?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
perf stat: Update walltime_nsecs_stats in interval mode
[linux/fpc-iii.git] / tools / perf / builtin-stat.c
blob88f1d5fbdb48823d930f402579dc33ad9ba70ff6
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/group.h"
67 #include "util/string2.h"
68 #include "util/metricgroup.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
85 #include "sane_ctype.h"
86
87 #define DEFAULT_SEPARATOR " "
88 #define CNTR_NOT_SUPPORTED "<not supported>"
89 #define CNTR_NOT_COUNTED "<not counted>"
90 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
91
92 static void print_counters(struct timespec *ts, int argc, const char **argv);
93
94 /* Default events used for perf stat -T */
95 static const char *transaction_attrs = {
96 "task-clock,"
97 "{"
98 "instructions,"
99 "cycles,"
100 "cpu/cycles-t/,"
101 "cpu/tx-start/,"
102 "cpu/el-start/,"
103 "cpu/cycles-ct/"
104 "}"
105 };
106
107 /* More limited version when the CPU does not have all events. */
108 static const char * transaction_limited_attrs = {
109 "task-clock,"
110 "{"
111 "instructions,"
112 "cycles,"
113 "cpu/cycles-t/,"
114 "cpu/tx-start/"
115 "}"
116 };
117
118 static const char * topdown_attrs[] = {
119 "topdown-total-slots",
120 "topdown-slots-retired",
121 "topdown-recovery-bubbles",
122 "topdown-fetch-bubbles",
123 "topdown-slots-issued",
124 NULL,
125 };
126
127 static const char *smi_cost_attrs = {
128 "{"
129 "msr/aperf/,"
130 "msr/smi/,"
131 "cycles"
132 "}"
133 };
134
135 static struct perf_evlist *evsel_list;
136
137 static struct rblist metric_events;
138
139 static struct target target = {
140 .uid = UINT_MAX,
141 };
142
143 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
144
145 static int run_count = 1;
146 static bool no_inherit = false;
147 static volatile pid_t child_pid = -1;
148 static bool null_run = false;
149 static int detailed_run = 0;
150 static bool transaction_run;
151 static bool topdown_run = false;
152 static bool smi_cost = false;
153 static bool smi_reset = false;
154 static bool big_num = true;
155 static int big_num_opt = -1;
156 static const char *csv_sep = NULL;
157 static bool csv_output = false;
158 static bool group = false;
159 static const char *pre_cmd = NULL;
160 static const char *post_cmd = NULL;
161 static bool sync_run = false;
162 static unsigned int initial_delay = 0;
163 static unsigned int unit_width = 4; /* strlen("unit") */
164 static bool forever = false;
165 static bool metric_only = false;
166 static bool force_metric_only = false;
167 static bool no_merge = false;
168 static struct timespec ref_time;
169 static struct cpu_map *aggr_map;
170 static aggr_get_id_t aggr_get_id;
171 static bool append_file;
172 static const char *output_name;
173 static int output_fd;
174 static int print_free_counters_hint;
175
176 struct perf_stat {
177 bool record;
178 struct perf_data_file file;
179 struct perf_session *session;
180 u64 bytes_written;
181 struct perf_tool tool;
182 bool maps_allocated;
183 struct cpu_map *cpus;
184 struct thread_map *threads;
185 enum aggr_mode aggr_mode;
186 };
187
188 static struct perf_stat perf_stat;
189 #define STAT_RECORD perf_stat.record
190
191 static volatile int done = 0;
192
193 static struct perf_stat_config stat_config = {
194 .aggr_mode = AGGR_GLOBAL,
195 .scale = true,
196 };
197
198 static bool is_duration_time(struct perf_evsel *evsel)
199 {
200 return !strcmp(evsel->name, "duration_time");
201 }
202
203 static inline void diff_timespec(struct timespec *r, struct timespec *a,
204 struct timespec *b)
205 {
206 r->tv_sec = a->tv_sec - b->tv_sec;
207 if (a->tv_nsec < b->tv_nsec) {
208 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
209 r->tv_sec--;
210 } else {
211 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
212 }
213 }
214
215 static void perf_stat__reset_stats(void)
216 {
217 perf_evlist__reset_stats(evsel_list);
218 perf_stat__reset_shadow_stats();
219 }
220
221 static int create_perf_stat_counter(struct perf_evsel *evsel)
222 {
223 struct perf_event_attr *attr = &evsel->attr;
224 struct perf_evsel *leader = evsel->leader;
225
226 if (stat_config.scale) {
227 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
228 PERF_FORMAT_TOTAL_TIME_RUNNING;
229 }
230
231 /*
232 * The event is part of non trivial group, let's enable
233 * the group read (for leader) and ID retrieval for all
234 * members.
235 */
236 if (leader->nr_members > 1)
237 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
238
239 attr->inherit = !no_inherit;
240
241 /*
242 * Some events get initialized with sample_(period/type) set,
243 * like tracepoints. Clear it up for counting.
244 */
245 attr->sample_period = 0;
246
247 /*
248 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
249 * while avoiding that older tools show confusing messages.
250 *
251 * However for pipe sessions we need to keep it zero,
252 * because script's perf_evsel__check_attr is triggered
253 * by attr->sample_type != 0, and we can't run it on
254 * stat sessions.
255 */
256 if (!(STAT_RECORD && perf_stat.file.is_pipe))
257 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
258
259 /*
260 * Disabling all counters initially, they will be enabled
261 * either manually by us or by kernel via enable_on_exec
262 * set later.
263 */
264 if (perf_evsel__is_group_leader(evsel)) {
265 attr->disabled = 1;
266
267 /*
268 * In case of initial_delay we enable tracee
269 * events manually.
270 */
271 if (target__none(&target) && !initial_delay)
272 attr->enable_on_exec = 1;
273 }
274
275 if (target__has_cpu(&target))
276 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
277
278 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
279 }
280
281 /*
282 * Does the counter have nsecs as a unit?
283 */
284 static inline int nsec_counter(struct perf_evsel *evsel)
285 {
286 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
287 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
288 return 1;
289
290 return 0;
291 }
292
293 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
294 union perf_event *event,
295 struct perf_sample *sample __maybe_unused,
296 struct machine *machine __maybe_unused)
297 {
298 if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
299 pr_err("failed to write perf data, error: %m\n");
300 return -1;
301 }
302
303 perf_stat.bytes_written += event->header.size;
304 return 0;
305 }
306
307 static int write_stat_round_event(u64 tm, u64 type)
308 {
309 return perf_event__synthesize_stat_round(NULL, tm, type,
310 process_synthesized_event,
311 NULL);
312 }
313
314 #define WRITE_STAT_ROUND_EVENT(time, interval) \
315 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
316
317 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
318
319 static int
320 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
321 struct perf_counts_values *count)
322 {
323 struct perf_sample_id *sid = SID(counter, cpu, thread);
324
325 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
326 process_synthesized_event, NULL);
327 }
328
329 /*
330 * Read out the results of a single counter:
331 * do not aggregate counts across CPUs in system-wide mode
332 */
333 static int read_counter(struct perf_evsel *counter)
334 {
335 int nthreads = thread_map__nr(evsel_list->threads);
336 int ncpus, cpu, thread;
337
338 if (target__has_cpu(&target))
339 ncpus = perf_evsel__nr_cpus(counter);
340 else
341 ncpus = 1;
342
343 if (!counter->supported)
344 return -ENOENT;
345
346 if (counter->system_wide)
347 nthreads = 1;
348
349 for (thread = 0; thread < nthreads; thread++) {
350 for (cpu = 0; cpu < ncpus; cpu++) {
351 struct perf_counts_values *count;
352
353 count = perf_counts(counter->counts, cpu, thread);
354
355 /*
356 * The leader's group read loads data into its group members
357 * (via perf_evsel__read_counter) and sets threir count->loaded.
358 */
359 if (!count->loaded &&
360 perf_evsel__read_counter(counter, cpu, thread)) {
361 counter->counts->scaled = -1;
362 perf_counts(counter->counts, cpu, thread)->ena = 0;
363 perf_counts(counter->counts, cpu, thread)->run = 0;
364 return -1;
365 }
366
367 count->loaded = false;
368
369 if (STAT_RECORD) {
370 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
371 pr_err("failed to write stat event\n");
372 return -1;
373 }
374 }
375
376 if (verbose > 1) {
377 fprintf(stat_config.output,
378 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
379 perf_evsel__name(counter),
380 cpu,
381 count->val, count->ena, count->run);
382 }
383 }
384 }
385
386 return 0;
387 }
388
389 static void read_counters(void)
390 {
391 struct perf_evsel *counter;
392 int ret;
393
394 evlist__for_each_entry(evsel_list, counter) {
395 ret = read_counter(counter);
396 if (ret)
397 pr_debug("failed to read counter %s\n", counter->name);
398
399 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
400 pr_warning("failed to process counter %s\n", counter->name);
401 }
402 }
403
404 static void process_interval(void)
405 {
406 struct timespec ts, rs;
407
408 read_counters();
409
410 clock_gettime(CLOCK_MONOTONIC, &ts);
411 diff_timespec(&rs, &ts, &ref_time);
412
413 if (STAT_RECORD) {
414 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
415 pr_err("failed to write stat round event\n");
416 }
417
418 init_stats(&walltime_nsecs_stats);
419 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
420 print_counters(&rs, 0, NULL);
421 }
422
423 static void enable_counters(void)
424 {
425 if (initial_delay)
426 usleep(initial_delay * USEC_PER_MSEC);
427
428 /*
429 * We need to enable counters only if:
430 * - we don't have tracee (attaching to task or cpu)
431 * - we have initial delay configured
432 */
433 if (!target__none(&target) || initial_delay)
434 perf_evlist__enable(evsel_list);
435 }
436
437 static void disable_counters(void)
438 {
439 /*
440 * If we don't have tracee (attaching to task or cpu), counters may
441 * still be running. To get accurate group ratios, we must stop groups
442 * from counting before reading their constituent counters.
443 */
444 if (!target__none(&target))
445 perf_evlist__disable(evsel_list);
446 }
447
448 static volatile int workload_exec_errno;
449
450 /*
451 * perf_evlist__prepare_workload will send a SIGUSR1
452 * if the fork fails, since we asked by setting its
453 * want_signal to true.
454 */
455 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
456 void *ucontext __maybe_unused)
457 {
458 workload_exec_errno = info->si_value.sival_int;
459 }
460
461 static bool has_unit(struct perf_evsel *counter)
462 {
463 return counter->unit && *counter->unit;
464 }
465
466 static bool has_scale(struct perf_evsel *counter)
467 {
468 return counter->scale != 1;
469 }
470
471 static int perf_stat_synthesize_config(bool is_pipe)
472 {
473 struct perf_evsel *counter;
474 int err;
475
476 if (is_pipe) {
477 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
478 process_synthesized_event);
479 if (err < 0) {
480 pr_err("Couldn't synthesize attrs.\n");
481 return err;
482 }
483 }
484
485 /*
486 * Synthesize other events stuff not carried within
487 * attr event - unit, scale, name
488 */
489 evlist__for_each_entry(evsel_list, counter) {
490 if (!counter->supported)
491 continue;
492
493 /*
494 * Synthesize unit and scale only if it's defined.
495 */
496 if (has_unit(counter)) {
497 err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
498 if (err < 0) {
499 pr_err("Couldn't synthesize evsel unit.\n");
500 return err;
501 }
502 }
503
504 if (has_scale(counter)) {
505 err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
506 if (err < 0) {
507 pr_err("Couldn't synthesize evsel scale.\n");
508 return err;
509 }
510 }
511
512 if (counter->own_cpus) {
513 err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
514 if (err < 0) {
515 pr_err("Couldn't synthesize evsel scale.\n");
516 return err;
517 }
518 }
519
520 /*
521 * Name is needed only for pipe output,
522 * perf.data carries event names.
523 */
524 if (is_pipe) {
525 err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
526 if (err < 0) {
527 pr_err("Couldn't synthesize evsel name.\n");
528 return err;
529 }
530 }
531 }
532
533 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
534 process_synthesized_event,
535 NULL);
536 if (err < 0) {
537 pr_err("Couldn't synthesize thread map.\n");
538 return err;
539 }
540
541 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
542 process_synthesized_event, NULL);
543 if (err < 0) {
544 pr_err("Couldn't synthesize thread map.\n");
545 return err;
546 }
547
548 err = perf_event__synthesize_stat_config(NULL, &stat_config,
549 process_synthesized_event, NULL);
550 if (err < 0) {
551 pr_err("Couldn't synthesize config.\n");
552 return err;
553 }
554
555 return 0;
556 }
557
558 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
559
560 static int __store_counter_ids(struct perf_evsel *counter,
561 struct cpu_map *cpus,
562 struct thread_map *threads)
563 {
564 int cpu, thread;
565
566 for (cpu = 0; cpu < cpus->nr; cpu++) {
567 for (thread = 0; thread < threads->nr; thread++) {
568 int fd = FD(counter, cpu, thread);
569
570 if (perf_evlist__id_add_fd(evsel_list, counter,
571 cpu, thread, fd) < 0)
572 return -1;
573 }
574 }
575
576 return 0;
577 }
578
579 static int store_counter_ids(struct perf_evsel *counter)
580 {
581 struct cpu_map *cpus = counter->cpus;
582 struct thread_map *threads = counter->threads;
583
584 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
585 return -ENOMEM;
586
587 return __store_counter_ids(counter, cpus, threads);
588 }
589
590 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
591 {
592 return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
593 }
594
595 static struct perf_evsel *perf_evsel__reset_weak_group(struct perf_evsel *evsel)
596 {
597 struct perf_evsel *c2, *leader;
598 bool is_open = true;
599
600 leader = evsel->leader;
601 pr_debug("Weak group for %s/%d failed\n",
602 leader->name, leader->nr_members);
603
604 /*
605 * for_each_group_member doesn't work here because it doesn't
606 * include the first entry.
607 */
608 evlist__for_each_entry(evsel_list, c2) {
609 if (c2 == evsel)
610 is_open = false;
611 if (c2->leader == leader) {
612 if (is_open)
613 perf_evsel__close(c2);
614 c2->leader = c2;
615 c2->nr_members = 0;
616 }
617 }
618 return leader;
619 }
620
621 static int __run_perf_stat(int argc, const char **argv)
622 {
623 int interval = stat_config.interval;
624 char msg[BUFSIZ];
625 unsigned long long t0, t1;
626 struct perf_evsel *counter;
627 struct timespec ts;
628 size_t l;
629 int status = 0;
630 const bool forks = (argc > 0);
631 bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
632 struct perf_evsel_config_term *err_term;
633
634 if (interval) {
635 ts.tv_sec = interval / USEC_PER_MSEC;
636 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
637 } else {
638 ts.tv_sec = 1;
639 ts.tv_nsec = 0;
640 }
641
642 if (forks) {
643 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
644 workload_exec_failed_signal) < 0) {
645 perror("failed to prepare workload");
646 return -1;
647 }
648 child_pid = evsel_list->workload.pid;
649 }
650
651 if (group)
652 perf_evlist__set_leader(evsel_list);
653
654 evlist__for_each_entry(evsel_list, counter) {
655 try_again:
656 if (create_perf_stat_counter(counter) < 0) {
657
658 /* Weak group failed. Reset the group. */
659 if (errno == EINVAL &&
660 counter->leader != counter &&
661 counter->weak_group) {
662 counter = perf_evsel__reset_weak_group(counter);
663 goto try_again;
664 }
665
666 /*
667 * PPC returns ENXIO for HW counters until 2.6.37
668 * (behavior changed with commit b0a873e).
669 */
670 if (errno == EINVAL || errno == ENOSYS ||
671 errno == ENOENT || errno == EOPNOTSUPP ||
672 errno == ENXIO) {
673 if (verbose > 0)
674 ui__warning("%s event is not supported by the kernel.\n",
675 perf_evsel__name(counter));
676 counter->supported = false;
677
678 if ((counter->leader != counter) ||
679 !(counter->leader->nr_members > 1))
680 continue;
681 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
682 if (verbose > 0)
683 ui__warning("%s\n", msg);
684 goto try_again;
685 }
686
687 perf_evsel__open_strerror(counter, &target,
688 errno, msg, sizeof(msg));
689 ui__error("%s\n", msg);
690
691 if (child_pid != -1)
692 kill(child_pid, SIGTERM);
693
694 return -1;
695 }
696 counter->supported = true;
697
698 l = strlen(counter->unit);
699 if (l > unit_width)
700 unit_width = l;
701
702 if (perf_evsel__should_store_id(counter) &&
703 store_counter_ids(counter))
704 return -1;
705 }
706
707 if (perf_evlist__apply_filters(evsel_list, &counter)) {
708 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
709 counter->filter, perf_evsel__name(counter), errno,
710 str_error_r(errno, msg, sizeof(msg)));
711 return -1;
712 }
713
714 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
715 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
716 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
717 str_error_r(errno, msg, sizeof(msg)));
718 return -1;
719 }
720
721 if (STAT_RECORD) {
722 int err, fd = perf_data_file__fd(&perf_stat.file);
723
724 if (is_pipe) {
725 err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
726 } else {
727 err = perf_session__write_header(perf_stat.session, evsel_list,
728 fd, false);
729 }
730
731 if (err < 0)
732 return err;
733
734 err = perf_stat_synthesize_config(is_pipe);
735 if (err < 0)
736 return err;
737 }
738
739 /*
740 * Enable counters and exec the command:
741 */
742 t0 = rdclock();
743 clock_gettime(CLOCK_MONOTONIC, &ref_time);
744
745 if (forks) {
746 perf_evlist__start_workload(evsel_list);
747 enable_counters();
748
749 if (interval) {
750 while (!waitpid(child_pid, &status, WNOHANG)) {
751 nanosleep(&ts, NULL);
752 process_interval();
753 }
754 }
755 waitpid(child_pid, &status, 0);
756
757 if (workload_exec_errno) {
758 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
759 pr_err("Workload failed: %s\n", emsg);
760 return -1;
761 }
762
763 if (WIFSIGNALED(status))
764 psignal(WTERMSIG(status), argv[0]);
765 } else {
766 enable_counters();
767 while (!done) {
768 nanosleep(&ts, NULL);
769 if (interval)
770 process_interval();
771 }
772 }
773
774 disable_counters();
775
776 t1 = rdclock();
777
778 update_stats(&walltime_nsecs_stats, t1 - t0);
779
780 /*
781 * Closing a group leader splits the group, and as we only disable
782 * group leaders, results in remaining events becoming enabled. To
783 * avoid arbitrary skew, we must read all counters before closing any
784 * group leaders.
785 */
786 read_counters();
787 perf_evlist__close(evsel_list);
788
789 return WEXITSTATUS(status);
790 }
791
792 static int run_perf_stat(int argc, const char **argv)
793 {
794 int ret;
795
796 if (pre_cmd) {
797 ret = system(pre_cmd);
798 if (ret)
799 return ret;
800 }
801
802 if (sync_run)
803 sync();
804
805 ret = __run_perf_stat(argc, argv);
806 if (ret)
807 return ret;
808
809 if (post_cmd) {
810 ret = system(post_cmd);
811 if (ret)
812 return ret;
813 }
814
815 return ret;
816 }
817
818 static void print_running(u64 run, u64 ena)
819 {
820 if (csv_output) {
821 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
822 csv_sep,
823 run,
824 csv_sep,
825 ena ? 100.0 * run / ena : 100.0);
826 } else if (run != ena) {
827 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
828 }
829 }
830
831 static void print_noise_pct(double total, double avg)
832 {
833 double pct = rel_stddev_stats(total, avg);
834
835 if (csv_output)
836 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
837 else if (pct)
838 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
839 }
840
841 static void print_noise(struct perf_evsel *evsel, double avg)
842 {
843 struct perf_stat_evsel *ps;
844
845 if (run_count == 1)
846 return;
847
848 ps = evsel->priv;
849 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
850 }
851
852 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
853 {
854 switch (stat_config.aggr_mode) {
855 case AGGR_CORE:
856 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
857 cpu_map__id_to_socket(id),
858 csv_output ? 0 : -8,
859 cpu_map__id_to_cpu(id),
860 csv_sep,
861 csv_output ? 0 : 4,
862 nr,
863 csv_sep);
864 break;
865 case AGGR_SOCKET:
866 fprintf(stat_config.output, "S%*d%s%*d%s",
867 csv_output ? 0 : -5,
868 id,
869 csv_sep,
870 csv_output ? 0 : 4,
871 nr,
872 csv_sep);
873 break;
874 case AGGR_NONE:
875 fprintf(stat_config.output, "CPU%*d%s",
876 csv_output ? 0 : -4,
877 perf_evsel__cpus(evsel)->map[id], csv_sep);
878 break;
879 case AGGR_THREAD:
880 fprintf(stat_config.output, "%*s-%*d%s",
881 csv_output ? 0 : 16,
882 thread_map__comm(evsel->threads, id),
883 csv_output ? 0 : -8,
884 thread_map__pid(evsel->threads, id),
885 csv_sep);
886 break;
887 case AGGR_GLOBAL:
888 case AGGR_UNSET:
889 default:
890 break;
891 }
892 }
893
894 struct outstate {
895 FILE *fh;
896 bool newline;
897 const char *prefix;
898 int nfields;
899 int id, nr;
900 struct perf_evsel *evsel;
901 };
902
903 #define METRIC_LEN 35
904
905 static void new_line_std(void *ctx)
906 {
907 struct outstate *os = ctx;
908
909 os->newline = true;
910 }
911
912 static void do_new_line_std(struct outstate *os)
913 {
914 fputc('\n', os->fh);
915 fputs(os->prefix, os->fh);
916 aggr_printout(os->evsel, os->id, os->nr);
917 if (stat_config.aggr_mode == AGGR_NONE)
918 fprintf(os->fh, " ");
919 fprintf(os->fh, " ");
920 }
921
922 static void print_metric_std(void *ctx, const char *color, const char *fmt,
923 const char *unit, double val)
924 {
925 struct outstate *os = ctx;
926 FILE *out = os->fh;
927 int n;
928 bool newline = os->newline;
929
930 os->newline = false;
931
932 if (unit == NULL || fmt == NULL) {
933 fprintf(out, "%-*s", METRIC_LEN, "");
934 return;
935 }
936
937 if (newline)
938 do_new_line_std(os);
939
940 n = fprintf(out, " # ");
941 if (color)
942 n += color_fprintf(out, color, fmt, val);
943 else
944 n += fprintf(out, fmt, val);
945 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
946 }
947
948 static void new_line_csv(void *ctx)
949 {
950 struct outstate *os = ctx;
951 int i;
952
953 fputc('\n', os->fh);
954 if (os->prefix)
955 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
956 aggr_printout(os->evsel, os->id, os->nr);
957 for (i = 0; i < os->nfields; i++)
958 fputs(csv_sep, os->fh);
959 }
960
961 static void print_metric_csv(void *ctx,
962 const char *color __maybe_unused,
963 const char *fmt, const char *unit, double val)
964 {
965 struct outstate *os = ctx;
966 FILE *out = os->fh;
967 char buf[64], *vals, *ends;
968
969 if (unit == NULL || fmt == NULL) {
970 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
971 return;
972 }
973 snprintf(buf, sizeof(buf), fmt, val);
974 ends = vals = ltrim(buf);
975 while (isdigit(*ends) || *ends == '.')
976 ends++;
977 *ends = 0;
978 while (isspace(*unit))
979 unit++;
980 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
981 }
982
983 #define METRIC_ONLY_LEN 20
984
985 /* Filter out some columns that don't work well in metrics only mode */
986
987 static bool valid_only_metric(const char *unit)
988 {
989 if (!unit)
990 return false;
991 if (strstr(unit, "/sec") ||
992 strstr(unit, "hz") ||
993 strstr(unit, "Hz") ||
994 strstr(unit, "CPUs utilized"))
995 return false;
996 return true;
997 }
998
999 static const char *fixunit(char *buf, struct perf_evsel *evsel,
1000 const char *unit)
1001 {
1002 if (!strncmp(unit, "of all", 6)) {
1003 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
1004 unit);
1005 return buf;
1006 }
1007 return unit;
1008 }
1009
1010 static void print_metric_only(void *ctx, const char *color, const char *fmt,
1011 const char *unit, double val)
1012 {
1013 struct outstate *os = ctx;
1014 FILE *out = os->fh;
1015 int n;
1016 char buf[1024];
1017 unsigned mlen = METRIC_ONLY_LEN;
1018
1019 if (!valid_only_metric(unit))
1020 return;
1021 unit = fixunit(buf, os->evsel, unit);
1022 if (color)
1023 n = color_fprintf(out, color, fmt, val);
1024 else
1025 n = fprintf(out, fmt, val);
1026 if (n > METRIC_ONLY_LEN)
1027 n = METRIC_ONLY_LEN;
1028 if (mlen < strlen(unit))
1029 mlen = strlen(unit) + 1;
1030 fprintf(out, "%*s", mlen - n, "");
1031 }
1032
1033 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
1034 const char *fmt,
1035 const char *unit, double val)
1036 {
1037 struct outstate *os = ctx;
1038 FILE *out = os->fh;
1039 char buf[64], *vals, *ends;
1040 char tbuf[1024];
1041
1042 if (!valid_only_metric(unit))
1043 return;
1044 unit = fixunit(tbuf, os->evsel, unit);
1045 snprintf(buf, sizeof buf, fmt, val);
1046 ends = vals = ltrim(buf);
1047 while (isdigit(*ends) || *ends == '.')
1048 ends++;
1049 *ends = 0;
1050 fprintf(out, "%s%s", vals, csv_sep);
1051 }
1052
1053 static void new_line_metric(void *ctx __maybe_unused)
1054 {
1055 }
1056
1057 static void print_metric_header(void *ctx, const char *color __maybe_unused,
1058 const char *fmt __maybe_unused,
1059 const char *unit, double val __maybe_unused)
1060 {
1061 struct outstate *os = ctx;
1062 char tbuf[1024];
1063
1064 if (!valid_only_metric(unit))
1065 return;
1066 unit = fixunit(tbuf, os->evsel, unit);
1067 if (csv_output)
1068 fprintf(os->fh, "%s%s", unit, csv_sep);
1069 else
1070 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
1071 }
1072
1073 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1074 {
1075 FILE *output = stat_config.output;
1076 double msecs = avg / NSEC_PER_MSEC;
1077 const char *fmt_v, *fmt_n;
1078 char name[25];
1079
1080 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
1081 fmt_n = csv_output ? "%s" : "%-25s";
1082
1083 aggr_printout(evsel, id, nr);
1084
1085 scnprintf(name, sizeof(name), "%s%s",
1086 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
1087
1088 fprintf(output, fmt_v, msecs, csv_sep);
1089
1090 if (csv_output)
1091 fprintf(output, "%s%s", evsel->unit, csv_sep);
1092 else
1093 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
1094
1095 fprintf(output, fmt_n, name);
1096
1097 if (evsel->cgrp)
1098 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1099 }
1100
1101 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1102 {
1103 int i;
1104
1105 if (!aggr_get_id)
1106 return 0;
1107
1108 if (stat_config.aggr_mode == AGGR_NONE)
1109 return id;
1110
1111 if (stat_config.aggr_mode == AGGR_GLOBAL)
1112 return 0;
1113
1114 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1115 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1116
1117 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1118 return cpu2;
1119 }
1120 return 0;
1121 }
1122
1123 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1124 {
1125 FILE *output = stat_config.output;
1126 double sc = evsel->scale;
1127 const char *fmt;
1128
1129 if (csv_output) {
1130 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
1131 } else {
1132 if (big_num)
1133 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1134 else
1135 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1136 }
1137
1138 aggr_printout(evsel, id, nr);
1139
1140 fprintf(output, fmt, avg, csv_sep);
1141
1142 if (evsel->unit)
1143 fprintf(output, "%-*s%s",
1144 csv_output ? 0 : unit_width,
1145 evsel->unit, csv_sep);
1146
1147 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1148
1149 if (evsel->cgrp)
1150 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1151 }
1152
1153 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1154 char *prefix, u64 run, u64 ena, double noise)
1155 {
1156 struct perf_stat_output_ctx out;
1157 struct outstate os = {
1158 .fh = stat_config.output,
1159 .prefix = prefix ? prefix : "",
1160 .id = id,
1161 .nr = nr,
1162 .evsel = counter,
1163 };
1164 print_metric_t pm = print_metric_std;
1165 void (*nl)(void *);
1166
1167 if (metric_only) {
1168 nl = new_line_metric;
1169 if (csv_output)
1170 pm = print_metric_only_csv;
1171 else
1172 pm = print_metric_only;
1173 } else
1174 nl = new_line_std;
1175
1176 if (csv_output && !metric_only) {
1177 static int aggr_fields[] = {
1178 [AGGR_GLOBAL] = 0,
1179 [AGGR_THREAD] = 1,
1180 [AGGR_NONE] = 1,
1181 [AGGR_SOCKET] = 2,
1182 [AGGR_CORE] = 2,
1183 };
1184
1185 pm = print_metric_csv;
1186 nl = new_line_csv;
1187 os.nfields = 3;
1188 os.nfields += aggr_fields[stat_config.aggr_mode];
1189 if (counter->cgrp)
1190 os.nfields++;
1191 }
1192 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1193 if (metric_only) {
1194 pm(&os, NULL, "", "", 0);
1195 return;
1196 }
1197 aggr_printout(counter, id, nr);
1198
1199 fprintf(stat_config.output, "%*s%s",
1200 csv_output ? 0 : 18,
1201 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1202 csv_sep);
1203
1204 if (counter->supported)
1205 print_free_counters_hint = 1;
1206
1207 fprintf(stat_config.output, "%-*s%s",
1208 csv_output ? 0 : unit_width,
1209 counter->unit, csv_sep);
1210
1211 fprintf(stat_config.output, "%*s",
1212 csv_output ? 0 : -25,
1213 perf_evsel__name(counter));
1214
1215 if (counter->cgrp)
1216 fprintf(stat_config.output, "%s%s",
1217 csv_sep, counter->cgrp->name);
1218
1219 if (!csv_output)
1220 pm(&os, NULL, NULL, "", 0);
1221 print_noise(counter, noise);
1222 print_running(run, ena);
1223 if (csv_output)
1224 pm(&os, NULL, NULL, "", 0);
1225 return;
1226 }
1227
1228 if (metric_only)
1229 /* nothing */;
1230 else if (nsec_counter(counter))
1231 nsec_printout(id, nr, counter, uval);
1232 else
1233 abs_printout(id, nr, counter, uval);
1234
1235 out.print_metric = pm;
1236 out.new_line = nl;
1237 out.ctx = &os;
1238 out.force_header = false;
1239
1240 if (csv_output && !metric_only) {
1241 print_noise(counter, noise);
1242 print_running(run, ena);
1243 }
1244
1245 perf_stat__print_shadow_stats(counter, uval,
1246 first_shadow_cpu(counter, id),
1247 &out, &metric_events);
1248 if (!csv_output && !metric_only) {
1249 print_noise(counter, noise);
1250 print_running(run, ena);
1251 }
1252 }
1253
1254 static void aggr_update_shadow(void)
1255 {
1256 int cpu, s2, id, s;
1257 u64 val;
1258 struct perf_evsel *counter;
1259
1260 for (s = 0; s < aggr_map->nr; s++) {
1261 id = aggr_map->map[s];
1262 evlist__for_each_entry(evsel_list, counter) {
1263 val = 0;
1264 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1265 s2 = aggr_get_id(evsel_list->cpus, cpu);
1266 if (s2 != id)
1267 continue;
1268 val += perf_counts(counter->counts, cpu, 0)->val;
1269 }
1270 val = val * counter->scale;
1271 perf_stat__update_shadow_stats(counter, &val,
1272 first_shadow_cpu(counter, id));
1273 }
1274 }
1275 }
1276
1277 static void collect_all_aliases(struct perf_evsel *counter,
1278 void (*cb)(struct perf_evsel *counter, void *data,
1279 bool first),
1280 void *data)
1281 {
1282 struct perf_evsel *alias;
1283
1284 alias = list_prepare_entry(counter, &(evsel_list->entries), node);
1285 list_for_each_entry_continue (alias, &evsel_list->entries, node) {
1286 if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) ||
1287 alias->scale != counter->scale ||
1288 alias->cgrp != counter->cgrp ||
1289 strcmp(alias->unit, counter->unit) ||
1290 nsec_counter(alias) != nsec_counter(counter))
1291 break;
1292 alias->merged_stat = true;
1293 cb(alias, data, false);
1294 }
1295 }
1296
1297 static bool collect_data(struct perf_evsel *counter,
1298 void (*cb)(struct perf_evsel *counter, void *data,
1299 bool first),
1300 void *data)
1301 {
1302 if (counter->merged_stat)
1303 return false;
1304 cb(counter, data, true);
1305 if (!no_merge && counter->auto_merge_stats)
1306 collect_all_aliases(counter, cb, data);
1307 return true;
1308 }
1309
1310 struct aggr_data {
1311 u64 ena, run, val;
1312 int id;
1313 int nr;
1314 int cpu;
1315 };
1316
1317 static void aggr_cb(struct perf_evsel *counter, void *data, bool first)
1318 {
1319 struct aggr_data *ad = data;
1320 int cpu, s2;
1321
1322 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1323 struct perf_counts_values *counts;
1324
1325 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1326 if (s2 != ad->id)
1327 continue;
1328 if (first)
1329 ad->nr++;
1330 counts = perf_counts(counter->counts, cpu, 0);
1331 /*
1332 * When any result is bad, make them all to give
1333 * consistent output in interval mode.
1334 */
1335 if (counts->ena == 0 || counts->run == 0 ||
1336 counter->counts->scaled == -1) {
1337 ad->ena = 0;
1338 ad->run = 0;
1339 break;
1340 }
1341 ad->val += counts->val;
1342 ad->ena += counts->ena;
1343 ad->run += counts->run;
1344 }
1345 }
1346
1347 static void print_aggr(char *prefix)
1348 {
1349 FILE *output = stat_config.output;
1350 struct perf_evsel *counter;
1351 int s, id, nr;
1352 double uval;
1353 u64 ena, run, val;
1354 bool first;
1355
1356 if (!(aggr_map || aggr_get_id))
1357 return;
1358
1359 aggr_update_shadow();
1360
1361 /*
1362 * With metric_only everything is on a single line.
1363 * Without each counter has its own line.
1364 */
1365 for (s = 0; s < aggr_map->nr; s++) {
1366 struct aggr_data ad;
1367 if (prefix && metric_only)
1368 fprintf(output, "%s", prefix);
1369
1370 ad.id = id = aggr_map->map[s];
1371 first = true;
1372 evlist__for_each_entry(evsel_list, counter) {
1373 if (is_duration_time(counter))
1374 continue;
1375
1376 ad.val = ad.ena = ad.run = 0;
1377 ad.nr = 0;
1378 if (!collect_data(counter, aggr_cb, &ad))
1379 continue;
1380 nr = ad.nr;
1381 ena = ad.ena;
1382 run = ad.run;
1383 val = ad.val;
1384 if (first && metric_only) {
1385 first = false;
1386 aggr_printout(counter, id, nr);
1387 }
1388 if (prefix && !metric_only)
1389 fprintf(output, "%s", prefix);
1390
1391 uval = val * counter->scale;
1392 printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1393 if (!metric_only)
1394 fputc('\n', output);
1395 }
1396 if (metric_only)
1397 fputc('\n', output);
1398 }
1399 }
1400
1401 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1402 {
1403 FILE *output = stat_config.output;
1404 int nthreads = thread_map__nr(counter->threads);
1405 int ncpus = cpu_map__nr(counter->cpus);
1406 int cpu, thread;
1407 double uval;
1408
1409 for (thread = 0; thread < nthreads; thread++) {
1410 u64 ena = 0, run = 0, val = 0;
1411
1412 for (cpu = 0; cpu < ncpus; cpu++) {
1413 val += perf_counts(counter->counts, cpu, thread)->val;
1414 ena += perf_counts(counter->counts, cpu, thread)->ena;
1415 run += perf_counts(counter->counts, cpu, thread)->run;
1416 }
1417
1418 if (prefix)
1419 fprintf(output, "%s", prefix);
1420
1421 uval = val * counter->scale;
1422 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1423 fputc('\n', output);
1424 }
1425 }
1426
1427 struct caggr_data {
1428 double avg, avg_enabled, avg_running;
1429 };
1430
1431 static void counter_aggr_cb(struct perf_evsel *counter, void *data,
1432 bool first __maybe_unused)
1433 {
1434 struct caggr_data *cd = data;
1435 struct perf_stat_evsel *ps = counter->priv;
1436
1437 cd->avg += avg_stats(&ps->res_stats[0]);
1438 cd->avg_enabled += avg_stats(&ps->res_stats[1]);
1439 cd->avg_running += avg_stats(&ps->res_stats[2]);
1440 }
1441
1442 /*
1443 * Print out the results of a single counter:
1444 * aggregated counts in system-wide mode
1445 */
1446 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1447 {
1448 FILE *output = stat_config.output;
1449 double uval;
1450 struct caggr_data cd = { .avg = 0.0 };
1451
1452 if (!collect_data(counter, counter_aggr_cb, &cd))
1453 return;
1454
1455 if (prefix && !metric_only)
1456 fprintf(output, "%s", prefix);
1457
1458 uval = cd.avg * counter->scale;
1459 printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled, cd.avg);
1460 if (!metric_only)
1461 fprintf(output, "\n");
1462 }
1463
1464 static void counter_cb(struct perf_evsel *counter, void *data,
1465 bool first __maybe_unused)
1466 {
1467 struct aggr_data *ad = data;
1468
1469 ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
1470 ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
1471 ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
1472 }
1473
1474 /*
1475 * Print out the results of a single counter:
1476 * does not use aggregated count in system-wide
1477 */
1478 static void print_counter(struct perf_evsel *counter, char *prefix)
1479 {
1480 FILE *output = stat_config.output;
1481 u64 ena, run, val;
1482 double uval;
1483 int cpu;
1484
1485 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1486 struct aggr_data ad = { .cpu = cpu };
1487
1488 if (!collect_data(counter, counter_cb, &ad))
1489 return;
1490 val = ad.val;
1491 ena = ad.ena;
1492 run = ad.run;
1493
1494 if (prefix)
1495 fprintf(output, "%s", prefix);
1496
1497 uval = val * counter->scale;
1498 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1499
1500 fputc('\n', output);
1501 }
1502 }
1503
1504 static void print_no_aggr_metric(char *prefix)
1505 {
1506 int cpu;
1507 int nrcpus = 0;
1508 struct perf_evsel *counter;
1509 u64 ena, run, val;
1510 double uval;
1511
1512 nrcpus = evsel_list->cpus->nr;
1513 for (cpu = 0; cpu < nrcpus; cpu++) {
1514 bool first = true;
1515
1516 if (prefix)
1517 fputs(prefix, stat_config.output);
1518 evlist__for_each_entry(evsel_list, counter) {
1519 if (is_duration_time(counter))
1520 continue;
1521 if (first) {
1522 aggr_printout(counter, cpu, 0);
1523 first = false;
1524 }
1525 val = perf_counts(counter->counts, cpu, 0)->val;
1526 ena = perf_counts(counter->counts, cpu, 0)->ena;
1527 run = perf_counts(counter->counts, cpu, 0)->run;
1528
1529 uval = val * counter->scale;
1530 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1531 }
1532 fputc('\n', stat_config.output);
1533 }
1534 }
1535
1536 static int aggr_header_lens[] = {
1537 [AGGR_CORE] = 18,
1538 [AGGR_SOCKET] = 12,
1539 [AGGR_NONE] = 6,
1540 [AGGR_THREAD] = 24,
1541 [AGGR_GLOBAL] = 0,
1542 };
1543
1544 static const char *aggr_header_csv[] = {
1545 [AGGR_CORE] = "core,cpus,",
1546 [AGGR_SOCKET] = "socket,cpus",
1547 [AGGR_NONE] = "cpu,",
1548 [AGGR_THREAD] = "comm-pid,",
1549 [AGGR_GLOBAL] = ""
1550 };
1551
1552 static void print_metric_headers(const char *prefix, bool no_indent)
1553 {
1554 struct perf_stat_output_ctx out;
1555 struct perf_evsel *counter;
1556 struct outstate os = {
1557 .fh = stat_config.output
1558 };
1559
1560 if (prefix)
1561 fprintf(stat_config.output, "%s", prefix);
1562
1563 if (!csv_output && !no_indent)
1564 fprintf(stat_config.output, "%*s",
1565 aggr_header_lens[stat_config.aggr_mode], "");
1566 if (csv_output) {
1567 if (stat_config.interval)
1568 fputs("time,", stat_config.output);
1569 fputs(aggr_header_csv[stat_config.aggr_mode],
1570 stat_config.output);
1571 }
1572
1573 /* Print metrics headers only */
1574 evlist__for_each_entry(evsel_list, counter) {
1575 if (is_duration_time(counter))
1576 continue;
1577 os.evsel = counter;
1578 out.ctx = &os;
1579 out.print_metric = print_metric_header;
1580 out.new_line = new_line_metric;
1581 out.force_header = true;
1582 os.evsel = counter;
1583 perf_stat__print_shadow_stats(counter, 0,
1584 0,
1585 &out,
1586 &metric_events);
1587 }
1588 fputc('\n', stat_config.output);
1589 }
1590
1591 static void print_interval(char *prefix, struct timespec *ts)
1592 {
1593 FILE *output = stat_config.output;
1594 static int num_print_interval;
1595
1596 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1597
1598 if (num_print_interval == 0 && !csv_output) {
1599 switch (stat_config.aggr_mode) {
1600 case AGGR_SOCKET:
1601 fprintf(output, "# time socket cpus");
1602 if (!metric_only)
1603 fprintf(output, " counts %*s events\n", unit_width, "unit");
1604 break;
1605 case AGGR_CORE:
1606 fprintf(output, "# time core cpus");
1607 if (!metric_only)
1608 fprintf(output, " counts %*s events\n", unit_width, "unit");
1609 break;
1610 case AGGR_NONE:
1611 fprintf(output, "# time CPU");
1612 if (!metric_only)
1613 fprintf(output, " counts %*s events\n", unit_width, "unit");
1614 break;
1615 case AGGR_THREAD:
1616 fprintf(output, "# time comm-pid");
1617 if (!metric_only)
1618 fprintf(output, " counts %*s events\n", unit_width, "unit");
1619 break;
1620 case AGGR_GLOBAL:
1621 default:
1622 fprintf(output, "# time");
1623 if (!metric_only)
1624 fprintf(output, " counts %*s events\n", unit_width, "unit");
1625 case AGGR_UNSET:
1626 break;
1627 }
1628 }
1629
1630 if (num_print_interval == 0 && metric_only)
1631 print_metric_headers(" ", true);
1632 if (++num_print_interval == 25)
1633 num_print_interval = 0;
1634 }
1635
1636 static void print_header(int argc, const char **argv)
1637 {
1638 FILE *output = stat_config.output;
1639 int i;
1640
1641 fflush(stdout);
1642
1643 if (!csv_output) {
1644 fprintf(output, "\n");
1645 fprintf(output, " Performance counter stats for ");
1646 if (target.system_wide)
1647 fprintf(output, "\'system wide");
1648 else if (target.cpu_list)
1649 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1650 else if (!target__has_task(&target)) {
1651 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1652 for (i = 1; argv && (i < argc); i++)
1653 fprintf(output, " %s", argv[i]);
1654 } else if (target.pid)
1655 fprintf(output, "process id \'%s", target.pid);
1656 else
1657 fprintf(output, "thread id \'%s", target.tid);
1658
1659 fprintf(output, "\'");
1660 if (run_count > 1)
1661 fprintf(output, " (%d runs)", run_count);
1662 fprintf(output, ":\n\n");
1663 }
1664 }
1665
1666 static void print_footer(void)
1667 {
1668 FILE *output = stat_config.output;
1669 int n;
1670
1671 if (!null_run)
1672 fprintf(output, "\n");
1673 fprintf(output, " %17.9f seconds time elapsed",
1674 avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC);
1675 if (run_count > 1) {
1676 fprintf(output, " ");
1677 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1678 avg_stats(&walltime_nsecs_stats));
1679 }
1680 fprintf(output, "\n\n");
1681
1682 if (print_free_counters_hint &&
1683 sysctl__read_int("kernel/nmi_watchdog", &n) >= 0 &&
1684 n > 0)
1685 fprintf(output,
1686 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1687 " echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1688 " perf stat ...\n"
1689 " echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1690 }
1691
1692 static void print_counters(struct timespec *ts, int argc, const char **argv)
1693 {
1694 int interval = stat_config.interval;
1695 struct perf_evsel *counter;
1696 char buf[64], *prefix = NULL;
1697
1698 /* Do not print anything if we record to the pipe. */
1699 if (STAT_RECORD && perf_stat.file.is_pipe)
1700 return;
1701
1702 if (interval)
1703 print_interval(prefix = buf, ts);
1704 else
1705 print_header(argc, argv);
1706
1707 if (metric_only) {
1708 static int num_print_iv;
1709
1710 if (num_print_iv == 0 && !interval)
1711 print_metric_headers(prefix, false);
1712 if (num_print_iv++ == 25)
1713 num_print_iv = 0;
1714 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1715 fprintf(stat_config.output, "%s", prefix);
1716 }
1717
1718 switch (stat_config.aggr_mode) {
1719 case AGGR_CORE:
1720 case AGGR_SOCKET:
1721 print_aggr(prefix);
1722 break;
1723 case AGGR_THREAD:
1724 evlist__for_each_entry(evsel_list, counter) {
1725 if (is_duration_time(counter))
1726 continue;
1727 print_aggr_thread(counter, prefix);
1728 }
1729 break;
1730 case AGGR_GLOBAL:
1731 evlist__for_each_entry(evsel_list, counter) {
1732 if (is_duration_time(counter))
1733 continue;
1734 print_counter_aggr(counter, prefix);
1735 }
1736 if (metric_only)
1737 fputc('\n', stat_config.output);
1738 break;
1739 case AGGR_NONE:
1740 if (metric_only)
1741 print_no_aggr_metric(prefix);
1742 else {
1743 evlist__for_each_entry(evsel_list, counter) {
1744 if (is_duration_time(counter))
1745 continue;
1746 print_counter(counter, prefix);
1747 }
1748 }
1749 break;
1750 case AGGR_UNSET:
1751 default:
1752 break;
1753 }
1754
1755 if (!interval && !csv_output)
1756 print_footer();
1757
1758 fflush(stat_config.output);
1759 }
1760
1761 static volatile int signr = -1;
1762
1763 static void skip_signal(int signo)
1764 {
1765 if ((child_pid == -1) || stat_config.interval)
1766 done = 1;
1767
1768 signr = signo;
1769 /*
1770 * render child_pid harmless
1771 * won't send SIGTERM to a random
1772 * process in case of race condition
1773 * and fast PID recycling
1774 */
1775 child_pid = -1;
1776 }
1777
1778 static void sig_atexit(void)
1779 {
1780 sigset_t set, oset;
1781
1782 /*
1783 * avoid race condition with SIGCHLD handler
1784 * in skip_signal() which is modifying child_pid
1785 * goal is to avoid send SIGTERM to a random
1786 * process
1787 */
1788 sigemptyset(&set);
1789 sigaddset(&set, SIGCHLD);
1790 sigprocmask(SIG_BLOCK, &set, &oset);
1791
1792 if (child_pid != -1)
1793 kill(child_pid, SIGTERM);
1794
1795 sigprocmask(SIG_SETMASK, &oset, NULL);
1796
1797 if (signr == -1)
1798 return;
1799
1800 signal(signr, SIG_DFL);
1801 kill(getpid(), signr);
1802 }
1803
1804 static int stat__set_big_num(const struct option *opt __maybe_unused,
1805 const char *s __maybe_unused, int unset)
1806 {
1807 big_num_opt = unset ? 0 : 1;
1808 return 0;
1809 }
1810
1811 static int enable_metric_only(const struct option *opt __maybe_unused,
1812 const char *s __maybe_unused, int unset)
1813 {
1814 force_metric_only = true;
1815 metric_only = !unset;
1816 return 0;
1817 }
1818
1819 static int parse_metric_groups(const struct option *opt,
1820 const char *str,
1821 int unset __maybe_unused)
1822 {
1823 return metricgroup__parse_groups(opt, str, &metric_events);
1824 }
1825
1826 static const struct option stat_options[] = {
1827 OPT_BOOLEAN('T', "transaction", &transaction_run,
1828 "hardware transaction statistics"),
1829 OPT_CALLBACK('e', "event", &evsel_list, "event",
1830 "event selector. use 'perf list' to list available events",
1831 parse_events_option),
1832 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1833 "event filter", parse_filter),
1834 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1835 "child tasks do not inherit counters"),
1836 OPT_STRING('p', "pid", &target.pid, "pid",
1837 "stat events on existing process id"),
1838 OPT_STRING('t', "tid", &target.tid, "tid",
1839 "stat events on existing thread id"),
1840 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1841 "system-wide collection from all CPUs"),
1842 OPT_BOOLEAN('g', "group", &group,
1843 "put the counters into a counter group"),
1844 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1845 OPT_INCR('v', "verbose", &verbose,
1846 "be more verbose (show counter open errors, etc)"),
1847 OPT_INTEGER('r', "repeat", &run_count,
1848 "repeat command and print average + stddev (max: 100, forever: 0)"),
1849 OPT_BOOLEAN('n', "null", &null_run,
1850 "null run - dont start any counters"),
1851 OPT_INCR('d', "detailed", &detailed_run,
1852 "detailed run - start a lot of events"),
1853 OPT_BOOLEAN('S', "sync", &sync_run,
1854 "call sync() before starting a run"),
1855 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1856 "print large numbers with thousands\' separators",
1857 stat__set_big_num),
1858 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1859 "list of cpus to monitor in system-wide"),
1860 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1861 "disable CPU count aggregation", AGGR_NONE),
1862 OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"),
1863 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1864 "print counts with custom separator"),
1865 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1866 "monitor event in cgroup name only", parse_cgroups),
1867 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1868 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1869 OPT_INTEGER(0, "log-fd", &output_fd,
1870 "log output to fd, instead of stderr"),
1871 OPT_STRING(0, "pre", &pre_cmd, "command",
1872 "command to run prior to the measured command"),
1873 OPT_STRING(0, "post", &post_cmd, "command",
1874 "command to run after to the measured command"),
1875 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1876 "print counts at regular interval in ms (>= 10)"),
1877 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1878 "aggregate counts per processor socket", AGGR_SOCKET),
1879 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1880 "aggregate counts per physical processor core", AGGR_CORE),
1881 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1882 "aggregate counts per thread", AGGR_THREAD),
1883 OPT_UINTEGER('D', "delay", &initial_delay,
1884 "ms to wait before starting measurement after program start"),
1885 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1886 "Only print computed metrics. No raw values", enable_metric_only),
1887 OPT_BOOLEAN(0, "topdown", &topdown_run,
1888 "measure topdown level 1 statistics"),
1889 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1890 "measure SMI cost"),
1891 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1892 "monitor specified metrics or metric groups (separated by ,)",
1893 parse_metric_groups),
1894 OPT_END()
1895 };
1896
1897 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1898 {
1899 return cpu_map__get_socket(map, cpu, NULL);
1900 }
1901
1902 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1903 {
1904 return cpu_map__get_core(map, cpu, NULL);
1905 }
1906
1907 static int cpu_map__get_max(struct cpu_map *map)
1908 {
1909 int i, max = -1;
1910
1911 for (i = 0; i < map->nr; i++) {
1912 if (map->map[i] > max)
1913 max = map->map[i];
1914 }
1915
1916 return max;
1917 }
1918
1919 static struct cpu_map *cpus_aggr_map;
1920
1921 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1922 {
1923 int cpu;
1924
1925 if (idx >= map->nr)
1926 return -1;
1927
1928 cpu = map->map[idx];
1929
1930 if (cpus_aggr_map->map[cpu] == -1)
1931 cpus_aggr_map->map[cpu] = get_id(map, idx);
1932
1933 return cpus_aggr_map->map[cpu];
1934 }
1935
1936 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1937 {
1938 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1939 }
1940
1941 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1942 {
1943 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1944 }
1945
1946 static int perf_stat_init_aggr_mode(void)
1947 {
1948 int nr;
1949
1950 switch (stat_config.aggr_mode) {
1951 case AGGR_SOCKET:
1952 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1953 perror("cannot build socket map");
1954 return -1;
1955 }
1956 aggr_get_id = perf_stat__get_socket_cached;
1957 break;
1958 case AGGR_CORE:
1959 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1960 perror("cannot build core map");
1961 return -1;
1962 }
1963 aggr_get_id = perf_stat__get_core_cached;
1964 break;
1965 case AGGR_NONE:
1966 case AGGR_GLOBAL:
1967 case AGGR_THREAD:
1968 case AGGR_UNSET:
1969 default:
1970 break;
1971 }
1972
1973 /*
1974 * The evsel_list->cpus is the base we operate on,
1975 * taking the highest cpu number to be the size of
1976 * the aggregation translate cpumap.
1977 */
1978 nr = cpu_map__get_max(evsel_list->cpus);
1979 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1980 return cpus_aggr_map ? 0 : -ENOMEM;
1981 }
1982
1983 static void perf_stat__exit_aggr_mode(void)
1984 {
1985 cpu_map__put(aggr_map);
1986 cpu_map__put(cpus_aggr_map);
1987 aggr_map = NULL;
1988 cpus_aggr_map = NULL;
1989 }
1990
1991 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1992 {
1993 int cpu;
1994
1995 if (idx > map->nr)
1996 return -1;
1997
1998 cpu = map->map[idx];
1999
2000 if (cpu >= env->nr_cpus_avail)
2001 return -1;
2002
2003 return cpu;
2004 }
2005
2006 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
2007 {
2008 struct perf_env *env = data;
2009 int cpu = perf_env__get_cpu(env, map, idx);
2010
2011 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
2012 }
2013
2014 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
2015 {
2016 struct perf_env *env = data;
2017 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
2018
2019 if (cpu != -1) {
2020 int socket_id = env->cpu[cpu].socket_id;
2021
2022 /*
2023 * Encode socket in upper 16 bits
2024 * core_id is relative to socket, and
2025 * we need a global id. So we combine
2026 * socket + core id.
2027 */
2028 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
2029 }
2030
2031 return core;
2032 }
2033
2034 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
2035 struct cpu_map **sockp)
2036 {
2037 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
2038 }
2039
2040 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
2041 struct cpu_map **corep)
2042 {
2043 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
2044 }
2045
2046 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
2047 {
2048 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
2049 }
2050
2051 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
2052 {
2053 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
2054 }
2055
2056 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
2057 {
2058 struct perf_env *env = &st->session->header.env;
2059
2060 switch (stat_config.aggr_mode) {
2061 case AGGR_SOCKET:
2062 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
2063 perror("cannot build socket map");
2064 return -1;
2065 }
2066 aggr_get_id = perf_stat__get_socket_file;
2067 break;
2068 case AGGR_CORE:
2069 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
2070 perror("cannot build core map");
2071 return -1;
2072 }
2073 aggr_get_id = perf_stat__get_core_file;
2074 break;
2075 case AGGR_NONE:
2076 case AGGR_GLOBAL:
2077 case AGGR_THREAD:
2078 case AGGR_UNSET:
2079 default:
2080 break;
2081 }
2082
2083 return 0;
2084 }
2085
2086 static int topdown_filter_events(const char **attr, char **str, bool use_group)
2087 {
2088 int off = 0;
2089 int i;
2090 int len = 0;
2091 char *s;
2092
2093 for (i = 0; attr[i]; i++) {
2094 if (pmu_have_event("cpu", attr[i])) {
2095 len += strlen(attr[i]) + 1;
2096 attr[i - off] = attr[i];
2097 } else
2098 off++;
2099 }
2100 attr[i - off] = NULL;
2101
2102 *str = malloc(len + 1 + 2);
2103 if (!*str)
2104 return -1;
2105 s = *str;
2106 if (i - off == 0) {
2107 *s = 0;
2108 return 0;
2109 }
2110 if (use_group)
2111 *s++ = '{';
2112 for (i = 0; attr[i]; i++) {
2113 strcpy(s, attr[i]);
2114 s += strlen(s);
2115 *s++ = ',';
2116 }
2117 if (use_group) {
2118 s[-1] = '}';
2119 *s = 0;
2120 } else
2121 s[-1] = 0;
2122 return 0;
2123 }
2124
2125 __weak bool arch_topdown_check_group(bool *warn)
2126 {
2127 *warn = false;
2128 return false;
2129 }
2130
2131 __weak void arch_topdown_group_warn(void)
2132 {
2133 }
2134
2135 /*
2136 * Add default attributes, if there were no attributes specified or
2137 * if -d/--detailed, -d -d or -d -d -d is used:
2138 */
2139 static int add_default_attributes(void)
2140 {
2141 int err;
2142 struct perf_event_attr default_attrs0[] = {
2143
2144 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
2145 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
2146 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
2147 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
2148
2149 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
2150 };
2151 struct perf_event_attr frontend_attrs[] = {
2152 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
2153 };
2154 struct perf_event_attr backend_attrs[] = {
2155 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
2156 };
2157 struct perf_event_attr default_attrs1[] = {
2158 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
2159 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
2160 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
2161
2162 };
2163
2164 /*
2165 * Detailed stats (-d), covering the L1 and last level data caches:
2166 */
2167 struct perf_event_attr detailed_attrs[] = {
2168
2169 { .type = PERF_TYPE_HW_CACHE,
2170 .config =
2171 PERF_COUNT_HW_CACHE_L1D << 0 |
2172 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2173 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2174
2175 { .type = PERF_TYPE_HW_CACHE,
2176 .config =
2177 PERF_COUNT_HW_CACHE_L1D << 0 |
2178 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2179 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2180
2181 { .type = PERF_TYPE_HW_CACHE,
2182 .config =
2183 PERF_COUNT_HW_CACHE_LL << 0 |
2184 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2185 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2186
2187 { .type = PERF_TYPE_HW_CACHE,
2188 .config =
2189 PERF_COUNT_HW_CACHE_LL << 0 |
2190 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2191 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2192 };
2193
2194 /*
2195 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2196 */
2197 struct perf_event_attr very_detailed_attrs[] = {
2198
2199 { .type = PERF_TYPE_HW_CACHE,
2200 .config =
2201 PERF_COUNT_HW_CACHE_L1I << 0 |
2202 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2203 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2204
2205 { .type = PERF_TYPE_HW_CACHE,
2206 .config =
2207 PERF_COUNT_HW_CACHE_L1I << 0 |
2208 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2209 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2210
2211 { .type = PERF_TYPE_HW_CACHE,
2212 .config =
2213 PERF_COUNT_HW_CACHE_DTLB << 0 |
2214 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2215 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2216
2217 { .type = PERF_TYPE_HW_CACHE,
2218 .config =
2219 PERF_COUNT_HW_CACHE_DTLB << 0 |
2220 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2221 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2222
2223 { .type = PERF_TYPE_HW_CACHE,
2224 .config =
2225 PERF_COUNT_HW_CACHE_ITLB << 0 |
2226 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2227 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2228
2229 { .type = PERF_TYPE_HW_CACHE,
2230 .config =
2231 PERF_COUNT_HW_CACHE_ITLB << 0 |
2232 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2233 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2234
2235 };
2236
2237 /*
2238 * Very, very detailed stats (-d -d -d), adding prefetch events:
2239 */
2240 struct perf_event_attr very_very_detailed_attrs[] = {
2241
2242 { .type = PERF_TYPE_HW_CACHE,
2243 .config =
2244 PERF_COUNT_HW_CACHE_L1D << 0 |
2245 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2246 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2247
2248 { .type = PERF_TYPE_HW_CACHE,
2249 .config =
2250 PERF_COUNT_HW_CACHE_L1D << 0 |
2251 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2252 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2253 };
2254
2255 /* Set attrs if no event is selected and !null_run: */
2256 if (null_run)
2257 return 0;
2258
2259 if (transaction_run) {
2260 if (pmu_have_event("cpu", "cycles-ct") &&
2261 pmu_have_event("cpu", "el-start"))
2262 err = parse_events(evsel_list, transaction_attrs, NULL);
2263 else
2264 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
2265 if (err) {
2266 fprintf(stderr, "Cannot set up transaction events\n");
2267 return -1;
2268 }
2269 return 0;
2270 }
2271
2272 if (smi_cost) {
2273 int smi;
2274
2275 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2276 fprintf(stderr, "freeze_on_smi is not supported.\n");
2277 return -1;
2278 }
2279
2280 if (!smi) {
2281 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2282 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2283 return -1;
2284 }
2285 smi_reset = true;
2286 }
2287
2288 if (pmu_have_event("msr", "aperf") &&
2289 pmu_have_event("msr", "smi")) {
2290 if (!force_metric_only)
2291 metric_only = true;
2292 err = parse_events(evsel_list, smi_cost_attrs, NULL);
2293 } else {
2294 fprintf(stderr, "To measure SMI cost, it needs "
2295 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
2296 return -1;
2297 }
2298 if (err) {
2299 fprintf(stderr, "Cannot set up SMI cost events\n");
2300 return -1;
2301 }
2302 return 0;
2303 }
2304
2305 if (topdown_run) {
2306 char *str = NULL;
2307 bool warn = false;
2308
2309 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2310 stat_config.aggr_mode != AGGR_CORE) {
2311 pr_err("top down event configuration requires --per-core mode\n");
2312 return -1;
2313 }
2314 stat_config.aggr_mode = AGGR_CORE;
2315 if (nr_cgroups || !target__has_cpu(&target)) {
2316 pr_err("top down event configuration requires system-wide mode (-a)\n");
2317 return -1;
2318 }
2319
2320 if (!force_metric_only)
2321 metric_only = true;
2322 if (topdown_filter_events(topdown_attrs, &str,
2323 arch_topdown_check_group(&warn)) < 0) {
2324 pr_err("Out of memory\n");
2325 return -1;
2326 }
2327 if (topdown_attrs[0] && str) {
2328 if (warn)
2329 arch_topdown_group_warn();
2330 err = parse_events(evsel_list, str, NULL);
2331 if (err) {
2332 fprintf(stderr,
2333 "Cannot set up top down events %s: %d\n",
2334 str, err);
2335 free(str);
2336 return -1;
2337 }
2338 } else {
2339 fprintf(stderr, "System does not support topdown\n");
2340 return -1;
2341 }
2342 free(str);
2343 }
2344
2345 if (!evsel_list->nr_entries) {
2346 if (target__has_cpu(&target))
2347 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2348
2349 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2350 return -1;
2351 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2352 if (perf_evlist__add_default_attrs(evsel_list,
2353 frontend_attrs) < 0)
2354 return -1;
2355 }
2356 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2357 if (perf_evlist__add_default_attrs(evsel_list,
2358 backend_attrs) < 0)
2359 return -1;
2360 }
2361 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2362 return -1;
2363 }
2364
2365 /* Detailed events get appended to the event list: */
2366
2367 if (detailed_run < 1)
2368 return 0;
2369
2370 /* Append detailed run extra attributes: */
2371 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2372 return -1;
2373
2374 if (detailed_run < 2)
2375 return 0;
2376
2377 /* Append very detailed run extra attributes: */
2378 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2379 return -1;
2380
2381 if (detailed_run < 3)
2382 return 0;
2383
2384 /* Append very, very detailed run extra attributes: */
2385 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2386 }
2387
2388 static const char * const stat_record_usage[] = {
2389 "perf stat record [<options>]",
2390 NULL,
2391 };
2392
2393 static void init_features(struct perf_session *session)
2394 {
2395 int feat;
2396
2397 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2398 perf_header__set_feat(&session->header, feat);
2399
2400 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2401 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2402 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2403 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2404 }
2405
2406 static int __cmd_record(int argc, const char **argv)
2407 {
2408 struct perf_session *session;
2409 struct perf_data_file *file = &perf_stat.file;
2410
2411 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2412 PARSE_OPT_STOP_AT_NON_OPTION);
2413
2414 if (output_name)
2415 file->path = output_name;
2416
2417 if (run_count != 1 || forever) {
2418 pr_err("Cannot use -r option with perf stat record.\n");
2419 return -1;
2420 }
2421
2422 session = perf_session__new(file, false, NULL);
2423 if (session == NULL) {
2424 pr_err("Perf session creation failed.\n");
2425 return -1;
2426 }
2427
2428 init_features(session);
2429
2430 session->evlist = evsel_list;
2431 perf_stat.session = session;
2432 perf_stat.record = true;
2433 return argc;
2434 }
2435
2436 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2437 union perf_event *event,
2438 struct perf_session *session)
2439 {
2440 struct stat_round_event *stat_round = &event->stat_round;
2441 struct perf_evsel *counter;
2442 struct timespec tsh, *ts = NULL;
2443 const char **argv = session->header.env.cmdline_argv;
2444 int argc = session->header.env.nr_cmdline;
2445
2446 evlist__for_each_entry(evsel_list, counter)
2447 perf_stat_process_counter(&stat_config, counter);
2448
2449 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2450 update_stats(&walltime_nsecs_stats, stat_round->time);
2451
2452 if (stat_config.interval && stat_round->time) {
2453 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2454 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2455 ts = &tsh;
2456 }
2457
2458 print_counters(ts, argc, argv);
2459 return 0;
2460 }
2461
2462 static
2463 int process_stat_config_event(struct perf_tool *tool,
2464 union perf_event *event,
2465 struct perf_session *session __maybe_unused)
2466 {
2467 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2468
2469 perf_event__read_stat_config(&stat_config, &event->stat_config);
2470
2471 if (cpu_map__empty(st->cpus)) {
2472 if (st->aggr_mode != AGGR_UNSET)
2473 pr_warning("warning: processing task data, aggregation mode not set\n");
2474 return 0;
2475 }
2476
2477 if (st->aggr_mode != AGGR_UNSET)
2478 stat_config.aggr_mode = st->aggr_mode;
2479
2480 if (perf_stat.file.is_pipe)
2481 perf_stat_init_aggr_mode();
2482 else
2483 perf_stat_init_aggr_mode_file(st);
2484
2485 return 0;
2486 }
2487
2488 static int set_maps(struct perf_stat *st)
2489 {
2490 if (!st->cpus || !st->threads)
2491 return 0;
2492
2493 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2494 return -EINVAL;
2495
2496 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2497
2498 if (perf_evlist__alloc_stats(evsel_list, true))
2499 return -ENOMEM;
2500
2501 st->maps_allocated = true;
2502 return 0;
2503 }
2504
2505 static
2506 int process_thread_map_event(struct perf_tool *tool,
2507 union perf_event *event,
2508 struct perf_session *session __maybe_unused)
2509 {
2510 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2511
2512 if (st->threads) {
2513 pr_warning("Extra thread map event, ignoring.\n");
2514 return 0;
2515 }
2516
2517 st->threads = thread_map__new_event(&event->thread_map);
2518 if (!st->threads)
2519 return -ENOMEM;
2520
2521 return set_maps(st);
2522 }
2523
2524 static
2525 int process_cpu_map_event(struct perf_tool *tool,
2526 union perf_event *event,
2527 struct perf_session *session __maybe_unused)
2528 {
2529 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2530 struct cpu_map *cpus;
2531
2532 if (st->cpus) {
2533 pr_warning("Extra cpu map event, ignoring.\n");
2534 return 0;
2535 }
2536
2537 cpus = cpu_map__new_data(&event->cpu_map.data);
2538 if (!cpus)
2539 return -ENOMEM;
2540
2541 st->cpus = cpus;
2542 return set_maps(st);
2543 }
2544
2545 static const char * const stat_report_usage[] = {
2546 "perf stat report [<options>]",
2547 NULL,
2548 };
2549
2550 static struct perf_stat perf_stat = {
2551 .tool = {
2552 .attr = perf_event__process_attr,
2553 .event_update = perf_event__process_event_update,
2554 .thread_map = process_thread_map_event,
2555 .cpu_map = process_cpu_map_event,
2556 .stat_config = process_stat_config_event,
2557 .stat = perf_event__process_stat_event,
2558 .stat_round = process_stat_round_event,
2559 },
2560 .aggr_mode = AGGR_UNSET,
2561 };
2562
2563 static int __cmd_report(int argc, const char **argv)
2564 {
2565 struct perf_session *session;
2566 const struct option options[] = {
2567 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2568 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2569 "aggregate counts per processor socket", AGGR_SOCKET),
2570 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2571 "aggregate counts per physical processor core", AGGR_CORE),
2572 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2573 "disable CPU count aggregation", AGGR_NONE),
2574 OPT_END()
2575 };
2576 struct stat st;
2577 int ret;
2578
2579 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2580
2581 if (!input_name || !strlen(input_name)) {
2582 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2583 input_name = "-";
2584 else
2585 input_name = "perf.data";
2586 }
2587
2588 perf_stat.file.path = input_name;
2589 perf_stat.file.mode = PERF_DATA_MODE_READ;
2590
2591 session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
2592 if (session == NULL)
2593 return -1;
2594
2595 perf_stat.session = session;
2596 stat_config.output = stderr;
2597 evsel_list = session->evlist;
2598
2599 ret = perf_session__process_events(session);
2600 if (ret)
2601 return ret;
2602
2603 perf_session__delete(session);
2604 return 0;
2605 }
2606
2607 static void setup_system_wide(int forks)
2608 {
2609 /*
2610 * Make system wide (-a) the default target if
2611 * no target was specified and one of following
2612 * conditions is met:
2613 *
2614 * - there's no workload specified
2615 * - there is workload specified but all requested
2616 * events are system wide events
2617 */
2618 if (!target__none(&target))
2619 return;
2620
2621 if (!forks)
2622 target.system_wide = true;
2623 else {
2624 struct perf_evsel *counter;
2625
2626 evlist__for_each_entry(evsel_list, counter) {
2627 if (!counter->system_wide)
2628 return;
2629 }
2630
2631 if (evsel_list->nr_entries)
2632 target.system_wide = true;
2633 }
2634 }
2635
2636 int cmd_stat(int argc, const char **argv)
2637 {
2638 const char * const stat_usage[] = {
2639 "perf stat [<options>] [<command>]",
2640 NULL
2641 };
2642 int status = -EINVAL, run_idx;
2643 const char *mode;
2644 FILE *output = stderr;
2645 unsigned int interval;
2646 const char * const stat_subcommands[] = { "record", "report" };
2647
2648 setlocale(LC_ALL, "");
2649
2650 evsel_list = perf_evlist__new();
2651 if (evsel_list == NULL)
2652 return -ENOMEM;
2653
2654 parse_events__shrink_config_terms();
2655 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2656 (const char **) stat_usage,
2657 PARSE_OPT_STOP_AT_NON_OPTION);
2658 perf_stat__collect_metric_expr(evsel_list);
2659 perf_stat__init_shadow_stats();
2660
2661 if (csv_sep) {
2662 csv_output = true;
2663 if (!strcmp(csv_sep, "\\t"))
2664 csv_sep = "\t";
2665 } else
2666 csv_sep = DEFAULT_SEPARATOR;
2667
2668 if (argc && !strncmp(argv[0], "rec", 3)) {
2669 argc = __cmd_record(argc, argv);
2670 if (argc < 0)
2671 return -1;
2672 } else if (argc && !strncmp(argv[0], "rep", 3))
2673 return __cmd_report(argc, argv);
2674
2675 interval = stat_config.interval;
2676
2677 /*
2678 * For record command the -o is already taken care of.
2679 */
2680 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2681 output = NULL;
2682
2683 if (output_name && output_fd) {
2684 fprintf(stderr, "cannot use both --output and --log-fd\n");
2685 parse_options_usage(stat_usage, stat_options, "o", 1);
2686 parse_options_usage(NULL, stat_options, "log-fd", 0);
2687 goto out;
2688 }
2689
2690 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2691 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2692 goto out;
2693 }
2694
2695 if (metric_only && run_count > 1) {
2696 fprintf(stderr, "--metric-only is not supported with -r\n");
2697 goto out;
2698 }
2699
2700 if (output_fd < 0) {
2701 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2702 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2703 goto out;
2704 }
2705
2706 if (!output) {
2707 struct timespec tm;
2708 mode = append_file ? "a" : "w";
2709
2710 output = fopen(output_name, mode);
2711 if (!output) {
2712 perror("failed to create output file");
2713 return -1;
2714 }
2715 clock_gettime(CLOCK_REALTIME, &tm);
2716 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2717 } else if (output_fd > 0) {
2718 mode = append_file ? "a" : "w";
2719 output = fdopen(output_fd, mode);
2720 if (!output) {
2721 perror("Failed opening logfd");
2722 return -errno;
2723 }
2724 }
2725
2726 stat_config.output = output;
2727
2728 /*
2729 * let the spreadsheet do the pretty-printing
2730 */
2731 if (csv_output) {
2732 /* User explicitly passed -B? */
2733 if (big_num_opt == 1) {
2734 fprintf(stderr, "-B option not supported with -x\n");
2735 parse_options_usage(stat_usage, stat_options, "B", 1);
2736 parse_options_usage(NULL, stat_options, "x", 1);
2737 goto out;
2738 } else /* Nope, so disable big number formatting */
2739 big_num = false;
2740 } else if (big_num_opt == 0) /* User passed --no-big-num */
2741 big_num = false;
2742
2743 setup_system_wide(argc);
2744
2745 if (run_count < 0) {
2746 pr_err("Run count must be a positive number\n");
2747 parse_options_usage(stat_usage, stat_options, "r", 1);
2748 goto out;
2749 } else if (run_count == 0) {
2750 forever = true;
2751 run_count = 1;
2752 }
2753
2754 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2755 fprintf(stderr, "The --per-thread option is only available "
2756 "when monitoring via -p -t options.\n");
2757 parse_options_usage(NULL, stat_options, "p", 1);
2758 parse_options_usage(NULL, stat_options, "t", 1);
2759 goto out;
2760 }
2761
2762 /*
2763 * no_aggr, cgroup are for system-wide only
2764 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2765 */
2766 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2767 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2768 !target__has_cpu(&target)) {
2769 fprintf(stderr, "both cgroup and no-aggregation "
2770 "modes only available in system-wide mode\n");
2771
2772 parse_options_usage(stat_usage, stat_options, "G", 1);
2773 parse_options_usage(NULL, stat_options, "A", 1);
2774 parse_options_usage(NULL, stat_options, "a", 1);
2775 goto out;
2776 }
2777
2778 if (add_default_attributes())
2779 goto out;
2780
2781 target__validate(&target);
2782
2783 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2784 if (target__has_task(&target)) {
2785 pr_err("Problems finding threads of monitor\n");
2786 parse_options_usage(stat_usage, stat_options, "p", 1);
2787 parse_options_usage(NULL, stat_options, "t", 1);
2788 } else if (target__has_cpu(&target)) {
2789 perror("failed to parse CPUs map");
2790 parse_options_usage(stat_usage, stat_options, "C", 1);
2791 parse_options_usage(NULL, stat_options, "a", 1);
2792 }
2793 goto out;
2794 }
2795
2796 /*
2797 * Initialize thread_map with comm names,
2798 * so we could print it out on output.
2799 */
2800 if (stat_config.aggr_mode == AGGR_THREAD)
2801 thread_map__read_comms(evsel_list->threads);
2802
2803 if (interval && interval < 100) {
2804 if (interval < 10) {
2805 pr_err("print interval must be >= 10ms\n");
2806 parse_options_usage(stat_usage, stat_options, "I", 1);
2807 goto out;
2808 } else
2809 pr_warning("print interval < 100ms. "
2810 "The overhead percentage could be high in some cases. "
2811 "Please proceed with caution.\n");
2812 }
2813
2814 if (perf_evlist__alloc_stats(evsel_list, interval))
2815 goto out;
2816
2817 if (perf_stat_init_aggr_mode())
2818 goto out;
2819
2820 /*
2821 * We dont want to block the signals - that would cause
2822 * child tasks to inherit that and Ctrl-C would not work.
2823 * What we want is for Ctrl-C to work in the exec()-ed
2824 * task, but being ignored by perf stat itself:
2825 */
2826 atexit(sig_atexit);
2827 if (!forever)
2828 signal(SIGINT, skip_signal);
2829 signal(SIGCHLD, skip_signal);
2830 signal(SIGALRM, skip_signal);
2831 signal(SIGABRT, skip_signal);
2832
2833 status = 0;
2834 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2835 if (run_count != 1 && verbose > 0)
2836 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2837 run_idx + 1);
2838
2839 status = run_perf_stat(argc, argv);
2840 if (forever && status != -1) {
2841 print_counters(NULL, argc, argv);
2842 perf_stat__reset_stats();
2843 }
2844 }
2845
2846 if (!forever && status != -1 && !interval)
2847 print_counters(NULL, argc, argv);
2848
2849 if (STAT_RECORD) {
2850 /*
2851 * We synthesize the kernel mmap record just so that older tools
2852 * don't emit warnings about not being able to resolve symbols
2853 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2854 * a saner message about no samples being in the perf.data file.
2855 *
2856 * This also serves to suppress a warning about f_header.data.size == 0
2857 * in header.c at the moment 'perf stat record' gets introduced, which
2858 * is not really needed once we start adding the stat specific PERF_RECORD_
2859 * records, but the need to suppress the kptr_restrict messages in older
2860 * tools remain -acme
2861 */
2862 int fd = perf_data_file__fd(&perf_stat.file);
2863 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2864 process_synthesized_event,
2865 &perf_stat.session->machines.host);
2866 if (err) {
2867 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2868 "older tools may produce warnings about this file\n.");
2869 }
2870
2871 if (!interval) {
2872 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2873 pr_err("failed to write stat round event\n");
2874 }
2875
2876 if (!perf_stat.file.is_pipe) {
2877 perf_stat.session->header.data_size += perf_stat.bytes_written;
2878 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2879 }
2880
2881 perf_session__delete(perf_stat.session);
2882 }
2883
2884 perf_stat__exit_aggr_mode();
2885 perf_evlist__free_stats(evsel_list);
2886 out:
2887 if (smi_cost && smi_reset)
2888 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2889
2890 perf_evlist__delete(evsel_list);
2891 return status;
2892 }
Linux with added FPC-III support