?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Revert "ALSA: hda: Flush interrupts on disabling"
[linux/fpc-iii.git] / tools / perf / builtin-stat.c
blob5cb58f3afa3559bc4c68926fcc0854a18bd927c8
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 "asm/bug.h"
68
69 #include <linux/time64.h>
70 #include <api/fs/fs.h>
71 #include <stdlib.h>
72 #include <sys/prctl.h>
73 #include <locale.h>
74 #include <math.h>
75
76 #define DEFAULT_SEPARATOR " "
77 #define CNTR_NOT_SUPPORTED "<not supported>"
78 #define CNTR_NOT_COUNTED "<not counted>"
79
80 static void print_counters(struct timespec *ts, int argc, const char **argv);
81
82 /* Default events used for perf stat -T */
83 static const char *transaction_attrs = {
84 "task-clock,"
85 "{"
86 "instructions,"
87 "cycles,"
88 "cpu/cycles-t/,"
89 "cpu/tx-start/,"
90 "cpu/el-start/,"
91 "cpu/cycles-ct/"
92 "}"
93 };
94
95 /* More limited version when the CPU does not have all events. */
96 static const char * transaction_limited_attrs = {
97 "task-clock,"
98 "{"
99 "instructions,"
100 "cycles,"
101 "cpu/cycles-t/,"
102 "cpu/tx-start/"
103 "}"
104 };
105
106 static const char * topdown_attrs[] = {
107 "topdown-total-slots",
108 "topdown-slots-retired",
109 "topdown-recovery-bubbles",
110 "topdown-fetch-bubbles",
111 "topdown-slots-issued",
112 NULL,
113 };
114
115 static struct perf_evlist *evsel_list;
116
117 static struct target target = {
118 .uid = UINT_MAX,
119 };
120
121 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
122
123 static int run_count = 1;
124 static bool no_inherit = false;
125 static volatile pid_t child_pid = -1;
126 static bool null_run = false;
127 static int detailed_run = 0;
128 static bool transaction_run;
129 static bool topdown_run = false;
130 static bool big_num = true;
131 static int big_num_opt = -1;
132 static const char *csv_sep = NULL;
133 static bool csv_output = false;
134 static bool group = false;
135 static const char *pre_cmd = NULL;
136 static const char *post_cmd = NULL;
137 static bool sync_run = false;
138 static unsigned int initial_delay = 0;
139 static unsigned int unit_width = 4; /* strlen("unit") */
140 static bool forever = false;
141 static bool metric_only = false;
142 static bool force_metric_only = false;
143 static struct timespec ref_time;
144 static struct cpu_map *aggr_map;
145 static aggr_get_id_t aggr_get_id;
146 static bool append_file;
147 static const char *output_name;
148 static int output_fd;
149 static int print_free_counters_hint;
150
151 struct perf_stat {
152 bool record;
153 struct perf_data_file file;
154 struct perf_session *session;
155 u64 bytes_written;
156 struct perf_tool tool;
157 bool maps_allocated;
158 struct cpu_map *cpus;
159 struct thread_map *threads;
160 enum aggr_mode aggr_mode;
161 };
162
163 static struct perf_stat perf_stat;
164 #define STAT_RECORD perf_stat.record
165
166 static volatile int done = 0;
167
168 static struct perf_stat_config stat_config = {
169 .aggr_mode = AGGR_GLOBAL,
170 .scale = true,
171 };
172
173 static inline void diff_timespec(struct timespec *r, struct timespec *a,
174 struct timespec *b)
175 {
176 r->tv_sec = a->tv_sec - b->tv_sec;
177 if (a->tv_nsec < b->tv_nsec) {
178 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
179 r->tv_sec--;
180 } else {
181 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
182 }
183 }
184
185 static void perf_stat__reset_stats(void)
186 {
187 perf_evlist__reset_stats(evsel_list);
188 perf_stat__reset_shadow_stats();
189 }
190
191 static int create_perf_stat_counter(struct perf_evsel *evsel)
192 {
193 struct perf_event_attr *attr = &evsel->attr;
194
195 if (stat_config.scale)
196 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
197 PERF_FORMAT_TOTAL_TIME_RUNNING;
198
199 attr->inherit = !no_inherit;
200
201 /*
202 * Some events get initialized with sample_(period/type) set,
203 * like tracepoints. Clear it up for counting.
204 */
205 attr->sample_period = 0;
206
207 /*
208 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
209 * while avoiding that older tools show confusing messages.
210 *
211 * However for pipe sessions we need to keep it zero,
212 * because script's perf_evsel__check_attr is triggered
213 * by attr->sample_type != 0, and we can't run it on
214 * stat sessions.
215 */
216 if (!(STAT_RECORD && perf_stat.file.is_pipe))
217 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
218
219 /*
220 * Disabling all counters initially, they will be enabled
221 * either manually by us or by kernel via enable_on_exec
222 * set later.
223 */
224 if (perf_evsel__is_group_leader(evsel)) {
225 attr->disabled = 1;
226
227 /*
228 * In case of initial_delay we enable tracee
229 * events manually.
230 */
231 if (target__none(&target) && !initial_delay)
232 attr->enable_on_exec = 1;
233 }
234
235 if (target__has_cpu(&target))
236 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
237
238 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
239 }
240
241 /*
242 * Does the counter have nsecs as a unit?
243 */
244 static inline int nsec_counter(struct perf_evsel *evsel)
245 {
246 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
247 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
248 return 1;
249
250 return 0;
251 }
252
253 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
254 union perf_event *event,
255 struct perf_sample *sample __maybe_unused,
256 struct machine *machine __maybe_unused)
257 {
258 if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
259 pr_err("failed to write perf data, error: %m\n");
260 return -1;
261 }
262
263 perf_stat.bytes_written += event->header.size;
264 return 0;
265 }
266
267 static int write_stat_round_event(u64 tm, u64 type)
268 {
269 return perf_event__synthesize_stat_round(NULL, tm, type,
270 process_synthesized_event,
271 NULL);
272 }
273
274 #define WRITE_STAT_ROUND_EVENT(time, interval) \
275 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
276
277 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
278
279 static int
280 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
281 struct perf_counts_values *count)
282 {
283 struct perf_sample_id *sid = SID(counter, cpu, thread);
284
285 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
286 process_synthesized_event, NULL);
287 }
288
289 /*
290 * Read out the results of a single counter:
291 * do not aggregate counts across CPUs in system-wide mode
292 */
293 static int read_counter(struct perf_evsel *counter)
294 {
295 int nthreads = thread_map__nr(evsel_list->threads);
296 int ncpus, cpu, thread;
297
298 if (target__has_cpu(&target))
299 ncpus = perf_evsel__nr_cpus(counter);
300 else
301 ncpus = 1;
302
303 if (!counter->supported)
304 return -ENOENT;
305
306 if (counter->system_wide)
307 nthreads = 1;
308
309 for (thread = 0; thread < nthreads; thread++) {
310 for (cpu = 0; cpu < ncpus; cpu++) {
311 struct perf_counts_values *count;
312
313 count = perf_counts(counter->counts, cpu, thread);
314 if (perf_evsel__read(counter, cpu, thread, count)) {
315 counter->counts->scaled = -1;
316 perf_counts(counter->counts, cpu, thread)->ena = 0;
317 perf_counts(counter->counts, cpu, thread)->run = 0;
318 return -1;
319 }
320
321 if (STAT_RECORD) {
322 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
323 pr_err("failed to write stat event\n");
324 return -1;
325 }
326 }
327
328 if (verbose > 1) {
329 fprintf(stat_config.output,
330 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
331 perf_evsel__name(counter),
332 cpu,
333 count->val, count->ena, count->run);
334 }
335 }
336 }
337
338 return 0;
339 }
340
341 static void read_counters(void)
342 {
343 struct perf_evsel *counter;
344 int ret;
345
346 evlist__for_each_entry(evsel_list, counter) {
347 ret = read_counter(counter);
348 if (ret)
349 pr_debug("failed to read counter %s\n", counter->name);
350
351 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
352 pr_warning("failed to process counter %s\n", counter->name);
353 }
354 }
355
356 static void process_interval(void)
357 {
358 struct timespec ts, rs;
359
360 read_counters();
361
362 clock_gettime(CLOCK_MONOTONIC, &ts);
363 diff_timespec(&rs, &ts, &ref_time);
364
365 if (STAT_RECORD) {
366 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
367 pr_err("failed to write stat round event\n");
368 }
369
370 print_counters(&rs, 0, NULL);
371 }
372
373 static void enable_counters(void)
374 {
375 if (initial_delay)
376 usleep(initial_delay * USEC_PER_MSEC);
377
378 /*
379 * We need to enable counters only if:
380 * - we don't have tracee (attaching to task or cpu)
381 * - we have initial delay configured
382 */
383 if (!target__none(&target) || initial_delay)
384 perf_evlist__enable(evsel_list);
385 }
386
387 static void disable_counters(void)
388 {
389 /*
390 * If we don't have tracee (attaching to task or cpu), counters may
391 * still be running. To get accurate group ratios, we must stop groups
392 * from counting before reading their constituent counters.
393 */
394 if (!target__none(&target))
395 perf_evlist__disable(evsel_list);
396 }
397
398 static volatile int workload_exec_errno;
399
400 /*
401 * perf_evlist__prepare_workload will send a SIGUSR1
402 * if the fork fails, since we asked by setting its
403 * want_signal to true.
404 */
405 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
406 void *ucontext __maybe_unused)
407 {
408 workload_exec_errno = info->si_value.sival_int;
409 }
410
411 static bool has_unit(struct perf_evsel *counter)
412 {
413 return counter->unit && *counter->unit;
414 }
415
416 static bool has_scale(struct perf_evsel *counter)
417 {
418 return counter->scale != 1;
419 }
420
421 static int perf_stat_synthesize_config(bool is_pipe)
422 {
423 struct perf_evsel *counter;
424 int err;
425
426 if (is_pipe) {
427 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
428 process_synthesized_event);
429 if (err < 0) {
430 pr_err("Couldn't synthesize attrs.\n");
431 return err;
432 }
433 }
434
435 /*
436 * Synthesize other events stuff not carried within
437 * attr event - unit, scale, name
438 */
439 evlist__for_each_entry(evsel_list, counter) {
440 if (!counter->supported)
441 continue;
442
443 /*
444 * Synthesize unit and scale only if it's defined.
445 */
446 if (has_unit(counter)) {
447 err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
448 if (err < 0) {
449 pr_err("Couldn't synthesize evsel unit.\n");
450 return err;
451 }
452 }
453
454 if (has_scale(counter)) {
455 err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
456 if (err < 0) {
457 pr_err("Couldn't synthesize evsel scale.\n");
458 return err;
459 }
460 }
461
462 if (counter->own_cpus) {
463 err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
464 if (err < 0) {
465 pr_err("Couldn't synthesize evsel scale.\n");
466 return err;
467 }
468 }
469
470 /*
471 * Name is needed only for pipe output,
472 * perf.data carries event names.
473 */
474 if (is_pipe) {
475 err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
476 if (err < 0) {
477 pr_err("Couldn't synthesize evsel name.\n");
478 return err;
479 }
480 }
481 }
482
483 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
484 process_synthesized_event,
485 NULL);
486 if (err < 0) {
487 pr_err("Couldn't synthesize thread map.\n");
488 return err;
489 }
490
491 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
492 process_synthesized_event, NULL);
493 if (err < 0) {
494 pr_err("Couldn't synthesize thread map.\n");
495 return err;
496 }
497
498 err = perf_event__synthesize_stat_config(NULL, &stat_config,
499 process_synthesized_event, NULL);
500 if (err < 0) {
501 pr_err("Couldn't synthesize config.\n");
502 return err;
503 }
504
505 return 0;
506 }
507
508 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
509
510 static int __store_counter_ids(struct perf_evsel *counter,
511 struct cpu_map *cpus,
512 struct thread_map *threads)
513 {
514 int cpu, thread;
515
516 for (cpu = 0; cpu < cpus->nr; cpu++) {
517 for (thread = 0; thread < threads->nr; thread++) {
518 int fd = FD(counter, cpu, thread);
519
520 if (perf_evlist__id_add_fd(evsel_list, counter,
521 cpu, thread, fd) < 0)
522 return -1;
523 }
524 }
525
526 return 0;
527 }
528
529 static int store_counter_ids(struct perf_evsel *counter)
530 {
531 struct cpu_map *cpus = counter->cpus;
532 struct thread_map *threads = counter->threads;
533
534 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
535 return -ENOMEM;
536
537 return __store_counter_ids(counter, cpus, threads);
538 }
539
540 static int __run_perf_stat(int argc, const char **argv)
541 {
542 int interval = stat_config.interval;
543 char msg[512];
544 unsigned long long t0, t1;
545 struct perf_evsel *counter;
546 struct timespec ts;
547 size_t l;
548 int status = 0;
549 const bool forks = (argc > 0);
550 bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
551 struct perf_evsel_config_term *err_term;
552
553 if (interval) {
554 ts.tv_sec = interval / USEC_PER_MSEC;
555 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
556 } else {
557 ts.tv_sec = 1;
558 ts.tv_nsec = 0;
559 }
560
561 if (forks) {
562 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
563 workload_exec_failed_signal) < 0) {
564 perror("failed to prepare workload");
565 return -1;
566 }
567 child_pid = evsel_list->workload.pid;
568 }
569
570 if (group)
571 perf_evlist__set_leader(evsel_list);
572
573 evlist__for_each_entry(evsel_list, counter) {
574 try_again:
575 if (create_perf_stat_counter(counter) < 0) {
576 /*
577 * PPC returns ENXIO for HW counters until 2.6.37
578 * (behavior changed with commit b0a873e).
579 */
580 if (errno == EINVAL || errno == ENOSYS ||
581 errno == ENOENT || errno == EOPNOTSUPP ||
582 errno == ENXIO) {
583 if (verbose)
584 ui__warning("%s event is not supported by the kernel.\n",
585 perf_evsel__name(counter));
586 counter->supported = false;
587
588 if ((counter->leader != counter) ||
589 !(counter->leader->nr_members > 1))
590 continue;
591 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
592 if (verbose)
593 ui__warning("%s\n", msg);
594 goto try_again;
595 }
596
597 perf_evsel__open_strerror(counter, &target,
598 errno, msg, sizeof(msg));
599 ui__error("%s\n", msg);
600
601 if (child_pid != -1)
602 kill(child_pid, SIGTERM);
603
604 return -1;
605 }
606 counter->supported = true;
607
608 l = strlen(counter->unit);
609 if (l > unit_width)
610 unit_width = l;
611
612 if (STAT_RECORD && store_counter_ids(counter))
613 return -1;
614 }
615
616 if (perf_evlist__apply_filters(evsel_list, &counter)) {
617 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
618 counter->filter, perf_evsel__name(counter), errno,
619 str_error_r(errno, msg, sizeof(msg)));
620 return -1;
621 }
622
623 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
624 error("failed to set config \"%s\" on event %s with %d (%s)\n",
625 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
626 str_error_r(errno, msg, sizeof(msg)));
627 return -1;
628 }
629
630 if (STAT_RECORD) {
631 int err, fd = perf_data_file__fd(&perf_stat.file);
632
633 if (is_pipe) {
634 err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
635 } else {
636 err = perf_session__write_header(perf_stat.session, evsel_list,
637 fd, false);
638 }
639
640 if (err < 0)
641 return err;
642
643 err = perf_stat_synthesize_config(is_pipe);
644 if (err < 0)
645 return err;
646 }
647
648 /*
649 * Enable counters and exec the command:
650 */
651 t0 = rdclock();
652 clock_gettime(CLOCK_MONOTONIC, &ref_time);
653
654 if (forks) {
655 perf_evlist__start_workload(evsel_list);
656 enable_counters();
657
658 if (interval) {
659 while (!waitpid(child_pid, &status, WNOHANG)) {
660 nanosleep(&ts, NULL);
661 process_interval();
662 }
663 }
664 wait(&status);
665
666 if (workload_exec_errno) {
667 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
668 pr_err("Workload failed: %s\n", emsg);
669 return -1;
670 }
671
672 if (WIFSIGNALED(status))
673 psignal(WTERMSIG(status), argv[0]);
674 } else {
675 enable_counters();
676 while (!done) {
677 nanosleep(&ts, NULL);
678 if (interval)
679 process_interval();
680 }
681 }
682
683 disable_counters();
684
685 t1 = rdclock();
686
687 update_stats(&walltime_nsecs_stats, t1 - t0);
688
689 /*
690 * Closing a group leader splits the group, and as we only disable
691 * group leaders, results in remaining events becoming enabled. To
692 * avoid arbitrary skew, we must read all counters before closing any
693 * group leaders.
694 */
695 read_counters();
696 perf_evlist__close(evsel_list);
697
698 return WEXITSTATUS(status);
699 }
700
701 static int run_perf_stat(int argc, const char **argv)
702 {
703 int ret;
704
705 if (pre_cmd) {
706 ret = system(pre_cmd);
707 if (ret)
708 return ret;
709 }
710
711 if (sync_run)
712 sync();
713
714 ret = __run_perf_stat(argc, argv);
715 if (ret)
716 return ret;
717
718 if (post_cmd) {
719 ret = system(post_cmd);
720 if (ret)
721 return ret;
722 }
723
724 return ret;
725 }
726
727 static void print_running(u64 run, u64 ena)
728 {
729 if (csv_output) {
730 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
731 csv_sep,
732 run,
733 csv_sep,
734 ena ? 100.0 * run / ena : 100.0);
735 } else if (run != ena) {
736 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
737 }
738 }
739
740 static void print_noise_pct(double total, double avg)
741 {
742 double pct = rel_stddev_stats(total, avg);
743
744 if (csv_output)
745 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
746 else if (pct)
747 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
748 }
749
750 static void print_noise(struct perf_evsel *evsel, double avg)
751 {
752 struct perf_stat_evsel *ps;
753
754 if (run_count == 1)
755 return;
756
757 ps = evsel->priv;
758 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
759 }
760
761 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
762 {
763 switch (stat_config.aggr_mode) {
764 case AGGR_CORE:
765 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
766 cpu_map__id_to_socket(id),
767 csv_output ? 0 : -8,
768 cpu_map__id_to_cpu(id),
769 csv_sep,
770 csv_output ? 0 : 4,
771 nr,
772 csv_sep);
773 break;
774 case AGGR_SOCKET:
775 fprintf(stat_config.output, "S%*d%s%*d%s",
776 csv_output ? 0 : -5,
777 id,
778 csv_sep,
779 csv_output ? 0 : 4,
780 nr,
781 csv_sep);
782 break;
783 case AGGR_NONE:
784 fprintf(stat_config.output, "CPU%*d%s",
785 csv_output ? 0 : -4,
786 perf_evsel__cpus(evsel)->map[id], csv_sep);
787 break;
788 case AGGR_THREAD:
789 fprintf(stat_config.output, "%*s-%*d%s",
790 csv_output ? 0 : 16,
791 thread_map__comm(evsel->threads, id),
792 csv_output ? 0 : -8,
793 thread_map__pid(evsel->threads, id),
794 csv_sep);
795 break;
796 case AGGR_GLOBAL:
797 case AGGR_UNSET:
798 default:
799 break;
800 }
801 }
802
803 struct outstate {
804 FILE *fh;
805 bool newline;
806 const char *prefix;
807 int nfields;
808 int id, nr;
809 struct perf_evsel *evsel;
810 };
811
812 #define METRIC_LEN 35
813
814 static void new_line_std(void *ctx)
815 {
816 struct outstate *os = ctx;
817
818 os->newline = true;
819 }
820
821 static void do_new_line_std(struct outstate *os)
822 {
823 fputc('\n', os->fh);
824 fputs(os->prefix, os->fh);
825 aggr_printout(os->evsel, os->id, os->nr);
826 if (stat_config.aggr_mode == AGGR_NONE)
827 fprintf(os->fh, " ");
828 fprintf(os->fh, " ");
829 }
830
831 static void print_metric_std(void *ctx, const char *color, const char *fmt,
832 const char *unit, double val)
833 {
834 struct outstate *os = ctx;
835 FILE *out = os->fh;
836 int n;
837 bool newline = os->newline;
838
839 os->newline = false;
840
841 if (unit == NULL || fmt == NULL) {
842 fprintf(out, "%-*s", METRIC_LEN, "");
843 return;
844 }
845
846 if (newline)
847 do_new_line_std(os);
848
849 n = fprintf(out, " # ");
850 if (color)
851 n += color_fprintf(out, color, fmt, val);
852 else
853 n += fprintf(out, fmt, val);
854 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
855 }
856
857 static void new_line_csv(void *ctx)
858 {
859 struct outstate *os = ctx;
860 int i;
861
862 fputc('\n', os->fh);
863 if (os->prefix)
864 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
865 aggr_printout(os->evsel, os->id, os->nr);
866 for (i = 0; i < os->nfields; i++)
867 fputs(csv_sep, os->fh);
868 }
869
870 static void print_metric_csv(void *ctx,
871 const char *color __maybe_unused,
872 const char *fmt, const char *unit, double val)
873 {
874 struct outstate *os = ctx;
875 FILE *out = os->fh;
876 char buf[64], *vals, *ends;
877
878 if (unit == NULL || fmt == NULL) {
879 fprintf(out, "%s%s", csv_sep, csv_sep);
880 return;
881 }
882 snprintf(buf, sizeof(buf), fmt, val);
883 vals = buf;
884 while (isspace(*vals))
885 vals++;
886 ends = vals;
887 while (isdigit(*ends) || *ends == '.')
888 ends++;
889 *ends = 0;
890 while (isspace(*unit))
891 unit++;
892 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
893 }
894
895 #define METRIC_ONLY_LEN 20
896
897 /* Filter out some columns that don't work well in metrics only mode */
898
899 static bool valid_only_metric(const char *unit)
900 {
901 if (!unit)
902 return false;
903 if (strstr(unit, "/sec") ||
904 strstr(unit, "hz") ||
905 strstr(unit, "Hz") ||
906 strstr(unit, "CPUs utilized"))
907 return false;
908 return true;
909 }
910
911 static const char *fixunit(char *buf, struct perf_evsel *evsel,
912 const char *unit)
913 {
914 if (!strncmp(unit, "of all", 6)) {
915 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
916 unit);
917 return buf;
918 }
919 return unit;
920 }
921
922 static void print_metric_only(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 char buf[1024];
929 unsigned mlen = METRIC_ONLY_LEN;
930
931 if (!valid_only_metric(unit))
932 return;
933 unit = fixunit(buf, os->evsel, unit);
934 if (color)
935 n = color_fprintf(out, color, fmt, val);
936 else
937 n = fprintf(out, fmt, val);
938 if (n > METRIC_ONLY_LEN)
939 n = METRIC_ONLY_LEN;
940 if (mlen < strlen(unit))
941 mlen = strlen(unit) + 1;
942 fprintf(out, "%*s", mlen - n, "");
943 }
944
945 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
946 const char *fmt,
947 const char *unit, double val)
948 {
949 struct outstate *os = ctx;
950 FILE *out = os->fh;
951 char buf[64], *vals, *ends;
952 char tbuf[1024];
953
954 if (!valid_only_metric(unit))
955 return;
956 unit = fixunit(tbuf, os->evsel, unit);
957 snprintf(buf, sizeof buf, fmt, val);
958 vals = buf;
959 while (isspace(*vals))
960 vals++;
961 ends = vals;
962 while (isdigit(*ends) || *ends == '.')
963 ends++;
964 *ends = 0;
965 fprintf(out, "%s%s", vals, csv_sep);
966 }
967
968 static void new_line_metric(void *ctx __maybe_unused)
969 {
970 }
971
972 static void print_metric_header(void *ctx, const char *color __maybe_unused,
973 const char *fmt __maybe_unused,
974 const char *unit, double val __maybe_unused)
975 {
976 struct outstate *os = ctx;
977 char tbuf[1024];
978
979 if (!valid_only_metric(unit))
980 return;
981 unit = fixunit(tbuf, os->evsel, unit);
982 if (csv_output)
983 fprintf(os->fh, "%s%s", unit, csv_sep);
984 else
985 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
986 }
987
988 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
989 {
990 FILE *output = stat_config.output;
991 double msecs = avg / NSEC_PER_MSEC;
992 const char *fmt_v, *fmt_n;
993 char name[25];
994
995 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
996 fmt_n = csv_output ? "%s" : "%-25s";
997
998 aggr_printout(evsel, id, nr);
999
1000 scnprintf(name, sizeof(name), "%s%s",
1001 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
1002
1003 fprintf(output, fmt_v, msecs, csv_sep);
1004
1005 if (csv_output)
1006 fprintf(output, "%s%s", evsel->unit, csv_sep);
1007 else
1008 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
1009
1010 fprintf(output, fmt_n, name);
1011
1012 if (evsel->cgrp)
1013 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1014 }
1015
1016 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1017 {
1018 int i;
1019
1020 if (!aggr_get_id)
1021 return 0;
1022
1023 if (stat_config.aggr_mode == AGGR_NONE)
1024 return id;
1025
1026 if (stat_config.aggr_mode == AGGR_GLOBAL)
1027 return 0;
1028
1029 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1030 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1031
1032 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1033 return cpu2;
1034 }
1035 return 0;
1036 }
1037
1038 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1039 {
1040 FILE *output = stat_config.output;
1041 double sc = evsel->scale;
1042 const char *fmt;
1043
1044 if (csv_output) {
1045 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
1046 } else {
1047 if (big_num)
1048 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1049 else
1050 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1051 }
1052
1053 aggr_printout(evsel, id, nr);
1054
1055 fprintf(output, fmt, avg, csv_sep);
1056
1057 if (evsel->unit)
1058 fprintf(output, "%-*s%s",
1059 csv_output ? 0 : unit_width,
1060 evsel->unit, csv_sep);
1061
1062 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1063
1064 if (evsel->cgrp)
1065 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1066 }
1067
1068 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1069 char *prefix, u64 run, u64 ena, double noise)
1070 {
1071 struct perf_stat_output_ctx out;
1072 struct outstate os = {
1073 .fh = stat_config.output,
1074 .prefix = prefix ? prefix : "",
1075 .id = id,
1076 .nr = nr,
1077 .evsel = counter,
1078 };
1079 print_metric_t pm = print_metric_std;
1080 void (*nl)(void *);
1081
1082 if (metric_only) {
1083 nl = new_line_metric;
1084 if (csv_output)
1085 pm = print_metric_only_csv;
1086 else
1087 pm = print_metric_only;
1088 } else
1089 nl = new_line_std;
1090
1091 if (csv_output && !metric_only) {
1092 static int aggr_fields[] = {
1093 [AGGR_GLOBAL] = 0,
1094 [AGGR_THREAD] = 1,
1095 [AGGR_NONE] = 1,
1096 [AGGR_SOCKET] = 2,
1097 [AGGR_CORE] = 2,
1098 };
1099
1100 pm = print_metric_csv;
1101 nl = new_line_csv;
1102 os.nfields = 3;
1103 os.nfields += aggr_fields[stat_config.aggr_mode];
1104 if (counter->cgrp)
1105 os.nfields++;
1106 }
1107 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1108 if (metric_only) {
1109 pm(&os, NULL, "", "", 0);
1110 return;
1111 }
1112 aggr_printout(counter, id, nr);
1113
1114 fprintf(stat_config.output, "%*s%s",
1115 csv_output ? 0 : 18,
1116 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1117 csv_sep);
1118
1119 if (counter->supported)
1120 print_free_counters_hint = 1;
1121
1122 fprintf(stat_config.output, "%-*s%s",
1123 csv_output ? 0 : unit_width,
1124 counter->unit, csv_sep);
1125
1126 fprintf(stat_config.output, "%*s",
1127 csv_output ? 0 : -25,
1128 perf_evsel__name(counter));
1129
1130 if (counter->cgrp)
1131 fprintf(stat_config.output, "%s%s",
1132 csv_sep, counter->cgrp->name);
1133
1134 if (!csv_output)
1135 pm(&os, NULL, NULL, "", 0);
1136 print_noise(counter, noise);
1137 print_running(run, ena);
1138 if (csv_output)
1139 pm(&os, NULL, NULL, "", 0);
1140 return;
1141 }
1142
1143 if (metric_only)
1144 /* nothing */;
1145 else if (nsec_counter(counter))
1146 nsec_printout(id, nr, counter, uval);
1147 else
1148 abs_printout(id, nr, counter, uval);
1149
1150 out.print_metric = pm;
1151 out.new_line = nl;
1152 out.ctx = &os;
1153
1154 if (csv_output && !metric_only) {
1155 print_noise(counter, noise);
1156 print_running(run, ena);
1157 }
1158
1159 perf_stat__print_shadow_stats(counter, uval,
1160 first_shadow_cpu(counter, id),
1161 &out);
1162 if (!csv_output && !metric_only) {
1163 print_noise(counter, noise);
1164 print_running(run, ena);
1165 }
1166 }
1167
1168 static void aggr_update_shadow(void)
1169 {
1170 int cpu, s2, id, s;
1171 u64 val;
1172 struct perf_evsel *counter;
1173
1174 for (s = 0; s < aggr_map->nr; s++) {
1175 id = aggr_map->map[s];
1176 evlist__for_each_entry(evsel_list, counter) {
1177 val = 0;
1178 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1179 s2 = aggr_get_id(evsel_list->cpus, cpu);
1180 if (s2 != id)
1181 continue;
1182 val += perf_counts(counter->counts, cpu, 0)->val;
1183 }
1184 val = val * counter->scale;
1185 perf_stat__update_shadow_stats(counter, &val,
1186 first_shadow_cpu(counter, id));
1187 }
1188 }
1189 }
1190
1191 static void print_aggr(char *prefix)
1192 {
1193 FILE *output = stat_config.output;
1194 struct perf_evsel *counter;
1195 int cpu, s, s2, id, nr;
1196 double uval;
1197 u64 ena, run, val;
1198 bool first;
1199
1200 if (!(aggr_map || aggr_get_id))
1201 return;
1202
1203 aggr_update_shadow();
1204
1205 /*
1206 * With metric_only everything is on a single line.
1207 * Without each counter has its own line.
1208 */
1209 for (s = 0; s < aggr_map->nr; s++) {
1210 if (prefix && metric_only)
1211 fprintf(output, "%s", prefix);
1212
1213 id = aggr_map->map[s];
1214 first = true;
1215 evlist__for_each_entry(evsel_list, counter) {
1216 val = ena = run = 0;
1217 nr = 0;
1218 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1219 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1220 if (s2 != id)
1221 continue;
1222 val += perf_counts(counter->counts, cpu, 0)->val;
1223 ena += perf_counts(counter->counts, cpu, 0)->ena;
1224 run += perf_counts(counter->counts, cpu, 0)->run;
1225 nr++;
1226 }
1227 if (first && metric_only) {
1228 first = false;
1229 aggr_printout(counter, id, nr);
1230 }
1231 if (prefix && !metric_only)
1232 fprintf(output, "%s", prefix);
1233
1234 uval = val * counter->scale;
1235 printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1236 if (!metric_only)
1237 fputc('\n', output);
1238 }
1239 if (metric_only)
1240 fputc('\n', output);
1241 }
1242 }
1243
1244 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1245 {
1246 FILE *output = stat_config.output;
1247 int nthreads = thread_map__nr(counter->threads);
1248 int ncpus = cpu_map__nr(counter->cpus);
1249 int cpu, thread;
1250 double uval;
1251
1252 for (thread = 0; thread < nthreads; thread++) {
1253 u64 ena = 0, run = 0, val = 0;
1254
1255 for (cpu = 0; cpu < ncpus; cpu++) {
1256 val += perf_counts(counter->counts, cpu, thread)->val;
1257 ena += perf_counts(counter->counts, cpu, thread)->ena;
1258 run += perf_counts(counter->counts, cpu, thread)->run;
1259 }
1260
1261 if (prefix)
1262 fprintf(output, "%s", prefix);
1263
1264 uval = val * counter->scale;
1265 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1266 fputc('\n', output);
1267 }
1268 }
1269
1270 /*
1271 * Print out the results of a single counter:
1272 * aggregated counts in system-wide mode
1273 */
1274 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1275 {
1276 FILE *output = stat_config.output;
1277 struct perf_stat_evsel *ps = counter->priv;
1278 double avg = avg_stats(&ps->res_stats[0]);
1279 double uval;
1280 double avg_enabled, avg_running;
1281
1282 avg_enabled = avg_stats(&ps->res_stats[1]);
1283 avg_running = avg_stats(&ps->res_stats[2]);
1284
1285 if (prefix && !metric_only)
1286 fprintf(output, "%s", prefix);
1287
1288 uval = avg * counter->scale;
1289 printout(-1, 0, counter, uval, prefix, avg_running, avg_enabled, avg);
1290 if (!metric_only)
1291 fprintf(output, "\n");
1292 }
1293
1294 /*
1295 * Print out the results of a single counter:
1296 * does not use aggregated count in system-wide
1297 */
1298 static void print_counter(struct perf_evsel *counter, char *prefix)
1299 {
1300 FILE *output = stat_config.output;
1301 u64 ena, run, val;
1302 double uval;
1303 int cpu;
1304
1305 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1306 val = perf_counts(counter->counts, cpu, 0)->val;
1307 ena = perf_counts(counter->counts, cpu, 0)->ena;
1308 run = perf_counts(counter->counts, cpu, 0)->run;
1309
1310 if (prefix)
1311 fprintf(output, "%s", prefix);
1312
1313 uval = val * counter->scale;
1314 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1315
1316 fputc('\n', output);
1317 }
1318 }
1319
1320 static void print_no_aggr_metric(char *prefix)
1321 {
1322 int cpu;
1323 int nrcpus = 0;
1324 struct perf_evsel *counter;
1325 u64 ena, run, val;
1326 double uval;
1327
1328 nrcpus = evsel_list->cpus->nr;
1329 for (cpu = 0; cpu < nrcpus; cpu++) {
1330 bool first = true;
1331
1332 if (prefix)
1333 fputs(prefix, stat_config.output);
1334 evlist__for_each_entry(evsel_list, counter) {
1335 if (first) {
1336 aggr_printout(counter, cpu, 0);
1337 first = false;
1338 }
1339 val = perf_counts(counter->counts, cpu, 0)->val;
1340 ena = perf_counts(counter->counts, cpu, 0)->ena;
1341 run = perf_counts(counter->counts, cpu, 0)->run;
1342
1343 uval = val * counter->scale;
1344 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1345 }
1346 fputc('\n', stat_config.output);
1347 }
1348 }
1349
1350 static int aggr_header_lens[] = {
1351 [AGGR_CORE] = 18,
1352 [AGGR_SOCKET] = 12,
1353 [AGGR_NONE] = 6,
1354 [AGGR_THREAD] = 24,
1355 [AGGR_GLOBAL] = 0,
1356 };
1357
1358 static const char *aggr_header_csv[] = {
1359 [AGGR_CORE] = "core,cpus,",
1360 [AGGR_SOCKET] = "socket,cpus",
1361 [AGGR_NONE] = "cpu,",
1362 [AGGR_THREAD] = "comm-pid,",
1363 [AGGR_GLOBAL] = ""
1364 };
1365
1366 static void print_metric_headers(const char *prefix, bool no_indent)
1367 {
1368 struct perf_stat_output_ctx out;
1369 struct perf_evsel *counter;
1370 struct outstate os = {
1371 .fh = stat_config.output
1372 };
1373
1374 if (prefix)
1375 fprintf(stat_config.output, "%s", prefix);
1376
1377 if (!csv_output && !no_indent)
1378 fprintf(stat_config.output, "%*s",
1379 aggr_header_lens[stat_config.aggr_mode], "");
1380 if (csv_output) {
1381 if (stat_config.interval)
1382 fputs("time,", stat_config.output);
1383 fputs(aggr_header_csv[stat_config.aggr_mode],
1384 stat_config.output);
1385 }
1386
1387 /* Print metrics headers only */
1388 evlist__for_each_entry(evsel_list, counter) {
1389 os.evsel = counter;
1390 out.ctx = &os;
1391 out.print_metric = print_metric_header;
1392 out.new_line = new_line_metric;
1393 os.evsel = counter;
1394 perf_stat__print_shadow_stats(counter, 0,
1395 0,
1396 &out);
1397 }
1398 fputc('\n', stat_config.output);
1399 }
1400
1401 static void print_interval(char *prefix, struct timespec *ts)
1402 {
1403 FILE *output = stat_config.output;
1404 static int num_print_interval;
1405
1406 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1407
1408 if (num_print_interval == 0 && !csv_output) {
1409 switch (stat_config.aggr_mode) {
1410 case AGGR_SOCKET:
1411 fprintf(output, "# time socket cpus");
1412 if (!metric_only)
1413 fprintf(output, " counts %*s events\n", unit_width, "unit");
1414 break;
1415 case AGGR_CORE:
1416 fprintf(output, "# time core cpus");
1417 if (!metric_only)
1418 fprintf(output, " counts %*s events\n", unit_width, "unit");
1419 break;
1420 case AGGR_NONE:
1421 fprintf(output, "# time CPU");
1422 if (!metric_only)
1423 fprintf(output, " counts %*s events\n", unit_width, "unit");
1424 break;
1425 case AGGR_THREAD:
1426 fprintf(output, "# time comm-pid");
1427 if (!metric_only)
1428 fprintf(output, " counts %*s events\n", unit_width, "unit");
1429 break;
1430 case AGGR_GLOBAL:
1431 default:
1432 fprintf(output, "# time");
1433 if (!metric_only)
1434 fprintf(output, " counts %*s events\n", unit_width, "unit");
1435 case AGGR_UNSET:
1436 break;
1437 }
1438 }
1439
1440 if (num_print_interval == 0 && metric_only)
1441 print_metric_headers(" ", true);
1442 if (++num_print_interval == 25)
1443 num_print_interval = 0;
1444 }
1445
1446 static void print_header(int argc, const char **argv)
1447 {
1448 FILE *output = stat_config.output;
1449 int i;
1450
1451 fflush(stdout);
1452
1453 if (!csv_output) {
1454 fprintf(output, "\n");
1455 fprintf(output, " Performance counter stats for ");
1456 if (target.system_wide)
1457 fprintf(output, "\'system wide");
1458 else if (target.cpu_list)
1459 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1460 else if (!target__has_task(&target)) {
1461 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1462 for (i = 1; argv && (i < argc); i++)
1463 fprintf(output, " %s", argv[i]);
1464 } else if (target.pid)
1465 fprintf(output, "process id \'%s", target.pid);
1466 else
1467 fprintf(output, "thread id \'%s", target.tid);
1468
1469 fprintf(output, "\'");
1470 if (run_count > 1)
1471 fprintf(output, " (%d runs)", run_count);
1472 fprintf(output, ":\n\n");
1473 }
1474 }
1475
1476 static void print_footer(void)
1477 {
1478 FILE *output = stat_config.output;
1479
1480 if (!null_run)
1481 fprintf(output, "\n");
1482 fprintf(output, " %17.9f seconds time elapsed",
1483 avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC);
1484 if (run_count > 1) {
1485 fprintf(output, " ");
1486 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1487 avg_stats(&walltime_nsecs_stats));
1488 }
1489 fprintf(output, "\n\n");
1490
1491 if (print_free_counters_hint)
1492 fprintf(output,
1493 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1494 " echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1495 " perf stat ...\n"
1496 " echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1497 }
1498
1499 static void print_counters(struct timespec *ts, int argc, const char **argv)
1500 {
1501 int interval = stat_config.interval;
1502 struct perf_evsel *counter;
1503 char buf[64], *prefix = NULL;
1504
1505 /* Do not print anything if we record to the pipe. */
1506 if (STAT_RECORD && perf_stat.file.is_pipe)
1507 return;
1508
1509 if (interval)
1510 print_interval(prefix = buf, ts);
1511 else
1512 print_header(argc, argv);
1513
1514 if (metric_only) {
1515 static int num_print_iv;
1516
1517 if (num_print_iv == 0 && !interval)
1518 print_metric_headers(prefix, false);
1519 if (num_print_iv++ == 25)
1520 num_print_iv = 0;
1521 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1522 fprintf(stat_config.output, "%s", prefix);
1523 }
1524
1525 switch (stat_config.aggr_mode) {
1526 case AGGR_CORE:
1527 case AGGR_SOCKET:
1528 print_aggr(prefix);
1529 break;
1530 case AGGR_THREAD:
1531 evlist__for_each_entry(evsel_list, counter)
1532 print_aggr_thread(counter, prefix);
1533 break;
1534 case AGGR_GLOBAL:
1535 evlist__for_each_entry(evsel_list, counter)
1536 print_counter_aggr(counter, prefix);
1537 if (metric_only)
1538 fputc('\n', stat_config.output);
1539 break;
1540 case AGGR_NONE:
1541 if (metric_only)
1542 print_no_aggr_metric(prefix);
1543 else {
1544 evlist__for_each_entry(evsel_list, counter)
1545 print_counter(counter, prefix);
1546 }
1547 break;
1548 case AGGR_UNSET:
1549 default:
1550 break;
1551 }
1552
1553 if (!interval && !csv_output)
1554 print_footer();
1555
1556 fflush(stat_config.output);
1557 }
1558
1559 static volatile int signr = -1;
1560
1561 static void skip_signal(int signo)
1562 {
1563 if ((child_pid == -1) || stat_config.interval)
1564 done = 1;
1565
1566 signr = signo;
1567 /*
1568 * render child_pid harmless
1569 * won't send SIGTERM to a random
1570 * process in case of race condition
1571 * and fast PID recycling
1572 */
1573 child_pid = -1;
1574 }
1575
1576 static void sig_atexit(void)
1577 {
1578 sigset_t set, oset;
1579
1580 /*
1581 * avoid race condition with SIGCHLD handler
1582 * in skip_signal() which is modifying child_pid
1583 * goal is to avoid send SIGTERM to a random
1584 * process
1585 */
1586 sigemptyset(&set);
1587 sigaddset(&set, SIGCHLD);
1588 sigprocmask(SIG_BLOCK, &set, &oset);
1589
1590 if (child_pid != -1)
1591 kill(child_pid, SIGTERM);
1592
1593 sigprocmask(SIG_SETMASK, &oset, NULL);
1594
1595 if (signr == -1)
1596 return;
1597
1598 signal(signr, SIG_DFL);
1599 kill(getpid(), signr);
1600 }
1601
1602 static int stat__set_big_num(const struct option *opt __maybe_unused,
1603 const char *s __maybe_unused, int unset)
1604 {
1605 big_num_opt = unset ? 0 : 1;
1606 return 0;
1607 }
1608
1609 static int enable_metric_only(const struct option *opt __maybe_unused,
1610 const char *s __maybe_unused, int unset)
1611 {
1612 force_metric_only = true;
1613 metric_only = !unset;
1614 return 0;
1615 }
1616
1617 static const struct option stat_options[] = {
1618 OPT_BOOLEAN('T', "transaction", &transaction_run,
1619 "hardware transaction statistics"),
1620 OPT_CALLBACK('e', "event", &evsel_list, "event",
1621 "event selector. use 'perf list' to list available events",
1622 parse_events_option),
1623 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1624 "event filter", parse_filter),
1625 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1626 "child tasks do not inherit counters"),
1627 OPT_STRING('p', "pid", &target.pid, "pid",
1628 "stat events on existing process id"),
1629 OPT_STRING('t', "tid", &target.tid, "tid",
1630 "stat events on existing thread id"),
1631 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1632 "system-wide collection from all CPUs"),
1633 OPT_BOOLEAN('g', "group", &group,
1634 "put the counters into a counter group"),
1635 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1636 OPT_INCR('v', "verbose", &verbose,
1637 "be more verbose (show counter open errors, etc)"),
1638 OPT_INTEGER('r', "repeat", &run_count,
1639 "repeat command and print average + stddev (max: 100, forever: 0)"),
1640 OPT_BOOLEAN('n', "null", &null_run,
1641 "null run - dont start any counters"),
1642 OPT_INCR('d', "detailed", &detailed_run,
1643 "detailed run - start a lot of events"),
1644 OPT_BOOLEAN('S', "sync", &sync_run,
1645 "call sync() before starting a run"),
1646 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1647 "print large numbers with thousands\' separators",
1648 stat__set_big_num),
1649 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1650 "list of cpus to monitor in system-wide"),
1651 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1652 "disable CPU count aggregation", AGGR_NONE),
1653 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1654 "print counts with custom separator"),
1655 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1656 "monitor event in cgroup name only", parse_cgroups),
1657 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1658 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1659 OPT_INTEGER(0, "log-fd", &output_fd,
1660 "log output to fd, instead of stderr"),
1661 OPT_STRING(0, "pre", &pre_cmd, "command",
1662 "command to run prior to the measured command"),
1663 OPT_STRING(0, "post", &post_cmd, "command",
1664 "command to run after to the measured command"),
1665 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1666 "print counts at regular interval in ms (>= 10)"),
1667 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1668 "aggregate counts per processor socket", AGGR_SOCKET),
1669 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1670 "aggregate counts per physical processor core", AGGR_CORE),
1671 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1672 "aggregate counts per thread", AGGR_THREAD),
1673 OPT_UINTEGER('D', "delay", &initial_delay,
1674 "ms to wait before starting measurement after program start"),
1675 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1676 "Only print computed metrics. No raw values", enable_metric_only),
1677 OPT_BOOLEAN(0, "topdown", &topdown_run,
1678 "measure topdown level 1 statistics"),
1679 OPT_END()
1680 };
1681
1682 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1683 {
1684 return cpu_map__get_socket(map, cpu, NULL);
1685 }
1686
1687 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1688 {
1689 return cpu_map__get_core(map, cpu, NULL);
1690 }
1691
1692 static int cpu_map__get_max(struct cpu_map *map)
1693 {
1694 int i, max = -1;
1695
1696 for (i = 0; i < map->nr; i++) {
1697 if (map->map[i] > max)
1698 max = map->map[i];
1699 }
1700
1701 return max;
1702 }
1703
1704 static struct cpu_map *cpus_aggr_map;
1705
1706 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1707 {
1708 int cpu;
1709
1710 if (idx >= map->nr)
1711 return -1;
1712
1713 cpu = map->map[idx];
1714
1715 if (cpus_aggr_map->map[cpu] == -1)
1716 cpus_aggr_map->map[cpu] = get_id(map, idx);
1717
1718 return cpus_aggr_map->map[cpu];
1719 }
1720
1721 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1722 {
1723 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1724 }
1725
1726 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1727 {
1728 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1729 }
1730
1731 static int perf_stat_init_aggr_mode(void)
1732 {
1733 int nr;
1734
1735 switch (stat_config.aggr_mode) {
1736 case AGGR_SOCKET:
1737 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1738 perror("cannot build socket map");
1739 return -1;
1740 }
1741 aggr_get_id = perf_stat__get_socket_cached;
1742 break;
1743 case AGGR_CORE:
1744 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1745 perror("cannot build core map");
1746 return -1;
1747 }
1748 aggr_get_id = perf_stat__get_core_cached;
1749 break;
1750 case AGGR_NONE:
1751 case AGGR_GLOBAL:
1752 case AGGR_THREAD:
1753 case AGGR_UNSET:
1754 default:
1755 break;
1756 }
1757
1758 /*
1759 * The evsel_list->cpus is the base we operate on,
1760 * taking the highest cpu number to be the size of
1761 * the aggregation translate cpumap.
1762 */
1763 nr = cpu_map__get_max(evsel_list->cpus);
1764 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1765 return cpus_aggr_map ? 0 : -ENOMEM;
1766 }
1767
1768 static void perf_stat__exit_aggr_mode(void)
1769 {
1770 cpu_map__put(aggr_map);
1771 cpu_map__put(cpus_aggr_map);
1772 aggr_map = NULL;
1773 cpus_aggr_map = NULL;
1774 }
1775
1776 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1777 {
1778 int cpu;
1779
1780 if (idx > map->nr)
1781 return -1;
1782
1783 cpu = map->map[idx];
1784
1785 if (cpu >= env->nr_cpus_online)
1786 return -1;
1787
1788 return cpu;
1789 }
1790
1791 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1792 {
1793 struct perf_env *env = data;
1794 int cpu = perf_env__get_cpu(env, map, idx);
1795
1796 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1797 }
1798
1799 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1800 {
1801 struct perf_env *env = data;
1802 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1803
1804 if (cpu != -1) {
1805 int socket_id = env->cpu[cpu].socket_id;
1806
1807 /*
1808 * Encode socket in upper 16 bits
1809 * core_id is relative to socket, and
1810 * we need a global id. So we combine
1811 * socket + core id.
1812 */
1813 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1814 }
1815
1816 return core;
1817 }
1818
1819 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1820 struct cpu_map **sockp)
1821 {
1822 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1823 }
1824
1825 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1826 struct cpu_map **corep)
1827 {
1828 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1829 }
1830
1831 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1832 {
1833 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1834 }
1835
1836 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1837 {
1838 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1839 }
1840
1841 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1842 {
1843 struct perf_env *env = &st->session->header.env;
1844
1845 switch (stat_config.aggr_mode) {
1846 case AGGR_SOCKET:
1847 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1848 perror("cannot build socket map");
1849 return -1;
1850 }
1851 aggr_get_id = perf_stat__get_socket_file;
1852 break;
1853 case AGGR_CORE:
1854 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1855 perror("cannot build core map");
1856 return -1;
1857 }
1858 aggr_get_id = perf_stat__get_core_file;
1859 break;
1860 case AGGR_NONE:
1861 case AGGR_GLOBAL:
1862 case AGGR_THREAD:
1863 case AGGR_UNSET:
1864 default:
1865 break;
1866 }
1867
1868 return 0;
1869 }
1870
1871 static int topdown_filter_events(const char **attr, char **str, bool use_group)
1872 {
1873 int off = 0;
1874 int i;
1875 int len = 0;
1876 char *s;
1877
1878 for (i = 0; attr[i]; i++) {
1879 if (pmu_have_event("cpu", attr[i])) {
1880 len += strlen(attr[i]) + 1;
1881 attr[i - off] = attr[i];
1882 } else
1883 off++;
1884 }
1885 attr[i - off] = NULL;
1886
1887 *str = malloc(len + 1 + 2);
1888 if (!*str)
1889 return -1;
1890 s = *str;
1891 if (i - off == 0) {
1892 *s = 0;
1893 return 0;
1894 }
1895 if (use_group)
1896 *s++ = '{';
1897 for (i = 0; attr[i]; i++) {
1898 strcpy(s, attr[i]);
1899 s += strlen(s);
1900 *s++ = ',';
1901 }
1902 if (use_group) {
1903 s[-1] = '}';
1904 *s = 0;
1905 } else
1906 s[-1] = 0;
1907 return 0;
1908 }
1909
1910 __weak bool arch_topdown_check_group(bool *warn)
1911 {
1912 *warn = false;
1913 return false;
1914 }
1915
1916 __weak void arch_topdown_group_warn(void)
1917 {
1918 }
1919
1920 /*
1921 * Add default attributes, if there were no attributes specified or
1922 * if -d/--detailed, -d -d or -d -d -d is used:
1923 */
1924 static int add_default_attributes(void)
1925 {
1926 int err;
1927 struct perf_event_attr default_attrs0[] = {
1928
1929 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1930 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1931 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1932 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1933
1934 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1935 };
1936 struct perf_event_attr frontend_attrs[] = {
1937 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1938 };
1939 struct perf_event_attr backend_attrs[] = {
1940 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1941 };
1942 struct perf_event_attr default_attrs1[] = {
1943 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1944 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1945 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1946
1947 };
1948
1949 /*
1950 * Detailed stats (-d), covering the L1 and last level data caches:
1951 */
1952 struct perf_event_attr detailed_attrs[] = {
1953
1954 { .type = PERF_TYPE_HW_CACHE,
1955 .config =
1956 PERF_COUNT_HW_CACHE_L1D << 0 |
1957 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1958 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1959
1960 { .type = PERF_TYPE_HW_CACHE,
1961 .config =
1962 PERF_COUNT_HW_CACHE_L1D << 0 |
1963 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1964 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1965
1966 { .type = PERF_TYPE_HW_CACHE,
1967 .config =
1968 PERF_COUNT_HW_CACHE_LL << 0 |
1969 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1970 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1971
1972 { .type = PERF_TYPE_HW_CACHE,
1973 .config =
1974 PERF_COUNT_HW_CACHE_LL << 0 |
1975 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1976 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1977 };
1978
1979 /*
1980 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1981 */
1982 struct perf_event_attr very_detailed_attrs[] = {
1983
1984 { .type = PERF_TYPE_HW_CACHE,
1985 .config =
1986 PERF_COUNT_HW_CACHE_L1I << 0 |
1987 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1988 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1989
1990 { .type = PERF_TYPE_HW_CACHE,
1991 .config =
1992 PERF_COUNT_HW_CACHE_L1I << 0 |
1993 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1994 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1995
1996 { .type = PERF_TYPE_HW_CACHE,
1997 .config =
1998 PERF_COUNT_HW_CACHE_DTLB << 0 |
1999 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2000 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2001
2002 { .type = PERF_TYPE_HW_CACHE,
2003 .config =
2004 PERF_COUNT_HW_CACHE_DTLB << 0 |
2005 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2006 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2007
2008 { .type = PERF_TYPE_HW_CACHE,
2009 .config =
2010 PERF_COUNT_HW_CACHE_ITLB << 0 |
2011 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2012 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2013
2014 { .type = PERF_TYPE_HW_CACHE,
2015 .config =
2016 PERF_COUNT_HW_CACHE_ITLB << 0 |
2017 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2018 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2019
2020 };
2021
2022 /*
2023 * Very, very detailed stats (-d -d -d), adding prefetch events:
2024 */
2025 struct perf_event_attr very_very_detailed_attrs[] = {
2026
2027 { .type = PERF_TYPE_HW_CACHE,
2028 .config =
2029 PERF_COUNT_HW_CACHE_L1D << 0 |
2030 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2031 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2032
2033 { .type = PERF_TYPE_HW_CACHE,
2034 .config =
2035 PERF_COUNT_HW_CACHE_L1D << 0 |
2036 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2037 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2038 };
2039
2040 /* Set attrs if no event is selected and !null_run: */
2041 if (null_run)
2042 return 0;
2043
2044 if (transaction_run) {
2045 struct parse_events_error errinfo;
2046
2047 if (pmu_have_event("cpu", "cycles-ct") &&
2048 pmu_have_event("cpu", "el-start"))
2049 err = parse_events(evsel_list, transaction_attrs,
2050 &errinfo);
2051 else
2052 err = parse_events(evsel_list,
2053 transaction_limited_attrs,
2054 &errinfo);
2055 if (err) {
2056 fprintf(stderr, "Cannot set up transaction events\n");
2057 return -1;
2058 }
2059 return 0;
2060 }
2061
2062 if (topdown_run) {
2063 char *str = NULL;
2064 bool warn = false;
2065
2066 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2067 stat_config.aggr_mode != AGGR_CORE) {
2068 pr_err("top down event configuration requires --per-core mode\n");
2069 return -1;
2070 }
2071 stat_config.aggr_mode = AGGR_CORE;
2072 if (nr_cgroups || !target__has_cpu(&target)) {
2073 pr_err("top down event configuration requires system-wide mode (-a)\n");
2074 return -1;
2075 }
2076
2077 if (!force_metric_only)
2078 metric_only = true;
2079 if (topdown_filter_events(topdown_attrs, &str,
2080 arch_topdown_check_group(&warn)) < 0) {
2081 pr_err("Out of memory\n");
2082 return -1;
2083 }
2084 if (topdown_attrs[0] && str) {
2085 if (warn)
2086 arch_topdown_group_warn();
2087 err = parse_events(evsel_list, str, NULL);
2088 if (err) {
2089 fprintf(stderr,
2090 "Cannot set up top down events %s: %d\n",
2091 str, err);
2092 free(str);
2093 return -1;
2094 }
2095 } else {
2096 fprintf(stderr, "System does not support topdown\n");
2097 return -1;
2098 }
2099 free(str);
2100 }
2101
2102 if (!evsel_list->nr_entries) {
2103 if (target__has_cpu(&target))
2104 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2105
2106 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2107 return -1;
2108 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2109 if (perf_evlist__add_default_attrs(evsel_list,
2110 frontend_attrs) < 0)
2111 return -1;
2112 }
2113 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2114 if (perf_evlist__add_default_attrs(evsel_list,
2115 backend_attrs) < 0)
2116 return -1;
2117 }
2118 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2119 return -1;
2120 }
2121
2122 /* Detailed events get appended to the event list: */
2123
2124 if (detailed_run < 1)
2125 return 0;
2126
2127 /* Append detailed run extra attributes: */
2128 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2129 return -1;
2130
2131 if (detailed_run < 2)
2132 return 0;
2133
2134 /* Append very detailed run extra attributes: */
2135 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2136 return -1;
2137
2138 if (detailed_run < 3)
2139 return 0;
2140
2141 /* Append very, very detailed run extra attributes: */
2142 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2143 }
2144
2145 static const char * const stat_record_usage[] = {
2146 "perf stat record [<options>]",
2147 NULL,
2148 };
2149
2150 static void init_features(struct perf_session *session)
2151 {
2152 int feat;
2153
2154 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2155 perf_header__set_feat(&session->header, feat);
2156
2157 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2158 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2159 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2160 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2161 }
2162
2163 static int __cmd_record(int argc, const char **argv)
2164 {
2165 struct perf_session *session;
2166 struct perf_data_file *file = &perf_stat.file;
2167
2168 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2169 PARSE_OPT_STOP_AT_NON_OPTION);
2170
2171 if (output_name)
2172 file->path = output_name;
2173
2174 if (run_count != 1 || forever) {
2175 pr_err("Cannot use -r option with perf stat record.\n");
2176 return -1;
2177 }
2178
2179 session = perf_session__new(file, false, NULL);
2180 if (session == NULL) {
2181 pr_err("Perf session creation failed.\n");
2182 return -1;
2183 }
2184
2185 init_features(session);
2186
2187 session->evlist = evsel_list;
2188 perf_stat.session = session;
2189 perf_stat.record = true;
2190 return argc;
2191 }
2192
2193 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2194 union perf_event *event,
2195 struct perf_session *session)
2196 {
2197 struct stat_round_event *stat_round = &event->stat_round;
2198 struct perf_evsel *counter;
2199 struct timespec tsh, *ts = NULL;
2200 const char **argv = session->header.env.cmdline_argv;
2201 int argc = session->header.env.nr_cmdline;
2202
2203 evlist__for_each_entry(evsel_list, counter)
2204 perf_stat_process_counter(&stat_config, counter);
2205
2206 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2207 update_stats(&walltime_nsecs_stats, stat_round->time);
2208
2209 if (stat_config.interval && stat_round->time) {
2210 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2211 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2212 ts = &tsh;
2213 }
2214
2215 print_counters(ts, argc, argv);
2216 return 0;
2217 }
2218
2219 static
2220 int process_stat_config_event(struct perf_tool *tool __maybe_unused,
2221 union perf_event *event,
2222 struct perf_session *session __maybe_unused)
2223 {
2224 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2225
2226 perf_event__read_stat_config(&stat_config, &event->stat_config);
2227
2228 if (cpu_map__empty(st->cpus)) {
2229 if (st->aggr_mode != AGGR_UNSET)
2230 pr_warning("warning: processing task data, aggregation mode not set\n");
2231 return 0;
2232 }
2233
2234 if (st->aggr_mode != AGGR_UNSET)
2235 stat_config.aggr_mode = st->aggr_mode;
2236
2237 if (perf_stat.file.is_pipe)
2238 perf_stat_init_aggr_mode();
2239 else
2240 perf_stat_init_aggr_mode_file(st);
2241
2242 return 0;
2243 }
2244
2245 static int set_maps(struct perf_stat *st)
2246 {
2247 if (!st->cpus || !st->threads)
2248 return 0;
2249
2250 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2251 return -EINVAL;
2252
2253 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2254
2255 if (perf_evlist__alloc_stats(evsel_list, true))
2256 return -ENOMEM;
2257
2258 st->maps_allocated = true;
2259 return 0;
2260 }
2261
2262 static
2263 int process_thread_map_event(struct perf_tool *tool __maybe_unused,
2264 union perf_event *event,
2265 struct perf_session *session __maybe_unused)
2266 {
2267 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2268
2269 if (st->threads) {
2270 pr_warning("Extra thread map event, ignoring.\n");
2271 return 0;
2272 }
2273
2274 st->threads = thread_map__new_event(&event->thread_map);
2275 if (!st->threads)
2276 return -ENOMEM;
2277
2278 return set_maps(st);
2279 }
2280
2281 static
2282 int process_cpu_map_event(struct perf_tool *tool __maybe_unused,
2283 union perf_event *event,
2284 struct perf_session *session __maybe_unused)
2285 {
2286 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2287 struct cpu_map *cpus;
2288
2289 if (st->cpus) {
2290 pr_warning("Extra cpu map event, ignoring.\n");
2291 return 0;
2292 }
2293
2294 cpus = cpu_map__new_data(&event->cpu_map.data);
2295 if (!cpus)
2296 return -ENOMEM;
2297
2298 st->cpus = cpus;
2299 return set_maps(st);
2300 }
2301
2302 static const char * const stat_report_usage[] = {
2303 "perf stat report [<options>]",
2304 NULL,
2305 };
2306
2307 static struct perf_stat perf_stat = {
2308 .tool = {
2309 .attr = perf_event__process_attr,
2310 .event_update = perf_event__process_event_update,
2311 .thread_map = process_thread_map_event,
2312 .cpu_map = process_cpu_map_event,
2313 .stat_config = process_stat_config_event,
2314 .stat = perf_event__process_stat_event,
2315 .stat_round = process_stat_round_event,
2316 },
2317 .aggr_mode = AGGR_UNSET,
2318 };
2319
2320 static int __cmd_report(int argc, const char **argv)
2321 {
2322 struct perf_session *session;
2323 const struct option options[] = {
2324 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2325 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2326 "aggregate counts per processor socket", AGGR_SOCKET),
2327 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2328 "aggregate counts per physical processor core", AGGR_CORE),
2329 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2330 "disable CPU count aggregation", AGGR_NONE),
2331 OPT_END()
2332 };
2333 struct stat st;
2334 int ret;
2335
2336 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2337
2338 if (!input_name || !strlen(input_name)) {
2339 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2340 input_name = "-";
2341 else
2342 input_name = "perf.data";
2343 }
2344
2345 perf_stat.file.path = input_name;
2346 perf_stat.file.mode = PERF_DATA_MODE_READ;
2347
2348 session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
2349 if (session == NULL)
2350 return -1;
2351
2352 perf_stat.session = session;
2353 stat_config.output = stderr;
2354 evsel_list = session->evlist;
2355
2356 ret = perf_session__process_events(session);
2357 if (ret)
2358 return ret;
2359
2360 perf_session__delete(session);
2361 return 0;
2362 }
2363
2364 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
2365 {
2366 const char * const stat_usage[] = {
2367 "perf stat [<options>] [<command>]",
2368 NULL
2369 };
2370 int status = -EINVAL, run_idx;
2371 const char *mode;
2372 FILE *output = stderr;
2373 unsigned int interval;
2374 const char * const stat_subcommands[] = { "record", "report" };
2375
2376 setlocale(LC_ALL, "");
2377
2378 evsel_list = perf_evlist__new();
2379 if (evsel_list == NULL)
2380 return -ENOMEM;
2381
2382 parse_events__shrink_config_terms();
2383 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2384 (const char **) stat_usage,
2385 PARSE_OPT_STOP_AT_NON_OPTION);
2386 perf_stat__init_shadow_stats();
2387
2388 if (csv_sep) {
2389 csv_output = true;
2390 if (!strcmp(csv_sep, "\\t"))
2391 csv_sep = "\t";
2392 } else
2393 csv_sep = DEFAULT_SEPARATOR;
2394
2395 if (argc && !strncmp(argv[0], "rec", 3)) {
2396 argc = __cmd_record(argc, argv);
2397 if (argc < 0)
2398 return -1;
2399 } else if (argc && !strncmp(argv[0], "rep", 3))
2400 return __cmd_report(argc, argv);
2401
2402 interval = stat_config.interval;
2403
2404 /*
2405 * For record command the -o is already taken care of.
2406 */
2407 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2408 output = NULL;
2409
2410 if (output_name && output_fd) {
2411 fprintf(stderr, "cannot use both --output and --log-fd\n");
2412 parse_options_usage(stat_usage, stat_options, "o", 1);
2413 parse_options_usage(NULL, stat_options, "log-fd", 0);
2414 goto out;
2415 }
2416
2417 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2418 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2419 goto out;
2420 }
2421
2422 if (metric_only && run_count > 1) {
2423 fprintf(stderr, "--metric-only is not supported with -r\n");
2424 goto out;
2425 }
2426
2427 if (output_fd < 0) {
2428 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2429 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2430 goto out;
2431 }
2432
2433 if (!output) {
2434 struct timespec tm;
2435 mode = append_file ? "a" : "w";
2436
2437 output = fopen(output_name, mode);
2438 if (!output) {
2439 perror("failed to create output file");
2440 return -1;
2441 }
2442 clock_gettime(CLOCK_REALTIME, &tm);
2443 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2444 } else if (output_fd > 0) {
2445 mode = append_file ? "a" : "w";
2446 output = fdopen(output_fd, mode);
2447 if (!output) {
2448 perror("Failed opening logfd");
2449 return -errno;
2450 }
2451 }
2452
2453 stat_config.output = output;
2454
2455 /*
2456 * let the spreadsheet do the pretty-printing
2457 */
2458 if (csv_output) {
2459 /* User explicitly passed -B? */
2460 if (big_num_opt == 1) {
2461 fprintf(stderr, "-B option not supported with -x\n");
2462 parse_options_usage(stat_usage, stat_options, "B", 1);
2463 parse_options_usage(NULL, stat_options, "x", 1);
2464 goto out;
2465 } else /* Nope, so disable big number formatting */
2466 big_num = false;
2467 } else if (big_num_opt == 0) /* User passed --no-big-num */
2468 big_num = false;
2469
2470 if (!argc && target__none(&target))
2471 usage_with_options(stat_usage, stat_options);
2472
2473 if (run_count < 0) {
2474 pr_err("Run count must be a positive number\n");
2475 parse_options_usage(stat_usage, stat_options, "r", 1);
2476 goto out;
2477 } else if (run_count == 0) {
2478 forever = true;
2479 run_count = 1;
2480 }
2481
2482 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2483 fprintf(stderr, "The --per-thread option is only available "
2484 "when monitoring via -p -t options.\n");
2485 parse_options_usage(NULL, stat_options, "p", 1);
2486 parse_options_usage(NULL, stat_options, "t", 1);
2487 goto out;
2488 }
2489
2490 /*
2491 * no_aggr, cgroup are for system-wide only
2492 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2493 */
2494 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2495 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2496 !target__has_cpu(&target)) {
2497 fprintf(stderr, "both cgroup and no-aggregation "
2498 "modes only available in system-wide mode\n");
2499
2500 parse_options_usage(stat_usage, stat_options, "G", 1);
2501 parse_options_usage(NULL, stat_options, "A", 1);
2502 parse_options_usage(NULL, stat_options, "a", 1);
2503 goto out;
2504 }
2505
2506 if (add_default_attributes())
2507 goto out;
2508
2509 target__validate(&target);
2510
2511 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2512 if (target__has_task(&target)) {
2513 pr_err("Problems finding threads of monitor\n");
2514 parse_options_usage(stat_usage, stat_options, "p", 1);
2515 parse_options_usage(NULL, stat_options, "t", 1);
2516 } else if (target__has_cpu(&target)) {
2517 perror("failed to parse CPUs map");
2518 parse_options_usage(stat_usage, stat_options, "C", 1);
2519 parse_options_usage(NULL, stat_options, "a", 1);
2520 }
2521 goto out;
2522 }
2523
2524 /*
2525 * Initialize thread_map with comm names,
2526 * so we could print it out on output.
2527 */
2528 if (stat_config.aggr_mode == AGGR_THREAD)
2529 thread_map__read_comms(evsel_list->threads);
2530
2531 if (interval && interval < 100) {
2532 if (interval < 10) {
2533 pr_err("print interval must be >= 10ms\n");
2534 parse_options_usage(stat_usage, stat_options, "I", 1);
2535 goto out;
2536 } else
2537 pr_warning("print interval < 100ms. "
2538 "The overhead percentage could be high in some cases. "
2539 "Please proceed with caution.\n");
2540 }
2541
2542 if (perf_evlist__alloc_stats(evsel_list, interval))
2543 goto out;
2544
2545 if (perf_stat_init_aggr_mode())
2546 goto out;
2547
2548 /*
2549 * We dont want to block the signals - that would cause
2550 * child tasks to inherit that and Ctrl-C would not work.
2551 * What we want is for Ctrl-C to work in the exec()-ed
2552 * task, but being ignored by perf stat itself:
2553 */
2554 atexit(sig_atexit);
2555 if (!forever)
2556 signal(SIGINT, skip_signal);
2557 signal(SIGCHLD, skip_signal);
2558 signal(SIGALRM, skip_signal);
2559 signal(SIGABRT, skip_signal);
2560
2561 status = 0;
2562 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2563 if (run_count != 1 && verbose)
2564 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2565 run_idx + 1);
2566
2567 if (run_idx != 0)
2568 perf_evlist__reset_prev_raw_counts(evsel_list);
2569
2570 status = run_perf_stat(argc, argv);
2571 if (forever && status != -1 && !interval) {
2572 print_counters(NULL, argc, argv);
2573 perf_stat__reset_stats();
2574 }
2575 }
2576
2577 if (!forever && status != -1 && !interval)
2578 print_counters(NULL, argc, argv);
2579
2580 if (STAT_RECORD) {
2581 /*
2582 * We synthesize the kernel mmap record just so that older tools
2583 * don't emit warnings about not being able to resolve symbols
2584 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2585 * a saner message about no samples being in the perf.data file.
2586 *
2587 * This also serves to suppress a warning about f_header.data.size == 0
2588 * in header.c at the moment 'perf stat record' gets introduced, which
2589 * is not really needed once we start adding the stat specific PERF_RECORD_
2590 * records, but the need to suppress the kptr_restrict messages in older
2591 * tools remain -acme
2592 */
2593 int fd = perf_data_file__fd(&perf_stat.file);
2594 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2595 process_synthesized_event,
2596 &perf_stat.session->machines.host);
2597 if (err) {
2598 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2599 "older tools may produce warnings about this file\n.");
2600 }
2601
2602 if (!interval) {
2603 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2604 pr_err("failed to write stat round event\n");
2605 }
2606
2607 if (!perf_stat.file.is_pipe) {
2608 perf_stat.session->header.data_size += perf_stat.bytes_written;
2609 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2610 }
2611
2612 perf_session__delete(perf_stat.session);
2613 }
2614
2615 perf_stat__exit_aggr_mode();
2616 perf_evlist__free_stats(evsel_list);
2617 out:
2618 perf_evlist__delete(evsel_list);
2619 return status;
2620 }
Linux with added FPC-III support