
782 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
783 	OPT_INTEGER(0, "log-fd", &output_fd,
784 		    "log output to fd, instead of stderr"),
785 	OPT_STRING(0, "pre", &pre_cmd, "command",
786 			"command to run prior to the measured command"),
787 	OPT_STRING(0, "post", &post_cmd, "command",
788 			"command to run after to the measured command"),
789 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
790 		    "print counts at regular interval in ms "
791 		    "(overhead is possible for values <= 100ms)"),
792 	OPT_INTEGER(0, "interval-count", &stat_config.times,
793 		    "print counts for fixed number of times"),
794 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
795 		    "clear screen in between new interval"),
796 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
797 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
798 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
799 		     "aggregate counts per processor socket", AGGR_SOCKET),
800 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
801 		     "aggregate counts per physical processor core", AGGR_CORE),
802 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
803 		     "aggregate counts per thread", AGGR_THREAD),
804 	OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
805 		     "ms to wait before starting measurement after program start"),
806 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
807 			"Only print computed metrics. No raw values", enable_metric_only),
808 	OPT_BOOLEAN(0, "topdown", &topdown_run,
809 			"measure topdown level 1 statistics"),
810 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
811 			"measure SMI cost"),
812 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
813 		     "monitor specified metrics or metric groups (separated by ,)",
814 		     parse_metric_groups),
815 	OPT_END()
816 };
817 
818 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
819 				 struct cpu_map *map, int cpu)
820 {
821 	return cpu_map__get_socket(map, cpu, NULL);
822 }
823 
824 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
825 			       struct cpu_map *map, int cpu)
826 {
827 	return cpu_map__get_core(map, cpu, NULL);
828 }
829 
830 static int cpu_map__get_max(struct cpu_map *map)
831 {
832 	int i, max = -1;
833 
834 	for (i = 0; i < map->nr; i++) {
835 		if (map->map[i] > max)
836 			max = map->map[i];
837 	}
838 
839 	return max;
840 }
841 
842 static int perf_stat__get_aggr(struct perf_stat_config *config,
843 			       aggr_get_id_t get_id, struct cpu_map *map, int idx)
844 {
845 	int cpu;
846 
847 	if (idx >= map->nr)
848 		return -1;
849 
850 	cpu = map->map[idx];
851 
852 	if (config->cpus_aggr_map->map[cpu] == -1)
853 		config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
854 
855 	return config->cpus_aggr_map->map[cpu];
856 }
857 
858 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
859 					struct cpu_map *map, int idx)
860 {
861 	return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
862 }
863 
864 static int perf_stat__get_core_cached(struct perf_stat_config *config,
865 				      struct cpu_map *map, int idx)
866 {
867 	return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
868 }
869 
870 static int perf_stat_init_aggr_mode(void)
871 {
872 	int nr;
873 
874 	switch (stat_config.aggr_mode) {
875 	case AGGR_SOCKET:
876 		if (cpu_map__build_socket_map(evsel_list->cpus, &stat_config.aggr_map)) {
877 			perror("cannot build socket map");
878 			return -1;
879 		}
880 		stat_config.aggr_get_id = perf_stat__get_socket_cached;
881 		break;
882 	case AGGR_CORE:
883 		if (cpu_map__build_core_map(evsel_list->cpus, &stat_config.aggr_map)) {
884 			perror("cannot build core map");
885 			return -1;
886 		}
887 		stat_config.aggr_get_id = perf_stat__get_core_cached;
888 		break;
889 	case AGGR_NONE:
890 	case AGGR_GLOBAL:
891 	case AGGR_THREAD:
892 	case AGGR_UNSET:
893 	default:
894 		break;
895 	}
896 
897 	/*
898 	 * The evsel_list->cpus is the base we operate on,
899 	 * taking the highest cpu number to be the size of
900 	 * the aggregation translate cpumap.
901 	 */
902 	nr = cpu_map__get_max(evsel_list->cpus);
903 	stat_config.cpus_aggr_map = cpu_map__empty_new(nr + 1);
904 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
905 }
906 
907 static void perf_stat__exit_aggr_mode(void)
908 {
909 	cpu_map__put(stat_config.aggr_map);
910 	cpu_map__put(stat_config.cpus_aggr_map);
911 	stat_config.aggr_map = NULL;
912 	stat_config.cpus_aggr_map = NULL;
913 }
914 
915 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
916 {
917 	int cpu;
918 
919 	if (idx > map->nr)
920 		return -1;
921 
922 	cpu = map->map[idx];
923 
924 	if (cpu >= env->nr_cpus_avail)
925 		return -1;
926 
927 	return cpu;
928 }
929 
930 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
931 {
932 	struct perf_env *env = data;
933 	int cpu = perf_env__get_cpu(env, map, idx);
934 
935 	return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
936 }
937 
938 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
939 {
940 	struct perf_env *env = data;
941 	int core = -1, cpu = perf_env__get_cpu(env, map, idx);
942 
943 	if (cpu != -1) {
944 		int socket_id = env->cpu[cpu].socket_id;
945 
946 		/*
947 		 * Encode socket in upper 16 bits
948 		 * core_id is relative to socket, and
949 		 * we need a global id. So we combine
950 		 * socket + core id.
951 		 */
952 		core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
953 	}
954 
955 	return core;
956 }
957 
958 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
959 				      struct cpu_map **sockp)
960 {
961 	return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
962 }
963 
964 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
965 				    struct cpu_map **corep)
966 {
967 	return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
968 }
969 
970 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
971 				      struct cpu_map *map, int idx)
972 {
973 	return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
974 }
975 
976 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
977 				    struct cpu_map *map, int idx)
978 {
979 	return perf_env__get_core(map, idx, &perf_stat.session->header.env);
980 }
981 
982 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
983 {
984 	struct perf_env *env = &st->session->header.env;
985 
986 	switch (stat_config.aggr_mode) {
987 	case AGGR_SOCKET:
988 		if (perf_env__build_socket_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
989 			perror("cannot build socket map");
990 			return -1;
991 		}
992 		stat_config.aggr_get_id = perf_stat__get_socket_file;
993 		break;
994 	case AGGR_CORE:
995 		if (perf_env__build_core_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
996 			perror("cannot build core map");
997 			return -1;
998 		}
999 		stat_config.aggr_get_id = perf_stat__get_core_file;
1000 		break;
1001 	case AGGR_NONE:
1002 	case AGGR_GLOBAL:
1003 	case AGGR_THREAD:
1004 	case AGGR_UNSET:
1005 	default:
1006 		break;
1007 	}
1008 
1009 	return 0;
1010 }
1011 
1012 static int topdown_filter_events(const char **attr, char **str, bool use_group)
1013 {
1014 	int off = 0;
1015 	int i;
1016 	int len = 0;
1017 	char *s;
1018 
1019 	for (i = 0; attr[i]; i++) {
1020 		if (pmu_have_event("cpu", attr[i])) {
1021 			len += strlen(attr[i]) + 1;
1022 			attr[i - off] = attr[i];
1023 		} else
1024 			off++;
1025 	}
1026 	attr[i - off] = NULL;
1027 
1028 	*str = malloc(len + 1 + 2);
1029 	if (!*str)
1030 		return -1;
1031 	s = *str;
1032 	if (i - off == 0) {
1033 		*s = 0;
1034 		return 0;
1035 	}
1036 	if (use_group)
1037 		*s++ = '{';
1038 	for (i = 0; attr[i]; i++) {
1039 		strcpy(s, attr[i]);
1040 		s += strlen(s);
1041 		*s++ = ',';
1042 	}
1043 	if (use_group) {
1044 		s[-1] = '}';
1045 		*s = 0;
1046 	} else
1047 		s[-1] = 0;
1048 	return 0;
1049 }
1050 
1051 __weak bool arch_topdown_check_group(bool *warn)
1052 {
1053 	*warn = false;
1054 	return false;
1055 }
1056 
1057 __weak void arch_topdown_group_warn(void)
1058 {
1059 }
1060 
1061 /*
1062  * Add default attributes, if there were no attributes specified or
1063  * if -d/--detailed, -d -d or -d -d -d is used:
1064  */
1065 static int add_default_attributes(void)
1066 {
1067 	int err;
1068 	struct perf_event_attr default_attrs0[] = {
1069 
1070   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1071   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1072   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1073   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1074 
1075   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1076 };
1077 	struct perf_event_attr frontend_attrs[] = {
1078   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1079 };
1080 	struct perf_event_attr backend_attrs[] = {
1081   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1082 };
1083 	struct perf_event_attr default_attrs1[] = {
1084   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1085   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1086   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1087 
1088 };
1089 
1090 /*
1091  * Detailed stats (-d), covering the L1 and last level data caches:
1092  */
1093 	struct perf_event_attr detailed_attrs[] = {
1094 
1095   { .type = PERF_TYPE_HW_CACHE,
1096     .config =
1097 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1098 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1099 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1100 
1101   { .type = PERF_TYPE_HW_CACHE,
1102     .config =
1103 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1104 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1105 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1106 
1107   { .type = PERF_TYPE_HW_CACHE,
1108     .config =
1109 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1110 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1111 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1112 
1113   { .type = PERF_TYPE_HW_CACHE,
1114     .config =
1115 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1116 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1117 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1118 };
1119 
1120 /*
1121  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1122  */
1123 	struct perf_event_attr very_detailed_attrs[] = {
1124 
1125   { .type = PERF_TYPE_HW_CACHE,
1126     .config =
1127 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1128 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1129 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1130 
1131   { .type = PERF_TYPE_HW_CACHE,
1132     .config =
1133 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1134 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1135 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1136 
1137   { .type = PERF_TYPE_HW_CACHE,
1138     .config =
1139 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1140 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1141 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1142 
1143   { .type = PERF_TYPE_HW_CACHE,
1144     .config =
1145 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1146 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1147 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1148 
1149   { .type = PERF_TYPE_HW_CACHE,
1150     .config =
1151 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1152 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1153 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1154 
1155   { .type = PERF_TYPE_HW_CACHE,
1156     .config =
1157 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1158 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1159 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1160 
1161 };
1162 
1163 /*
1164  * Very, very detailed stats (-d -d -d), adding prefetch events:
1165  */
1166 	struct perf_event_attr very_very_detailed_attrs[] = {
1167 
1168   { .type = PERF_TYPE_HW_CACHE,
1169     .config =
1170 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1171 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1172 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1173 
1174   { .type = PERF_TYPE_HW_CACHE,
1175     .config =
1176 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1177 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1178 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1179 };
1180 	struct parse_events_error errinfo;
1181 
1182 	/* Set attrs if no event is selected and !null_run: */
1183 	if (stat_config.null_run)
1184 		return 0;
1185 
1186 	if (transaction_run) {
1187 		/* Handle -T as -M transaction. Once platform specific metrics
1188 		 * support has been added to the json files, all archictures
1189 		 * will use this approach. To determine transaction support
1190 		 * on an architecture test for such a metric name.
1191 		 */
1192 		if (metricgroup__has_metric("transaction")) {
1193 			struct option opt = { .value = &evsel_list };
1194 
1195 			return metricgroup__parse_groups(&opt, "transaction",
1196 							 &stat_config.metric_events);
1197 		}
1198 
1199 		if (pmu_have_event("cpu", "cycles-ct") &&
1200 		    pmu_have_event("cpu", "el-start"))
1201 			err = parse_events(evsel_list, transaction_attrs,
1202 					   &errinfo);
1203 		else
1204 			err = parse_events(evsel_list,
1205 					   transaction_limited_attrs,
1206 					   &errinfo);
1207 		if (err) {
1208 			fprintf(stderr, "Cannot set up transaction events\n");
1209 			parse_events_print_error(&errinfo, transaction_attrs);
1210 			return -1;
1211 		}
1212 		return 0;
1213 	}
1214 
1215 	if (smi_cost) {
1216 		int smi;
1217 
1218 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1219 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1220 			return -1;
1221 		}
1222 
1223 		if (!smi) {
1224 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1225 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1226 				return -1;
1227 			}
1228 			smi_reset = true;
1229 		}
1230 
1231 		if (pmu_have_event("msr", "aperf") &&
1232 		    pmu_have_event("msr", "smi")) {
1233 			if (!force_metric_only)
1234 				stat_config.metric_only = true;
1235 			err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1236 		} else {
1237 			fprintf(stderr, "To measure SMI cost, it needs "
1238 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1239 			parse_events_print_error(&errinfo, smi_cost_attrs);
1240 			return -1;
1241 		}
1242 		if (err) {
1243 			fprintf(stderr, "Cannot set up SMI cost events\n");
1244 			return -1;
1245 		}
1246 		return 0;
1247 	}
1248 
1249 	if (topdown_run) {
1250 		char *str = NULL;
1251 		bool warn = false;
1252 
1253 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1254 		    stat_config.aggr_mode != AGGR_CORE) {
1255 			pr_err("top down event configuration requires --per-core mode\n");
1256 			return -1;
1257 		}
1258 		stat_config.aggr_mode = AGGR_CORE;
1259 		if (nr_cgroups || !target__has_cpu(&target)) {
1260 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1261 			return -1;
1262 		}
1263 
1264 		if (!force_metric_only)
1265 			stat_config.metric_only = true;
1266 		if (topdown_filter_events(topdown_attrs, &str,
1267 				arch_topdown_check_group(&warn)) < 0) {
1268 			pr_err("Out of memory\n");
1269 			return -1;
1270 		}
1271 		if (topdown_attrs[0] && str) {
1272 			if (warn)
1273 				arch_topdown_group_warn();
1274 			err = parse_events(evsel_list, str, &errinfo);
1275 			if (err) {
1276 				fprintf(stderr,
1277 					"Cannot set up top down events %s: %d\n",
1278 					str, err);
1279 				free(str);
1280 				parse_events_print_error(&errinfo, str);
1281 				return -1;
1282 			}
1283 		} else {
1284 			fprintf(stderr, "System does not support topdown\n");
1285 			return -1;
1286 		}
1287 		free(str);
1288 	}
1289 
1290 	if (!evsel_list->nr_entries) {
1291 		if (target__has_cpu(&target))
1292 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1293 
1294 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1295 			return -1;
1296 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1297 			if (perf_evlist__add_default_attrs(evsel_list,
1298 						frontend_attrs) < 0)
1299 				return -1;
1300 		}
1301 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1302 			if (perf_evlist__add_default_attrs(evsel_list,
1303 						backend_attrs) < 0)
1304 				return -1;
1305 		}
1306 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1307 			return -1;
1308 	}
1309 
1310 	/* Detailed events get appended to the event list: */
1311 
1312 	if (detailed_run <  1)
1313 		return 0;
1314 
1315 	/* Append detailed run extra attributes: */
1316 	if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1317 		return -1;
1318 
1319 	if (detailed_run < 2)
1320 		return 0;
1321 
1322 	/* Append very detailed run extra attributes: */
1323 	if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1324 		return -1;
1325 
1326 	if (detailed_run < 3)
1327 		return 0;
1328 
1329 	/* Append very, very detailed run extra attributes: */
1330 	return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1331 }
1332 
1333 static const char * const stat_record_usage[] = {
1334 	"perf stat record [<options>]",
1335 	NULL,
1336 };
1337 
1338 static void init_features(struct perf_session *session)
1339 {
1340 	int feat;
1341 
1342 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1343 		perf_header__set_feat(&session->header, feat);
1344 
1345 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1346 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1347 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1348 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1349 }
1350 
1351 static int __cmd_record(int argc, const char **argv)
1352 {
1353 	struct perf_session *session;
1354 	struct perf_data *data = &perf_stat.data;
1355 
1356 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
1357 			     PARSE_OPT_STOP_AT_NON_OPTION);
1358 
1359 	if (output_name)
1360 		data->file.path = output_name;
1361 
1362 	if (stat_config.run_count != 1 || forever) {
1363 		pr_err("Cannot use -r option with perf stat record.\n");
1364 		return -1;
1365 	}
1366 
1367 	session = perf_session__new(data, false, NULL);
1368 	if (session == NULL) {
1369 		pr_err("Perf session creation failed.\n");
1370 		return -1;
1371 	}
1372 
1373 	init_features(session);
1374 
1375 	session->evlist   = evsel_list;
1376 	perf_stat.session = session;
1377 	perf_stat.record  = true;
1378 	return argc;
1379 }
1380 
1381 static int process_stat_round_event(struct perf_session *session,
1382 				    union perf_event *event)
1383 {
1384 	struct stat_round_event *stat_round = &event->stat_round;
1385 	struct perf_evsel *counter;
1386 	struct timespec tsh, *ts = NULL;
1387 	const char **argv = session->header.env.cmdline_argv;
1388 	int argc = session->header.env.nr_cmdline;
1389 
1390 	evlist__for_each_entry(evsel_list, counter)
1391 		perf_stat_process_counter(&stat_config, counter);
1392 
1393 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1394 		update_stats(&walltime_nsecs_stats, stat_round->time);
1395 
1396 	if (stat_config.interval && stat_round->time) {
1397 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
1398 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1399 		ts = &tsh;
1400 	}
1401 
1402 	print_counters(ts, argc, argv);
1403 	return 0;
1404 }
1405 
1406 static
1407 int process_stat_config_event(struct perf_session *session,
1408 			      union perf_event *event)
1409 {
1410 	struct perf_tool *tool = session->tool;
1411 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1412 
1413 	perf_event__read_stat_config(&stat_config, &event->stat_config);
1414 
1415 	if (cpu_map__empty(st->cpus)) {
1416 		if (st->aggr_mode != AGGR_UNSET)
1417 			pr_warning("warning: processing task data, aggregation mode not set\n");
1418 		return 0;
1419 	}
1420 
1421 	if (st->aggr_mode != AGGR_UNSET)
1422 		stat_config.aggr_mode = st->aggr_mode;
1423 
1424 	if (perf_stat.data.is_pipe)
1425 		perf_stat_init_aggr_mode();
1426 	else
1427 		perf_stat_init_aggr_mode_file(st);
1428 
1429 	return 0;
1430 }
1431 
1432 static int set_maps(struct perf_stat *st)
1433 {
1434 	if (!st->cpus || !st->threads)
1435 		return 0;
1436 
1437 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1438 		return -EINVAL;
1439 
1440 	perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
1441 
1442 	if (perf_evlist__alloc_stats(evsel_list, true))
1443 		return -ENOMEM;
1444 
1445 	st->maps_allocated = true;
1446 	return 0;
1447 }
1448 
1449 static
1450 int process_thread_map_event(struct perf_session *session,
1451 			     union perf_event *event)
1452 {
1453 	struct perf_tool *tool = session->tool;
1454 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1455 
1456 	if (st->threads) {
1457 		pr_warning("Extra thread map event, ignoring.\n");
1458 		return 0;
1459 	}
1460 
1461 	st->threads = thread_map__new_event(&event->thread_map);
1462 	if (!st->threads)
1463 		return -ENOMEM;
1464 
1465 	return set_maps(st);
1466 }
1467 
1468 static
1469 int process_cpu_map_event(struct perf_session *session,
1470 			  union perf_event *event)
1471 {
1472 	struct perf_tool *tool = session->tool;
1473 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1474 	struct cpu_map *cpus;
1475 
1476 	if (st->cpus) {
1477 		pr_warning("Extra cpu map event, ignoring.\n");
1478 		return 0;
1479 	}
1480 
1481 	cpus = cpu_map__new_data(&event->cpu_map.data);
1482 	if (!cpus)
1483 		return -ENOMEM;
1484 
1485 	st->cpus = cpus;
1486 	return set_maps(st);
1487 }
1488 
1489 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
1490 {
1491 	int i;
1492 
1493 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
1494 	if (!config->stats)
1495 		return -1;
1496 
1497 	config->stats_num = nthreads;
1498 
1499 	for (i = 0; i < nthreads; i++)
1500 		runtime_stat__init(&config->stats[i]);
1501 
1502 	return 0;
1503 }
1504 
1505 static void runtime_stat_delete(struct perf_stat_config *config)
1506 {
1507 	int i;
1508 
1509 	if (!config->stats)
1510 		return;
1511 
1512 	for (i = 0; i < config->stats_num; i++)
1513 		runtime_stat__exit(&config->stats[i]);
1514 
1515 	free(config->stats);
1516 }
1517 
1518 static const char * const stat_report_usage[] = {
1519 	"perf stat report [<options>]",
1520 	NULL,
1521 };
1522 
1523 static struct perf_stat perf_stat = {
1524 	.tool = {
1525 		.attr		= perf_event__process_attr,
1526 		.event_update	= perf_event__process_event_update,
1527 		.thread_map	= process_thread_map_event,
1528 		.cpu_map	= process_cpu_map_event,
1529 		.stat_config	= process_stat_config_event,
1530 		.stat		= perf_event__process_stat_event,
1531 		.stat_round	= process_stat_round_event,
1532 	},
1533 	.aggr_mode = AGGR_UNSET,
1534 };
1535 
1536 static int __cmd_report(int argc, const char **argv)
1537 {
1538 	struct perf_session *session;
1539 	const struct option options[] = {
1540 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1541 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1542 		     "aggregate counts per processor socket", AGGR_SOCKET),
1543 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1544 		     "aggregate counts per physical processor core", AGGR_CORE),
1545 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1546 		     "disable CPU count aggregation", AGGR_NONE),
1547 	OPT_END()
1548 	};
1549 	struct stat st;
1550 	int ret;
1551 
1552 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
1553 
1554 	if (!input_name || !strlen(input_name)) {
1555 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1556 			input_name = "-";
1557 		else
1558 			input_name = "perf.data";
1559 	}
1560 
1561 	perf_stat.data.file.path = input_name;
1562 	perf_stat.data.mode      = PERF_DATA_MODE_READ;
1563 
1564 	session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
1565 	if (session == NULL)
1566 		return -1;
1567 
1568 	perf_stat.session  = session;
1569 	stat_config.output = stderr;
1570 	evsel_list         = session->evlist;
1571 
1572 	ret = perf_session__process_events(session);
1573 	if (ret)
1574 		return ret;
1575 
1576 	perf_session__delete(session);
1577 	return 0;
1578 }
1579 
1580 static void setup_system_wide(int forks)
1581 {
1582 	/*
1583 	 * Make system wide (-a) the default target if
1584 	 * no target was specified and one of following
1585 	 * conditions is met:
1586 	 *
1587 	 *   - there's no workload specified
1588 	 *   - there is workload specified but all requested
1589 	 *     events are system wide events
1590 	 */
1591 	if (!target__none(&target))
1592 		return;
1593 
1594 	if (!forks)
1595 		target.system_wide = true;
1596 	else {
1597 		struct perf_evsel *counter;
1598 
1599 		evlist__for_each_entry(evsel_list, counter) {
1600 			if (!counter->system_wide)
1601 				return;
1602 		}
1603 
1604 		if (evsel_list->nr_entries)
1605 			target.system_wide = true;
1606 	}
1607 }
1608 
1609 int cmd_stat(int argc, const char **argv)
1610 {
1611 	const char * const stat_usage[] = {
1612 		"perf stat [<options>] [<command>]",
1613 		NULL
1614 	};
1615 	int status = -EINVAL, run_idx;
1616 	const char *mode;
1617 	FILE *output = stderr;
1618 	unsigned int interval, timeout;
1619 	const char * const stat_subcommands[] = { "record", "report" };
1620 
1621 	setlocale(LC_ALL, "");
1622 
1623 	evsel_list = perf_evlist__new();
1624 	if (evsel_list == NULL)
1625 		return -ENOMEM;
1626 
1627 	parse_events__shrink_config_terms();
1628 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1629 					(const char **) stat_usage,
1630 					PARSE_OPT_STOP_AT_NON_OPTION);
1631 	perf_stat__collect_metric_expr(evsel_list);
1632 	perf_stat__init_shadow_stats();
1633 
1634 	if (stat_config.csv_sep) {
1635 		stat_config.csv_output = true;
1636 		if (!strcmp(stat_config.csv_sep, "\\t"))
1637 			stat_config.csv_sep = "\t";
1638 	} else
1639 		stat_config.csv_sep = DEFAULT_SEPARATOR;
1640 
1641 	if (argc && !strncmp(argv[0], "rec", 3)) {
1642 		argc = __cmd_record(argc, argv);
1643 		if (argc < 0)
1644 			return -1;
1645 	} else if (argc && !strncmp(argv[0], "rep", 3))
1646 		return __cmd_report(argc, argv);
1647 
1648 	interval = stat_config.interval;
1649 	timeout = stat_config.timeout;
1650 
1651 	/*
1652 	 * For record command the -o is already taken care of.
1653 	 */
1654 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1655 		output = NULL;
1656 
1657 	if (output_name && output_fd) {
1658 		fprintf(stderr, "cannot use both --output and --log-fd\n");
1659 		parse_options_usage(stat_usage, stat_options, "o", 1);
1660 		parse_options_usage(NULL, stat_options, "log-fd", 0);
1661 		goto out;
1662 	}
1663 
1664 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
1665 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
1666 		goto out;
1667 	}
1668 
1669 	if (stat_config.metric_only && stat_config.run_count > 1) {
1670 		fprintf(stderr, "--metric-only is not supported with -r\n");
1671 		goto out;
1672 	}
1673 
1674 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
1675 		fprintf(stderr, "--table is only supported with -r\n");
1676 		parse_options_usage(stat_usage, stat_options, "r", 1);
1677 		parse_options_usage(NULL, stat_options, "table", 0);
1678 		goto out;
1679 	}
1680 
1681 	if (output_fd < 0) {
1682 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
1683 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1684 		goto out;
1685 	}
1686 
1687 	if (!output) {
1688 		struct timespec tm;
1689 		mode = append_file ? "a" : "w";
1690 
1691 		output = fopen(output_name, mode);
1692 		if (!output) {
1693 			perror("failed to create output file");
1694 			return -1;
1695 		}
1696 		clock_gettime(CLOCK_REALTIME, &tm);
1697 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1698 	} else if (output_fd > 0) {
1699 		mode = append_file ? "a" : "w";
1700 		output = fdopen(output_fd, mode);
1701 		if (!output) {
1702 			perror("Failed opening logfd");
1703 			return -errno;
1704 		}
1705 	}
1706 
1707 	stat_config.output = output;
1708 
1709 	/*
1710 	 * let the spreadsheet do the pretty-printing
1711 	 */
1712 	if (stat_config.csv_output) {
1713 		/* User explicitly passed -B? */
1714 		if (big_num_opt == 1) {
1715 			fprintf(stderr, "-B option not supported with -x\n");
1716 			parse_options_usage(stat_usage, stat_options, "B", 1);
1717 			parse_options_usage(NULL, stat_options, "x", 1);
1718 			goto out;
1719 		} else /* Nope, so disable big number formatting */
1720 			stat_config.big_num = false;
1721 	} else if (big_num_opt == 0) /* User passed --no-big-num */
1722 		stat_config.big_num = false;
1723 
1724 	setup_system_wide(argc);
1725 
1726 	/*
1727 	 * Display user/system times only for single
1728 	 * run and when there's specified tracee.
1729 	 */
1730 	if ((stat_config.run_count == 1) && target__none(&target))
1731 		stat_config.ru_display = true;
1732 
1733 	if (stat_config.run_count < 0) {
1734 		pr_err("Run count must be a positive number\n");
1735 		parse_options_usage(stat_usage, stat_options, "r", 1);
1736 		goto out;
1737 	} else if (stat_config.run_count == 0) {
1738 		forever = true;
1739 		stat_config.run_count = 1;
1740 	}
1741 
1742 	if (stat_config.walltime_run_table) {
1743 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
1744 		if (!stat_config.walltime_run) {
1745 			pr_err("failed to setup -r option");
1746 			goto out;
1747 		}
1748 	}
1749 
1750 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
1751 		!target__has_task(&target)) {
1752 		if (!target.system_wide || target.cpu_list) {
1753 			fprintf(stderr, "The --per-thread option is only "
1754 				"available when monitoring via -p -t -a "
1755 				"options or only --per-thread.\n");
1756 			parse_options_usage(NULL, stat_options, "p", 1);
1757 			parse_options_usage(NULL, stat_options, "t", 1);
1758 			goto out;
1759 		}
1760 	}
1761 
1762 	/*
1763 	 * no_aggr, cgroup are for system-wide only
1764 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1765 	 */
1766 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1767 	      stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1768 	    !target__has_cpu(&target)) {
1769 		fprintf(stderr, "both cgroup and no-aggregation "
1770 			"modes only available in system-wide mode\n");
1771 
1772 		parse_options_usage(stat_usage, stat_options, "G", 1);
1773 		parse_options_usage(NULL, stat_options, "A", 1);
1774 		parse_options_usage(NULL, stat_options, "a", 1);
1775 		goto out;
1776 	}
1777 
1778 	if (add_default_attributes())
1779 		goto out;
1780 
1781 	target__validate(&target);
1782 
1783 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
1784 		target.per_thread = true;
1785 
1786 	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1787 		if (target__has_task(&target)) {
1788 			pr_err("Problems finding threads of monitor\n");
1789 			parse_options_usage(stat_usage, stat_options, "p", 1);
1790 			parse_options_usage(NULL, stat_options, "t", 1);
1791 		} else if (target__has_cpu(&target)) {
1792 			perror("failed to parse CPUs map");
1793 			parse_options_usage(stat_usage, stat_options, "C", 1);
1794 			parse_options_usage(NULL, stat_options, "a", 1);
1795 		}
1796 		goto out;
1797 	}
1798 
1799 	/*
1800 	 * Initialize thread_map with comm names,
1801 	 * so we could print it out on output.
1802 	 */
1803 	if (stat_config.aggr_mode == AGGR_THREAD) {
1804 		thread_map__read_comms(evsel_list->threads);
1805 		if (target.system_wide) {
1806 			if (runtime_stat_new(&stat_config,
1807 				thread_map__nr(evsel_list->threads))) {
1808 				goto out;
1809 			}
1810 		}
1811 	}
1812 
1813 	if (stat_config.times && interval)
1814 		interval_count = true;
1815 	else if (stat_config.times && !interval) {
1816 		pr_err("interval-count option should be used together with "
1817 				"interval-print.\n");
1818 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
1819 		parse_options_usage(stat_usage, stat_options, "I", 1);
1820 		goto out;
1821 	}
1822 
1823 	if (timeout && timeout < 100) {
1824 		if (timeout < 10) {
1825 			pr_err("timeout must be >= 10ms.\n");
1826 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
1827 			goto out;
1828 		} else
1829 			pr_warning("timeout < 100ms. "
1830 				   "The overhead percentage could be high in some cases. "
1831 				   "Please proceed with caution.\n");
1832 	}
1833 	if (timeout && interval) {
1834 		pr_err("timeout option is not supported with interval-print.\n");
1835 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
1836 		parse_options_usage(stat_usage, stat_options, "I", 1);
1837 		goto out;
1838 	}
1839 
1840 	if (perf_evlist__alloc_stats(evsel_list, interval))
1841 		goto out;
1842 
1843 	if (perf_stat_init_aggr_mode())
1844 		goto out;
1845 
1846 	/*
1847 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
1848 	 * while avoiding that older tools show confusing messages.
1849 	 *
1850 	 * However for pipe sessions we need to keep it zero,
1851 	 * because script's perf_evsel__check_attr is triggered
1852 	 * by attr->sample_type != 0, and we can't run it on
1853 	 * stat sessions.
1854 	 */
1855 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
1856 
1857 	/*
1858 	 * We dont want to block the signals - that would cause
1859 	 * child tasks to inherit that and Ctrl-C would not work.
1860 	 * What we want is for Ctrl-C to work in the exec()-ed
1861 	 * task, but being ignored by perf stat itself:
1862 	 */
1863 	atexit(sig_atexit);
1864 	if (!forever)
1865 		signal(SIGINT,  skip_signal);
1866 	signal(SIGCHLD, skip_signal);
1867 	signal(SIGALRM, skip_signal);
1868 	signal(SIGABRT, skip_signal);
1869 
1870 	status = 0;
1871 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
1872 		if (stat_config.run_count != 1 && verbose > 0)
1873 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1874 				run_idx + 1);
1875 
1876 		status = run_perf_stat(argc, argv, run_idx);
1877 		if (forever && status != -1) {
1878 			print_counters(NULL, argc, argv);
1879 			perf_stat__reset_stats();
1880 		}
1881 	}
1882 
1883 	if (!forever && status != -1 && !interval)
1884 		print_counters(NULL, argc, argv);
1885 
1886 	if (STAT_RECORD) {
1887 		/*
1888 		 * We synthesize the kernel mmap record just so that older tools
1889 		 * don't emit warnings about not being able to resolve symbols
1890 		 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
1891 		 * a saner message about no samples being in the perf.data file.
1892 		 *
1893 		 * This also serves to suppress a warning about f_header.data.size == 0
1894 		 * in header.c at the moment 'perf stat record' gets introduced, which
1895 		 * is not really needed once we start adding the stat specific PERF_RECORD_
1896 		 * records, but the need to suppress the kptr_restrict messages in older
1897 		 * tools remain  -acme
1898 		 */
1899 		int fd = perf_data__fd(&perf_stat.data);
1900 		int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
1901 							     process_synthesized_event,
1902 							     &perf_stat.session->machines.host);
1903 		if (err) {
1904 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
1905 				   "older tools may produce warnings about this file\n.");
1906 		}
1907 
1908 		if (!interval) {
1909 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
1910 				pr_err("failed to write stat round event\n");
1911 		}
1912 
1913 		if (!perf_stat.data.is_pipe) {
1914 			perf_stat.session->header.data_size += perf_stat.bytes_written;
1915 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
1916 		}
1917 
1918 		perf_session__delete(perf_stat.session);
1919 	}
1920 
1921 	perf_stat__exit_aggr_mode();
1922 	perf_evlist__free_stats(evsel_list);
1923 out:
1924 	free(stat_config.walltime_run);
1925 
1926 	if (smi_cost && smi_reset)
1927 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
1928 
1929 	perf_evlist__delete(evsel_list);
1930 
1931 	runtime_stat_delete(&stat_config);
1932 
1933 	return status;
1934 }
1935