?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
perf stat: Move perf_stat_synthesize_config() to stat.c
[linux/fpc-iii.git] / tools / perf / builtin-stat.c
blob54768ec15dbc3623706c7b398d8ad7aab4ce8c1b
1 /*
2 * builtin-stat.c
3 *
4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
6 *
7 * Sample output:
8
9 $ perf stat ./hackbench 10
10
11 Time: 0.118
12
13 Performance counter stats for './hackbench 10':
14
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
26
27 0.154822978 seconds time elapsed
28
29 *
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
31 *
32 * Improvements and fixes by:
33 *
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
40 *
41 * Released under the GPL v2. (and only v2, not any later version)
42 */
43
44 #include "perf.h"
45 #include "builtin.h"
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include <subcmd/parse-options.h>
49 #include "util/parse-events.h"
50 #include "util/pmu.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/drv_configs.h"
56 #include "util/color.h"
57 #include "util/stat.h"
58 #include "util/header.h"
59 #include "util/cpumap.h"
60 #include "util/thread.h"
61 #include "util/thread_map.h"
62 #include "util/counts.h"
63 #include "util/group.h"
64 #include "util/session.h"
65 #include "util/tool.h"
66 #include "util/string2.h"
67 #include "util/metricgroup.h"
68 #include "util/top.h"
69 #include "asm/bug.h"
70
71 #include <linux/time64.h>
72 #include <api/fs/fs.h>
73 #include <errno.h>
74 #include <signal.h>
75 #include <stdlib.h>
76 #include <sys/prctl.h>
77 #include <inttypes.h>
78 #include <locale.h>
79 #include <math.h>
80 #include <sys/types.h>
81 #include <sys/stat.h>
82 #include <sys/wait.h>
83 #include <unistd.h>
84 #include <sys/time.h>
85 #include <sys/resource.h>
86 #include <sys/wait.h>
87
88 #include "sane_ctype.h"
89
90 #define DEFAULT_SEPARATOR " "
91 #define CNTR_NOT_SUPPORTED "<not supported>"
92 #define CNTR_NOT_COUNTED "<not counted>"
93 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
94
95 static void print_counters(struct timespec *ts, int argc, const char **argv);
96
97 /* Default events used for perf stat -T */
98 static const char *transaction_attrs = {
99 "task-clock,"
100 "{"
101 "instructions,"
102 "cycles,"
103 "cpu/cycles-t/,"
104 "cpu/tx-start/,"
105 "cpu/el-start/,"
106 "cpu/cycles-ct/"
107 "}"
108 };
109
110 /* More limited version when the CPU does not have all events. */
111 static const char * transaction_limited_attrs = {
112 "task-clock,"
113 "{"
114 "instructions,"
115 "cycles,"
116 "cpu/cycles-t/,"
117 "cpu/tx-start/"
118 "}"
119 };
120
121 static const char * topdown_attrs[] = {
122 "topdown-total-slots",
123 "topdown-slots-retired",
124 "topdown-recovery-bubbles",
125 "topdown-fetch-bubbles",
126 "topdown-slots-issued",
127 NULL,
128 };
129
130 static const char *smi_cost_attrs = {
131 "{"
132 "msr/aperf/,"
133 "msr/smi/,"
134 "cycles"
135 "}"
136 };
137
138 static struct perf_evlist *evsel_list;
139
140 static struct rblist metric_events;
141
142 static struct target target = {
143 .uid = UINT_MAX,
144 };
145
146 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
147
148 #define METRIC_ONLY_LEN 20
149
150 static int run_count = 1;
151 static volatile pid_t child_pid = -1;
152 static bool null_run = false;
153 static int detailed_run = 0;
154 static bool transaction_run;
155 static bool topdown_run = false;
156 static bool smi_cost = false;
157 static bool smi_reset = false;
158 static bool big_num = true;
159 static int big_num_opt = -1;
160 static const char *csv_sep = NULL;
161 static bool csv_output = false;
162 static bool group = false;
163 static const char *pre_cmd = NULL;
164 static const char *post_cmd = NULL;
165 static bool sync_run = false;
166 static unsigned int unit_width = 4; /* strlen("unit") */
167 static bool forever = false;
168 static bool metric_only = false;
169 static bool force_metric_only = false;
170 static bool no_merge = false;
171 static bool walltime_run_table = false;
172 static struct timespec ref_time;
173 static struct cpu_map *aggr_map;
174 static aggr_get_id_t aggr_get_id;
175 static bool append_file;
176 static bool interval_count;
177 static bool interval_clear;
178 static const char *output_name;
179 static int output_fd;
180 static int print_free_counters_hint;
181 static int print_mixed_hw_group_error;
182 static u64 *walltime_run;
183 static bool ru_display = false;
184 static struct rusage ru_data;
185 static unsigned int metric_only_len = METRIC_ONLY_LEN;
186
187 struct perf_stat {
188 bool record;
189 struct perf_data data;
190 struct perf_session *session;
191 u64 bytes_written;
192 struct perf_tool tool;
193 bool maps_allocated;
194 struct cpu_map *cpus;
195 struct thread_map *threads;
196 enum aggr_mode aggr_mode;
197 };
198
199 static struct perf_stat perf_stat;
200 #define STAT_RECORD perf_stat.record
201
202 static volatile int done = 0;
203
204 static struct perf_stat_config stat_config = {
205 .aggr_mode = AGGR_GLOBAL,
206 .scale = true,
207 };
208
209 static bool is_duration_time(struct perf_evsel *evsel)
210 {
211 return !strcmp(evsel->name, "duration_time");
212 }
213
214 static inline void diff_timespec(struct timespec *r, struct timespec *a,
215 struct timespec *b)
216 {
217 r->tv_sec = a->tv_sec - b->tv_sec;
218 if (a->tv_nsec < b->tv_nsec) {
219 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
220 r->tv_sec--;
221 } else {
222 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
223 }
224 }
225
226 static void perf_stat__reset_stats(void)
227 {
228 int i;
229
230 perf_evlist__reset_stats(evsel_list);
231 perf_stat__reset_shadow_stats();
232
233 for (i = 0; i < stat_config.stats_num; i++)
234 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
235 }
236
237 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
238 union perf_event *event,
239 struct perf_sample *sample __maybe_unused,
240 struct machine *machine __maybe_unused)
241 {
242 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
243 pr_err("failed to write perf data, error: %m\n");
244 return -1;
245 }
246
247 perf_stat.bytes_written += event->header.size;
248 return 0;
249 }
250
251 static int write_stat_round_event(u64 tm, u64 type)
252 {
253 return perf_event__synthesize_stat_round(NULL, tm, type,
254 process_synthesized_event,
255 NULL);
256 }
257
258 #define WRITE_STAT_ROUND_EVENT(time, interval) \
259 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
260
261 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
262
263 static int
264 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
265 struct perf_counts_values *count)
266 {
267 struct perf_sample_id *sid = SID(counter, cpu, thread);
268
269 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
270 process_synthesized_event, NULL);
271 }
272
273 /*
274 * Read out the results of a single counter:
275 * do not aggregate counts across CPUs in system-wide mode
276 */
277 static int read_counter(struct perf_evsel *counter)
278 {
279 int nthreads = thread_map__nr(evsel_list->threads);
280 int ncpus, cpu, thread;
281
282 if (target__has_cpu(&target) && !target__has_per_thread(&target))
283 ncpus = perf_evsel__nr_cpus(counter);
284 else
285 ncpus = 1;
286
287 if (!counter->supported)
288 return -ENOENT;
289
290 if (counter->system_wide)
291 nthreads = 1;
292
293 for (thread = 0; thread < nthreads; thread++) {
294 for (cpu = 0; cpu < ncpus; cpu++) {
295 struct perf_counts_values *count;
296
297 count = perf_counts(counter->counts, cpu, thread);
298
299 /*
300 * The leader's group read loads data into its group members
301 * (via perf_evsel__read_counter) and sets threir count->loaded.
302 */
303 if (!count->loaded &&
304 perf_evsel__read_counter(counter, cpu, thread)) {
305 counter->counts->scaled = -1;
306 perf_counts(counter->counts, cpu, thread)->ena = 0;
307 perf_counts(counter->counts, cpu, thread)->run = 0;
308 return -1;
309 }
310
311 count->loaded = false;
312
313 if (STAT_RECORD) {
314 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
315 pr_err("failed to write stat event\n");
316 return -1;
317 }
318 }
319
320 if (verbose > 1) {
321 fprintf(stat_config.output,
322 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
323 perf_evsel__name(counter),
324 cpu,
325 count->val, count->ena, count->run);
326 }
327 }
328 }
329
330 return 0;
331 }
332
333 static void read_counters(void)
334 {
335 struct perf_evsel *counter;
336 int ret;
337
338 evlist__for_each_entry(evsel_list, counter) {
339 ret = read_counter(counter);
340 if (ret)
341 pr_debug("failed to read counter %s\n", counter->name);
342
343 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
344 pr_warning("failed to process counter %s\n", counter->name);
345 }
346 }
347
348 static void process_interval(void)
349 {
350 struct timespec ts, rs;
351
352 read_counters();
353
354 clock_gettime(CLOCK_MONOTONIC, &ts);
355 diff_timespec(&rs, &ts, &ref_time);
356
357 if (STAT_RECORD) {
358 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
359 pr_err("failed to write stat round event\n");
360 }
361
362 init_stats(&walltime_nsecs_stats);
363 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
364 print_counters(&rs, 0, NULL);
365 }
366
367 static void enable_counters(void)
368 {
369 if (stat_config.initial_delay)
370 usleep(stat_config.initial_delay * USEC_PER_MSEC);
371
372 /*
373 * We need to enable counters only if:
374 * - we don't have tracee (attaching to task or cpu)
375 * - we have initial delay configured
376 */
377 if (!target__none(&target) || stat_config.initial_delay)
378 perf_evlist__enable(evsel_list);
379 }
380
381 static void disable_counters(void)
382 {
383 /*
384 * If we don't have tracee (attaching to task or cpu), counters may
385 * still be running. To get accurate group ratios, we must stop groups
386 * from counting before reading their constituent counters.
387 */
388 if (!target__none(&target))
389 perf_evlist__disable(evsel_list);
390 }
391
392 static volatile int workload_exec_errno;
393
394 /*
395 * perf_evlist__prepare_workload will send a SIGUSR1
396 * if the fork fails, since we asked by setting its
397 * want_signal to true.
398 */
399 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
400 void *ucontext __maybe_unused)
401 {
402 workload_exec_errno = info->si_value.sival_int;
403 }
404
405 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
406 {
407 return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
408 }
409
410 static struct perf_evsel *perf_evsel__reset_weak_group(struct perf_evsel *evsel)
411 {
412 struct perf_evsel *c2, *leader;
413 bool is_open = true;
414
415 leader = evsel->leader;
416 pr_debug("Weak group for %s/%d failed\n",
417 leader->name, leader->nr_members);
418
419 /*
420 * for_each_group_member doesn't work here because it doesn't
421 * include the first entry.
422 */
423 evlist__for_each_entry(evsel_list, c2) {
424 if (c2 == evsel)
425 is_open = false;
426 if (c2->leader == leader) {
427 if (is_open)
428 perf_evsel__close(c2);
429 c2->leader = c2;
430 c2->nr_members = 0;
431 }
432 }
433 return leader;
434 }
435
436 static int __run_perf_stat(int argc, const char **argv, int run_idx)
437 {
438 int interval = stat_config.interval;
439 int times = stat_config.times;
440 int timeout = stat_config.timeout;
441 char msg[BUFSIZ];
442 unsigned long long t0, t1;
443 struct perf_evsel *counter;
444 struct timespec ts;
445 size_t l;
446 int status = 0;
447 const bool forks = (argc > 0);
448 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
449 struct perf_evsel_config_term *err_term;
450
451 if (interval) {
452 ts.tv_sec = interval / USEC_PER_MSEC;
453 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
454 } else if (timeout) {
455 ts.tv_sec = timeout / USEC_PER_MSEC;
456 ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
457 } else {
458 ts.tv_sec = 1;
459 ts.tv_nsec = 0;
460 }
461
462 if (forks) {
463 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
464 workload_exec_failed_signal) < 0) {
465 perror("failed to prepare workload");
466 return -1;
467 }
468 child_pid = evsel_list->workload.pid;
469 }
470
471 if (group)
472 perf_evlist__set_leader(evsel_list);
473
474 evlist__for_each_entry(evsel_list, counter) {
475 try_again:
476 if (create_perf_stat_counter(counter, &stat_config, &target) < 0) {
477
478 /* Weak group failed. Reset the group. */
479 if ((errno == EINVAL || errno == EBADF) &&
480 counter->leader != counter &&
481 counter->weak_group) {
482 counter = perf_evsel__reset_weak_group(counter);
483 goto try_again;
484 }
485
486 /*
487 * PPC returns ENXIO for HW counters until 2.6.37
488 * (behavior changed with commit b0a873e).
489 */
490 if (errno == EINVAL || errno == ENOSYS ||
491 errno == ENOENT || errno == EOPNOTSUPP ||
492 errno == ENXIO) {
493 if (verbose > 0)
494 ui__warning("%s event is not supported by the kernel.\n",
495 perf_evsel__name(counter));
496 counter->supported = false;
497
498 if ((counter->leader != counter) ||
499 !(counter->leader->nr_members > 1))
500 continue;
501 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
502 if (verbose > 0)
503 ui__warning("%s\n", msg);
504 goto try_again;
505 } else if (target__has_per_thread(&target) &&
506 evsel_list->threads &&
507 evsel_list->threads->err_thread != -1) {
508 /*
509 * For global --per-thread case, skip current
510 * error thread.
511 */
512 if (!thread_map__remove(evsel_list->threads,
513 evsel_list->threads->err_thread)) {
514 evsel_list->threads->err_thread = -1;
515 goto try_again;
516 }
517 }
518
519 perf_evsel__open_strerror(counter, &target,
520 errno, msg, sizeof(msg));
521 ui__error("%s\n", msg);
522
523 if (child_pid != -1)
524 kill(child_pid, SIGTERM);
525
526 return -1;
527 }
528 counter->supported = true;
529
530 l = strlen(counter->unit);
531 if (l > unit_width)
532 unit_width = l;
533
534 if (perf_evsel__should_store_id(counter) &&
535 perf_evsel__store_ids(counter, evsel_list))
536 return -1;
537 }
538
539 if (perf_evlist__apply_filters(evsel_list, &counter)) {
540 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
541 counter->filter, perf_evsel__name(counter), errno,
542 str_error_r(errno, msg, sizeof(msg)));
543 return -1;
544 }
545
546 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
547 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
548 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
549 str_error_r(errno, msg, sizeof(msg)));
550 return -1;
551 }
552
553 if (STAT_RECORD) {
554 int err, fd = perf_data__fd(&perf_stat.data);
555
556 if (is_pipe) {
557 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
558 } else {
559 err = perf_session__write_header(perf_stat.session, evsel_list,
560 fd, false);
561 }
562
563 if (err < 0)
564 return err;
565
566 err = perf_stat_synthesize_config(&stat_config, NULL, evsel_list,
567 process_synthesized_event, is_pipe);
568 if (err < 0)
569 return err;
570 }
571
572 /*
573 * Enable counters and exec the command:
574 */
575 t0 = rdclock();
576 clock_gettime(CLOCK_MONOTONIC, &ref_time);
577
578 if (forks) {
579 perf_evlist__start_workload(evsel_list);
580 enable_counters();
581
582 if (interval || timeout) {
583 while (!waitpid(child_pid, &status, WNOHANG)) {
584 nanosleep(&ts, NULL);
585 if (timeout)
586 break;
587 process_interval();
588 if (interval_count && !(--times))
589 break;
590 }
591 }
592 wait4(child_pid, &status, 0, &ru_data);
593
594 if (workload_exec_errno) {
595 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
596 pr_err("Workload failed: %s\n", emsg);
597 return -1;
598 }
599
600 if (WIFSIGNALED(status))
601 psignal(WTERMSIG(status), argv[0]);
602 } else {
603 enable_counters();
604 while (!done) {
605 nanosleep(&ts, NULL);
606 if (timeout)
607 break;
608 if (interval) {
609 process_interval();
610 if (interval_count && !(--times))
611 break;
612 }
613 }
614 }
615
616 disable_counters();
617
618 t1 = rdclock();
619
620 if (walltime_run_table)
621 walltime_run[run_idx] = t1 - t0;
622
623 update_stats(&walltime_nsecs_stats, t1 - t0);
624
625 /*
626 * Closing a group leader splits the group, and as we only disable
627 * group leaders, results in remaining events becoming enabled. To
628 * avoid arbitrary skew, we must read all counters before closing any
629 * group leaders.
630 */
631 read_counters();
632 perf_evlist__close(evsel_list);
633
634 return WEXITSTATUS(status);
635 }
636
637 static int run_perf_stat(int argc, const char **argv, int run_idx)
638 {
639 int ret;
640
641 if (pre_cmd) {
642 ret = system(pre_cmd);
643 if (ret)
644 return ret;
645 }
646
647 if (sync_run)
648 sync();
649
650 ret = __run_perf_stat(argc, argv, run_idx);
651 if (ret)
652 return ret;
653
654 if (post_cmd) {
655 ret = system(post_cmd);
656 if (ret)
657 return ret;
658 }
659
660 return ret;
661 }
662
663 static void print_running(u64 run, u64 ena)
664 {
665 if (csv_output) {
666 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
667 csv_sep,
668 run,
669 csv_sep,
670 ena ? 100.0 * run / ena : 100.0);
671 } else if (run != ena) {
672 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
673 }
674 }
675
676 static void print_noise_pct(double total, double avg)
677 {
678 double pct = rel_stddev_stats(total, avg);
679
680 if (csv_output)
681 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
682 else if (pct)
683 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
684 }
685
686 static void print_noise(struct perf_evsel *evsel, double avg)
687 {
688 struct perf_stat_evsel *ps;
689
690 if (run_count == 1)
691 return;
692
693 ps = evsel->stats;
694 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
695 }
696
697 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
698 {
699 switch (stat_config.aggr_mode) {
700 case AGGR_CORE:
701 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
702 cpu_map__id_to_socket(id),
703 csv_output ? 0 : -8,
704 cpu_map__id_to_cpu(id),
705 csv_sep,
706 csv_output ? 0 : 4,
707 nr,
708 csv_sep);
709 break;
710 case AGGR_SOCKET:
711 fprintf(stat_config.output, "S%*d%s%*d%s",
712 csv_output ? 0 : -5,
713 id,
714 csv_sep,
715 csv_output ? 0 : 4,
716 nr,
717 csv_sep);
718 break;
719 case AGGR_NONE:
720 fprintf(stat_config.output, "CPU%*d%s",
721 csv_output ? 0 : -4,
722 perf_evsel__cpus(evsel)->map[id], csv_sep);
723 break;
724 case AGGR_THREAD:
725 fprintf(stat_config.output, "%*s-%*d%s",
726 csv_output ? 0 : 16,
727 thread_map__comm(evsel->threads, id),
728 csv_output ? 0 : -8,
729 thread_map__pid(evsel->threads, id),
730 csv_sep);
731 break;
732 case AGGR_GLOBAL:
733 case AGGR_UNSET:
734 default:
735 break;
736 }
737 }
738
739 struct outstate {
740 FILE *fh;
741 bool newline;
742 const char *prefix;
743 int nfields;
744 int id, nr;
745 struct perf_evsel *evsel;
746 };
747
748 #define METRIC_LEN 35
749
750 static void new_line_std(void *ctx)
751 {
752 struct outstate *os = ctx;
753
754 os->newline = true;
755 }
756
757 static void do_new_line_std(struct outstate *os)
758 {
759 fputc('\n', os->fh);
760 fputs(os->prefix, os->fh);
761 aggr_printout(os->evsel, os->id, os->nr);
762 if (stat_config.aggr_mode == AGGR_NONE)
763 fprintf(os->fh, " ");
764 fprintf(os->fh, " ");
765 }
766
767 static void print_metric_std(void *ctx, const char *color, const char *fmt,
768 const char *unit, double val)
769 {
770 struct outstate *os = ctx;
771 FILE *out = os->fh;
772 int n;
773 bool newline = os->newline;
774
775 os->newline = false;
776
777 if (unit == NULL || fmt == NULL) {
778 fprintf(out, "%-*s", METRIC_LEN, "");
779 return;
780 }
781
782 if (newline)
783 do_new_line_std(os);
784
785 n = fprintf(out, " # ");
786 if (color)
787 n += color_fprintf(out, color, fmt, val);
788 else
789 n += fprintf(out, fmt, val);
790 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
791 }
792
793 static void new_line_csv(void *ctx)
794 {
795 struct outstate *os = ctx;
796 int i;
797
798 fputc('\n', os->fh);
799 if (os->prefix)
800 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
801 aggr_printout(os->evsel, os->id, os->nr);
802 for (i = 0; i < os->nfields; i++)
803 fputs(csv_sep, os->fh);
804 }
805
806 static void print_metric_csv(void *ctx,
807 const char *color __maybe_unused,
808 const char *fmt, const char *unit, double val)
809 {
810 struct outstate *os = ctx;
811 FILE *out = os->fh;
812 char buf[64], *vals, *ends;
813
814 if (unit == NULL || fmt == NULL) {
815 fprintf(out, "%s%s", csv_sep, csv_sep);
816 return;
817 }
818 snprintf(buf, sizeof(buf), fmt, val);
819 ends = vals = ltrim(buf);
820 while (isdigit(*ends) || *ends == '.')
821 ends++;
822 *ends = 0;
823 while (isspace(*unit))
824 unit++;
825 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
826 }
827
828 /* Filter out some columns that don't work well in metrics only mode */
829
830 static bool valid_only_metric(const char *unit)
831 {
832 if (!unit)
833 return false;
834 if (strstr(unit, "/sec") ||
835 strstr(unit, "hz") ||
836 strstr(unit, "Hz") ||
837 strstr(unit, "CPUs utilized"))
838 return false;
839 return true;
840 }
841
842 static const char *fixunit(char *buf, struct perf_evsel *evsel,
843 const char *unit)
844 {
845 if (!strncmp(unit, "of all", 6)) {
846 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
847 unit);
848 return buf;
849 }
850 return unit;
851 }
852
853 static void print_metric_only(void *ctx, const char *color, const char *fmt,
854 const char *unit, double val)
855 {
856 struct outstate *os = ctx;
857 FILE *out = os->fh;
858 char buf[1024], str[1024];
859 unsigned mlen = metric_only_len;
860
861 if (!valid_only_metric(unit))
862 return;
863 unit = fixunit(buf, os->evsel, unit);
864 if (mlen < strlen(unit))
865 mlen = strlen(unit) + 1;
866
867 if (color)
868 mlen += strlen(color) + sizeof(PERF_COLOR_RESET) - 1;
869
870 color_snprintf(str, sizeof(str), color ?: "", fmt, val);
871 fprintf(out, "%*s ", mlen, str);
872 }
873
874 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
875 const char *fmt,
876 const char *unit, double val)
877 {
878 struct outstate *os = ctx;
879 FILE *out = os->fh;
880 char buf[64], *vals, *ends;
881 char tbuf[1024];
882
883 if (!valid_only_metric(unit))
884 return;
885 unit = fixunit(tbuf, os->evsel, unit);
886 snprintf(buf, sizeof buf, fmt, val);
887 ends = vals = ltrim(buf);
888 while (isdigit(*ends) || *ends == '.')
889 ends++;
890 *ends = 0;
891 fprintf(out, "%s%s", vals, csv_sep);
892 }
893
894 static void new_line_metric(void *ctx __maybe_unused)
895 {
896 }
897
898 static void print_metric_header(void *ctx, const char *color __maybe_unused,
899 const char *fmt __maybe_unused,
900 const char *unit, double val __maybe_unused)
901 {
902 struct outstate *os = ctx;
903 char tbuf[1024];
904
905 if (!valid_only_metric(unit))
906 return;
907 unit = fixunit(tbuf, os->evsel, unit);
908 if (csv_output)
909 fprintf(os->fh, "%s%s", unit, csv_sep);
910 else
911 fprintf(os->fh, "%*s ", metric_only_len, unit);
912 }
913
914 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
915 {
916 int i;
917
918 if (!aggr_get_id)
919 return 0;
920
921 if (stat_config.aggr_mode == AGGR_NONE)
922 return id;
923
924 if (stat_config.aggr_mode == AGGR_GLOBAL)
925 return 0;
926
927 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
928 int cpu2 = perf_evsel__cpus(evsel)->map[i];
929
930 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
931 return cpu2;
932 }
933 return 0;
934 }
935
936 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
937 {
938 FILE *output = stat_config.output;
939 double sc = evsel->scale;
940 const char *fmt;
941
942 if (csv_output) {
943 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
944 } else {
945 if (big_num)
946 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
947 else
948 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
949 }
950
951 aggr_printout(evsel, id, nr);
952
953 fprintf(output, fmt, avg, csv_sep);
954
955 if (evsel->unit)
956 fprintf(output, "%-*s%s",
957 csv_output ? 0 : unit_width,
958 evsel->unit, csv_sep);
959
960 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
961
962 if (evsel->cgrp)
963 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
964 }
965
966 static bool is_mixed_hw_group(struct perf_evsel *counter)
967 {
968 struct perf_evlist *evlist = counter->evlist;
969 u32 pmu_type = counter->attr.type;
970 struct perf_evsel *pos;
971
972 if (counter->nr_members < 2)
973 return false;
974
975 evlist__for_each_entry(evlist, pos) {
976 /* software events can be part of any hardware group */
977 if (pos->attr.type == PERF_TYPE_SOFTWARE)
978 continue;
979 if (pmu_type == PERF_TYPE_SOFTWARE) {
980 pmu_type = pos->attr.type;
981 continue;
982 }
983 if (pmu_type != pos->attr.type)
984 return true;
985 }
986
987 return false;
988 }
989
990 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
991 char *prefix, u64 run, u64 ena, double noise,
992 struct runtime_stat *st)
993 {
994 struct perf_stat_output_ctx out;
995 struct outstate os = {
996 .fh = stat_config.output,
997 .prefix = prefix ? prefix : "",
998 .id = id,
999 .nr = nr,
1000 .evsel = counter,
1001 };
1002 print_metric_t pm = print_metric_std;
1003 void (*nl)(void *);
1004
1005 if (metric_only) {
1006 nl = new_line_metric;
1007 if (csv_output)
1008 pm = print_metric_only_csv;
1009 else
1010 pm = print_metric_only;
1011 } else
1012 nl = new_line_std;
1013
1014 if (csv_output && !metric_only) {
1015 static int aggr_fields[] = {
1016 [AGGR_GLOBAL] = 0,
1017 [AGGR_THREAD] = 1,
1018 [AGGR_NONE] = 1,
1019 [AGGR_SOCKET] = 2,
1020 [AGGR_CORE] = 2,
1021 };
1022
1023 pm = print_metric_csv;
1024 nl = new_line_csv;
1025 os.nfields = 3;
1026 os.nfields += aggr_fields[stat_config.aggr_mode];
1027 if (counter->cgrp)
1028 os.nfields++;
1029 }
1030 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1031 if (metric_only) {
1032 pm(&os, NULL, "", "", 0);
1033 return;
1034 }
1035 aggr_printout(counter, id, nr);
1036
1037 fprintf(stat_config.output, "%*s%s",
1038 csv_output ? 0 : 18,
1039 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1040 csv_sep);
1041
1042 if (counter->supported) {
1043 print_free_counters_hint = 1;
1044 if (is_mixed_hw_group(counter))
1045 print_mixed_hw_group_error = 1;
1046 }
1047
1048 fprintf(stat_config.output, "%-*s%s",
1049 csv_output ? 0 : unit_width,
1050 counter->unit, csv_sep);
1051
1052 fprintf(stat_config.output, "%*s",
1053 csv_output ? 0 : -25,
1054 perf_evsel__name(counter));
1055
1056 if (counter->cgrp)
1057 fprintf(stat_config.output, "%s%s",
1058 csv_sep, counter->cgrp->name);
1059
1060 if (!csv_output)
1061 pm(&os, NULL, NULL, "", 0);
1062 print_noise(counter, noise);
1063 print_running(run, ena);
1064 if (csv_output)
1065 pm(&os, NULL, NULL, "", 0);
1066 return;
1067 }
1068
1069 if (!metric_only)
1070 abs_printout(id, nr, counter, uval);
1071
1072 out.print_metric = pm;
1073 out.new_line = nl;
1074 out.ctx = &os;
1075 out.force_header = false;
1076
1077 if (csv_output && !metric_only) {
1078 print_noise(counter, noise);
1079 print_running(run, ena);
1080 }
1081
1082 perf_stat__print_shadow_stats(counter, uval,
1083 first_shadow_cpu(counter, id),
1084 &out, &metric_events, st);
1085 if (!csv_output && !metric_only) {
1086 print_noise(counter, noise);
1087 print_running(run, ena);
1088 }
1089 }
1090
1091 static void aggr_update_shadow(void)
1092 {
1093 int cpu, s2, id, s;
1094 u64 val;
1095 struct perf_evsel *counter;
1096
1097 for (s = 0; s < aggr_map->nr; s++) {
1098 id = aggr_map->map[s];
1099 evlist__for_each_entry(evsel_list, counter) {
1100 val = 0;
1101 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1102 s2 = aggr_get_id(evsel_list->cpus, cpu);
1103 if (s2 != id)
1104 continue;
1105 val += perf_counts(counter->counts, cpu, 0)->val;
1106 }
1107 perf_stat__update_shadow_stats(counter, val,
1108 first_shadow_cpu(counter, id),
1109 &rt_stat);
1110 }
1111 }
1112 }
1113
1114 static void uniquify_event_name(struct perf_evsel *counter)
1115 {
1116 char *new_name;
1117 char *config;
1118
1119 if (counter->uniquified_name ||
1120 !counter->pmu_name || !strncmp(counter->name, counter->pmu_name,
1121 strlen(counter->pmu_name)))
1122 return;
1123
1124 config = strchr(counter->name, '/');
1125 if (config) {
1126 if (asprintf(&new_name,
1127 "%s%s", counter->pmu_name, config) > 0) {
1128 free(counter->name);
1129 counter->name = new_name;
1130 }
1131 } else {
1132 if (asprintf(&new_name,
1133 "%s [%s]", counter->name, counter->pmu_name) > 0) {
1134 free(counter->name);
1135 counter->name = new_name;
1136 }
1137 }
1138
1139 counter->uniquified_name = true;
1140 }
1141
1142 static void collect_all_aliases(struct perf_evsel *counter,
1143 void (*cb)(struct perf_evsel *counter, void *data,
1144 bool first),
1145 void *data)
1146 {
1147 struct perf_evsel *alias;
1148
1149 alias = list_prepare_entry(counter, &(evsel_list->entries), node);
1150 list_for_each_entry_continue (alias, &evsel_list->entries, node) {
1151 if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) ||
1152 alias->scale != counter->scale ||
1153 alias->cgrp != counter->cgrp ||
1154 strcmp(alias->unit, counter->unit) ||
1155 perf_evsel__is_clock(alias) != perf_evsel__is_clock(counter))
1156 break;
1157 alias->merged_stat = true;
1158 cb(alias, data, false);
1159 }
1160 }
1161
1162 static bool collect_data(struct perf_evsel *counter,
1163 void (*cb)(struct perf_evsel *counter, void *data,
1164 bool first),
1165 void *data)
1166 {
1167 if (counter->merged_stat)
1168 return false;
1169 cb(counter, data, true);
1170 if (no_merge)
1171 uniquify_event_name(counter);
1172 else if (counter->auto_merge_stats)
1173 collect_all_aliases(counter, cb, data);
1174 return true;
1175 }
1176
1177 struct aggr_data {
1178 u64 ena, run, val;
1179 int id;
1180 int nr;
1181 int cpu;
1182 };
1183
1184 static void aggr_cb(struct perf_evsel *counter, void *data, bool first)
1185 {
1186 struct aggr_data *ad = data;
1187 int cpu, s2;
1188
1189 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1190 struct perf_counts_values *counts;
1191
1192 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1193 if (s2 != ad->id)
1194 continue;
1195 if (first)
1196 ad->nr++;
1197 counts = perf_counts(counter->counts, cpu, 0);
1198 /*
1199 * When any result is bad, make them all to give
1200 * consistent output in interval mode.
1201 */
1202 if (counts->ena == 0 || counts->run == 0 ||
1203 counter->counts->scaled == -1) {
1204 ad->ena = 0;
1205 ad->run = 0;
1206 break;
1207 }
1208 ad->val += counts->val;
1209 ad->ena += counts->ena;
1210 ad->run += counts->run;
1211 }
1212 }
1213
1214 static void print_aggr(char *prefix)
1215 {
1216 FILE *output = stat_config.output;
1217 struct perf_evsel *counter;
1218 int s, id, nr;
1219 double uval;
1220 u64 ena, run, val;
1221 bool first;
1222
1223 if (!(aggr_map || aggr_get_id))
1224 return;
1225
1226 aggr_update_shadow();
1227
1228 /*
1229 * With metric_only everything is on a single line.
1230 * Without each counter has its own line.
1231 */
1232 for (s = 0; s < aggr_map->nr; s++) {
1233 struct aggr_data ad;
1234 if (prefix && metric_only)
1235 fprintf(output, "%s", prefix);
1236
1237 ad.id = id = aggr_map->map[s];
1238 first = true;
1239 evlist__for_each_entry(evsel_list, counter) {
1240 if (is_duration_time(counter))
1241 continue;
1242
1243 ad.val = ad.ena = ad.run = 0;
1244 ad.nr = 0;
1245 if (!collect_data(counter, aggr_cb, &ad))
1246 continue;
1247 nr = ad.nr;
1248 ena = ad.ena;
1249 run = ad.run;
1250 val = ad.val;
1251 if (first && metric_only) {
1252 first = false;
1253 aggr_printout(counter, id, nr);
1254 }
1255 if (prefix && !metric_only)
1256 fprintf(output, "%s", prefix);
1257
1258 uval = val * counter->scale;
1259 printout(id, nr, counter, uval, prefix, run, ena, 1.0,
1260 &rt_stat);
1261 if (!metric_only)
1262 fputc('\n', output);
1263 }
1264 if (metric_only)
1265 fputc('\n', output);
1266 }
1267 }
1268
1269 static int cmp_val(const void *a, const void *b)
1270 {
1271 return ((struct perf_aggr_thread_value *)b)->val -
1272 ((struct perf_aggr_thread_value *)a)->val;
1273 }
1274
1275 static struct perf_aggr_thread_value *sort_aggr_thread(
1276 struct perf_evsel *counter,
1277 int nthreads, int ncpus,
1278 int *ret)
1279 {
1280 int cpu, thread, i = 0;
1281 double uval;
1282 struct perf_aggr_thread_value *buf;
1283
1284 buf = calloc(nthreads, sizeof(struct perf_aggr_thread_value));
1285 if (!buf)
1286 return NULL;
1287
1288 for (thread = 0; thread < nthreads; thread++) {
1289 u64 ena = 0, run = 0, val = 0;
1290
1291 for (cpu = 0; cpu < ncpus; cpu++) {
1292 val += perf_counts(counter->counts, cpu, thread)->val;
1293 ena += perf_counts(counter->counts, cpu, thread)->ena;
1294 run += perf_counts(counter->counts, cpu, thread)->run;
1295 }
1296
1297 uval = val * counter->scale;
1298
1299 /*
1300 * Skip value 0 when enabling --per-thread globally,
1301 * otherwise too many 0 output.
1302 */
1303 if (uval == 0.0 && target__has_per_thread(&target))
1304 continue;
1305
1306 buf[i].counter = counter;
1307 buf[i].id = thread;
1308 buf[i].uval = uval;
1309 buf[i].val = val;
1310 buf[i].run = run;
1311 buf[i].ena = ena;
1312 i++;
1313 }
1314
1315 qsort(buf, i, sizeof(struct perf_aggr_thread_value), cmp_val);
1316
1317 if (ret)
1318 *ret = i;
1319
1320 return buf;
1321 }
1322
1323 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1324 {
1325 FILE *output = stat_config.output;
1326 int nthreads = thread_map__nr(counter->threads);
1327 int ncpus = cpu_map__nr(counter->cpus);
1328 int thread, sorted_threads, id;
1329 struct perf_aggr_thread_value *buf;
1330
1331 buf = sort_aggr_thread(counter, nthreads, ncpus, &sorted_threads);
1332 if (!buf) {
1333 perror("cannot sort aggr thread");
1334 return;
1335 }
1336
1337 for (thread = 0; thread < sorted_threads; thread++) {
1338 if (prefix)
1339 fprintf(output, "%s", prefix);
1340
1341 id = buf[thread].id;
1342 if (stat_config.stats)
1343 printout(id, 0, buf[thread].counter, buf[thread].uval,
1344 prefix, buf[thread].run, buf[thread].ena, 1.0,
1345 &stat_config.stats[id]);
1346 else
1347 printout(id, 0, buf[thread].counter, buf[thread].uval,
1348 prefix, buf[thread].run, buf[thread].ena, 1.0,
1349 &rt_stat);
1350 fputc('\n', output);
1351 }
1352
1353 free(buf);
1354 }
1355
1356 struct caggr_data {
1357 double avg, avg_enabled, avg_running;
1358 };
1359
1360 static void counter_aggr_cb(struct perf_evsel *counter, void *data,
1361 bool first __maybe_unused)
1362 {
1363 struct caggr_data *cd = data;
1364 struct perf_stat_evsel *ps = counter->stats;
1365
1366 cd->avg += avg_stats(&ps->res_stats[0]);
1367 cd->avg_enabled += avg_stats(&ps->res_stats[1]);
1368 cd->avg_running += avg_stats(&ps->res_stats[2]);
1369 }
1370
1371 /*
1372 * Print out the results of a single counter:
1373 * aggregated counts in system-wide mode
1374 */
1375 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1376 {
1377 FILE *output = stat_config.output;
1378 double uval;
1379 struct caggr_data cd = { .avg = 0.0 };
1380
1381 if (!collect_data(counter, counter_aggr_cb, &cd))
1382 return;
1383
1384 if (prefix && !metric_only)
1385 fprintf(output, "%s", prefix);
1386
1387 uval = cd.avg * counter->scale;
1388 printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled,
1389 cd.avg, &rt_stat);
1390 if (!metric_only)
1391 fprintf(output, "\n");
1392 }
1393
1394 static void counter_cb(struct perf_evsel *counter, void *data,
1395 bool first __maybe_unused)
1396 {
1397 struct aggr_data *ad = data;
1398
1399 ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
1400 ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
1401 ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
1402 }
1403
1404 /*
1405 * Print out the results of a single counter:
1406 * does not use aggregated count in system-wide
1407 */
1408 static void print_counter(struct perf_evsel *counter, char *prefix)
1409 {
1410 FILE *output = stat_config.output;
1411 u64 ena, run, val;
1412 double uval;
1413 int cpu;
1414
1415 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1416 struct aggr_data ad = { .cpu = cpu };
1417
1418 if (!collect_data(counter, counter_cb, &ad))
1419 return;
1420 val = ad.val;
1421 ena = ad.ena;
1422 run = ad.run;
1423
1424 if (prefix)
1425 fprintf(output, "%s", prefix);
1426
1427 uval = val * counter->scale;
1428 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0,
1429 &rt_stat);
1430
1431 fputc('\n', output);
1432 }
1433 }
1434
1435 static void print_no_aggr_metric(char *prefix)
1436 {
1437 int cpu;
1438 int nrcpus = 0;
1439 struct perf_evsel *counter;
1440 u64 ena, run, val;
1441 double uval;
1442
1443 nrcpus = evsel_list->cpus->nr;
1444 for (cpu = 0; cpu < nrcpus; cpu++) {
1445 bool first = true;
1446
1447 if (prefix)
1448 fputs(prefix, stat_config.output);
1449 evlist__for_each_entry(evsel_list, counter) {
1450 if (is_duration_time(counter))
1451 continue;
1452 if (first) {
1453 aggr_printout(counter, cpu, 0);
1454 first = false;
1455 }
1456 val = perf_counts(counter->counts, cpu, 0)->val;
1457 ena = perf_counts(counter->counts, cpu, 0)->ena;
1458 run = perf_counts(counter->counts, cpu, 0)->run;
1459
1460 uval = val * counter->scale;
1461 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0,
1462 &rt_stat);
1463 }
1464 fputc('\n', stat_config.output);
1465 }
1466 }
1467
1468 static int aggr_header_lens[] = {
1469 [AGGR_CORE] = 18,
1470 [AGGR_SOCKET] = 12,
1471 [AGGR_NONE] = 6,
1472 [AGGR_THREAD] = 24,
1473 [AGGR_GLOBAL] = 0,
1474 };
1475
1476 static const char *aggr_header_csv[] = {
1477 [AGGR_CORE] = "core,cpus,",
1478 [AGGR_SOCKET] = "socket,cpus",
1479 [AGGR_NONE] = "cpu,",
1480 [AGGR_THREAD] = "comm-pid,",
1481 [AGGR_GLOBAL] = ""
1482 };
1483
1484 static void print_metric_headers(const char *prefix, bool no_indent)
1485 {
1486 struct perf_stat_output_ctx out;
1487 struct perf_evsel *counter;
1488 struct outstate os = {
1489 .fh = stat_config.output
1490 };
1491
1492 if (prefix)
1493 fprintf(stat_config.output, "%s", prefix);
1494
1495 if (!csv_output && !no_indent)
1496 fprintf(stat_config.output, "%*s",
1497 aggr_header_lens[stat_config.aggr_mode], "");
1498 if (csv_output) {
1499 if (stat_config.interval)
1500 fputs("time,", stat_config.output);
1501 fputs(aggr_header_csv[stat_config.aggr_mode],
1502 stat_config.output);
1503 }
1504
1505 /* Print metrics headers only */
1506 evlist__for_each_entry(evsel_list, counter) {
1507 if (is_duration_time(counter))
1508 continue;
1509 os.evsel = counter;
1510 out.ctx = &os;
1511 out.print_metric = print_metric_header;
1512 out.new_line = new_line_metric;
1513 out.force_header = true;
1514 os.evsel = counter;
1515 perf_stat__print_shadow_stats(counter, 0,
1516 0,
1517 &out,
1518 &metric_events,
1519 &rt_stat);
1520 }
1521 fputc('\n', stat_config.output);
1522 }
1523
1524 static void print_interval(char *prefix, struct timespec *ts)
1525 {
1526 FILE *output = stat_config.output;
1527 static int num_print_interval;
1528
1529 if (interval_clear)
1530 puts(CONSOLE_CLEAR);
1531
1532 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1533
1534 if ((num_print_interval == 0 && !csv_output) || interval_clear) {
1535 switch (stat_config.aggr_mode) {
1536 case AGGR_SOCKET:
1537 fprintf(output, "# time socket cpus");
1538 if (!metric_only)
1539 fprintf(output, " counts %*s events\n", unit_width, "unit");
1540 break;
1541 case AGGR_CORE:
1542 fprintf(output, "# time core cpus");
1543 if (!metric_only)
1544 fprintf(output, " counts %*s events\n", unit_width, "unit");
1545 break;
1546 case AGGR_NONE:
1547 fprintf(output, "# time CPU ");
1548 if (!metric_only)
1549 fprintf(output, " counts %*s events\n", unit_width, "unit");
1550 break;
1551 case AGGR_THREAD:
1552 fprintf(output, "# time comm-pid");
1553 if (!metric_only)
1554 fprintf(output, " counts %*s events\n", unit_width, "unit");
1555 break;
1556 case AGGR_GLOBAL:
1557 default:
1558 fprintf(output, "# time");
1559 if (!metric_only)
1560 fprintf(output, " counts %*s events\n", unit_width, "unit");
1561 case AGGR_UNSET:
1562 break;
1563 }
1564 }
1565
1566 if ((num_print_interval == 0 || interval_clear) && metric_only)
1567 print_metric_headers(" ", true);
1568 if (++num_print_interval == 25)
1569 num_print_interval = 0;
1570 }
1571
1572 static void print_header(int argc, const char **argv)
1573 {
1574 FILE *output = stat_config.output;
1575 int i;
1576
1577 fflush(stdout);
1578
1579 if (!csv_output) {
1580 fprintf(output, "\n");
1581 fprintf(output, " Performance counter stats for ");
1582 if (target.system_wide)
1583 fprintf(output, "\'system wide");
1584 else if (target.cpu_list)
1585 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1586 else if (!target__has_task(&target)) {
1587 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1588 for (i = 1; argv && (i < argc); i++)
1589 fprintf(output, " %s", argv[i]);
1590 } else if (target.pid)
1591 fprintf(output, "process id \'%s", target.pid);
1592 else
1593 fprintf(output, "thread id \'%s", target.tid);
1594
1595 fprintf(output, "\'");
1596 if (run_count > 1)
1597 fprintf(output, " (%d runs)", run_count);
1598 fprintf(output, ":\n\n");
1599 }
1600 }
1601
1602 static int get_precision(double num)
1603 {
1604 if (num > 1)
1605 return 0;
1606
1607 return lround(ceil(-log10(num)));
1608 }
1609
1610 static void print_table(FILE *output, int precision, double avg)
1611 {
1612 char tmp[64];
1613 int idx, indent = 0;
1614
1615 scnprintf(tmp, 64, " %17.*f", precision, avg);
1616 while (tmp[indent] == ' ')
1617 indent++;
1618
1619 fprintf(output, "%*s# Table of individual measurements:\n", indent, "");
1620
1621 for (idx = 0; idx < run_count; idx++) {
1622 double run = (double) walltime_run[idx] / NSEC_PER_SEC;
1623 int h, n = 1 + abs((int) (100.0 * (run - avg)/run) / 5);
1624
1625 fprintf(output, " %17.*f (%+.*f) ",
1626 precision, run, precision, run - avg);
1627
1628 for (h = 0; h < n; h++)
1629 fprintf(output, "#");
1630
1631 fprintf(output, "\n");
1632 }
1633
1634 fprintf(output, "\n%*s# Final result:\n", indent, "");
1635 }
1636
1637 static double timeval2double(struct timeval *t)
1638 {
1639 return t->tv_sec + (double) t->tv_usec/USEC_PER_SEC;
1640 }
1641
1642 static void print_footer(void)
1643 {
1644 double avg = avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC;
1645 FILE *output = stat_config.output;
1646 int n;
1647
1648 if (!null_run)
1649 fprintf(output, "\n");
1650
1651 if (run_count == 1) {
1652 fprintf(output, " %17.9f seconds time elapsed", avg);
1653
1654 if (ru_display) {
1655 double ru_utime = timeval2double(&ru_data.ru_utime);
1656 double ru_stime = timeval2double(&ru_data.ru_stime);
1657
1658 fprintf(output, "\n\n");
1659 fprintf(output, " %17.9f seconds user\n", ru_utime);
1660 fprintf(output, " %17.9f seconds sys\n", ru_stime);
1661 }
1662 } else {
1663 double sd = stddev_stats(&walltime_nsecs_stats) / NSEC_PER_SEC;
1664 /*
1665 * Display at most 2 more significant
1666 * digits than the stddev inaccuracy.
1667 */
1668 int precision = get_precision(sd) + 2;
1669
1670 if (walltime_run_table)
1671 print_table(output, precision, avg);
1672
1673 fprintf(output, " %17.*f +- %.*f seconds time elapsed",
1674 precision, avg, precision, sd);
1675
1676 print_noise_pct(sd, avg);
1677 }
1678 fprintf(output, "\n\n");
1679
1680 if (print_free_counters_hint &&
1681 sysctl__read_int("kernel/nmi_watchdog", &n) >= 0 &&
1682 n > 0)
1683 fprintf(output,
1684 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1685 " echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1686 " perf stat ...\n"
1687 " echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1688
1689 if (print_mixed_hw_group_error)
1690 fprintf(output,
1691 "The events in group usually have to be from "
1692 "the same PMU. Try reorganizing the group.\n");
1693 }
1694
1695 static void print_counters(struct timespec *ts, int argc, const char **argv)
1696 {
1697 int interval = stat_config.interval;
1698 struct perf_evsel *counter;
1699 char buf[64], *prefix = NULL;
1700
1701 /* Do not print anything if we record to the pipe. */
1702 if (STAT_RECORD && perf_stat.data.is_pipe)
1703 return;
1704
1705 if (interval)
1706 print_interval(prefix = buf, ts);
1707 else
1708 print_header(argc, argv);
1709
1710 if (metric_only) {
1711 static int num_print_iv;
1712
1713 if (num_print_iv == 0 && !interval)
1714 print_metric_headers(prefix, false);
1715 if (num_print_iv++ == 25)
1716 num_print_iv = 0;
1717 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1718 fprintf(stat_config.output, "%s", prefix);
1719 }
1720
1721 switch (stat_config.aggr_mode) {
1722 case AGGR_CORE:
1723 case AGGR_SOCKET:
1724 print_aggr(prefix);
1725 break;
1726 case AGGR_THREAD:
1727 evlist__for_each_entry(evsel_list, counter) {
1728 if (is_duration_time(counter))
1729 continue;
1730 print_aggr_thread(counter, prefix);
1731 }
1732 break;
1733 case AGGR_GLOBAL:
1734 evlist__for_each_entry(evsel_list, counter) {
1735 if (is_duration_time(counter))
1736 continue;
1737 print_counter_aggr(counter, prefix);
1738 }
1739 if (metric_only)
1740 fputc('\n', stat_config.output);
1741 break;
1742 case AGGR_NONE:
1743 if (metric_only)
1744 print_no_aggr_metric(prefix);
1745 else {
1746 evlist__for_each_entry(evsel_list, counter) {
1747 if (is_duration_time(counter))
1748 continue;
1749 print_counter(counter, prefix);
1750 }
1751 }
1752 break;
1753 case AGGR_UNSET:
1754 default:
1755 break;
1756 }
1757
1758 if (!interval && !csv_output)
1759 print_footer();
1760
1761 fflush(stat_config.output);
1762 }
1763
1764 static volatile int signr = -1;
1765
1766 static void skip_signal(int signo)
1767 {
1768 if ((child_pid == -1) || stat_config.interval)
1769 done = 1;
1770
1771 signr = signo;
1772 /*
1773 * render child_pid harmless
1774 * won't send SIGTERM to a random
1775 * process in case of race condition
1776 * and fast PID recycling
1777 */
1778 child_pid = -1;
1779 }
1780
1781 static void sig_atexit(void)
1782 {
1783 sigset_t set, oset;
1784
1785 /*
1786 * avoid race condition with SIGCHLD handler
1787 * in skip_signal() which is modifying child_pid
1788 * goal is to avoid send SIGTERM to a random
1789 * process
1790 */
1791 sigemptyset(&set);
1792 sigaddset(&set, SIGCHLD);
1793 sigprocmask(SIG_BLOCK, &set, &oset);
1794
1795 if (child_pid != -1)
1796 kill(child_pid, SIGTERM);
1797
1798 sigprocmask(SIG_SETMASK, &oset, NULL);
1799
1800 if (signr == -1)
1801 return;
1802
1803 signal(signr, SIG_DFL);
1804 kill(getpid(), signr);
1805 }
1806
1807 static int stat__set_big_num(const struct option *opt __maybe_unused,
1808 const char *s __maybe_unused, int unset)
1809 {
1810 big_num_opt = unset ? 0 : 1;
1811 return 0;
1812 }
1813
1814 static int enable_metric_only(const struct option *opt __maybe_unused,
1815 const char *s __maybe_unused, int unset)
1816 {
1817 force_metric_only = true;
1818 metric_only = !unset;
1819 return 0;
1820 }
1821
1822 static int parse_metric_groups(const struct option *opt,
1823 const char *str,
1824 int unset __maybe_unused)
1825 {
1826 return metricgroup__parse_groups(opt, str, &metric_events);
1827 }
1828
1829 static const struct option stat_options[] = {
1830 OPT_BOOLEAN('T', "transaction", &transaction_run,
1831 "hardware transaction statistics"),
1832 OPT_CALLBACK('e', "event", &evsel_list, "event",
1833 "event selector. use 'perf list' to list available events",
1834 parse_events_option),
1835 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1836 "event filter", parse_filter),
1837 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1838 "child tasks do not inherit counters"),
1839 OPT_STRING('p', "pid", &target.pid, "pid",
1840 "stat events on existing process id"),
1841 OPT_STRING('t', "tid", &target.tid, "tid",
1842 "stat events on existing thread id"),
1843 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1844 "system-wide collection from all CPUs"),
1845 OPT_BOOLEAN('g', "group", &group,
1846 "put the counters into a counter group"),
1847 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1848 OPT_INCR('v', "verbose", &verbose,
1849 "be more verbose (show counter open errors, etc)"),
1850 OPT_INTEGER('r', "repeat", &run_count,
1851 "repeat command and print average + stddev (max: 100, forever: 0)"),
1852 OPT_BOOLEAN(0, "table", &walltime_run_table,
1853 "display details about each run (only with -r option)"),
1854 OPT_BOOLEAN('n', "null", &null_run,
1855 "null run - dont start any counters"),
1856 OPT_INCR('d', "detailed", &detailed_run,
1857 "detailed run - start a lot of events"),
1858 OPT_BOOLEAN('S', "sync", &sync_run,
1859 "call sync() before starting a run"),
1860 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1861 "print large numbers with thousands\' separators",
1862 stat__set_big_num),
1863 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1864 "list of cpus to monitor in system-wide"),
1865 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1866 "disable CPU count aggregation", AGGR_NONE),
1867 OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"),
1868 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1869 "print counts with custom separator"),
1870 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1871 "monitor event in cgroup name only", parse_cgroups),
1872 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1873 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1874 OPT_INTEGER(0, "log-fd", &output_fd,
1875 "log output to fd, instead of stderr"),
1876 OPT_STRING(0, "pre", &pre_cmd, "command",
1877 "command to run prior to the measured command"),
1878 OPT_STRING(0, "post", &post_cmd, "command",
1879 "command to run after to the measured command"),
1880 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1881 "print counts at regular interval in ms "
1882 "(overhead is possible for values <= 100ms)"),
1883 OPT_INTEGER(0, "interval-count", &stat_config.times,
1884 "print counts for fixed number of times"),
1885 OPT_BOOLEAN(0, "interval-clear", &interval_clear,
1886 "clear screen in between new interval"),
1887 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1888 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1889 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1890 "aggregate counts per processor socket", AGGR_SOCKET),
1891 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1892 "aggregate counts per physical processor core", AGGR_CORE),
1893 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1894 "aggregate counts per thread", AGGR_THREAD),
1895 OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
1896 "ms to wait before starting measurement after program start"),
1897 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1898 "Only print computed metrics. No raw values", enable_metric_only),
1899 OPT_BOOLEAN(0, "topdown", &topdown_run,
1900 "measure topdown level 1 statistics"),
1901 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1902 "measure SMI cost"),
1903 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1904 "monitor specified metrics or metric groups (separated by ,)",
1905 parse_metric_groups),
1906 OPT_END()
1907 };
1908
1909 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1910 {
1911 return cpu_map__get_socket(map, cpu, NULL);
1912 }
1913
1914 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1915 {
1916 return cpu_map__get_core(map, cpu, NULL);
1917 }
1918
1919 static int cpu_map__get_max(struct cpu_map *map)
1920 {
1921 int i, max = -1;
1922
1923 for (i = 0; i < map->nr; i++) {
1924 if (map->map[i] > max)
1925 max = map->map[i];
1926 }
1927
1928 return max;
1929 }
1930
1931 static struct cpu_map *cpus_aggr_map;
1932
1933 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1934 {
1935 int cpu;
1936
1937 if (idx >= map->nr)
1938 return -1;
1939
1940 cpu = map->map[idx];
1941
1942 if (cpus_aggr_map->map[cpu] == -1)
1943 cpus_aggr_map->map[cpu] = get_id(map, idx);
1944
1945 return cpus_aggr_map->map[cpu];
1946 }
1947
1948 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1949 {
1950 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1951 }
1952
1953 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1954 {
1955 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1956 }
1957
1958 static int perf_stat_init_aggr_mode(void)
1959 {
1960 int nr;
1961
1962 switch (stat_config.aggr_mode) {
1963 case AGGR_SOCKET:
1964 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1965 perror("cannot build socket map");
1966 return -1;
1967 }
1968 aggr_get_id = perf_stat__get_socket_cached;
1969 break;
1970 case AGGR_CORE:
1971 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1972 perror("cannot build core map");
1973 return -1;
1974 }
1975 aggr_get_id = perf_stat__get_core_cached;
1976 break;
1977 case AGGR_NONE:
1978 case AGGR_GLOBAL:
1979 case AGGR_THREAD:
1980 case AGGR_UNSET:
1981 default:
1982 break;
1983 }
1984
1985 /*
1986 * The evsel_list->cpus is the base we operate on,
1987 * taking the highest cpu number to be the size of
1988 * the aggregation translate cpumap.
1989 */
1990 nr = cpu_map__get_max(evsel_list->cpus);
1991 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1992 return cpus_aggr_map ? 0 : -ENOMEM;
1993 }
1994
1995 static void perf_stat__exit_aggr_mode(void)
1996 {
1997 cpu_map__put(aggr_map);
1998 cpu_map__put(cpus_aggr_map);
1999 aggr_map = NULL;
2000 cpus_aggr_map = NULL;
2001 }
2002
2003 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
2004 {
2005 int cpu;
2006
2007 if (idx > map->nr)
2008 return -1;
2009
2010 cpu = map->map[idx];
2011
2012 if (cpu >= env->nr_cpus_avail)
2013 return -1;
2014
2015 return cpu;
2016 }
2017
2018 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
2019 {
2020 struct perf_env *env = data;
2021 int cpu = perf_env__get_cpu(env, map, idx);
2022
2023 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
2024 }
2025
2026 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
2027 {
2028 struct perf_env *env = data;
2029 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
2030
2031 if (cpu != -1) {
2032 int socket_id = env->cpu[cpu].socket_id;
2033
2034 /*
2035 * Encode socket in upper 16 bits
2036 * core_id is relative to socket, and
2037 * we need a global id. So we combine
2038 * socket + core id.
2039 */
2040 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
2041 }
2042
2043 return core;
2044 }
2045
2046 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
2047 struct cpu_map **sockp)
2048 {
2049 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
2050 }
2051
2052 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
2053 struct cpu_map **corep)
2054 {
2055 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
2056 }
2057
2058 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
2059 {
2060 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
2061 }
2062
2063 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
2064 {
2065 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
2066 }
2067
2068 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
2069 {
2070 struct perf_env *env = &st->session->header.env;
2071
2072 switch (stat_config.aggr_mode) {
2073 case AGGR_SOCKET:
2074 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
2075 perror("cannot build socket map");
2076 return -1;
2077 }
2078 aggr_get_id = perf_stat__get_socket_file;
2079 break;
2080 case AGGR_CORE:
2081 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
2082 perror("cannot build core map");
2083 return -1;
2084 }
2085 aggr_get_id = perf_stat__get_core_file;
2086 break;
2087 case AGGR_NONE:
2088 case AGGR_GLOBAL:
2089 case AGGR_THREAD:
2090 case AGGR_UNSET:
2091 default:
2092 break;
2093 }
2094
2095 return 0;
2096 }
2097
2098 static int topdown_filter_events(const char **attr, char **str, bool use_group)
2099 {
2100 int off = 0;
2101 int i;
2102 int len = 0;
2103 char *s;
2104
2105 for (i = 0; attr[i]; i++) {
2106 if (pmu_have_event("cpu", attr[i])) {
2107 len += strlen(attr[i]) + 1;
2108 attr[i - off] = attr[i];
2109 } else
2110 off++;
2111 }
2112 attr[i - off] = NULL;
2113
2114 *str = malloc(len + 1 + 2);
2115 if (!*str)
2116 return -1;
2117 s = *str;
2118 if (i - off == 0) {
2119 *s = 0;
2120 return 0;
2121 }
2122 if (use_group)
2123 *s++ = '{';
2124 for (i = 0; attr[i]; i++) {
2125 strcpy(s, attr[i]);
2126 s += strlen(s);
2127 *s++ = ',';
2128 }
2129 if (use_group) {
2130 s[-1] = '}';
2131 *s = 0;
2132 } else
2133 s[-1] = 0;
2134 return 0;
2135 }
2136
2137 __weak bool arch_topdown_check_group(bool *warn)
2138 {
2139 *warn = false;
2140 return false;
2141 }
2142
2143 __weak void arch_topdown_group_warn(void)
2144 {
2145 }
2146
2147 /*
2148 * Add default attributes, if there were no attributes specified or
2149 * if -d/--detailed, -d -d or -d -d -d is used:
2150 */
2151 static int add_default_attributes(void)
2152 {
2153 int err;
2154 struct perf_event_attr default_attrs0[] = {
2155
2156 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
2157 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
2158 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
2159 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
2160
2161 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
2162 };
2163 struct perf_event_attr frontend_attrs[] = {
2164 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
2165 };
2166 struct perf_event_attr backend_attrs[] = {
2167 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
2168 };
2169 struct perf_event_attr default_attrs1[] = {
2170 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
2171 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
2172 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
2173
2174 };
2175
2176 /*
2177 * Detailed stats (-d), covering the L1 and last level data caches:
2178 */
2179 struct perf_event_attr detailed_attrs[] = {
2180
2181 { .type = PERF_TYPE_HW_CACHE,
2182 .config =
2183 PERF_COUNT_HW_CACHE_L1D << 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_L1D << 0 |
2190 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2191 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2192
2193 { .type = PERF_TYPE_HW_CACHE,
2194 .config =
2195 PERF_COUNT_HW_CACHE_LL << 0 |
2196 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2197 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2198
2199 { .type = PERF_TYPE_HW_CACHE,
2200 .config =
2201 PERF_COUNT_HW_CACHE_LL << 0 |
2202 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2203 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2204 };
2205
2206 /*
2207 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2208 */
2209 struct perf_event_attr very_detailed_attrs[] = {
2210
2211 { .type = PERF_TYPE_HW_CACHE,
2212 .config =
2213 PERF_COUNT_HW_CACHE_L1I << 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_L1I << 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_DTLB << 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_DTLB << 0 |
2232 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2233 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2234
2235 { .type = PERF_TYPE_HW_CACHE,
2236 .config =
2237 PERF_COUNT_HW_CACHE_ITLB << 0 |
2238 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2239 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2240
2241 { .type = PERF_TYPE_HW_CACHE,
2242 .config =
2243 PERF_COUNT_HW_CACHE_ITLB << 0 |
2244 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2245 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2246
2247 };
2248
2249 /*
2250 * Very, very detailed stats (-d -d -d), adding prefetch events:
2251 */
2252 struct perf_event_attr very_very_detailed_attrs[] = {
2253
2254 { .type = PERF_TYPE_HW_CACHE,
2255 .config =
2256 PERF_COUNT_HW_CACHE_L1D << 0 |
2257 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2258 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2259
2260 { .type = PERF_TYPE_HW_CACHE,
2261 .config =
2262 PERF_COUNT_HW_CACHE_L1D << 0 |
2263 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2264 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2265 };
2266 struct parse_events_error errinfo;
2267
2268 /* Set attrs if no event is selected and !null_run: */
2269 if (null_run)
2270 return 0;
2271
2272 if (transaction_run) {
2273 /* Handle -T as -M transaction. Once platform specific metrics
2274 * support has been added to the json files, all archictures
2275 * will use this approach. To determine transaction support
2276 * on an architecture test for such a metric name.
2277 */
2278 if (metricgroup__has_metric("transaction")) {
2279 struct option opt = { .value = &evsel_list };
2280
2281 return metricgroup__parse_groups(&opt, "transaction",
2282 &metric_events);
2283 }
2284
2285 if (pmu_have_event("cpu", "cycles-ct") &&
2286 pmu_have_event("cpu", "el-start"))
2287 err = parse_events(evsel_list, transaction_attrs,
2288 &errinfo);
2289 else
2290 err = parse_events(evsel_list,
2291 transaction_limited_attrs,
2292 &errinfo);
2293 if (err) {
2294 fprintf(stderr, "Cannot set up transaction events\n");
2295 parse_events_print_error(&errinfo, transaction_attrs);
2296 return -1;
2297 }
2298 return 0;
2299 }
2300
2301 if (smi_cost) {
2302 int smi;
2303
2304 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2305 fprintf(stderr, "freeze_on_smi is not supported.\n");
2306 return -1;
2307 }
2308
2309 if (!smi) {
2310 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2311 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2312 return -1;
2313 }
2314 smi_reset = true;
2315 }
2316
2317 if (pmu_have_event("msr", "aperf") &&
2318 pmu_have_event("msr", "smi")) {
2319 if (!force_metric_only)
2320 metric_only = true;
2321 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
2322 } else {
2323 fprintf(stderr, "To measure SMI cost, it needs "
2324 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
2325 parse_events_print_error(&errinfo, smi_cost_attrs);
2326 return -1;
2327 }
2328 if (err) {
2329 fprintf(stderr, "Cannot set up SMI cost events\n");
2330 return -1;
2331 }
2332 return 0;
2333 }
2334
2335 if (topdown_run) {
2336 char *str = NULL;
2337 bool warn = false;
2338
2339 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2340 stat_config.aggr_mode != AGGR_CORE) {
2341 pr_err("top down event configuration requires --per-core mode\n");
2342 return -1;
2343 }
2344 stat_config.aggr_mode = AGGR_CORE;
2345 if (nr_cgroups || !target__has_cpu(&target)) {
2346 pr_err("top down event configuration requires system-wide mode (-a)\n");
2347 return -1;
2348 }
2349
2350 if (!force_metric_only)
2351 metric_only = true;
2352 if (topdown_filter_events(topdown_attrs, &str,
2353 arch_topdown_check_group(&warn)) < 0) {
2354 pr_err("Out of memory\n");
2355 return -1;
2356 }
2357 if (topdown_attrs[0] && str) {
2358 if (warn)
2359 arch_topdown_group_warn();
2360 err = parse_events(evsel_list, str, &errinfo);
2361 if (err) {
2362 fprintf(stderr,
2363 "Cannot set up top down events %s: %d\n",
2364 str, err);
2365 free(str);
2366 parse_events_print_error(&errinfo, str);
2367 return -1;
2368 }
2369 } else {
2370 fprintf(stderr, "System does not support topdown\n");
2371 return -1;
2372 }
2373 free(str);
2374 }
2375
2376 if (!evsel_list->nr_entries) {
2377 if (target__has_cpu(&target))
2378 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2379
2380 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2381 return -1;
2382 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2383 if (perf_evlist__add_default_attrs(evsel_list,
2384 frontend_attrs) < 0)
2385 return -1;
2386 }
2387 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2388 if (perf_evlist__add_default_attrs(evsel_list,
2389 backend_attrs) < 0)
2390 return -1;
2391 }
2392 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2393 return -1;
2394 }
2395
2396 /* Detailed events get appended to the event list: */
2397
2398 if (detailed_run < 1)
2399 return 0;
2400
2401 /* Append detailed run extra attributes: */
2402 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2403 return -1;
2404
2405 if (detailed_run < 2)
2406 return 0;
2407
2408 /* Append very detailed run extra attributes: */
2409 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2410 return -1;
2411
2412 if (detailed_run < 3)
2413 return 0;
2414
2415 /* Append very, very detailed run extra attributes: */
2416 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2417 }
2418
2419 static const char * const stat_record_usage[] = {
2420 "perf stat record [<options>]",
2421 NULL,
2422 };
2423
2424 static void init_features(struct perf_session *session)
2425 {
2426 int feat;
2427
2428 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2429 perf_header__set_feat(&session->header, feat);
2430
2431 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2432 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2433 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2434 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2435 }
2436
2437 static int __cmd_record(int argc, const char **argv)
2438 {
2439 struct perf_session *session;
2440 struct perf_data *data = &perf_stat.data;
2441
2442 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2443 PARSE_OPT_STOP_AT_NON_OPTION);
2444
2445 if (output_name)
2446 data->file.path = output_name;
2447
2448 if (run_count != 1 || forever) {
2449 pr_err("Cannot use -r option with perf stat record.\n");
2450 return -1;
2451 }
2452
2453 session = perf_session__new(data, false, NULL);
2454 if (session == NULL) {
2455 pr_err("Perf session creation failed.\n");
2456 return -1;
2457 }
2458
2459 init_features(session);
2460
2461 session->evlist = evsel_list;
2462 perf_stat.session = session;
2463 perf_stat.record = true;
2464 return argc;
2465 }
2466
2467 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2468 union perf_event *event,
2469 struct perf_session *session)
2470 {
2471 struct stat_round_event *stat_round = &event->stat_round;
2472 struct perf_evsel *counter;
2473 struct timespec tsh, *ts = NULL;
2474 const char **argv = session->header.env.cmdline_argv;
2475 int argc = session->header.env.nr_cmdline;
2476
2477 evlist__for_each_entry(evsel_list, counter)
2478 perf_stat_process_counter(&stat_config, counter);
2479
2480 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2481 update_stats(&walltime_nsecs_stats, stat_round->time);
2482
2483 if (stat_config.interval && stat_round->time) {
2484 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2485 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2486 ts = &tsh;
2487 }
2488
2489 print_counters(ts, argc, argv);
2490 return 0;
2491 }
2492
2493 static
2494 int process_stat_config_event(struct perf_tool *tool,
2495 union perf_event *event,
2496 struct perf_session *session __maybe_unused)
2497 {
2498 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2499
2500 perf_event__read_stat_config(&stat_config, &event->stat_config);
2501
2502 if (cpu_map__empty(st->cpus)) {
2503 if (st->aggr_mode != AGGR_UNSET)
2504 pr_warning("warning: processing task data, aggregation mode not set\n");
2505 return 0;
2506 }
2507
2508 if (st->aggr_mode != AGGR_UNSET)
2509 stat_config.aggr_mode = st->aggr_mode;
2510
2511 if (perf_stat.data.is_pipe)
2512 perf_stat_init_aggr_mode();
2513 else
2514 perf_stat_init_aggr_mode_file(st);
2515
2516 return 0;
2517 }
2518
2519 static int set_maps(struct perf_stat *st)
2520 {
2521 if (!st->cpus || !st->threads)
2522 return 0;
2523
2524 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2525 return -EINVAL;
2526
2527 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2528
2529 if (perf_evlist__alloc_stats(evsel_list, true))
2530 return -ENOMEM;
2531
2532 st->maps_allocated = true;
2533 return 0;
2534 }
2535
2536 static
2537 int process_thread_map_event(struct perf_tool *tool,
2538 union perf_event *event,
2539 struct perf_session *session __maybe_unused)
2540 {
2541 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2542
2543 if (st->threads) {
2544 pr_warning("Extra thread map event, ignoring.\n");
2545 return 0;
2546 }
2547
2548 st->threads = thread_map__new_event(&event->thread_map);
2549 if (!st->threads)
2550 return -ENOMEM;
2551
2552 return set_maps(st);
2553 }
2554
2555 static
2556 int process_cpu_map_event(struct perf_tool *tool,
2557 union perf_event *event,
2558 struct perf_session *session __maybe_unused)
2559 {
2560 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2561 struct cpu_map *cpus;
2562
2563 if (st->cpus) {
2564 pr_warning("Extra cpu map event, ignoring.\n");
2565 return 0;
2566 }
2567
2568 cpus = cpu_map__new_data(&event->cpu_map.data);
2569 if (!cpus)
2570 return -ENOMEM;
2571
2572 st->cpus = cpus;
2573 return set_maps(st);
2574 }
2575
2576 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
2577 {
2578 int i;
2579
2580 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
2581 if (!config->stats)
2582 return -1;
2583
2584 config->stats_num = nthreads;
2585
2586 for (i = 0; i < nthreads; i++)
2587 runtime_stat__init(&config->stats[i]);
2588
2589 return 0;
2590 }
2591
2592 static void runtime_stat_delete(struct perf_stat_config *config)
2593 {
2594 int i;
2595
2596 if (!config->stats)
2597 return;
2598
2599 for (i = 0; i < config->stats_num; i++)
2600 runtime_stat__exit(&config->stats[i]);
2601
2602 free(config->stats);
2603 }
2604
2605 static const char * const stat_report_usage[] = {
2606 "perf stat report [<options>]",
2607 NULL,
2608 };
2609
2610 static struct perf_stat perf_stat = {
2611 .tool = {
2612 .attr = perf_event__process_attr,
2613 .event_update = perf_event__process_event_update,
2614 .thread_map = process_thread_map_event,
2615 .cpu_map = process_cpu_map_event,
2616 .stat_config = process_stat_config_event,
2617 .stat = perf_event__process_stat_event,
2618 .stat_round = process_stat_round_event,
2619 },
2620 .aggr_mode = AGGR_UNSET,
2621 };
2622
2623 static int __cmd_report(int argc, const char **argv)
2624 {
2625 struct perf_session *session;
2626 const struct option options[] = {
2627 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2628 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2629 "aggregate counts per processor socket", AGGR_SOCKET),
2630 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2631 "aggregate counts per physical processor core", AGGR_CORE),
2632 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2633 "disable CPU count aggregation", AGGR_NONE),
2634 OPT_END()
2635 };
2636 struct stat st;
2637 int ret;
2638
2639 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2640
2641 if (!input_name || !strlen(input_name)) {
2642 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2643 input_name = "-";
2644 else
2645 input_name = "perf.data";
2646 }
2647
2648 perf_stat.data.file.path = input_name;
2649 perf_stat.data.mode = PERF_DATA_MODE_READ;
2650
2651 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2652 if (session == NULL)
2653 return -1;
2654
2655 perf_stat.session = session;
2656 stat_config.output = stderr;
2657 evsel_list = session->evlist;
2658
2659 ret = perf_session__process_events(session);
2660 if (ret)
2661 return ret;
2662
2663 perf_session__delete(session);
2664 return 0;
2665 }
2666
2667 static void setup_system_wide(int forks)
2668 {
2669 /*
2670 * Make system wide (-a) the default target if
2671 * no target was specified and one of following
2672 * conditions is met:
2673 *
2674 * - there's no workload specified
2675 * - there is workload specified but all requested
2676 * events are system wide events
2677 */
2678 if (!target__none(&target))
2679 return;
2680
2681 if (!forks)
2682 target.system_wide = true;
2683 else {
2684 struct perf_evsel *counter;
2685
2686 evlist__for_each_entry(evsel_list, counter) {
2687 if (!counter->system_wide)
2688 return;
2689 }
2690
2691 if (evsel_list->nr_entries)
2692 target.system_wide = true;
2693 }
2694 }
2695
2696 int cmd_stat(int argc, const char **argv)
2697 {
2698 const char * const stat_usage[] = {
2699 "perf stat [<options>] [<command>]",
2700 NULL
2701 };
2702 int status = -EINVAL, run_idx;
2703 const char *mode;
2704 FILE *output = stderr;
2705 unsigned int interval, timeout;
2706 const char * const stat_subcommands[] = { "record", "report" };
2707
2708 setlocale(LC_ALL, "");
2709
2710 evsel_list = perf_evlist__new();
2711 if (evsel_list == NULL)
2712 return -ENOMEM;
2713
2714 parse_events__shrink_config_terms();
2715 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2716 (const char **) stat_usage,
2717 PARSE_OPT_STOP_AT_NON_OPTION);
2718 perf_stat__collect_metric_expr(evsel_list);
2719 perf_stat__init_shadow_stats();
2720
2721 if (csv_sep) {
2722 csv_output = true;
2723 if (!strcmp(csv_sep, "\\t"))
2724 csv_sep = "\t";
2725 } else
2726 csv_sep = DEFAULT_SEPARATOR;
2727
2728 if (argc && !strncmp(argv[0], "rec", 3)) {
2729 argc = __cmd_record(argc, argv);
2730 if (argc < 0)
2731 return -1;
2732 } else if (argc && !strncmp(argv[0], "rep", 3))
2733 return __cmd_report(argc, argv);
2734
2735 interval = stat_config.interval;
2736 timeout = stat_config.timeout;
2737
2738 /*
2739 * For record command the -o is already taken care of.
2740 */
2741 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2742 output = NULL;
2743
2744 if (output_name && output_fd) {
2745 fprintf(stderr, "cannot use both --output and --log-fd\n");
2746 parse_options_usage(stat_usage, stat_options, "o", 1);
2747 parse_options_usage(NULL, stat_options, "log-fd", 0);
2748 goto out;
2749 }
2750
2751 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2752 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2753 goto out;
2754 }
2755
2756 if (metric_only && run_count > 1) {
2757 fprintf(stderr, "--metric-only is not supported with -r\n");
2758 goto out;
2759 }
2760
2761 if (walltime_run_table && run_count <= 1) {
2762 fprintf(stderr, "--table is only supported with -r\n");
2763 parse_options_usage(stat_usage, stat_options, "r", 1);
2764 parse_options_usage(NULL, stat_options, "table", 0);
2765 goto out;
2766 }
2767
2768 if (output_fd < 0) {
2769 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2770 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2771 goto out;
2772 }
2773
2774 if (!output) {
2775 struct timespec tm;
2776 mode = append_file ? "a" : "w";
2777
2778 output = fopen(output_name, mode);
2779 if (!output) {
2780 perror("failed to create output file");
2781 return -1;
2782 }
2783 clock_gettime(CLOCK_REALTIME, &tm);
2784 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2785 } else if (output_fd > 0) {
2786 mode = append_file ? "a" : "w";
2787 output = fdopen(output_fd, mode);
2788 if (!output) {
2789 perror("Failed opening logfd");
2790 return -errno;
2791 }
2792 }
2793
2794 stat_config.output = output;
2795
2796 /*
2797 * let the spreadsheet do the pretty-printing
2798 */
2799 if (csv_output) {
2800 /* User explicitly passed -B? */
2801 if (big_num_opt == 1) {
2802 fprintf(stderr, "-B option not supported with -x\n");
2803 parse_options_usage(stat_usage, stat_options, "B", 1);
2804 parse_options_usage(NULL, stat_options, "x", 1);
2805 goto out;
2806 } else /* Nope, so disable big number formatting */
2807 big_num = false;
2808 } else if (big_num_opt == 0) /* User passed --no-big-num */
2809 big_num = false;
2810
2811 setup_system_wide(argc);
2812
2813 /*
2814 * Display user/system times only for single
2815 * run and when there's specified tracee.
2816 */
2817 if ((run_count == 1) && target__none(&target))
2818 ru_display = true;
2819
2820 if (run_count < 0) {
2821 pr_err("Run count must be a positive number\n");
2822 parse_options_usage(stat_usage, stat_options, "r", 1);
2823 goto out;
2824 } else if (run_count == 0) {
2825 forever = true;
2826 run_count = 1;
2827 }
2828
2829 if (walltime_run_table) {
2830 walltime_run = zalloc(run_count * sizeof(walltime_run[0]));
2831 if (!walltime_run) {
2832 pr_err("failed to setup -r option");
2833 goto out;
2834 }
2835 }
2836
2837 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2838 !target__has_task(&target)) {
2839 if (!target.system_wide || target.cpu_list) {
2840 fprintf(stderr, "The --per-thread option is only "
2841 "available when monitoring via -p -t -a "
2842 "options or only --per-thread.\n");
2843 parse_options_usage(NULL, stat_options, "p", 1);
2844 parse_options_usage(NULL, stat_options, "t", 1);
2845 goto out;
2846 }
2847 }
2848
2849 /*
2850 * no_aggr, cgroup are for system-wide only
2851 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2852 */
2853 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2854 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2855 !target__has_cpu(&target)) {
2856 fprintf(stderr, "both cgroup and no-aggregation "
2857 "modes only available in system-wide mode\n");
2858
2859 parse_options_usage(stat_usage, stat_options, "G", 1);
2860 parse_options_usage(NULL, stat_options, "A", 1);
2861 parse_options_usage(NULL, stat_options, "a", 1);
2862 goto out;
2863 }
2864
2865 if (add_default_attributes())
2866 goto out;
2867
2868 target__validate(&target);
2869
2870 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2871 target.per_thread = true;
2872
2873 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2874 if (target__has_task(&target)) {
2875 pr_err("Problems finding threads of monitor\n");
2876 parse_options_usage(stat_usage, stat_options, "p", 1);
2877 parse_options_usage(NULL, stat_options, "t", 1);
2878 } else if (target__has_cpu(&target)) {
2879 perror("failed to parse CPUs map");
2880 parse_options_usage(stat_usage, stat_options, "C", 1);
2881 parse_options_usage(NULL, stat_options, "a", 1);
2882 }
2883 goto out;
2884 }
2885
2886 /*
2887 * Initialize thread_map with comm names,
2888 * so we could print it out on output.
2889 */
2890 if (stat_config.aggr_mode == AGGR_THREAD) {
2891 thread_map__read_comms(evsel_list->threads);
2892 if (target.system_wide) {
2893 if (runtime_stat_new(&stat_config,
2894 thread_map__nr(evsel_list->threads))) {
2895 goto out;
2896 }
2897 }
2898 }
2899
2900 if (stat_config.times && interval)
2901 interval_count = true;
2902 else if (stat_config.times && !interval) {
2903 pr_err("interval-count option should be used together with "
2904 "interval-print.\n");
2905 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2906 parse_options_usage(stat_usage, stat_options, "I", 1);
2907 goto out;
2908 }
2909
2910 if (timeout && timeout < 100) {
2911 if (timeout < 10) {
2912 pr_err("timeout must be >= 10ms.\n");
2913 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2914 goto out;
2915 } else
2916 pr_warning("timeout < 100ms. "
2917 "The overhead percentage could be high in some cases. "
2918 "Please proceed with caution.\n");
2919 }
2920 if (timeout && interval) {
2921 pr_err("timeout option is not supported with interval-print.\n");
2922 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2923 parse_options_usage(stat_usage, stat_options, "I", 1);
2924 goto out;
2925 }
2926
2927 if (perf_evlist__alloc_stats(evsel_list, interval))
2928 goto out;
2929
2930 if (perf_stat_init_aggr_mode())
2931 goto out;
2932
2933 /*
2934 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2935 * while avoiding that older tools show confusing messages.
2936 *
2937 * However for pipe sessions we need to keep it zero,
2938 * because script's perf_evsel__check_attr is triggered
2939 * by attr->sample_type != 0, and we can't run it on
2940 * stat sessions.
2941 */
2942 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2943
2944 /*
2945 * We dont want to block the signals - that would cause
2946 * child tasks to inherit that and Ctrl-C would not work.
2947 * What we want is for Ctrl-C to work in the exec()-ed
2948 * task, but being ignored by perf stat itself:
2949 */
2950 atexit(sig_atexit);
2951 if (!forever)
2952 signal(SIGINT, skip_signal);
2953 signal(SIGCHLD, skip_signal);
2954 signal(SIGALRM, skip_signal);
2955 signal(SIGABRT, skip_signal);
2956
2957 status = 0;
2958 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2959 if (run_count != 1 && verbose > 0)
2960 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2961 run_idx + 1);
2962
2963 status = run_perf_stat(argc, argv, run_idx);
2964 if (forever && status != -1) {
2965 print_counters(NULL, argc, argv);
2966 perf_stat__reset_stats();
2967 }
2968 }
2969
2970 if (!forever && status != -1 && !interval)
2971 print_counters(NULL, argc, argv);
2972
2973 if (STAT_RECORD) {
2974 /*
2975 * We synthesize the kernel mmap record just so that older tools
2976 * don't emit warnings about not being able to resolve symbols
2977 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2978 * a saner message about no samples being in the perf.data file.
2979 *
2980 * This also serves to suppress a warning about f_header.data.size == 0
2981 * in header.c at the moment 'perf stat record' gets introduced, which
2982 * is not really needed once we start adding the stat specific PERF_RECORD_
2983 * records, but the need to suppress the kptr_restrict messages in older
2984 * tools remain -acme
2985 */
2986 int fd = perf_data__fd(&perf_stat.data);
2987 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2988 process_synthesized_event,
2989 &perf_stat.session->machines.host);
2990 if (err) {
2991 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2992 "older tools may produce warnings about this file\n.");
2993 }
2994
2995 if (!interval) {
2996 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2997 pr_err("failed to write stat round event\n");
2998 }
2999
3000 if (!perf_stat.data.is_pipe) {
3001 perf_stat.session->header.data_size += perf_stat.bytes_written;
3002 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
3003 }
3004
3005 perf_session__delete(perf_stat.session);
3006 }
3007
3008 perf_stat__exit_aggr_mode();
3009 perf_evlist__free_stats(evsel_list);
3010 out:
3011 free(walltime_run);
3012
3013 if (smi_cost && smi_reset)
3014 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
3015
3016 perf_evlist__delete(evsel_list);
3017
3018 runtime_stat_delete(&stat_config);
3019
3020 return status;
3021 }
Linux with added FPC-III support