xref: /openbmc/linux/tools/perf/builtin-stat.c (revision b664e06d)
1 /*
2  * builtin-stat.c
3  *
4  * Builtin stat command: Give a precise performance counters summary
5  * overview about any workload, CPU or specific PID.
6  *
7  * Sample output:
8 
9    $ perf stat ./hackbench 10
10 
11   Time: 0.118
12 
13   Performance counter stats for './hackbench 10':
14 
15        1708.761321 task-clock                #   11.037 CPUs utilized
16             41,190 context-switches          #    0.024 M/sec
17              6,735 CPU-migrations            #    0.004 M/sec
18             17,318 page-faults               #    0.010 M/sec
19      5,205,202,243 cycles                    #    3.046 GHz
20      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
21      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
22      2,603,501,247 instructions              #    0.50  insns per cycle
23                                              #    1.48  stalled cycles per insn
24        484,357,498 branches                  #  283.455 M/sec
25          6,388,934 branch-misses             #    1.32% of all branches
26 
27         0.154822978  seconds time elapsed
28 
29  *
30  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
31  *
32  * Improvements and fixes by:
33  *
34  *   Arjan van de Ven <arjan@linux.intel.com>
35  *   Yanmin Zhang <yanmin.zhang@intel.com>
36  *   Wu Fengguang <fengguang.wu@intel.com>
37  *   Mike Galbraith <efault@gmx.de>
38  *   Paul Mackerras <paulus@samba.org>
39  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
40  *
41  * Released under the GPL v2. (and only v2, not any later version)
42  */
43 
44 #include "perf.h"
45 #include "builtin.h"
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include <subcmd/parse-options.h>
49 #include "util/parse-events.h"
50 #include "util/pmu.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/color.h"
56 #include "util/stat.h"
57 #include "util/header.h"
58 #include "util/cpumap.h"
59 #include "util/thread.h"
60 #include "util/thread_map.h"
61 #include "util/counts.h"
62 #include "util/group.h"
63 #include "util/session.h"
64 #include "util/tool.h"
65 #include "util/string2.h"
66 #include "util/metricgroup.h"
67 #include "util/top.h"
68 #include "asm/bug.h"
69 
70 #include <linux/time64.h>
71 #include <api/fs/fs.h>
72 #include <errno.h>
73 #include <signal.h>
74 #include <stdlib.h>
75 #include <sys/prctl.h>
76 #include <inttypes.h>
77 #include <locale.h>
78 #include <math.h>
79 #include <sys/types.h>
80 #include <sys/stat.h>
81 #include <sys/wait.h>
82 #include <unistd.h>
83 #include <sys/time.h>
84 #include <sys/resource.h>
85 
86 #include "sane_ctype.h"
87 
88 #define DEFAULT_SEPARATOR	" "
89 #define FREEZE_ON_SMI_PATH	"devices/cpu/freeze_on_smi"
90 
91 static void print_counters(struct timespec *ts, int argc, const char **argv);
92 
93 /* Default events used for perf stat -T */
94 static const char *transaction_attrs = {
95 	"task-clock,"
96 	"{"
97 	"instructions,"
98 	"cycles,"
99 	"cpu/cycles-t/,"
100 	"cpu/tx-start/,"
101 	"cpu/el-start/,"
102 	"cpu/cycles-ct/"
103 	"}"
104 };
105 
106 /* More limited version when the CPU does not have all events. */
107 static const char * transaction_limited_attrs = {
108 	"task-clock,"
109 	"{"
110 	"instructions,"
111 	"cycles,"
112 	"cpu/cycles-t/,"
113 	"cpu/tx-start/"
114 	"}"
115 };
116 
117 static const char * topdown_attrs[] = {
118 	"topdown-total-slots",
119 	"topdown-slots-retired",
120 	"topdown-recovery-bubbles",
121 	"topdown-fetch-bubbles",
122 	"topdown-slots-issued",
123 	NULL,
124 };
125 
126 static const char *smi_cost_attrs = {
127 	"{"
128 	"msr/aperf/,"
129 	"msr/smi/,"
130 	"cycles"
131 	"}"
132 };
133 
134 static struct perf_evlist	*evsel_list;
135 
136 static struct target target = {
137 	.uid	= UINT_MAX,
138 };
139 
140 #define METRIC_ONLY_LEN 20
141 
142 static volatile pid_t		child_pid			= -1;
143 static int			detailed_run			=  0;
144 static bool			transaction_run;
145 static bool			topdown_run			= false;
146 static bool			smi_cost			= false;
147 static bool			smi_reset			= false;
148 static int			big_num_opt			=  -1;
149 static bool			group				= false;
150 static const char		*pre_cmd			= NULL;
151 static const char		*post_cmd			= NULL;
152 static bool			sync_run			= false;
153 static bool			forever				= false;
154 static bool			force_metric_only		= false;
155 static struct timespec		ref_time;
156 static bool			append_file;
157 static bool			interval_count;
158 static const char		*output_name;
159 static int			output_fd;
160 
161 struct perf_stat {
162 	bool			 record;
163 	struct perf_data	 data;
164 	struct perf_session	*session;
165 	u64			 bytes_written;
166 	struct perf_tool	 tool;
167 	bool			 maps_allocated;
168 	struct cpu_map		*cpus;
169 	struct thread_map	*threads;
170 	enum aggr_mode		 aggr_mode;
171 };
172 
173 static struct perf_stat		perf_stat;
174 #define STAT_RECORD		perf_stat.record
175 
176 static volatile int done = 0;
177 
178 static struct perf_stat_config stat_config = {
179 	.aggr_mode		= AGGR_GLOBAL,
180 	.scale			= true,
181 	.unit_width		= 4, /* strlen("unit") */
182 	.run_count		= 1,
183 	.metric_only_len	= METRIC_ONLY_LEN,
184 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
185 	.big_num		= true,
186 };
187 
188 static inline void diff_timespec(struct timespec *r, struct timespec *a,
189 				 struct timespec *b)
190 {
191 	r->tv_sec = a->tv_sec - b->tv_sec;
192 	if (a->tv_nsec < b->tv_nsec) {
193 		r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
194 		r->tv_sec--;
195 	} else {
196 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
197 	}
198 }
199 
200 static void perf_stat__reset_stats(void)
201 {
202 	int i;
203 
204 	perf_evlist__reset_stats(evsel_list);
205 	perf_stat__reset_shadow_stats();
206 
207 	for (i = 0; i < stat_config.stats_num; i++)
208 		perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
209 }
210 
211 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
212 				     union perf_event *event,
213 				     struct perf_sample *sample __maybe_unused,
214 				     struct machine *machine __maybe_unused)
215 {
216 	if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
217 		pr_err("failed to write perf data, error: %m\n");
218 		return -1;
219 	}
220 
221 	perf_stat.bytes_written += event->header.size;
222 	return 0;
223 }
224 
225 static int write_stat_round_event(u64 tm, u64 type)
226 {
227 	return perf_event__synthesize_stat_round(NULL, tm, type,
228 						 process_synthesized_event,
229 						 NULL);
230 }
231 
232 #define WRITE_STAT_ROUND_EVENT(time, interval) \
233 	write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
234 
235 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
236 
237 static int
238 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
239 			     struct perf_counts_values *count)
240 {
241 	struct perf_sample_id *sid = SID(counter, cpu, thread);
242 
243 	return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
244 					   process_synthesized_event, NULL);
245 }
246 
247 static int read_single_counter(struct perf_evsel *counter, int cpu,
248 			       int thread, struct timespec *rs)
249 {
250 	if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
251 		u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
252 		struct perf_counts_values *count =
253 			perf_counts(counter->counts, cpu, thread);
254 		count->ena = count->run = val;
255 		count->val = val;
256 		return 0;
257 	}
258 	return perf_evsel__read_counter(counter, cpu, thread);
259 }
260 
261 /*
262  * Read out the results of a single counter:
263  * do not aggregate counts across CPUs in system-wide mode
264  */
265 static int read_counter(struct perf_evsel *counter, struct timespec *rs)
266 {
267 	int nthreads = thread_map__nr(evsel_list->threads);
268 	int ncpus, cpu, thread;
269 
270 	if (target__has_cpu(&target) && !target__has_per_thread(&target))
271 		ncpus = perf_evsel__nr_cpus(counter);
272 	else
273 		ncpus = 1;
274 
275 	if (!counter->supported)
276 		return -ENOENT;
277 
278 	if (counter->system_wide)
279 		nthreads = 1;
280 
281 	for (thread = 0; thread < nthreads; thread++) {
282 		for (cpu = 0; cpu < ncpus; cpu++) {
283 			struct perf_counts_values *count;
284 
285 			count = perf_counts(counter->counts, cpu, thread);
286 
287 			/*
288 			 * The leader's group read loads data into its group members
289 			 * (via perf_evsel__read_counter) and sets threir count->loaded.
290 			 */
291 			if (!count->loaded &&
292 			    read_single_counter(counter, cpu, thread, rs)) {
293 				counter->counts->scaled = -1;
294 				perf_counts(counter->counts, cpu, thread)->ena = 0;
295 				perf_counts(counter->counts, cpu, thread)->run = 0;
296 				return -1;
297 			}
298 
299 			count->loaded = false;
300 
301 			if (STAT_RECORD) {
302 				if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
303 					pr_err("failed to write stat event\n");
304 					return -1;
305 				}
306 			}
307 
308 			if (verbose > 1) {
309 				fprintf(stat_config.output,
310 					"%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
311 						perf_evsel__name(counter),
312 						cpu,
313 						count->val, count->ena, count->run);
314 			}
315 		}
316 	}
317 
318 	return 0;
319 }
320 
321 static void read_counters(struct timespec *rs)
322 {
323 	struct perf_evsel *counter;
324 	int ret;
325 
326 	evlist__for_each_entry(evsel_list, counter) {
327 		ret = read_counter(counter, rs);
328 		if (ret)
329 			pr_debug("failed to read counter %s\n", counter->name);
330 
331 		if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
332 			pr_warning("failed to process counter %s\n", counter->name);
333 	}
334 }
335 
336 static void process_interval(void)
337 {
338 	struct timespec ts, rs;
339 
340 	clock_gettime(CLOCK_MONOTONIC, &ts);
341 	diff_timespec(&rs, &ts, &ref_time);
342 
343 	read_counters(&rs);
344 
345 	if (STAT_RECORD) {
346 		if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
347 			pr_err("failed to write stat round event\n");
348 	}
349 
350 	init_stats(&walltime_nsecs_stats);
351 	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
352 	print_counters(&rs, 0, NULL);
353 }
354 
355 static void enable_counters(void)
356 {
357 	if (stat_config.initial_delay)
358 		usleep(stat_config.initial_delay * USEC_PER_MSEC);
359 
360 	/*
361 	 * We need to enable counters only if:
362 	 * - we don't have tracee (attaching to task or cpu)
363 	 * - we have initial delay configured
364 	 */
365 	if (!target__none(&target) || stat_config.initial_delay)
366 		perf_evlist__enable(evsel_list);
367 }
368 
369 static void disable_counters(void)
370 {
371 	/*
372 	 * If we don't have tracee (attaching to task or cpu), counters may
373 	 * still be running. To get accurate group ratios, we must stop groups
374 	 * from counting before reading their constituent counters.
375 	 */
376 	if (!target__none(&target))
377 		perf_evlist__disable(evsel_list);
378 }
379 
380 static volatile int workload_exec_errno;
381 
382 /*
383  * perf_evlist__prepare_workload will send a SIGUSR1
384  * if the fork fails, since we asked by setting its
385  * want_signal to true.
386  */
387 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
388 					void *ucontext __maybe_unused)
389 {
390 	workload_exec_errno = info->si_value.sival_int;
391 }
392 
393 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
394 {
395 	return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
396 }
397 
398 static bool is_target_alive(struct target *_target,
399 			    struct thread_map *threads)
400 {
401 	struct stat st;
402 	int i;
403 
404 	if (!target__has_task(_target))
405 		return true;
406 
407 	for (i = 0; i < threads->nr; i++) {
408 		char path[PATH_MAX];
409 
410 		scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
411 			  threads->map[i].pid);
412 
413 		if (!stat(path, &st))
414 			return true;
415 	}
416 
417 	return false;
418 }
419 
420 static int __run_perf_stat(int argc, const char **argv, int run_idx)
421 {
422 	int interval = stat_config.interval;
423 	int times = stat_config.times;
424 	int timeout = stat_config.timeout;
425 	char msg[BUFSIZ];
426 	unsigned long long t0, t1;
427 	struct perf_evsel *counter;
428 	struct timespec ts;
429 	size_t l;
430 	int status = 0;
431 	const bool forks = (argc > 0);
432 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
433 
434 	if (interval) {
435 		ts.tv_sec  = interval / USEC_PER_MSEC;
436 		ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
437 	} else if (timeout) {
438 		ts.tv_sec  = timeout / USEC_PER_MSEC;
439 		ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
440 	} else {
441 		ts.tv_sec  = 1;
442 		ts.tv_nsec = 0;
443 	}
444 
445 	if (forks) {
446 		if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
447 						  workload_exec_failed_signal) < 0) {
448 			perror("failed to prepare workload");
449 			return -1;
450 		}
451 		child_pid = evsel_list->workload.pid;
452 	}
453 
454 	if (group)
455 		perf_evlist__set_leader(evsel_list);
456 
457 	evlist__for_each_entry(evsel_list, counter) {
458 try_again:
459 		if (create_perf_stat_counter(counter, &stat_config, &target) < 0) {
460 
461 			/* Weak group failed. Reset the group. */
462 			if ((errno == EINVAL || errno == EBADF) &&
463 			    counter->leader != counter &&
464 			    counter->weak_group) {
465 				counter = perf_evlist__reset_weak_group(evsel_list, counter);
466 				goto try_again;
467 			}
468 
469 			/*
470 			 * PPC returns ENXIO for HW counters until 2.6.37
471 			 * (behavior changed with commit b0a873e).
472 			 */
473 			if (errno == EINVAL || errno == ENOSYS ||
474 			    errno == ENOENT || errno == EOPNOTSUPP ||
475 			    errno == ENXIO) {
476 				if (verbose > 0)
477 					ui__warning("%s event is not supported by the kernel.\n",
478 						    perf_evsel__name(counter));
479 				counter->supported = false;
480 
481 				if ((counter->leader != counter) ||
482 				    !(counter->leader->nr_members > 1))
483 					continue;
484 			} else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
485                                 if (verbose > 0)
486                                         ui__warning("%s\n", msg);
487                                 goto try_again;
488 			} else if (target__has_per_thread(&target) &&
489 				   evsel_list->threads &&
490 				   evsel_list->threads->err_thread != -1) {
491 				/*
492 				 * For global --per-thread case, skip current
493 				 * error thread.
494 				 */
495 				if (!thread_map__remove(evsel_list->threads,
496 							evsel_list->threads->err_thread)) {
497 					evsel_list->threads->err_thread = -1;
498 					goto try_again;
499 				}
500 			}
501 
502 			perf_evsel__open_strerror(counter, &target,
503 						  errno, msg, sizeof(msg));
504 			ui__error("%s\n", msg);
505 
506 			if (child_pid != -1)
507 				kill(child_pid, SIGTERM);
508 
509 			return -1;
510 		}
511 		counter->supported = true;
512 
513 		l = strlen(counter->unit);
514 		if (l > stat_config.unit_width)
515 			stat_config.unit_width = l;
516 
517 		if (perf_evsel__should_store_id(counter) &&
518 		    perf_evsel__store_ids(counter, evsel_list))
519 			return -1;
520 	}
521 
522 	if (perf_evlist__apply_filters(evsel_list, &counter)) {
523 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
524 			counter->filter, perf_evsel__name(counter), errno,
525 			str_error_r(errno, msg, sizeof(msg)));
526 		return -1;
527 	}
528 
529 	if (STAT_RECORD) {
530 		int err, fd = perf_data__fd(&perf_stat.data);
531 
532 		if (is_pipe) {
533 			err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
534 		} else {
535 			err = perf_session__write_header(perf_stat.session, evsel_list,
536 							 fd, false);
537 		}
538 
539 		if (err < 0)
540 			return err;
541 
542 		err = perf_stat_synthesize_config(&stat_config, NULL, evsel_list,
543 						  process_synthesized_event, is_pipe);
544 		if (err < 0)
545 			return err;
546 	}
547 
548 	/*
549 	 * Enable counters and exec the command:
550 	 */
551 	t0 = rdclock();
552 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
553 
554 	if (forks) {
555 		perf_evlist__start_workload(evsel_list);
556 		enable_counters();
557 
558 		if (interval || timeout) {
559 			while (!waitpid(child_pid, &status, WNOHANG)) {
560 				nanosleep(&ts, NULL);
561 				if (timeout)
562 					break;
563 				process_interval();
564 				if (interval_count && !(--times))
565 					break;
566 			}
567 		}
568 		if (child_pid != -1)
569 			wait4(child_pid, &status, 0, &stat_config.ru_data);
570 
571 		if (workload_exec_errno) {
572 			const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
573 			pr_err("Workload failed: %s\n", emsg);
574 			return -1;
575 		}
576 
577 		if (WIFSIGNALED(status))
578 			psignal(WTERMSIG(status), argv[0]);
579 	} else {
580 		enable_counters();
581 		while (!done) {
582 			nanosleep(&ts, NULL);
583 			if (!is_target_alive(&target, evsel_list->threads))
584 				break;
585 			if (timeout)
586 				break;
587 			if (interval) {
588 				process_interval();
589 				if (interval_count && !(--times))
590 					break;
591 			}
592 		}
593 	}
594 
595 	disable_counters();
596 
597 	t1 = rdclock();
598 
599 	if (stat_config.walltime_run_table)
600 		stat_config.walltime_run[run_idx] = t1 - t0;
601 
602 	update_stats(&walltime_nsecs_stats, t1 - t0);
603 
604 	/*
605 	 * Closing a group leader splits the group, and as we only disable
606 	 * group leaders, results in remaining events becoming enabled. To
607 	 * avoid arbitrary skew, we must read all counters before closing any
608 	 * group leaders.
609 	 */
610 	read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
611 	perf_evlist__close(evsel_list);
612 
613 	return WEXITSTATUS(status);
614 }
615 
616 static int run_perf_stat(int argc, const char **argv, int run_idx)
617 {
618 	int ret;
619 
620 	if (pre_cmd) {
621 		ret = system(pre_cmd);
622 		if (ret)
623 			return ret;
624 	}
625 
626 	if (sync_run)
627 		sync();
628 
629 	ret = __run_perf_stat(argc, argv, run_idx);
630 	if (ret)
631 		return ret;
632 
633 	if (post_cmd) {
634 		ret = system(post_cmd);
635 		if (ret)
636 			return ret;
637 	}
638 
639 	return ret;
640 }
641 
642 static void print_counters(struct timespec *ts, int argc, const char **argv)
643 {
644 	/* Do not print anything if we record to the pipe. */
645 	if (STAT_RECORD && perf_stat.data.is_pipe)
646 		return;
647 
648 	perf_evlist__print_counters(evsel_list, &stat_config, &target,
649 				    ts, argc, argv);
650 }
651 
652 static volatile int signr = -1;
653 
654 static void skip_signal(int signo)
655 {
656 	if ((child_pid == -1) || stat_config.interval)
657 		done = 1;
658 
659 	signr = signo;
660 	/*
661 	 * render child_pid harmless
662 	 * won't send SIGTERM to a random
663 	 * process in case of race condition
664 	 * and fast PID recycling
665 	 */
666 	child_pid = -1;
667 }
668 
669 static void sig_atexit(void)
670 {
671 	sigset_t set, oset;
672 
673 	/*
674 	 * avoid race condition with SIGCHLD handler
675 	 * in skip_signal() which is modifying child_pid
676 	 * goal is to avoid send SIGTERM to a random
677 	 * process
678 	 */
679 	sigemptyset(&set);
680 	sigaddset(&set, SIGCHLD);
681 	sigprocmask(SIG_BLOCK, &set, &oset);
682 
683 	if (child_pid != -1)
684 		kill(child_pid, SIGTERM);
685 
686 	sigprocmask(SIG_SETMASK, &oset, NULL);
687 
688 	if (signr == -1)
689 		return;
690 
691 	signal(signr, SIG_DFL);
692 	kill(getpid(), signr);
693 }
694 
695 static int stat__set_big_num(const struct option *opt __maybe_unused,
696 			     const char *s __maybe_unused, int unset)
697 {
698 	big_num_opt = unset ? 0 : 1;
699 	return 0;
700 }
701 
702 static int enable_metric_only(const struct option *opt __maybe_unused,
703 			      const char *s __maybe_unused, int unset)
704 {
705 	force_metric_only = true;
706 	stat_config.metric_only = !unset;
707 	return 0;
708 }
709 
710 static int parse_metric_groups(const struct option *opt,
711 			       const char *str,
712 			       int unset __maybe_unused)
713 {
714 	return metricgroup__parse_groups(opt, str, &stat_config.metric_events);
715 }
716 
717 static struct option stat_options[] = {
718 	OPT_BOOLEAN('T', "transaction", &transaction_run,
719 		    "hardware transaction statistics"),
720 	OPT_CALLBACK('e', "event", &evsel_list, "event",
721 		     "event selector. use 'perf list' to list available events",
722 		     parse_events_option),
723 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
724 		     "event filter", parse_filter),
725 	OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
726 		    "child tasks do not inherit counters"),
727 	OPT_STRING('p', "pid", &target.pid, "pid",
728 		   "stat events on existing process id"),
729 	OPT_STRING('t', "tid", &target.tid, "tid",
730 		   "stat events on existing thread id"),
731 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
732 		    "system-wide collection from all CPUs"),
733 	OPT_BOOLEAN('g', "group", &group,
734 		    "put the counters into a counter group"),
735 	OPT_BOOLEAN(0, "scale", &stat_config.scale,
736 		    "Use --no-scale to disable counter scaling for multiplexing"),
737 	OPT_INCR('v', "verbose", &verbose,
738 		    "be more verbose (show counter open errors, etc)"),
739 	OPT_INTEGER('r', "repeat", &stat_config.run_count,
740 		    "repeat command and print average + stddev (max: 100, forever: 0)"),
741 	OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
742 		    "display details about each run (only with -r option)"),
743 	OPT_BOOLEAN('n', "null", &stat_config.null_run,
744 		    "null run - dont start any counters"),
745 	OPT_INCR('d', "detailed", &detailed_run,
746 		    "detailed run - start a lot of events"),
747 	OPT_BOOLEAN('S', "sync", &sync_run,
748 		    "call sync() before starting a run"),
749 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
750 			   "print large numbers with thousands\' separators",
751 			   stat__set_big_num),
752 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
753 		    "list of cpus to monitor in system-wide"),
754 	OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
755 		    "disable CPU count aggregation", AGGR_NONE),
756 	OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
757 	OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
758 		   "print counts with custom separator"),
759 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
760 		     "monitor event in cgroup name only", parse_cgroups),
761 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
762 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
763 	OPT_INTEGER(0, "log-fd", &output_fd,
764 		    "log output to fd, instead of stderr"),
765 	OPT_STRING(0, "pre", &pre_cmd, "command",
766 			"command to run prior to the measured command"),
767 	OPT_STRING(0, "post", &post_cmd, "command",
768 			"command to run after to the measured command"),
769 	OPT_UINTEGER('I', "interval-print", &stat_config.interval,
770 		    "print counts at regular interval in ms "
771 		    "(overhead is possible for values <= 100ms)"),
772 	OPT_INTEGER(0, "interval-count", &stat_config.times,
773 		    "print counts for fixed number of times"),
774 	OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
775 		    "clear screen in between new interval"),
776 	OPT_UINTEGER(0, "timeout", &stat_config.timeout,
777 		    "stop workload and print counts after a timeout period in ms (>= 10ms)"),
778 	OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
779 		     "aggregate counts per processor socket", AGGR_SOCKET),
780 	OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
781 		     "aggregate counts per physical processor core", AGGR_CORE),
782 	OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
783 		     "aggregate counts per thread", AGGR_THREAD),
784 	OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
785 		     "ms to wait before starting measurement after program start"),
786 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
787 			"Only print computed metrics. No raw values", enable_metric_only),
788 	OPT_BOOLEAN(0, "topdown", &topdown_run,
789 			"measure topdown level 1 statistics"),
790 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
791 			"measure SMI cost"),
792 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
793 		     "monitor specified metrics or metric groups (separated by ,)",
794 		     parse_metric_groups),
795 	OPT_END()
796 };
797 
798 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
799 				 struct cpu_map *map, int cpu)
800 {
801 	return cpu_map__get_socket(map, cpu, NULL);
802 }
803 
804 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
805 			       struct cpu_map *map, int cpu)
806 {
807 	return cpu_map__get_core(map, cpu, NULL);
808 }
809 
810 static int cpu_map__get_max(struct cpu_map *map)
811 {
812 	int i, max = -1;
813 
814 	for (i = 0; i < map->nr; i++) {
815 		if (map->map[i] > max)
816 			max = map->map[i];
817 	}
818 
819 	return max;
820 }
821 
822 static int perf_stat__get_aggr(struct perf_stat_config *config,
823 			       aggr_get_id_t get_id, struct cpu_map *map, int idx)
824 {
825 	int cpu;
826 
827 	if (idx >= map->nr)
828 		return -1;
829 
830 	cpu = map->map[idx];
831 
832 	if (config->cpus_aggr_map->map[cpu] == -1)
833 		config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
834 
835 	return config->cpus_aggr_map->map[cpu];
836 }
837 
838 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
839 					struct cpu_map *map, int idx)
840 {
841 	return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
842 }
843 
844 static int perf_stat__get_core_cached(struct perf_stat_config *config,
845 				      struct cpu_map *map, int idx)
846 {
847 	return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
848 }
849 
850 static bool term_percore_set(void)
851 {
852 	struct perf_evsel *counter;
853 
854 	evlist__for_each_entry(evsel_list, counter) {
855 		if (counter->percore)
856 			return true;
857 	}
858 
859 	return false;
860 }
861 
862 static int perf_stat_init_aggr_mode(void)
863 {
864 	int nr;
865 
866 	switch (stat_config.aggr_mode) {
867 	case AGGR_SOCKET:
868 		if (cpu_map__build_socket_map(evsel_list->cpus, &stat_config.aggr_map)) {
869 			perror("cannot build socket map");
870 			return -1;
871 		}
872 		stat_config.aggr_get_id = perf_stat__get_socket_cached;
873 		break;
874 	case AGGR_CORE:
875 		if (cpu_map__build_core_map(evsel_list->cpus, &stat_config.aggr_map)) {
876 			perror("cannot build core map");
877 			return -1;
878 		}
879 		stat_config.aggr_get_id = perf_stat__get_core_cached;
880 		break;
881 	case AGGR_NONE:
882 		if (term_percore_set()) {
883 			if (cpu_map__build_core_map(evsel_list->cpus,
884 						    &stat_config.aggr_map)) {
885 				perror("cannot build core map");
886 				return -1;
887 			}
888 			stat_config.aggr_get_id = perf_stat__get_core_cached;
889 		}
890 		break;
891 	case AGGR_GLOBAL:
892 	case AGGR_THREAD:
893 	case AGGR_UNSET:
894 	default:
895 		break;
896 	}
897 
898 	/*
899 	 * The evsel_list->cpus is the base we operate on,
900 	 * taking the highest cpu number to be the size of
901 	 * the aggregation translate cpumap.
902 	 */
903 	nr = cpu_map__get_max(evsel_list->cpus);
904 	stat_config.cpus_aggr_map = cpu_map__empty_new(nr + 1);
905 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
906 }
907 
908 static void perf_stat__exit_aggr_mode(void)
909 {
910 	cpu_map__put(stat_config.aggr_map);
911 	cpu_map__put(stat_config.cpus_aggr_map);
912 	stat_config.aggr_map = NULL;
913 	stat_config.cpus_aggr_map = NULL;
914 }
915 
916 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
917 {
918 	int cpu;
919 
920 	if (idx > map->nr)
921 		return -1;
922 
923 	cpu = map->map[idx];
924 
925 	if (cpu >= env->nr_cpus_avail)
926 		return -1;
927 
928 	return cpu;
929 }
930 
931 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
932 {
933 	struct perf_env *env = data;
934 	int cpu = perf_env__get_cpu(env, map, idx);
935 
936 	return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
937 }
938 
939 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
940 {
941 	struct perf_env *env = data;
942 	int core = -1, cpu = perf_env__get_cpu(env, map, idx);
943 
944 	if (cpu != -1) {
945 		int socket_id = env->cpu[cpu].socket_id;
946 
947 		/*
948 		 * Encode socket in upper 16 bits
949 		 * core_id is relative to socket, and
950 		 * we need a global id. So we combine
951 		 * socket + core id.
952 		 */
953 		core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
954 	}
955 
956 	return core;
957 }
958 
959 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
960 				      struct cpu_map **sockp)
961 {
962 	return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
963 }
964 
965 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
966 				    struct cpu_map **corep)
967 {
968 	return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
969 }
970 
971 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
972 				      struct cpu_map *map, int idx)
973 {
974 	return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
975 }
976 
977 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
978 				    struct cpu_map *map, int idx)
979 {
980 	return perf_env__get_core(map, idx, &perf_stat.session->header.env);
981 }
982 
983 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
984 {
985 	struct perf_env *env = &st->session->header.env;
986 
987 	switch (stat_config.aggr_mode) {
988 	case AGGR_SOCKET:
989 		if (perf_env__build_socket_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
990 			perror("cannot build socket map");
991 			return -1;
992 		}
993 		stat_config.aggr_get_id = perf_stat__get_socket_file;
994 		break;
995 	case AGGR_CORE:
996 		if (perf_env__build_core_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
997 			perror("cannot build core map");
998 			return -1;
999 		}
1000 		stat_config.aggr_get_id = perf_stat__get_core_file;
1001 		break;
1002 	case AGGR_NONE:
1003 	case AGGR_GLOBAL:
1004 	case AGGR_THREAD:
1005 	case AGGR_UNSET:
1006 	default:
1007 		break;
1008 	}
1009 
1010 	return 0;
1011 }
1012 
1013 static int topdown_filter_events(const char **attr, char **str, bool use_group)
1014 {
1015 	int off = 0;
1016 	int i;
1017 	int len = 0;
1018 	char *s;
1019 
1020 	for (i = 0; attr[i]; i++) {
1021 		if (pmu_have_event("cpu", attr[i])) {
1022 			len += strlen(attr[i]) + 1;
1023 			attr[i - off] = attr[i];
1024 		} else
1025 			off++;
1026 	}
1027 	attr[i - off] = NULL;
1028 
1029 	*str = malloc(len + 1 + 2);
1030 	if (!*str)
1031 		return -1;
1032 	s = *str;
1033 	if (i - off == 0) {
1034 		*s = 0;
1035 		return 0;
1036 	}
1037 	if (use_group)
1038 		*s++ = '{';
1039 	for (i = 0; attr[i]; i++) {
1040 		strcpy(s, attr[i]);
1041 		s += strlen(s);
1042 		*s++ = ',';
1043 	}
1044 	if (use_group) {
1045 		s[-1] = '}';
1046 		*s = 0;
1047 	} else
1048 		s[-1] = 0;
1049 	return 0;
1050 }
1051 
1052 __weak bool arch_topdown_check_group(bool *warn)
1053 {
1054 	*warn = false;
1055 	return false;
1056 }
1057 
1058 __weak void arch_topdown_group_warn(void)
1059 {
1060 }
1061 
1062 /*
1063  * Add default attributes, if there were no attributes specified or
1064  * if -d/--detailed, -d -d or -d -d -d is used:
1065  */
1066 static int add_default_attributes(void)
1067 {
1068 	int err;
1069 	struct perf_event_attr default_attrs0[] = {
1070 
1071   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1072   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1073   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1074   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1075 
1076   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1077 };
1078 	struct perf_event_attr frontend_attrs[] = {
1079   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1080 };
1081 	struct perf_event_attr backend_attrs[] = {
1082   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1083 };
1084 	struct perf_event_attr default_attrs1[] = {
1085   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1086   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1087   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1088 
1089 };
1090 
1091 /*
1092  * Detailed stats (-d), covering the L1 and last level data caches:
1093  */
1094 	struct perf_event_attr detailed_attrs[] = {
1095 
1096   { .type = PERF_TYPE_HW_CACHE,
1097     .config =
1098 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1099 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1100 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1101 
1102   { .type = PERF_TYPE_HW_CACHE,
1103     .config =
1104 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1105 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1106 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1107 
1108   { .type = PERF_TYPE_HW_CACHE,
1109     .config =
1110 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1111 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1112 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1113 
1114   { .type = PERF_TYPE_HW_CACHE,
1115     .config =
1116 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1117 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1118 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1119 };
1120 
1121 /*
1122  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1123  */
1124 	struct perf_event_attr very_detailed_attrs[] = {
1125 
1126   { .type = PERF_TYPE_HW_CACHE,
1127     .config =
1128 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1129 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1130 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1131 
1132   { .type = PERF_TYPE_HW_CACHE,
1133     .config =
1134 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1135 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1136 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1137 
1138   { .type = PERF_TYPE_HW_CACHE,
1139     .config =
1140 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1141 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1142 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1143 
1144   { .type = PERF_TYPE_HW_CACHE,
1145     .config =
1146 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1147 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1148 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1149 
1150   { .type = PERF_TYPE_HW_CACHE,
1151     .config =
1152 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1153 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1154 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1155 
1156   { .type = PERF_TYPE_HW_CACHE,
1157     .config =
1158 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1159 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1160 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1161 
1162 };
1163 
1164 /*
1165  * Very, very detailed stats (-d -d -d), adding prefetch events:
1166  */
1167 	struct perf_event_attr very_very_detailed_attrs[] = {
1168 
1169   { .type = PERF_TYPE_HW_CACHE,
1170     .config =
1171 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1172 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1173 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1174 
1175   { .type = PERF_TYPE_HW_CACHE,
1176     .config =
1177 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1178 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1179 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1180 };
1181 	struct parse_events_error errinfo;
1182 
1183 	/* Set attrs if no event is selected and !null_run: */
1184 	if (stat_config.null_run)
1185 		return 0;
1186 
1187 	if (transaction_run) {
1188 		/* Handle -T as -M transaction. Once platform specific metrics
1189 		 * support has been added to the json files, all archictures
1190 		 * will use this approach. To determine transaction support
1191 		 * on an architecture test for such a metric name.
1192 		 */
1193 		if (metricgroup__has_metric("transaction")) {
1194 			struct option opt = { .value = &evsel_list };
1195 
1196 			return metricgroup__parse_groups(&opt, "transaction",
1197 							 &stat_config.metric_events);
1198 		}
1199 
1200 		if (pmu_have_event("cpu", "cycles-ct") &&
1201 		    pmu_have_event("cpu", "el-start"))
1202 			err = parse_events(evsel_list, transaction_attrs,
1203 					   &errinfo);
1204 		else
1205 			err = parse_events(evsel_list,
1206 					   transaction_limited_attrs,
1207 					   &errinfo);
1208 		if (err) {
1209 			fprintf(stderr, "Cannot set up transaction events\n");
1210 			parse_events_print_error(&errinfo, transaction_attrs);
1211 			return -1;
1212 		}
1213 		return 0;
1214 	}
1215 
1216 	if (smi_cost) {
1217 		int smi;
1218 
1219 		if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1220 			fprintf(stderr, "freeze_on_smi is not supported.\n");
1221 			return -1;
1222 		}
1223 
1224 		if (!smi) {
1225 			if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1226 				fprintf(stderr, "Failed to set freeze_on_smi.\n");
1227 				return -1;
1228 			}
1229 			smi_reset = true;
1230 		}
1231 
1232 		if (pmu_have_event("msr", "aperf") &&
1233 		    pmu_have_event("msr", "smi")) {
1234 			if (!force_metric_only)
1235 				stat_config.metric_only = true;
1236 			err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1237 		} else {
1238 			fprintf(stderr, "To measure SMI cost, it needs "
1239 				"msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1240 			parse_events_print_error(&errinfo, smi_cost_attrs);
1241 			return -1;
1242 		}
1243 		if (err) {
1244 			fprintf(stderr, "Cannot set up SMI cost events\n");
1245 			return -1;
1246 		}
1247 		return 0;
1248 	}
1249 
1250 	if (topdown_run) {
1251 		char *str = NULL;
1252 		bool warn = false;
1253 
1254 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
1255 		    stat_config.aggr_mode != AGGR_CORE) {
1256 			pr_err("top down event configuration requires --per-core mode\n");
1257 			return -1;
1258 		}
1259 		stat_config.aggr_mode = AGGR_CORE;
1260 		if (nr_cgroups || !target__has_cpu(&target)) {
1261 			pr_err("top down event configuration requires system-wide mode (-a)\n");
1262 			return -1;
1263 		}
1264 
1265 		if (!force_metric_only)
1266 			stat_config.metric_only = true;
1267 		if (topdown_filter_events(topdown_attrs, &str,
1268 				arch_topdown_check_group(&warn)) < 0) {
1269 			pr_err("Out of memory\n");
1270 			return -1;
1271 		}
1272 		if (topdown_attrs[0] && str) {
1273 			if (warn)
1274 				arch_topdown_group_warn();
1275 			err = parse_events(evsel_list, str, &errinfo);
1276 			if (err) {
1277 				fprintf(stderr,
1278 					"Cannot set up top down events %s: %d\n",
1279 					str, err);
1280 				free(str);
1281 				parse_events_print_error(&errinfo, str);
1282 				return -1;
1283 			}
1284 		} else {
1285 			fprintf(stderr, "System does not support topdown\n");
1286 			return -1;
1287 		}
1288 		free(str);
1289 	}
1290 
1291 	if (!evsel_list->nr_entries) {
1292 		if (target__has_cpu(&target))
1293 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1294 
1295 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1296 			return -1;
1297 		if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1298 			if (perf_evlist__add_default_attrs(evsel_list,
1299 						frontend_attrs) < 0)
1300 				return -1;
1301 		}
1302 		if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1303 			if (perf_evlist__add_default_attrs(evsel_list,
1304 						backend_attrs) < 0)
1305 				return -1;
1306 		}
1307 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1308 			return -1;
1309 	}
1310 
1311 	/* Detailed events get appended to the event list: */
1312 
1313 	if (detailed_run <  1)
1314 		return 0;
1315 
1316 	/* Append detailed run extra attributes: */
1317 	if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1318 		return -1;
1319 
1320 	if (detailed_run < 2)
1321 		return 0;
1322 
1323 	/* Append very detailed run extra attributes: */
1324 	if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1325 		return -1;
1326 
1327 	if (detailed_run < 3)
1328 		return 0;
1329 
1330 	/* Append very, very detailed run extra attributes: */
1331 	return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1332 }
1333 
1334 static const char * const stat_record_usage[] = {
1335 	"perf stat record [<options>]",
1336 	NULL,
1337 };
1338 
1339 static void init_features(struct perf_session *session)
1340 {
1341 	int feat;
1342 
1343 	for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1344 		perf_header__set_feat(&session->header, feat);
1345 
1346 	perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1347 	perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1348 	perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1349 	perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1350 	perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1351 }
1352 
1353 static int __cmd_record(int argc, const char **argv)
1354 {
1355 	struct perf_session *session;
1356 	struct perf_data *data = &perf_stat.data;
1357 
1358 	argc = parse_options(argc, argv, stat_options, stat_record_usage,
1359 			     PARSE_OPT_STOP_AT_NON_OPTION);
1360 
1361 	if (output_name)
1362 		data->path = output_name;
1363 
1364 	if (stat_config.run_count != 1 || forever) {
1365 		pr_err("Cannot use -r option with perf stat record.\n");
1366 		return -1;
1367 	}
1368 
1369 	session = perf_session__new(data, false, NULL);
1370 	if (session == NULL) {
1371 		pr_err("Perf session creation failed.\n");
1372 		return -1;
1373 	}
1374 
1375 	init_features(session);
1376 
1377 	session->evlist   = evsel_list;
1378 	perf_stat.session = session;
1379 	perf_stat.record  = true;
1380 	return argc;
1381 }
1382 
1383 static int process_stat_round_event(struct perf_session *session,
1384 				    union perf_event *event)
1385 {
1386 	struct stat_round_event *stat_round = &event->stat_round;
1387 	struct perf_evsel *counter;
1388 	struct timespec tsh, *ts = NULL;
1389 	const char **argv = session->header.env.cmdline_argv;
1390 	int argc = session->header.env.nr_cmdline;
1391 
1392 	evlist__for_each_entry(evsel_list, counter)
1393 		perf_stat_process_counter(&stat_config, counter);
1394 
1395 	if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1396 		update_stats(&walltime_nsecs_stats, stat_round->time);
1397 
1398 	if (stat_config.interval && stat_round->time) {
1399 		tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
1400 		tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1401 		ts = &tsh;
1402 	}
1403 
1404 	print_counters(ts, argc, argv);
1405 	return 0;
1406 }
1407 
1408 static
1409 int process_stat_config_event(struct perf_session *session,
1410 			      union perf_event *event)
1411 {
1412 	struct perf_tool *tool = session->tool;
1413 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1414 
1415 	perf_event__read_stat_config(&stat_config, &event->stat_config);
1416 
1417 	if (cpu_map__empty(st->cpus)) {
1418 		if (st->aggr_mode != AGGR_UNSET)
1419 			pr_warning("warning: processing task data, aggregation mode not set\n");
1420 		return 0;
1421 	}
1422 
1423 	if (st->aggr_mode != AGGR_UNSET)
1424 		stat_config.aggr_mode = st->aggr_mode;
1425 
1426 	if (perf_stat.data.is_pipe)
1427 		perf_stat_init_aggr_mode();
1428 	else
1429 		perf_stat_init_aggr_mode_file(st);
1430 
1431 	return 0;
1432 }
1433 
1434 static int set_maps(struct perf_stat *st)
1435 {
1436 	if (!st->cpus || !st->threads)
1437 		return 0;
1438 
1439 	if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1440 		return -EINVAL;
1441 
1442 	perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
1443 
1444 	if (perf_evlist__alloc_stats(evsel_list, true))
1445 		return -ENOMEM;
1446 
1447 	st->maps_allocated = true;
1448 	return 0;
1449 }
1450 
1451 static
1452 int process_thread_map_event(struct perf_session *session,
1453 			     union perf_event *event)
1454 {
1455 	struct perf_tool *tool = session->tool;
1456 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1457 
1458 	if (st->threads) {
1459 		pr_warning("Extra thread map event, ignoring.\n");
1460 		return 0;
1461 	}
1462 
1463 	st->threads = thread_map__new_event(&event->thread_map);
1464 	if (!st->threads)
1465 		return -ENOMEM;
1466 
1467 	return set_maps(st);
1468 }
1469 
1470 static
1471 int process_cpu_map_event(struct perf_session *session,
1472 			  union perf_event *event)
1473 {
1474 	struct perf_tool *tool = session->tool;
1475 	struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1476 	struct cpu_map *cpus;
1477 
1478 	if (st->cpus) {
1479 		pr_warning("Extra cpu map event, ignoring.\n");
1480 		return 0;
1481 	}
1482 
1483 	cpus = cpu_map__new_data(&event->cpu_map.data);
1484 	if (!cpus)
1485 		return -ENOMEM;
1486 
1487 	st->cpus = cpus;
1488 	return set_maps(st);
1489 }
1490 
1491 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
1492 {
1493 	int i;
1494 
1495 	config->stats = calloc(nthreads, sizeof(struct runtime_stat));
1496 	if (!config->stats)
1497 		return -1;
1498 
1499 	config->stats_num = nthreads;
1500 
1501 	for (i = 0; i < nthreads; i++)
1502 		runtime_stat__init(&config->stats[i]);
1503 
1504 	return 0;
1505 }
1506 
1507 static void runtime_stat_delete(struct perf_stat_config *config)
1508 {
1509 	int i;
1510 
1511 	if (!config->stats)
1512 		return;
1513 
1514 	for (i = 0; i < config->stats_num; i++)
1515 		runtime_stat__exit(&config->stats[i]);
1516 
1517 	free(config->stats);
1518 }
1519 
1520 static const char * const stat_report_usage[] = {
1521 	"perf stat report [<options>]",
1522 	NULL,
1523 };
1524 
1525 static struct perf_stat perf_stat = {
1526 	.tool = {
1527 		.attr		= perf_event__process_attr,
1528 		.event_update	= perf_event__process_event_update,
1529 		.thread_map	= process_thread_map_event,
1530 		.cpu_map	= process_cpu_map_event,
1531 		.stat_config	= process_stat_config_event,
1532 		.stat		= perf_event__process_stat_event,
1533 		.stat_round	= process_stat_round_event,
1534 	},
1535 	.aggr_mode = AGGR_UNSET,
1536 };
1537 
1538 static int __cmd_report(int argc, const char **argv)
1539 {
1540 	struct perf_session *session;
1541 	const struct option options[] = {
1542 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1543 	OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1544 		     "aggregate counts per processor socket", AGGR_SOCKET),
1545 	OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1546 		     "aggregate counts per physical processor core", AGGR_CORE),
1547 	OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1548 		     "disable CPU count aggregation", AGGR_NONE),
1549 	OPT_END()
1550 	};
1551 	struct stat st;
1552 	int ret;
1553 
1554 	argc = parse_options(argc, argv, options, stat_report_usage, 0);
1555 
1556 	if (!input_name || !strlen(input_name)) {
1557 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1558 			input_name = "-";
1559 		else
1560 			input_name = "perf.data";
1561 	}
1562 
1563 	perf_stat.data.path = input_name;
1564 	perf_stat.data.mode = PERF_DATA_MODE_READ;
1565 
1566 	session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
1567 	if (session == NULL)
1568 		return -1;
1569 
1570 	perf_stat.session  = session;
1571 	stat_config.output = stderr;
1572 	evsel_list         = session->evlist;
1573 
1574 	ret = perf_session__process_events(session);
1575 	if (ret)
1576 		return ret;
1577 
1578 	perf_session__delete(session);
1579 	return 0;
1580 }
1581 
1582 static void setup_system_wide(int forks)
1583 {
1584 	/*
1585 	 * Make system wide (-a) the default target if
1586 	 * no target was specified and one of following
1587 	 * conditions is met:
1588 	 *
1589 	 *   - there's no workload specified
1590 	 *   - there is workload specified but all requested
1591 	 *     events are system wide events
1592 	 */
1593 	if (!target__none(&target))
1594 		return;
1595 
1596 	if (!forks)
1597 		target.system_wide = true;
1598 	else {
1599 		struct perf_evsel *counter;
1600 
1601 		evlist__for_each_entry(evsel_list, counter) {
1602 			if (!counter->system_wide)
1603 				return;
1604 		}
1605 
1606 		if (evsel_list->nr_entries)
1607 			target.system_wide = true;
1608 	}
1609 }
1610 
1611 int cmd_stat(int argc, const char **argv)
1612 {
1613 	const char * const stat_usage[] = {
1614 		"perf stat [<options>] [<command>]",
1615 		NULL
1616 	};
1617 	int status = -EINVAL, run_idx;
1618 	const char *mode;
1619 	FILE *output = stderr;
1620 	unsigned int interval, timeout;
1621 	const char * const stat_subcommands[] = { "record", "report" };
1622 
1623 	setlocale(LC_ALL, "");
1624 
1625 	evsel_list = perf_evlist__new();
1626 	if (evsel_list == NULL)
1627 		return -ENOMEM;
1628 
1629 	parse_events__shrink_config_terms();
1630 
1631 	/* String-parsing callback-based options would segfault when negated */
1632 	set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
1633 	set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
1634 	set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
1635 
1636 	argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1637 					(const char **) stat_usage,
1638 					PARSE_OPT_STOP_AT_NON_OPTION);
1639 	perf_stat__collect_metric_expr(evsel_list);
1640 	perf_stat__init_shadow_stats();
1641 
1642 	if (stat_config.csv_sep) {
1643 		stat_config.csv_output = true;
1644 		if (!strcmp(stat_config.csv_sep, "\\t"))
1645 			stat_config.csv_sep = "\t";
1646 	} else
1647 		stat_config.csv_sep = DEFAULT_SEPARATOR;
1648 
1649 	if (argc && !strncmp(argv[0], "rec", 3)) {
1650 		argc = __cmd_record(argc, argv);
1651 		if (argc < 0)
1652 			return -1;
1653 	} else if (argc && !strncmp(argv[0], "rep", 3))
1654 		return __cmd_report(argc, argv);
1655 
1656 	interval = stat_config.interval;
1657 	timeout = stat_config.timeout;
1658 
1659 	/*
1660 	 * For record command the -o is already taken care of.
1661 	 */
1662 	if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1663 		output = NULL;
1664 
1665 	if (output_name && output_fd) {
1666 		fprintf(stderr, "cannot use both --output and --log-fd\n");
1667 		parse_options_usage(stat_usage, stat_options, "o", 1);
1668 		parse_options_usage(NULL, stat_options, "log-fd", 0);
1669 		goto out;
1670 	}
1671 
1672 	if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
1673 		fprintf(stderr, "--metric-only is not supported with --per-thread\n");
1674 		goto out;
1675 	}
1676 
1677 	if (stat_config.metric_only && stat_config.run_count > 1) {
1678 		fprintf(stderr, "--metric-only is not supported with -r\n");
1679 		goto out;
1680 	}
1681 
1682 	if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
1683 		fprintf(stderr, "--table is only supported with -r\n");
1684 		parse_options_usage(stat_usage, stat_options, "r", 1);
1685 		parse_options_usage(NULL, stat_options, "table", 0);
1686 		goto out;
1687 	}
1688 
1689 	if (output_fd < 0) {
1690 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
1691 		parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1692 		goto out;
1693 	}
1694 
1695 	if (!output) {
1696 		struct timespec tm;
1697 		mode = append_file ? "a" : "w";
1698 
1699 		output = fopen(output_name, mode);
1700 		if (!output) {
1701 			perror("failed to create output file");
1702 			return -1;
1703 		}
1704 		clock_gettime(CLOCK_REALTIME, &tm);
1705 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1706 	} else if (output_fd > 0) {
1707 		mode = append_file ? "a" : "w";
1708 		output = fdopen(output_fd, mode);
1709 		if (!output) {
1710 			perror("Failed opening logfd");
1711 			return -errno;
1712 		}
1713 	}
1714 
1715 	stat_config.output = output;
1716 
1717 	/*
1718 	 * let the spreadsheet do the pretty-printing
1719 	 */
1720 	if (stat_config.csv_output) {
1721 		/* User explicitly passed -B? */
1722 		if (big_num_opt == 1) {
1723 			fprintf(stderr, "-B option not supported with -x\n");
1724 			parse_options_usage(stat_usage, stat_options, "B", 1);
1725 			parse_options_usage(NULL, stat_options, "x", 1);
1726 			goto out;
1727 		} else /* Nope, so disable big number formatting */
1728 			stat_config.big_num = false;
1729 	} else if (big_num_opt == 0) /* User passed --no-big-num */
1730 		stat_config.big_num = false;
1731 
1732 	setup_system_wide(argc);
1733 
1734 	/*
1735 	 * Display user/system times only for single
1736 	 * run and when there's specified tracee.
1737 	 */
1738 	if ((stat_config.run_count == 1) && target__none(&target))
1739 		stat_config.ru_display = true;
1740 
1741 	if (stat_config.run_count < 0) {
1742 		pr_err("Run count must be a positive number\n");
1743 		parse_options_usage(stat_usage, stat_options, "r", 1);
1744 		goto out;
1745 	} else if (stat_config.run_count == 0) {
1746 		forever = true;
1747 		stat_config.run_count = 1;
1748 	}
1749 
1750 	if (stat_config.walltime_run_table) {
1751 		stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
1752 		if (!stat_config.walltime_run) {
1753 			pr_err("failed to setup -r option");
1754 			goto out;
1755 		}
1756 	}
1757 
1758 	if ((stat_config.aggr_mode == AGGR_THREAD) &&
1759 		!target__has_task(&target)) {
1760 		if (!target.system_wide || target.cpu_list) {
1761 			fprintf(stderr, "The --per-thread option is only "
1762 				"available when monitoring via -p -t -a "
1763 				"options or only --per-thread.\n");
1764 			parse_options_usage(NULL, stat_options, "p", 1);
1765 			parse_options_usage(NULL, stat_options, "t", 1);
1766 			goto out;
1767 		}
1768 	}
1769 
1770 	/*
1771 	 * no_aggr, cgroup are for system-wide only
1772 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1773 	 */
1774 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1775 	      stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1776 	    !target__has_cpu(&target)) {
1777 		fprintf(stderr, "both cgroup and no-aggregation "
1778 			"modes only available in system-wide mode\n");
1779 
1780 		parse_options_usage(stat_usage, stat_options, "G", 1);
1781 		parse_options_usage(NULL, stat_options, "A", 1);
1782 		parse_options_usage(NULL, stat_options, "a", 1);
1783 		goto out;
1784 	}
1785 
1786 	if (add_default_attributes())
1787 		goto out;
1788 
1789 	target__validate(&target);
1790 
1791 	if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
1792 		target.per_thread = true;
1793 
1794 	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1795 		if (target__has_task(&target)) {
1796 			pr_err("Problems finding threads of monitor\n");
1797 			parse_options_usage(stat_usage, stat_options, "p", 1);
1798 			parse_options_usage(NULL, stat_options, "t", 1);
1799 		} else if (target__has_cpu(&target)) {
1800 			perror("failed to parse CPUs map");
1801 			parse_options_usage(stat_usage, stat_options, "C", 1);
1802 			parse_options_usage(NULL, stat_options, "a", 1);
1803 		}
1804 		goto out;
1805 	}
1806 
1807 	/*
1808 	 * Initialize thread_map with comm names,
1809 	 * so we could print it out on output.
1810 	 */
1811 	if (stat_config.aggr_mode == AGGR_THREAD) {
1812 		thread_map__read_comms(evsel_list->threads);
1813 		if (target.system_wide) {
1814 			if (runtime_stat_new(&stat_config,
1815 				thread_map__nr(evsel_list->threads))) {
1816 				goto out;
1817 			}
1818 		}
1819 	}
1820 
1821 	if (stat_config.times && interval)
1822 		interval_count = true;
1823 	else if (stat_config.times && !interval) {
1824 		pr_err("interval-count option should be used together with "
1825 				"interval-print.\n");
1826 		parse_options_usage(stat_usage, stat_options, "interval-count", 0);
1827 		parse_options_usage(stat_usage, stat_options, "I", 1);
1828 		goto out;
1829 	}
1830 
1831 	if (timeout && timeout < 100) {
1832 		if (timeout < 10) {
1833 			pr_err("timeout must be >= 10ms.\n");
1834 			parse_options_usage(stat_usage, stat_options, "timeout", 0);
1835 			goto out;
1836 		} else
1837 			pr_warning("timeout < 100ms. "
1838 				   "The overhead percentage could be high in some cases. "
1839 				   "Please proceed with caution.\n");
1840 	}
1841 	if (timeout && interval) {
1842 		pr_err("timeout option is not supported with interval-print.\n");
1843 		parse_options_usage(stat_usage, stat_options, "timeout", 0);
1844 		parse_options_usage(stat_usage, stat_options, "I", 1);
1845 		goto out;
1846 	}
1847 
1848 	if (perf_evlist__alloc_stats(evsel_list, interval))
1849 		goto out;
1850 
1851 	if (perf_stat_init_aggr_mode())
1852 		goto out;
1853 
1854 	/*
1855 	 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
1856 	 * while avoiding that older tools show confusing messages.
1857 	 *
1858 	 * However for pipe sessions we need to keep it zero,
1859 	 * because script's perf_evsel__check_attr is triggered
1860 	 * by attr->sample_type != 0, and we can't run it on
1861 	 * stat sessions.
1862 	 */
1863 	stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
1864 
1865 	/*
1866 	 * We dont want to block the signals - that would cause
1867 	 * child tasks to inherit that and Ctrl-C would not work.
1868 	 * What we want is for Ctrl-C to work in the exec()-ed
1869 	 * task, but being ignored by perf stat itself:
1870 	 */
1871 	atexit(sig_atexit);
1872 	if (!forever)
1873 		signal(SIGINT,  skip_signal);
1874 	signal(SIGCHLD, skip_signal);
1875 	signal(SIGALRM, skip_signal);
1876 	signal(SIGABRT, skip_signal);
1877 
1878 	status = 0;
1879 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
1880 		if (stat_config.run_count != 1 && verbose > 0)
1881 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1882 				run_idx + 1);
1883 
1884 		status = run_perf_stat(argc, argv, run_idx);
1885 		if (forever && status != -1) {
1886 			print_counters(NULL, argc, argv);
1887 			perf_stat__reset_stats();
1888 		}
1889 	}
1890 
1891 	if (!forever && status != -1 && !interval)
1892 		print_counters(NULL, argc, argv);
1893 
1894 	if (STAT_RECORD) {
1895 		/*
1896 		 * We synthesize the kernel mmap record just so that older tools
1897 		 * don't emit warnings about not being able to resolve symbols
1898 		 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
1899 		 * a saner message about no samples being in the perf.data file.
1900 		 *
1901 		 * This also serves to suppress a warning about f_header.data.size == 0
1902 		 * in header.c at the moment 'perf stat record' gets introduced, which
1903 		 * is not really needed once we start adding the stat specific PERF_RECORD_
1904 		 * records, but the need to suppress the kptr_restrict messages in older
1905 		 * tools remain  -acme
1906 		 */
1907 		int fd = perf_data__fd(&perf_stat.data);
1908 		int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
1909 							     process_synthesized_event,
1910 							     &perf_stat.session->machines.host);
1911 		if (err) {
1912 			pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
1913 				   "older tools may produce warnings about this file\n.");
1914 		}
1915 
1916 		if (!interval) {
1917 			if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
1918 				pr_err("failed to write stat round event\n");
1919 		}
1920 
1921 		if (!perf_stat.data.is_pipe) {
1922 			perf_stat.session->header.data_size += perf_stat.bytes_written;
1923 			perf_session__write_header(perf_stat.session, evsel_list, fd, true);
1924 		}
1925 
1926 		perf_session__delete(perf_stat.session);
1927 	}
1928 
1929 	perf_stat__exit_aggr_mode();
1930 	perf_evlist__free_stats(evsel_list);
1931 out:
1932 	free(stat_config.walltime_run);
1933 
1934 	if (smi_cost && smi_reset)
1935 		sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
1936 
1937 	perf_evlist__delete(evsel_list);
1938 
1939 	runtime_stat_delete(&stat_config);
1940 
1941 	return status;
1942 }
1943