?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 4.19-rc7
[linux-2.6/btrfs-unstable.git] / tools / perf / builtin-stat.c
blobd097b5b47eb81e16b83a8a37f3385aeea124a0ed
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 bool no_inherit = false;
152 static volatile pid_t child_pid = -1;
153 static bool null_run = false;
154 static int detailed_run = 0;
155 static bool transaction_run;
156 static bool topdown_run = false;
157 static bool smi_cost = false;
158 static bool smi_reset = false;
159 static bool big_num = true;
160 static int big_num_opt = -1;
161 static const char *csv_sep = NULL;
162 static bool csv_output = false;
163 static bool group = false;
164 static const char *pre_cmd = NULL;
165 static const char *post_cmd = NULL;
166 static bool sync_run = false;
167 static unsigned int initial_delay = 0;
168 static unsigned int unit_width = 4; /* strlen("unit") */
169 static bool forever = false;
170 static bool metric_only = false;
171 static bool force_metric_only = false;
172 static bool no_merge = false;
173 static bool walltime_run_table = false;
174 static struct timespec ref_time;
175 static struct cpu_map *aggr_map;
176 static aggr_get_id_t aggr_get_id;
177 static bool append_file;
178 static bool interval_count;
179 static bool interval_clear;
180 static const char *output_name;
181 static int output_fd;
182 static int print_free_counters_hint;
183 static int print_mixed_hw_group_error;
184 static u64 *walltime_run;
185 static bool ru_display = false;
186 static struct rusage ru_data;
187 static unsigned int metric_only_len = METRIC_ONLY_LEN;
188
189 struct perf_stat {
190 bool record;
191 struct perf_data data;
192 struct perf_session *session;
193 u64 bytes_written;
194 struct perf_tool tool;
195 bool maps_allocated;
196 struct cpu_map *cpus;
197 struct thread_map *threads;
198 enum aggr_mode aggr_mode;
199 };
200
201 static struct perf_stat perf_stat;
202 #define STAT_RECORD perf_stat.record
203
204 static volatile int done = 0;
205
206 static struct perf_stat_config stat_config = {
207 .aggr_mode = AGGR_GLOBAL,
208 .scale = true,
209 };
210
211 static bool is_duration_time(struct perf_evsel *evsel)
212 {
213 return !strcmp(evsel->name, "duration_time");
214 }
215
216 static inline void diff_timespec(struct timespec *r, struct timespec *a,
217 struct timespec *b)
218 {
219 r->tv_sec = a->tv_sec - b->tv_sec;
220 if (a->tv_nsec < b->tv_nsec) {
221 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
222 r->tv_sec--;
223 } else {
224 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
225 }
226 }
227
228 static void perf_stat__reset_stats(void)
229 {
230 int i;
231
232 perf_evlist__reset_stats(evsel_list);
233 perf_stat__reset_shadow_stats();
234
235 for (i = 0; i < stat_config.stats_num; i++)
236 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
237 }
238
239 static int create_perf_stat_counter(struct perf_evsel *evsel)
240 {
241 struct perf_event_attr *attr = &evsel->attr;
242 struct perf_evsel *leader = evsel->leader;
243
244 if (stat_config.scale) {
245 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
246 PERF_FORMAT_TOTAL_TIME_RUNNING;
247 }
248
249 /*
250 * The event is part of non trivial group, let's enable
251 * the group read (for leader) and ID retrieval for all
252 * members.
253 */
254 if (leader->nr_members > 1)
255 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
256
257 attr->inherit = !no_inherit;
258
259 /*
260 * Some events get initialized with sample_(period/type) set,
261 * like tracepoints. Clear it up for counting.
262 */
263 attr->sample_period = 0;
264
265 /*
266 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
267 * while avoiding that older tools show confusing messages.
268 *
269 * However for pipe sessions we need to keep it zero,
270 * because script's perf_evsel__check_attr is triggered
271 * by attr->sample_type != 0, and we can't run it on
272 * stat sessions.
273 */
274 if (!(STAT_RECORD && perf_stat.data.is_pipe))
275 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
276
277 /*
278 * Disabling all counters initially, they will be enabled
279 * either manually by us or by kernel via enable_on_exec
280 * set later.
281 */
282 if (perf_evsel__is_group_leader(evsel)) {
283 attr->disabled = 1;
284
285 /*
286 * In case of initial_delay we enable tracee
287 * events manually.
288 */
289 if (target__none(&target) && !initial_delay)
290 attr->enable_on_exec = 1;
291 }
292
293 if (target__has_cpu(&target) && !target__has_per_thread(&target))
294 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
295
296 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
297 }
298
299 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
300 union perf_event *event,
301 struct perf_sample *sample __maybe_unused,
302 struct machine *machine __maybe_unused)
303 {
304 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
305 pr_err("failed to write perf data, error: %m\n");
306 return -1;
307 }
308
309 perf_stat.bytes_written += event->header.size;
310 return 0;
311 }
312
313 static int write_stat_round_event(u64 tm, u64 type)
314 {
315 return perf_event__synthesize_stat_round(NULL, tm, type,
316 process_synthesized_event,
317 NULL);
318 }
319
320 #define WRITE_STAT_ROUND_EVENT(time, interval) \
321 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
322
323 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
324
325 static int
326 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
327 struct perf_counts_values *count)
328 {
329 struct perf_sample_id *sid = SID(counter, cpu, thread);
330
331 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
332 process_synthesized_event, NULL);
333 }
334
335 /*
336 * Read out the results of a single counter:
337 * do not aggregate counts across CPUs in system-wide mode
338 */
339 static int read_counter(struct perf_evsel *counter)
340 {
341 int nthreads = thread_map__nr(evsel_list->threads);
342 int ncpus, cpu, thread;
343
344 if (target__has_cpu(&target) && !target__has_per_thread(&target))
345 ncpus = perf_evsel__nr_cpus(counter);
346 else
347 ncpus = 1;
348
349 if (!counter->supported)
350 return -ENOENT;
351
352 if (counter->system_wide)
353 nthreads = 1;
354
355 for (thread = 0; thread < nthreads; thread++) {
356 for (cpu = 0; cpu < ncpus; cpu++) {
357 struct perf_counts_values *count;
358
359 count = perf_counts(counter->counts, cpu, thread);
360
361 /*
362 * The leader's group read loads data into its group members
363 * (via perf_evsel__read_counter) and sets threir count->loaded.
364 */
365 if (!count->loaded &&
366 perf_evsel__read_counter(counter, cpu, thread)) {
367 counter->counts->scaled = -1;
368 perf_counts(counter->counts, cpu, thread)->ena = 0;
369 perf_counts(counter->counts, cpu, thread)->run = 0;
370 return -1;
371 }
372
373 count->loaded = false;
374
375 if (STAT_RECORD) {
376 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
377 pr_err("failed to write stat event\n");
378 return -1;
379 }
380 }
381
382 if (verbose > 1) {
383 fprintf(stat_config.output,
384 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
385 perf_evsel__name(counter),
386 cpu,
387 count->val, count->ena, count->run);
388 }
389 }
390 }
391
392 return 0;
393 }
394
395 static void read_counters(void)
396 {
397 struct perf_evsel *counter;
398 int ret;
399
400 evlist__for_each_entry(evsel_list, counter) {
401 ret = read_counter(counter);
402 if (ret)
403 pr_debug("failed to read counter %s\n", counter->name);
404
405 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
406 pr_warning("failed to process counter %s\n", counter->name);
407 }
408 }
409
410 static void process_interval(void)
411 {
412 struct timespec ts, rs;
413
414 read_counters();
415
416 clock_gettime(CLOCK_MONOTONIC, &ts);
417 diff_timespec(&rs, &ts, &ref_time);
418
419 if (STAT_RECORD) {
420 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
421 pr_err("failed to write stat round event\n");
422 }
423
424 init_stats(&walltime_nsecs_stats);
425 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
426 print_counters(&rs, 0, NULL);
427 }
428
429 static void enable_counters(void)
430 {
431 if (initial_delay)
432 usleep(initial_delay * USEC_PER_MSEC);
433
434 /*
435 * We need to enable counters only if:
436 * - we don't have tracee (attaching to task or cpu)
437 * - we have initial delay configured
438 */
439 if (!target__none(&target) || initial_delay)
440 perf_evlist__enable(evsel_list);
441 }
442
443 static void disable_counters(void)
444 {
445 /*
446 * If we don't have tracee (attaching to task or cpu), counters may
447 * still be running. To get accurate group ratios, we must stop groups
448 * from counting before reading their constituent counters.
449 */
450 if (!target__none(&target))
451 perf_evlist__disable(evsel_list);
452 }
453
454 static volatile int workload_exec_errno;
455
456 /*
457 * perf_evlist__prepare_workload will send a SIGUSR1
458 * if the fork fails, since we asked by setting its
459 * want_signal to true.
460 */
461 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
462 void *ucontext __maybe_unused)
463 {
464 workload_exec_errno = info->si_value.sival_int;
465 }
466
467 static int perf_stat_synthesize_config(bool is_pipe)
468 {
469 int err;
470
471 if (is_pipe) {
472 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
473 process_synthesized_event);
474 if (err < 0) {
475 pr_err("Couldn't synthesize attrs.\n");
476 return err;
477 }
478 }
479
480 err = perf_event__synthesize_extra_attr(NULL,
481 evsel_list,
482 process_synthesized_event,
483 is_pipe);
484
485 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
486 process_synthesized_event,
487 NULL);
488 if (err < 0) {
489 pr_err("Couldn't synthesize thread map.\n");
490 return err;
491 }
492
493 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
494 process_synthesized_event, NULL);
495 if (err < 0) {
496 pr_err("Couldn't synthesize thread map.\n");
497 return err;
498 }
499
500 err = perf_event__synthesize_stat_config(NULL, &stat_config,
501 process_synthesized_event, NULL);
502 if (err < 0) {
503 pr_err("Couldn't synthesize config.\n");
504 return err;
505 }
506
507 return 0;
508 }
509
510 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
511
512 static int __store_counter_ids(struct perf_evsel *counter)
513 {
514 int cpu, thread;
515
516 for (cpu = 0; cpu < xyarray__max_x(counter->fd); cpu++) {
517 for (thread = 0; thread < xyarray__max_y(counter->fd);
518 thread++) {
519 int fd = FD(counter, cpu, thread);
520
521 if (perf_evlist__id_add_fd(evsel_list, counter,
522 cpu, thread, fd) < 0)
523 return -1;
524 }
525 }
526
527 return 0;
528 }
529
530 static int store_counter_ids(struct perf_evsel *counter)
531 {
532 struct cpu_map *cpus = counter->cpus;
533 struct thread_map *threads = counter->threads;
534
535 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
536 return -ENOMEM;
537
538 return __store_counter_ids(counter);
539 }
540
541 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
542 {
543 return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
544 }
545
546 static struct perf_evsel *perf_evsel__reset_weak_group(struct perf_evsel *evsel)
547 {
548 struct perf_evsel *c2, *leader;
549 bool is_open = true;
550
551 leader = evsel->leader;
552 pr_debug("Weak group for %s/%d failed\n",
553 leader->name, leader->nr_members);
554
555 /*
556 * for_each_group_member doesn't work here because it doesn't
557 * include the first entry.
558 */
559 evlist__for_each_entry(evsel_list, c2) {
560 if (c2 == evsel)
561 is_open = false;
562 if (c2->leader == leader) {
563 if (is_open)
564 perf_evsel__close(c2);
565 c2->leader = c2;
566 c2->nr_members = 0;
567 }
568 }
569 return leader;
570 }
571
572 static int __run_perf_stat(int argc, const char **argv, int run_idx)
573 {
574 int interval = stat_config.interval;
575 int times = stat_config.times;
576 int timeout = stat_config.timeout;
577 char msg[BUFSIZ];
578 unsigned long long t0, t1;
579 struct perf_evsel *counter;
580 struct timespec ts;
581 size_t l;
582 int status = 0;
583 const bool forks = (argc > 0);
584 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
585 struct perf_evsel_config_term *err_term;
586
587 if (interval) {
588 ts.tv_sec = interval / USEC_PER_MSEC;
589 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
590 } else if (timeout) {
591 ts.tv_sec = timeout / USEC_PER_MSEC;
592 ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
593 } else {
594 ts.tv_sec = 1;
595 ts.tv_nsec = 0;
596 }
597
598 if (forks) {
599 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
600 workload_exec_failed_signal) < 0) {
601 perror("failed to prepare workload");
602 return -1;
603 }
604 child_pid = evsel_list->workload.pid;
605 }
606
607 if (group)
608 perf_evlist__set_leader(evsel_list);
609
610 evlist__for_each_entry(evsel_list, counter) {
611 try_again:
612 if (create_perf_stat_counter(counter) < 0) {
613
614 /* Weak group failed. Reset the group. */
615 if ((errno == EINVAL || errno == EBADF) &&
616 counter->leader != counter &&
617 counter->weak_group) {
618 counter = perf_evsel__reset_weak_group(counter);
619 goto try_again;
620 }
621
622 /*
623 * PPC returns ENXIO for HW counters until 2.6.37
624 * (behavior changed with commit b0a873e).
625 */
626 if (errno == EINVAL || errno == ENOSYS ||
627 errno == ENOENT || errno == EOPNOTSUPP ||
628 errno == ENXIO) {
629 if (verbose > 0)
630 ui__warning("%s event is not supported by the kernel.\n",
631 perf_evsel__name(counter));
632 counter->supported = false;
633
634 if ((counter->leader != counter) ||
635 !(counter->leader->nr_members > 1))
636 continue;
637 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
638 if (verbose > 0)
639 ui__warning("%s\n", msg);
640 goto try_again;
641 } else if (target__has_per_thread(&target) &&
642 evsel_list->threads &&
643 evsel_list->threads->err_thread != -1) {
644 /*
645 * For global --per-thread case, skip current
646 * error thread.
647 */
648 if (!thread_map__remove(evsel_list->threads,
649 evsel_list->threads->err_thread)) {
650 evsel_list->threads->err_thread = -1;
651 goto try_again;
652 }
653 }
654
655 perf_evsel__open_strerror(counter, &target,
656 errno, msg, sizeof(msg));
657 ui__error("%s\n", msg);
658
659 if (child_pid != -1)
660 kill(child_pid, SIGTERM);
661
662 return -1;
663 }
664 counter->supported = true;
665
666 l = strlen(counter->unit);
667 if (l > unit_width)
668 unit_width = l;
669
670 if (perf_evsel__should_store_id(counter) &&
671 store_counter_ids(counter))
672 return -1;
673 }
674
675 if (perf_evlist__apply_filters(evsel_list, &counter)) {
676 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
677 counter->filter, perf_evsel__name(counter), errno,
678 str_error_r(errno, msg, sizeof(msg)));
679 return -1;
680 }
681
682 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
683 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
684 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
685 str_error_r(errno, msg, sizeof(msg)));
686 return -1;
687 }
688
689 if (STAT_RECORD) {
690 int err, fd = perf_data__fd(&perf_stat.data);
691
692 if (is_pipe) {
693 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
694 } else {
695 err = perf_session__write_header(perf_stat.session, evsel_list,
696 fd, false);
697 }
698
699 if (err < 0)
700 return err;
701
702 err = perf_stat_synthesize_config(is_pipe);
703 if (err < 0)
704 return err;
705 }
706
707 /*
708 * Enable counters and exec the command:
709 */
710 t0 = rdclock();
711 clock_gettime(CLOCK_MONOTONIC, &ref_time);
712
713 if (forks) {
714 perf_evlist__start_workload(evsel_list);
715 enable_counters();
716
717 if (interval || timeout) {
718 while (!waitpid(child_pid, &status, WNOHANG)) {
719 nanosleep(&ts, NULL);
720 if (timeout)
721 break;
722 process_interval();
723 if (interval_count && !(--times))
724 break;
725 }
726 }
727 wait4(child_pid, &status, 0, &ru_data);
728
729 if (workload_exec_errno) {
730 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
731 pr_err("Workload failed: %s\n", emsg);
732 return -1;
733 }
734
735 if (WIFSIGNALED(status))
736 psignal(WTERMSIG(status), argv[0]);
737 } else {
738 enable_counters();
739 while (!done) {
740 nanosleep(&ts, NULL);
741 if (timeout)
742 break;
743 if (interval) {
744 process_interval();
745 if (interval_count && !(--times))
746 break;
747 }
748 }
749 }
750
751 disable_counters();
752
753 t1 = rdclock();
754
755 if (walltime_run_table)
756 walltime_run[run_idx] = t1 - t0;
757
758 update_stats(&walltime_nsecs_stats, t1 - t0);
759
760 /*
761 * Closing a group leader splits the group, and as we only disable
762 * group leaders, results in remaining events becoming enabled. To
763 * avoid arbitrary skew, we must read all counters before closing any
764 * group leaders.
765 */
766 read_counters();
767 perf_evlist__close(evsel_list);
768
769 return WEXITSTATUS(status);
770 }
771
772 static int run_perf_stat(int argc, const char **argv, int run_idx)
773 {
774 int ret;
775
776 if (pre_cmd) {
777 ret = system(pre_cmd);
778 if (ret)
779 return ret;
780 }
781
782 if (sync_run)
783 sync();
784
785 ret = __run_perf_stat(argc, argv, run_idx);
786 if (ret)
787 return ret;
788
789 if (post_cmd) {
790 ret = system(post_cmd);
791 if (ret)
792 return ret;
793 }
794
795 return ret;
796 }
797
798 static void print_running(u64 run, u64 ena)
799 {
800 if (csv_output) {
801 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
802 csv_sep,
803 run,
804 csv_sep,
805 ena ? 100.0 * run / ena : 100.0);
806 } else if (run != ena) {
807 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
808 }
809 }
810
811 static void print_noise_pct(double total, double avg)
812 {
813 double pct = rel_stddev_stats(total, avg);
814
815 if (csv_output)
816 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
817 else if (pct)
818 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
819 }
820
821 static void print_noise(struct perf_evsel *evsel, double avg)
822 {
823 struct perf_stat_evsel *ps;
824
825 if (run_count == 1)
826 return;
827
828 ps = evsel->stats;
829 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
830 }
831
832 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
833 {
834 switch (stat_config.aggr_mode) {
835 case AGGR_CORE:
836 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
837 cpu_map__id_to_socket(id),
838 csv_output ? 0 : -8,
839 cpu_map__id_to_cpu(id),
840 csv_sep,
841 csv_output ? 0 : 4,
842 nr,
843 csv_sep);
844 break;
845 case AGGR_SOCKET:
846 fprintf(stat_config.output, "S%*d%s%*d%s",
847 csv_output ? 0 : -5,
848 id,
849 csv_sep,
850 csv_output ? 0 : 4,
851 nr,
852 csv_sep);
853 break;
854 case AGGR_NONE:
855 fprintf(stat_config.output, "CPU%*d%s",
856 csv_output ? 0 : -4,
857 perf_evsel__cpus(evsel)->map[id], csv_sep);
858 break;
859 case AGGR_THREAD:
860 fprintf(stat_config.output, "%*s-%*d%s",
861 csv_output ? 0 : 16,
862 thread_map__comm(evsel->threads, id),
863 csv_output ? 0 : -8,
864 thread_map__pid(evsel->threads, id),
865 csv_sep);
866 break;
867 case AGGR_GLOBAL:
868 case AGGR_UNSET:
869 default:
870 break;
871 }
872 }
873
874 struct outstate {
875 FILE *fh;
876 bool newline;
877 const char *prefix;
878 int nfields;
879 int id, nr;
880 struct perf_evsel *evsel;
881 };
882
883 #define METRIC_LEN 35
884
885 static void new_line_std(void *ctx)
886 {
887 struct outstate *os = ctx;
888
889 os->newline = true;
890 }
891
892 static void do_new_line_std(struct outstate *os)
893 {
894 fputc('\n', os->fh);
895 fputs(os->prefix, os->fh);
896 aggr_printout(os->evsel, os->id, os->nr);
897 if (stat_config.aggr_mode == AGGR_NONE)
898 fprintf(os->fh, " ");
899 fprintf(os->fh, " ");
900 }
901
902 static void print_metric_std(void *ctx, const char *color, const char *fmt,
903 const char *unit, double val)
904 {
905 struct outstate *os = ctx;
906 FILE *out = os->fh;
907 int n;
908 bool newline = os->newline;
909
910 os->newline = false;
911
912 if (unit == NULL || fmt == NULL) {
913 fprintf(out, "%-*s", METRIC_LEN, "");
914 return;
915 }
916
917 if (newline)
918 do_new_line_std(os);
919
920 n = fprintf(out, " # ");
921 if (color)
922 n += color_fprintf(out, color, fmt, val);
923 else
924 n += fprintf(out, fmt, val);
925 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
926 }
927
928 static void new_line_csv(void *ctx)
929 {
930 struct outstate *os = ctx;
931 int i;
932
933 fputc('\n', os->fh);
934 if (os->prefix)
935 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
936 aggr_printout(os->evsel, os->id, os->nr);
937 for (i = 0; i < os->nfields; i++)
938 fputs(csv_sep, os->fh);
939 }
940
941 static void print_metric_csv(void *ctx,
942 const char *color __maybe_unused,
943 const char *fmt, const char *unit, double val)
944 {
945 struct outstate *os = ctx;
946 FILE *out = os->fh;
947 char buf[64], *vals, *ends;
948
949 if (unit == NULL || fmt == NULL) {
950 fprintf(out, "%s%s", csv_sep, csv_sep);
951 return;
952 }
953 snprintf(buf, sizeof(buf), fmt, val);
954 ends = vals = ltrim(buf);
955 while (isdigit(*ends) || *ends == '.')
956 ends++;
957 *ends = 0;
958 while (isspace(*unit))
959 unit++;
960 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
961 }
962
963 /* Filter out some columns that don't work well in metrics only mode */
964
965 static bool valid_only_metric(const char *unit)
966 {
967 if (!unit)
968 return false;
969 if (strstr(unit, "/sec") ||
970 strstr(unit, "hz") ||
971 strstr(unit, "Hz") ||
972 strstr(unit, "CPUs utilized"))
973 return false;
974 return true;
975 }
976
977 static const char *fixunit(char *buf, struct perf_evsel *evsel,
978 const char *unit)
979 {
980 if (!strncmp(unit, "of all", 6)) {
981 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
982 unit);
983 return buf;
984 }
985 return unit;
986 }
987
988 static void print_metric_only(void *ctx, const char *color, const char *fmt,
989 const char *unit, double val)
990 {
991 struct outstate *os = ctx;
992 FILE *out = os->fh;
993 char buf[1024], str[1024];
994 unsigned mlen = metric_only_len;
995
996 if (!valid_only_metric(unit))
997 return;
998 unit = fixunit(buf, os->evsel, unit);
999 if (mlen < strlen(unit))
1000 mlen = strlen(unit) + 1;
1001
1002 if (color)
1003 mlen += strlen(color) + sizeof(PERF_COLOR_RESET) - 1;
1004
1005 color_snprintf(str, sizeof(str), color ?: "", fmt, val);
1006 fprintf(out, "%*s ", mlen, str);
1007 }
1008
1009 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
1010 const char *fmt,
1011 const char *unit, double val)
1012 {
1013 struct outstate *os = ctx;
1014 FILE *out = os->fh;
1015 char buf[64], *vals, *ends;
1016 char tbuf[1024];
1017
1018 if (!valid_only_metric(unit))
1019 return;
1020 unit = fixunit(tbuf, os->evsel, unit);
1021 snprintf(buf, sizeof buf, fmt, val);
1022 ends = vals = ltrim(buf);
1023 while (isdigit(*ends) || *ends == '.')
1024 ends++;
1025 *ends = 0;
1026 fprintf(out, "%s%s", vals, csv_sep);
1027 }
1028
1029 static void new_line_metric(void *ctx __maybe_unused)
1030 {
1031 }
1032
1033 static void print_metric_header(void *ctx, const char *color __maybe_unused,
1034 const char *fmt __maybe_unused,
1035 const char *unit, double val __maybe_unused)
1036 {
1037 struct outstate *os = ctx;
1038 char tbuf[1024];
1039
1040 if (!valid_only_metric(unit))
1041 return;
1042 unit = fixunit(tbuf, os->evsel, unit);
1043 if (csv_output)
1044 fprintf(os->fh, "%s%s", unit, csv_sep);
1045 else
1046 fprintf(os->fh, "%*s ", metric_only_len, unit);
1047 }
1048
1049 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1050 {
1051 int i;
1052
1053 if (!aggr_get_id)
1054 return 0;
1055
1056 if (stat_config.aggr_mode == AGGR_NONE)
1057 return id;
1058
1059 if (stat_config.aggr_mode == AGGR_GLOBAL)
1060 return 0;
1061
1062 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1063 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1064
1065 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1066 return cpu2;
1067 }
1068 return 0;
1069 }
1070
1071 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1072 {
1073 FILE *output = stat_config.output;
1074 double sc = evsel->scale;
1075 const char *fmt;
1076
1077 if (csv_output) {
1078 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
1079 } else {
1080 if (big_num)
1081 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1082 else
1083 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1084 }
1085
1086 aggr_printout(evsel, id, nr);
1087
1088 fprintf(output, fmt, avg, csv_sep);
1089
1090 if (evsel->unit)
1091 fprintf(output, "%-*s%s",
1092 csv_output ? 0 : unit_width,
1093 evsel->unit, csv_sep);
1094
1095 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1096
1097 if (evsel->cgrp)
1098 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1099 }
1100
1101 static bool is_mixed_hw_group(struct perf_evsel *counter)
1102 {
1103 struct perf_evlist *evlist = counter->evlist;
1104 u32 pmu_type = counter->attr.type;
1105 struct perf_evsel *pos;
1106
1107 if (counter->nr_members < 2)
1108 return false;
1109
1110 evlist__for_each_entry(evlist, pos) {
1111 /* software events can be part of any hardware group */
1112 if (pos->attr.type == PERF_TYPE_SOFTWARE)
1113 continue;
1114 if (pmu_type == PERF_TYPE_SOFTWARE) {
1115 pmu_type = pos->attr.type;
1116 continue;
1117 }
1118 if (pmu_type != pos->attr.type)
1119 return true;
1120 }
1121
1122 return false;
1123 }
1124
1125 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1126 char *prefix, u64 run, u64 ena, double noise,
1127 struct runtime_stat *st)
1128 {
1129 struct perf_stat_output_ctx out;
1130 struct outstate os = {
1131 .fh = stat_config.output,
1132 .prefix = prefix ? prefix : "",
1133 .id = id,
1134 .nr = nr,
1135 .evsel = counter,
1136 };
1137 print_metric_t pm = print_metric_std;
1138 void (*nl)(void *);
1139
1140 if (metric_only) {
1141 nl = new_line_metric;
1142 if (csv_output)
1143 pm = print_metric_only_csv;
1144 else
1145 pm = print_metric_only;
1146 } else
1147 nl = new_line_std;
1148
1149 if (csv_output && !metric_only) {
1150 static int aggr_fields[] = {
1151 [AGGR_GLOBAL] = 0,
1152 [AGGR_THREAD] = 1,
1153 [AGGR_NONE] = 1,
1154 [AGGR_SOCKET] = 2,
1155 [AGGR_CORE] = 2,
1156 };
1157
1158 pm = print_metric_csv;
1159 nl = new_line_csv;
1160 os.nfields = 3;
1161 os.nfields += aggr_fields[stat_config.aggr_mode];
1162 if (counter->cgrp)
1163 os.nfields++;
1164 }
1165 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1166 if (metric_only) {
1167 pm(&os, NULL, "", "", 0);
1168 return;
1169 }
1170 aggr_printout(counter, id, nr);
1171
1172 fprintf(stat_config.output, "%*s%s",
1173 csv_output ? 0 : 18,
1174 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1175 csv_sep);
1176
1177 if (counter->supported) {
1178 print_free_counters_hint = 1;
1179 if (is_mixed_hw_group(counter))
1180 print_mixed_hw_group_error = 1;
1181 }
1182
1183 fprintf(stat_config.output, "%-*s%s",
1184 csv_output ? 0 : unit_width,
1185 counter->unit, csv_sep);
1186
1187 fprintf(stat_config.output, "%*s",
1188 csv_output ? 0 : -25,
1189 perf_evsel__name(counter));
1190
1191 if (counter->cgrp)
1192 fprintf(stat_config.output, "%s%s",
1193 csv_sep, counter->cgrp->name);
1194
1195 if (!csv_output)
1196 pm(&os, NULL, NULL, "", 0);
1197 print_noise(counter, noise);
1198 print_running(run, ena);
1199 if (csv_output)
1200 pm(&os, NULL, NULL, "", 0);
1201 return;
1202 }
1203
1204 if (!metric_only)
1205 abs_printout(id, nr, counter, uval);
1206
1207 out.print_metric = pm;
1208 out.new_line = nl;
1209 out.ctx = &os;
1210 out.force_header = false;
1211
1212 if (csv_output && !metric_only) {
1213 print_noise(counter, noise);
1214 print_running(run, ena);
1215 }
1216
1217 perf_stat__print_shadow_stats(counter, uval,
1218 first_shadow_cpu(counter, id),
1219 &out, &metric_events, st);
1220 if (!csv_output && !metric_only) {
1221 print_noise(counter, noise);
1222 print_running(run, ena);
1223 }
1224 }
1225
1226 static void aggr_update_shadow(void)
1227 {
1228 int cpu, s2, id, s;
1229 u64 val;
1230 struct perf_evsel *counter;
1231
1232 for (s = 0; s < aggr_map->nr; s++) {
1233 id = aggr_map->map[s];
1234 evlist__for_each_entry(evsel_list, counter) {
1235 val = 0;
1236 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1237 s2 = aggr_get_id(evsel_list->cpus, cpu);
1238 if (s2 != id)
1239 continue;
1240 val += perf_counts(counter->counts, cpu, 0)->val;
1241 }
1242 perf_stat__update_shadow_stats(counter, val,
1243 first_shadow_cpu(counter, id),
1244 &rt_stat);
1245 }
1246 }
1247 }
1248
1249 static void uniquify_event_name(struct perf_evsel *counter)
1250 {
1251 char *new_name;
1252 char *config;
1253
1254 if (counter->uniquified_name ||
1255 !counter->pmu_name || !strncmp(counter->name, counter->pmu_name,
1256 strlen(counter->pmu_name)))
1257 return;
1258
1259 config = strchr(counter->name, '/');
1260 if (config) {
1261 if (asprintf(&new_name,
1262 "%s%s", counter->pmu_name, config) > 0) {
1263 free(counter->name);
1264 counter->name = new_name;
1265 }
1266 } else {
1267 if (asprintf(&new_name,
1268 "%s [%s]", counter->name, counter->pmu_name) > 0) {
1269 free(counter->name);
1270 counter->name = new_name;
1271 }
1272 }
1273
1274 counter->uniquified_name = true;
1275 }
1276
1277 static void collect_all_aliases(struct perf_evsel *counter,
1278 void (*cb)(struct perf_evsel *counter, void *data,
1279 bool first),
1280 void *data)
1281 {
1282 struct perf_evsel *alias;
1283
1284 alias = list_prepare_entry(counter, &(evsel_list->entries), node);
1285 list_for_each_entry_continue (alias, &evsel_list->entries, node) {
1286 if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) ||
1287 alias->scale != counter->scale ||
1288 alias->cgrp != counter->cgrp ||
1289 strcmp(alias->unit, counter->unit) ||
1290 perf_evsel__is_clock(alias) != perf_evsel__is_clock(counter))
1291 break;
1292 alias->merged_stat = true;
1293 cb(alias, data, false);
1294 }
1295 }
1296
1297 static bool collect_data(struct perf_evsel *counter,
1298 void (*cb)(struct perf_evsel *counter, void *data,
1299 bool first),
1300 void *data)
1301 {
1302 if (counter->merged_stat)
1303 return false;
1304 cb(counter, data, true);
1305 if (no_merge)
1306 uniquify_event_name(counter);
1307 else if (counter->auto_merge_stats)
1308 collect_all_aliases(counter, cb, data);
1309 return true;
1310 }
1311
1312 struct aggr_data {
1313 u64 ena, run, val;
1314 int id;
1315 int nr;
1316 int cpu;
1317 };
1318
1319 static void aggr_cb(struct perf_evsel *counter, void *data, bool first)
1320 {
1321 struct aggr_data *ad = data;
1322 int cpu, s2;
1323
1324 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1325 struct perf_counts_values *counts;
1326
1327 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1328 if (s2 != ad->id)
1329 continue;
1330 if (first)
1331 ad->nr++;
1332 counts = perf_counts(counter->counts, cpu, 0);
1333 /*
1334 * When any result is bad, make them all to give
1335 * consistent output in interval mode.
1336 */
1337 if (counts->ena == 0 || counts->run == 0 ||
1338 counter->counts->scaled == -1) {
1339 ad->ena = 0;
1340 ad->run = 0;
1341 break;
1342 }
1343 ad->val += counts->val;
1344 ad->ena += counts->ena;
1345 ad->run += counts->run;
1346 }
1347 }
1348
1349 static void print_aggr(char *prefix)
1350 {
1351 FILE *output = stat_config.output;
1352 struct perf_evsel *counter;
1353 int s, id, nr;
1354 double uval;
1355 u64 ena, run, val;
1356 bool first;
1357
1358 if (!(aggr_map || aggr_get_id))
1359 return;
1360
1361 aggr_update_shadow();
1362
1363 /*
1364 * With metric_only everything is on a single line.
1365 * Without each counter has its own line.
1366 */
1367 for (s = 0; s < aggr_map->nr; s++) {
1368 struct aggr_data ad;
1369 if (prefix && metric_only)
1370 fprintf(output, "%s", prefix);
1371
1372 ad.id = id = aggr_map->map[s];
1373 first = true;
1374 evlist__for_each_entry(evsel_list, counter) {
1375 if (is_duration_time(counter))
1376 continue;
1377
1378 ad.val = ad.ena = ad.run = 0;
1379 ad.nr = 0;
1380 if (!collect_data(counter, aggr_cb, &ad))
1381 continue;
1382 nr = ad.nr;
1383 ena = ad.ena;
1384 run = ad.run;
1385 val = ad.val;
1386 if (first && metric_only) {
1387 first = false;
1388 aggr_printout(counter, id, nr);
1389 }
1390 if (prefix && !metric_only)
1391 fprintf(output, "%s", prefix);
1392
1393 uval = val * counter->scale;
1394 printout(id, nr, counter, uval, prefix, run, ena, 1.0,
1395 &rt_stat);
1396 if (!metric_only)
1397 fputc('\n', output);
1398 }
1399 if (metric_only)
1400 fputc('\n', output);
1401 }
1402 }
1403
1404 static int cmp_val(const void *a, const void *b)
1405 {
1406 return ((struct perf_aggr_thread_value *)b)->val -
1407 ((struct perf_aggr_thread_value *)a)->val;
1408 }
1409
1410 static struct perf_aggr_thread_value *sort_aggr_thread(
1411 struct perf_evsel *counter,
1412 int nthreads, int ncpus,
1413 int *ret)
1414 {
1415 int cpu, thread, i = 0;
1416 double uval;
1417 struct perf_aggr_thread_value *buf;
1418
1419 buf = calloc(nthreads, sizeof(struct perf_aggr_thread_value));
1420 if (!buf)
1421 return NULL;
1422
1423 for (thread = 0; thread < nthreads; thread++) {
1424 u64 ena = 0, run = 0, val = 0;
1425
1426 for (cpu = 0; cpu < ncpus; cpu++) {
1427 val += perf_counts(counter->counts, cpu, thread)->val;
1428 ena += perf_counts(counter->counts, cpu, thread)->ena;
1429 run += perf_counts(counter->counts, cpu, thread)->run;
1430 }
1431
1432 uval = val * counter->scale;
1433
1434 /*
1435 * Skip value 0 when enabling --per-thread globally,
1436 * otherwise too many 0 output.
1437 */
1438 if (uval == 0.0 && target__has_per_thread(&target))
1439 continue;
1440
1441 buf[i].counter = counter;
1442 buf[i].id = thread;
1443 buf[i].uval = uval;
1444 buf[i].val = val;
1445 buf[i].run = run;
1446 buf[i].ena = ena;
1447 i++;
1448 }
1449
1450 qsort(buf, i, sizeof(struct perf_aggr_thread_value), cmp_val);
1451
1452 if (ret)
1453 *ret = i;
1454
1455 return buf;
1456 }
1457
1458 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1459 {
1460 FILE *output = stat_config.output;
1461 int nthreads = thread_map__nr(counter->threads);
1462 int ncpus = cpu_map__nr(counter->cpus);
1463 int thread, sorted_threads, id;
1464 struct perf_aggr_thread_value *buf;
1465
1466 buf = sort_aggr_thread(counter, nthreads, ncpus, &sorted_threads);
1467 if (!buf) {
1468 perror("cannot sort aggr thread");
1469 return;
1470 }
1471
1472 for (thread = 0; thread < sorted_threads; thread++) {
1473 if (prefix)
1474 fprintf(output, "%s", prefix);
1475
1476 id = buf[thread].id;
1477 if (stat_config.stats)
1478 printout(id, 0, buf[thread].counter, buf[thread].uval,
1479 prefix, buf[thread].run, buf[thread].ena, 1.0,
1480 &stat_config.stats[id]);
1481 else
1482 printout(id, 0, buf[thread].counter, buf[thread].uval,
1483 prefix, buf[thread].run, buf[thread].ena, 1.0,
1484 &rt_stat);
1485 fputc('\n', output);
1486 }
1487
1488 free(buf);
1489 }
1490
1491 struct caggr_data {
1492 double avg, avg_enabled, avg_running;
1493 };
1494
1495 static void counter_aggr_cb(struct perf_evsel *counter, void *data,
1496 bool first __maybe_unused)
1497 {
1498 struct caggr_data *cd = data;
1499 struct perf_stat_evsel *ps = counter->stats;
1500
1501 cd->avg += avg_stats(&ps->res_stats[0]);
1502 cd->avg_enabled += avg_stats(&ps->res_stats[1]);
1503 cd->avg_running += avg_stats(&ps->res_stats[2]);
1504 }
1505
1506 /*
1507 * Print out the results of a single counter:
1508 * aggregated counts in system-wide mode
1509 */
1510 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1511 {
1512 FILE *output = stat_config.output;
1513 double uval;
1514 struct caggr_data cd = { .avg = 0.0 };
1515
1516 if (!collect_data(counter, counter_aggr_cb, &cd))
1517 return;
1518
1519 if (prefix && !metric_only)
1520 fprintf(output, "%s", prefix);
1521
1522 uval = cd.avg * counter->scale;
1523 printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled,
1524 cd.avg, &rt_stat);
1525 if (!metric_only)
1526 fprintf(output, "\n");
1527 }
1528
1529 static void counter_cb(struct perf_evsel *counter, void *data,
1530 bool first __maybe_unused)
1531 {
1532 struct aggr_data *ad = data;
1533
1534 ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
1535 ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
1536 ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
1537 }
1538
1539 /*
1540 * Print out the results of a single counter:
1541 * does not use aggregated count in system-wide
1542 */
1543 static void print_counter(struct perf_evsel *counter, char *prefix)
1544 {
1545 FILE *output = stat_config.output;
1546 u64 ena, run, val;
1547 double uval;
1548 int cpu;
1549
1550 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1551 struct aggr_data ad = { .cpu = cpu };
1552
1553 if (!collect_data(counter, counter_cb, &ad))
1554 return;
1555 val = ad.val;
1556 ena = ad.ena;
1557 run = ad.run;
1558
1559 if (prefix)
1560 fprintf(output, "%s", prefix);
1561
1562 uval = val * counter->scale;
1563 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0,
1564 &rt_stat);
1565
1566 fputc('\n', output);
1567 }
1568 }
1569
1570 static void print_no_aggr_metric(char *prefix)
1571 {
1572 int cpu;
1573 int nrcpus = 0;
1574 struct perf_evsel *counter;
1575 u64 ena, run, val;
1576 double uval;
1577
1578 nrcpus = evsel_list->cpus->nr;
1579 for (cpu = 0; cpu < nrcpus; cpu++) {
1580 bool first = true;
1581
1582 if (prefix)
1583 fputs(prefix, stat_config.output);
1584 evlist__for_each_entry(evsel_list, counter) {
1585 if (is_duration_time(counter))
1586 continue;
1587 if (first) {
1588 aggr_printout(counter, cpu, 0);
1589 first = false;
1590 }
1591 val = perf_counts(counter->counts, cpu, 0)->val;
1592 ena = perf_counts(counter->counts, cpu, 0)->ena;
1593 run = perf_counts(counter->counts, cpu, 0)->run;
1594
1595 uval = val * counter->scale;
1596 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0,
1597 &rt_stat);
1598 }
1599 fputc('\n', stat_config.output);
1600 }
1601 }
1602
1603 static int aggr_header_lens[] = {
1604 [AGGR_CORE] = 18,
1605 [AGGR_SOCKET] = 12,
1606 [AGGR_NONE] = 6,
1607 [AGGR_THREAD] = 24,
1608 [AGGR_GLOBAL] = 0,
1609 };
1610
1611 static const char *aggr_header_csv[] = {
1612 [AGGR_CORE] = "core,cpus,",
1613 [AGGR_SOCKET] = "socket,cpus",
1614 [AGGR_NONE] = "cpu,",
1615 [AGGR_THREAD] = "comm-pid,",
1616 [AGGR_GLOBAL] = ""
1617 };
1618
1619 static void print_metric_headers(const char *prefix, bool no_indent)
1620 {
1621 struct perf_stat_output_ctx out;
1622 struct perf_evsel *counter;
1623 struct outstate os = {
1624 .fh = stat_config.output
1625 };
1626
1627 if (prefix)
1628 fprintf(stat_config.output, "%s", prefix);
1629
1630 if (!csv_output && !no_indent)
1631 fprintf(stat_config.output, "%*s",
1632 aggr_header_lens[stat_config.aggr_mode], "");
1633 if (csv_output) {
1634 if (stat_config.interval)
1635 fputs("time,", stat_config.output);
1636 fputs(aggr_header_csv[stat_config.aggr_mode],
1637 stat_config.output);
1638 }
1639
1640 /* Print metrics headers only */
1641 evlist__for_each_entry(evsel_list, counter) {
1642 if (is_duration_time(counter))
1643 continue;
1644 os.evsel = counter;
1645 out.ctx = &os;
1646 out.print_metric = print_metric_header;
1647 out.new_line = new_line_metric;
1648 out.force_header = true;
1649 os.evsel = counter;
1650 perf_stat__print_shadow_stats(counter, 0,
1651 0,
1652 &out,
1653 &metric_events,
1654 &rt_stat);
1655 }
1656 fputc('\n', stat_config.output);
1657 }
1658
1659 static void print_interval(char *prefix, struct timespec *ts)
1660 {
1661 FILE *output = stat_config.output;
1662 static int num_print_interval;
1663
1664 if (interval_clear)
1665 puts(CONSOLE_CLEAR);
1666
1667 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1668
1669 if ((num_print_interval == 0 && !csv_output) || interval_clear) {
1670 switch (stat_config.aggr_mode) {
1671 case AGGR_SOCKET:
1672 fprintf(output, "# time socket cpus");
1673 if (!metric_only)
1674 fprintf(output, " counts %*s events\n", unit_width, "unit");
1675 break;
1676 case AGGR_CORE:
1677 fprintf(output, "# time core cpus");
1678 if (!metric_only)
1679 fprintf(output, " counts %*s events\n", unit_width, "unit");
1680 break;
1681 case AGGR_NONE:
1682 fprintf(output, "# time CPU ");
1683 if (!metric_only)
1684 fprintf(output, " counts %*s events\n", unit_width, "unit");
1685 break;
1686 case AGGR_THREAD:
1687 fprintf(output, "# time comm-pid");
1688 if (!metric_only)
1689 fprintf(output, " counts %*s events\n", unit_width, "unit");
1690 break;
1691 case AGGR_GLOBAL:
1692 default:
1693 fprintf(output, "# time");
1694 if (!metric_only)
1695 fprintf(output, " counts %*s events\n", unit_width, "unit");
1696 case AGGR_UNSET:
1697 break;
1698 }
1699 }
1700
1701 if ((num_print_interval == 0 || interval_clear) && metric_only)
1702 print_metric_headers(" ", true);
1703 if (++num_print_interval == 25)
1704 num_print_interval = 0;
1705 }
1706
1707 static void print_header(int argc, const char **argv)
1708 {
1709 FILE *output = stat_config.output;
1710 int i;
1711
1712 fflush(stdout);
1713
1714 if (!csv_output) {
1715 fprintf(output, "\n");
1716 fprintf(output, " Performance counter stats for ");
1717 if (target.system_wide)
1718 fprintf(output, "\'system wide");
1719 else if (target.cpu_list)
1720 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1721 else if (!target__has_task(&target)) {
1722 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1723 for (i = 1; argv && (i < argc); i++)
1724 fprintf(output, " %s", argv[i]);
1725 } else if (target.pid)
1726 fprintf(output, "process id \'%s", target.pid);
1727 else
1728 fprintf(output, "thread id \'%s", target.tid);
1729
1730 fprintf(output, "\'");
1731 if (run_count > 1)
1732 fprintf(output, " (%d runs)", run_count);
1733 fprintf(output, ":\n\n");
1734 }
1735 }
1736
1737 static int get_precision(double num)
1738 {
1739 if (num > 1)
1740 return 0;
1741
1742 return lround(ceil(-log10(num)));
1743 }
1744
1745 static void print_table(FILE *output, int precision, double avg)
1746 {
1747 char tmp[64];
1748 int idx, indent = 0;
1749
1750 scnprintf(tmp, 64, " %17.*f", precision, avg);
1751 while (tmp[indent] == ' ')
1752 indent++;
1753
1754 fprintf(output, "%*s# Table of individual measurements:\n", indent, "");
1755
1756 for (idx = 0; idx < run_count; idx++) {
1757 double run = (double) walltime_run[idx] / NSEC_PER_SEC;
1758 int h, n = 1 + abs((int) (100.0 * (run - avg)/run) / 5);
1759
1760 fprintf(output, " %17.*f (%+.*f) ",
1761 precision, run, precision, run - avg);
1762
1763 for (h = 0; h < n; h++)
1764 fprintf(output, "#");
1765
1766 fprintf(output, "\n");
1767 }
1768
1769 fprintf(output, "\n%*s# Final result:\n", indent, "");
1770 }
1771
1772 static double timeval2double(struct timeval *t)
1773 {
1774 return t->tv_sec + (double) t->tv_usec/USEC_PER_SEC;
1775 }
1776
1777 static void print_footer(void)
1778 {
1779 double avg = avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC;
1780 FILE *output = stat_config.output;
1781 int n;
1782
1783 if (!null_run)
1784 fprintf(output, "\n");
1785
1786 if (run_count == 1) {
1787 fprintf(output, " %17.9f seconds time elapsed", avg);
1788
1789 if (ru_display) {
1790 double ru_utime = timeval2double(&ru_data.ru_utime);
1791 double ru_stime = timeval2double(&ru_data.ru_stime);
1792
1793 fprintf(output, "\n\n");
1794 fprintf(output, " %17.9f seconds user\n", ru_utime);
1795 fprintf(output, " %17.9f seconds sys\n", ru_stime);
1796 }
1797 } else {
1798 double sd = stddev_stats(&walltime_nsecs_stats) / NSEC_PER_SEC;
1799 /*
1800 * Display at most 2 more significant
1801 * digits than the stddev inaccuracy.
1802 */
1803 int precision = get_precision(sd) + 2;
1804
1805 if (walltime_run_table)
1806 print_table(output, precision, avg);
1807
1808 fprintf(output, " %17.*f +- %.*f seconds time elapsed",
1809 precision, avg, precision, sd);
1810
1811 print_noise_pct(sd, avg);
1812 }
1813 fprintf(output, "\n\n");
1814
1815 if (print_free_counters_hint &&
1816 sysctl__read_int("kernel/nmi_watchdog", &n) >= 0 &&
1817 n > 0)
1818 fprintf(output,
1819 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1820 " echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1821 " perf stat ...\n"
1822 " echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1823
1824 if (print_mixed_hw_group_error)
1825 fprintf(output,
1826 "The events in group usually have to be from "
1827 "the same PMU. Try reorganizing the group.\n");
1828 }
1829
1830 static void print_counters(struct timespec *ts, int argc, const char **argv)
1831 {
1832 int interval = stat_config.interval;
1833 struct perf_evsel *counter;
1834 char buf[64], *prefix = NULL;
1835
1836 /* Do not print anything if we record to the pipe. */
1837 if (STAT_RECORD && perf_stat.data.is_pipe)
1838 return;
1839
1840 if (interval)
1841 print_interval(prefix = buf, ts);
1842 else
1843 print_header(argc, argv);
1844
1845 if (metric_only) {
1846 static int num_print_iv;
1847
1848 if (num_print_iv == 0 && !interval)
1849 print_metric_headers(prefix, false);
1850 if (num_print_iv++ == 25)
1851 num_print_iv = 0;
1852 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1853 fprintf(stat_config.output, "%s", prefix);
1854 }
1855
1856 switch (stat_config.aggr_mode) {
1857 case AGGR_CORE:
1858 case AGGR_SOCKET:
1859 print_aggr(prefix);
1860 break;
1861 case AGGR_THREAD:
1862 evlist__for_each_entry(evsel_list, counter) {
1863 if (is_duration_time(counter))
1864 continue;
1865 print_aggr_thread(counter, prefix);
1866 }
1867 break;
1868 case AGGR_GLOBAL:
1869 evlist__for_each_entry(evsel_list, counter) {
1870 if (is_duration_time(counter))
1871 continue;
1872 print_counter_aggr(counter, prefix);
1873 }
1874 if (metric_only)
1875 fputc('\n', stat_config.output);
1876 break;
1877 case AGGR_NONE:
1878 if (metric_only)
1879 print_no_aggr_metric(prefix);
1880 else {
1881 evlist__for_each_entry(evsel_list, counter) {
1882 if (is_duration_time(counter))
1883 continue;
1884 print_counter(counter, prefix);
1885 }
1886 }
1887 break;
1888 case AGGR_UNSET:
1889 default:
1890 break;
1891 }
1892
1893 if (!interval && !csv_output)
1894 print_footer();
1895
1896 fflush(stat_config.output);
1897 }
1898
1899 static volatile int signr = -1;
1900
1901 static void skip_signal(int signo)
1902 {
1903 if ((child_pid == -1) || stat_config.interval)
1904 done = 1;
1905
1906 signr = signo;
1907 /*
1908 * render child_pid harmless
1909 * won't send SIGTERM to a random
1910 * process in case of race condition
1911 * and fast PID recycling
1912 */
1913 child_pid = -1;
1914 }
1915
1916 static void sig_atexit(void)
1917 {
1918 sigset_t set, oset;
1919
1920 /*
1921 * avoid race condition with SIGCHLD handler
1922 * in skip_signal() which is modifying child_pid
1923 * goal is to avoid send SIGTERM to a random
1924 * process
1925 */
1926 sigemptyset(&set);
1927 sigaddset(&set, SIGCHLD);
1928 sigprocmask(SIG_BLOCK, &set, &oset);
1929
1930 if (child_pid != -1)
1931 kill(child_pid, SIGTERM);
1932
1933 sigprocmask(SIG_SETMASK, &oset, NULL);
1934
1935 if (signr == -1)
1936 return;
1937
1938 signal(signr, SIG_DFL);
1939 kill(getpid(), signr);
1940 }
1941
1942 static int stat__set_big_num(const struct option *opt __maybe_unused,
1943 const char *s __maybe_unused, int unset)
1944 {
1945 big_num_opt = unset ? 0 : 1;
1946 return 0;
1947 }
1948
1949 static int enable_metric_only(const struct option *opt __maybe_unused,
1950 const char *s __maybe_unused, int unset)
1951 {
1952 force_metric_only = true;
1953 metric_only = !unset;
1954 return 0;
1955 }
1956
1957 static int parse_metric_groups(const struct option *opt,
1958 const char *str,
1959 int unset __maybe_unused)
1960 {
1961 return metricgroup__parse_groups(opt, str, &metric_events);
1962 }
1963
1964 static const struct option stat_options[] = {
1965 OPT_BOOLEAN('T', "transaction", &transaction_run,
1966 "hardware transaction statistics"),
1967 OPT_CALLBACK('e', "event", &evsel_list, "event",
1968 "event selector. use 'perf list' to list available events",
1969 parse_events_option),
1970 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1971 "event filter", parse_filter),
1972 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1973 "child tasks do not inherit counters"),
1974 OPT_STRING('p', "pid", &target.pid, "pid",
1975 "stat events on existing process id"),
1976 OPT_STRING('t', "tid", &target.tid, "tid",
1977 "stat events on existing thread id"),
1978 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1979 "system-wide collection from all CPUs"),
1980 OPT_BOOLEAN('g', "group", &group,
1981 "put the counters into a counter group"),
1982 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1983 OPT_INCR('v', "verbose", &verbose,
1984 "be more verbose (show counter open errors, etc)"),
1985 OPT_INTEGER('r', "repeat", &run_count,
1986 "repeat command and print average + stddev (max: 100, forever: 0)"),
1987 OPT_BOOLEAN(0, "table", &walltime_run_table,
1988 "display details about each run (only with -r option)"),
1989 OPT_BOOLEAN('n', "null", &null_run,
1990 "null run - dont start any counters"),
1991 OPT_INCR('d', "detailed", &detailed_run,
1992 "detailed run - start a lot of events"),
1993 OPT_BOOLEAN('S', "sync", &sync_run,
1994 "call sync() before starting a run"),
1995 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1996 "print large numbers with thousands\' separators",
1997 stat__set_big_num),
1998 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1999 "list of cpus to monitor in system-wide"),
2000 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
2001 "disable CPU count aggregation", AGGR_NONE),
2002 OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"),
2003 OPT_STRING('x', "field-separator", &csv_sep, "separator",
2004 "print counts with custom separator"),
2005 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
2006 "monitor event in cgroup name only", parse_cgroups),
2007 OPT_STRING('o', "output", &output_name, "file", "output file name"),
2008 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
2009 OPT_INTEGER(0, "log-fd", &output_fd,
2010 "log output to fd, instead of stderr"),
2011 OPT_STRING(0, "pre", &pre_cmd, "command",
2012 "command to run prior to the measured command"),
2013 OPT_STRING(0, "post", &post_cmd, "command",
2014 "command to run after to the measured command"),
2015 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
2016 "print counts at regular interval in ms "
2017 "(overhead is possible for values <= 100ms)"),
2018 OPT_INTEGER(0, "interval-count", &stat_config.times,
2019 "print counts for fixed number of times"),
2020 OPT_BOOLEAN(0, "interval-clear", &interval_clear,
2021 "clear screen in between new interval"),
2022 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
2023 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
2024 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
2025 "aggregate counts per processor socket", AGGR_SOCKET),
2026 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
2027 "aggregate counts per physical processor core", AGGR_CORE),
2028 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
2029 "aggregate counts per thread", AGGR_THREAD),
2030 OPT_UINTEGER('D', "delay", &initial_delay,
2031 "ms to wait before starting measurement after program start"),
2032 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
2033 "Only print computed metrics. No raw values", enable_metric_only),
2034 OPT_BOOLEAN(0, "topdown", &topdown_run,
2035 "measure topdown level 1 statistics"),
2036 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
2037 "measure SMI cost"),
2038 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
2039 "monitor specified metrics or metric groups (separated by ,)",
2040 parse_metric_groups),
2041 OPT_END()
2042 };
2043
2044 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
2045 {
2046 return cpu_map__get_socket(map, cpu, NULL);
2047 }
2048
2049 static int perf_stat__get_core(struct cpu_map *map, int cpu)
2050 {
2051 return cpu_map__get_core(map, cpu, NULL);
2052 }
2053
2054 static int cpu_map__get_max(struct cpu_map *map)
2055 {
2056 int i, max = -1;
2057
2058 for (i = 0; i < map->nr; i++) {
2059 if (map->map[i] > max)
2060 max = map->map[i];
2061 }
2062
2063 return max;
2064 }
2065
2066 static struct cpu_map *cpus_aggr_map;
2067
2068 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
2069 {
2070 int cpu;
2071
2072 if (idx >= map->nr)
2073 return -1;
2074
2075 cpu = map->map[idx];
2076
2077 if (cpus_aggr_map->map[cpu] == -1)
2078 cpus_aggr_map->map[cpu] = get_id(map, idx);
2079
2080 return cpus_aggr_map->map[cpu];
2081 }
2082
2083 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
2084 {
2085 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
2086 }
2087
2088 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
2089 {
2090 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
2091 }
2092
2093 static int perf_stat_init_aggr_mode(void)
2094 {
2095 int nr;
2096
2097 switch (stat_config.aggr_mode) {
2098 case AGGR_SOCKET:
2099 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
2100 perror("cannot build socket map");
2101 return -1;
2102 }
2103 aggr_get_id = perf_stat__get_socket_cached;
2104 break;
2105 case AGGR_CORE:
2106 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
2107 perror("cannot build core map");
2108 return -1;
2109 }
2110 aggr_get_id = perf_stat__get_core_cached;
2111 break;
2112 case AGGR_NONE:
2113 case AGGR_GLOBAL:
2114 case AGGR_THREAD:
2115 case AGGR_UNSET:
2116 default:
2117 break;
2118 }
2119
2120 /*
2121 * The evsel_list->cpus is the base we operate on,
2122 * taking the highest cpu number to be the size of
2123 * the aggregation translate cpumap.
2124 */
2125 nr = cpu_map__get_max(evsel_list->cpus);
2126 cpus_aggr_map = cpu_map__empty_new(nr + 1);
2127 return cpus_aggr_map ? 0 : -ENOMEM;
2128 }
2129
2130 static void perf_stat__exit_aggr_mode(void)
2131 {
2132 cpu_map__put(aggr_map);
2133 cpu_map__put(cpus_aggr_map);
2134 aggr_map = NULL;
2135 cpus_aggr_map = NULL;
2136 }
2137
2138 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
2139 {
2140 int cpu;
2141
2142 if (idx > map->nr)
2143 return -1;
2144
2145 cpu = map->map[idx];
2146
2147 if (cpu >= env->nr_cpus_avail)
2148 return -1;
2149
2150 return cpu;
2151 }
2152
2153 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
2154 {
2155 struct perf_env *env = data;
2156 int cpu = perf_env__get_cpu(env, map, idx);
2157
2158 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
2159 }
2160
2161 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
2162 {
2163 struct perf_env *env = data;
2164 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
2165
2166 if (cpu != -1) {
2167 int socket_id = env->cpu[cpu].socket_id;
2168
2169 /*
2170 * Encode socket in upper 16 bits
2171 * core_id is relative to socket, and
2172 * we need a global id. So we combine
2173 * socket + core id.
2174 */
2175 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
2176 }
2177
2178 return core;
2179 }
2180
2181 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
2182 struct cpu_map **sockp)
2183 {
2184 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
2185 }
2186
2187 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
2188 struct cpu_map **corep)
2189 {
2190 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
2191 }
2192
2193 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
2194 {
2195 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
2196 }
2197
2198 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
2199 {
2200 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
2201 }
2202
2203 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
2204 {
2205 struct perf_env *env = &st->session->header.env;
2206
2207 switch (stat_config.aggr_mode) {
2208 case AGGR_SOCKET:
2209 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
2210 perror("cannot build socket map");
2211 return -1;
2212 }
2213 aggr_get_id = perf_stat__get_socket_file;
2214 break;
2215 case AGGR_CORE:
2216 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
2217 perror("cannot build core map");
2218 return -1;
2219 }
2220 aggr_get_id = perf_stat__get_core_file;
2221 break;
2222 case AGGR_NONE:
2223 case AGGR_GLOBAL:
2224 case AGGR_THREAD:
2225 case AGGR_UNSET:
2226 default:
2227 break;
2228 }
2229
2230 return 0;
2231 }
2232
2233 static int topdown_filter_events(const char **attr, char **str, bool use_group)
2234 {
2235 int off = 0;
2236 int i;
2237 int len = 0;
2238 char *s;
2239
2240 for (i = 0; attr[i]; i++) {
2241 if (pmu_have_event("cpu", attr[i])) {
2242 len += strlen(attr[i]) + 1;
2243 attr[i - off] = attr[i];
2244 } else
2245 off++;
2246 }
2247 attr[i - off] = NULL;
2248
2249 *str = malloc(len + 1 + 2);
2250 if (!*str)
2251 return -1;
2252 s = *str;
2253 if (i - off == 0) {
2254 *s = 0;
2255 return 0;
2256 }
2257 if (use_group)
2258 *s++ = '{';
2259 for (i = 0; attr[i]; i++) {
2260 strcpy(s, attr[i]);
2261 s += strlen(s);
2262 *s++ = ',';
2263 }
2264 if (use_group) {
2265 s[-1] = '}';
2266 *s = 0;
2267 } else
2268 s[-1] = 0;
2269 return 0;
2270 }
2271
2272 __weak bool arch_topdown_check_group(bool *warn)
2273 {
2274 *warn = false;
2275 return false;
2276 }
2277
2278 __weak void arch_topdown_group_warn(void)
2279 {
2280 }
2281
2282 /*
2283 * Add default attributes, if there were no attributes specified or
2284 * if -d/--detailed, -d -d or -d -d -d is used:
2285 */
2286 static int add_default_attributes(void)
2287 {
2288 int err;
2289 struct perf_event_attr default_attrs0[] = {
2290
2291 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
2292 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
2293 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
2294 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
2295
2296 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
2297 };
2298 struct perf_event_attr frontend_attrs[] = {
2299 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
2300 };
2301 struct perf_event_attr backend_attrs[] = {
2302 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
2303 };
2304 struct perf_event_attr default_attrs1[] = {
2305 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
2306 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
2307 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
2308
2309 };
2310
2311 /*
2312 * Detailed stats (-d), covering the L1 and last level data caches:
2313 */
2314 struct perf_event_attr detailed_attrs[] = {
2315
2316 { .type = PERF_TYPE_HW_CACHE,
2317 .config =
2318 PERF_COUNT_HW_CACHE_L1D << 0 |
2319 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2320 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2321
2322 { .type = PERF_TYPE_HW_CACHE,
2323 .config =
2324 PERF_COUNT_HW_CACHE_L1D << 0 |
2325 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2326 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2327
2328 { .type = PERF_TYPE_HW_CACHE,
2329 .config =
2330 PERF_COUNT_HW_CACHE_LL << 0 |
2331 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2332 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2333
2334 { .type = PERF_TYPE_HW_CACHE,
2335 .config =
2336 PERF_COUNT_HW_CACHE_LL << 0 |
2337 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2338 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2339 };
2340
2341 /*
2342 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2343 */
2344 struct perf_event_attr very_detailed_attrs[] = {
2345
2346 { .type = PERF_TYPE_HW_CACHE,
2347 .config =
2348 PERF_COUNT_HW_CACHE_L1I << 0 |
2349 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2350 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2351
2352 { .type = PERF_TYPE_HW_CACHE,
2353 .config =
2354 PERF_COUNT_HW_CACHE_L1I << 0 |
2355 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2356 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2357
2358 { .type = PERF_TYPE_HW_CACHE,
2359 .config =
2360 PERF_COUNT_HW_CACHE_DTLB << 0 |
2361 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2362 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2363
2364 { .type = PERF_TYPE_HW_CACHE,
2365 .config =
2366 PERF_COUNT_HW_CACHE_DTLB << 0 |
2367 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2368 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2369
2370 { .type = PERF_TYPE_HW_CACHE,
2371 .config =
2372 PERF_COUNT_HW_CACHE_ITLB << 0 |
2373 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2374 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2375
2376 { .type = PERF_TYPE_HW_CACHE,
2377 .config =
2378 PERF_COUNT_HW_CACHE_ITLB << 0 |
2379 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2380 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2381
2382 };
2383
2384 /*
2385 * Very, very detailed stats (-d -d -d), adding prefetch events:
2386 */
2387 struct perf_event_attr very_very_detailed_attrs[] = {
2388
2389 { .type = PERF_TYPE_HW_CACHE,
2390 .config =
2391 PERF_COUNT_HW_CACHE_L1D << 0 |
2392 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2393 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2394
2395 { .type = PERF_TYPE_HW_CACHE,
2396 .config =
2397 PERF_COUNT_HW_CACHE_L1D << 0 |
2398 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2399 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2400 };
2401 struct parse_events_error errinfo;
2402
2403 /* Set attrs if no event is selected and !null_run: */
2404 if (null_run)
2405 return 0;
2406
2407 if (transaction_run) {
2408 /* Handle -T as -M transaction. Once platform specific metrics
2409 * support has been added to the json files, all archictures
2410 * will use this approach. To determine transaction support
2411 * on an architecture test for such a metric name.
2412 */
2413 if (metricgroup__has_metric("transaction")) {
2414 struct option opt = { .value = &evsel_list };
2415
2416 return metricgroup__parse_groups(&opt, "transaction",
2417 &metric_events);
2418 }
2419
2420 if (pmu_have_event("cpu", "cycles-ct") &&
2421 pmu_have_event("cpu", "el-start"))
2422 err = parse_events(evsel_list, transaction_attrs,
2423 &errinfo);
2424 else
2425 err = parse_events(evsel_list,
2426 transaction_limited_attrs,
2427 &errinfo);
2428 if (err) {
2429 fprintf(stderr, "Cannot set up transaction events\n");
2430 parse_events_print_error(&errinfo, transaction_attrs);
2431 return -1;
2432 }
2433 return 0;
2434 }
2435
2436 if (smi_cost) {
2437 int smi;
2438
2439 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2440 fprintf(stderr, "freeze_on_smi is not supported.\n");
2441 return -1;
2442 }
2443
2444 if (!smi) {
2445 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2446 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2447 return -1;
2448 }
2449 smi_reset = true;
2450 }
2451
2452 if (pmu_have_event("msr", "aperf") &&
2453 pmu_have_event("msr", "smi")) {
2454 if (!force_metric_only)
2455 metric_only = true;
2456 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
2457 } else {
2458 fprintf(stderr, "To measure SMI cost, it needs "
2459 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
2460 parse_events_print_error(&errinfo, smi_cost_attrs);
2461 return -1;
2462 }
2463 if (err) {
2464 fprintf(stderr, "Cannot set up SMI cost events\n");
2465 return -1;
2466 }
2467 return 0;
2468 }
2469
2470 if (topdown_run) {
2471 char *str = NULL;
2472 bool warn = false;
2473
2474 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2475 stat_config.aggr_mode != AGGR_CORE) {
2476 pr_err("top down event configuration requires --per-core mode\n");
2477 return -1;
2478 }
2479 stat_config.aggr_mode = AGGR_CORE;
2480 if (nr_cgroups || !target__has_cpu(&target)) {
2481 pr_err("top down event configuration requires system-wide mode (-a)\n");
2482 return -1;
2483 }
2484
2485 if (!force_metric_only)
2486 metric_only = true;
2487 if (topdown_filter_events(topdown_attrs, &str,
2488 arch_topdown_check_group(&warn)) < 0) {
2489 pr_err("Out of memory\n");
2490 return -1;
2491 }
2492 if (topdown_attrs[0] && str) {
2493 if (warn)
2494 arch_topdown_group_warn();
2495 err = parse_events(evsel_list, str, &errinfo);
2496 if (err) {
2497 fprintf(stderr,
2498 "Cannot set up top down events %s: %d\n",
2499 str, err);
2500 free(str);
2501 parse_events_print_error(&errinfo, str);
2502 return -1;
2503 }
2504 } else {
2505 fprintf(stderr, "System does not support topdown\n");
2506 return -1;
2507 }
2508 free(str);
2509 }
2510
2511 if (!evsel_list->nr_entries) {
2512 if (target__has_cpu(&target))
2513 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2514
2515 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2516 return -1;
2517 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2518 if (perf_evlist__add_default_attrs(evsel_list,
2519 frontend_attrs) < 0)
2520 return -1;
2521 }
2522 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2523 if (perf_evlist__add_default_attrs(evsel_list,
2524 backend_attrs) < 0)
2525 return -1;
2526 }
2527 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2528 return -1;
2529 }
2530
2531 /* Detailed events get appended to the event list: */
2532
2533 if (detailed_run < 1)
2534 return 0;
2535
2536 /* Append detailed run extra attributes: */
2537 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2538 return -1;
2539
2540 if (detailed_run < 2)
2541 return 0;
2542
2543 /* Append very detailed run extra attributes: */
2544 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2545 return -1;
2546
2547 if (detailed_run < 3)
2548 return 0;
2549
2550 /* Append very, very detailed run extra attributes: */
2551 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2552 }
2553
2554 static const char * const stat_record_usage[] = {
2555 "perf stat record [<options>]",
2556 NULL,
2557 };
2558
2559 static void init_features(struct perf_session *session)
2560 {
2561 int feat;
2562
2563 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2564 perf_header__set_feat(&session->header, feat);
2565
2566 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2567 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2568 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2569 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2570 }
2571
2572 static int __cmd_record(int argc, const char **argv)
2573 {
2574 struct perf_session *session;
2575 struct perf_data *data = &perf_stat.data;
2576
2577 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2578 PARSE_OPT_STOP_AT_NON_OPTION);
2579
2580 if (output_name)
2581 data->file.path = output_name;
2582
2583 if (run_count != 1 || forever) {
2584 pr_err("Cannot use -r option with perf stat record.\n");
2585 return -1;
2586 }
2587
2588 session = perf_session__new(data, false, NULL);
2589 if (session == NULL) {
2590 pr_err("Perf session creation failed.\n");
2591 return -1;
2592 }
2593
2594 init_features(session);
2595
2596 session->evlist = evsel_list;
2597 perf_stat.session = session;
2598 perf_stat.record = true;
2599 return argc;
2600 }
2601
2602 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2603 union perf_event *event,
2604 struct perf_session *session)
2605 {
2606 struct stat_round_event *stat_round = &event->stat_round;
2607 struct perf_evsel *counter;
2608 struct timespec tsh, *ts = NULL;
2609 const char **argv = session->header.env.cmdline_argv;
2610 int argc = session->header.env.nr_cmdline;
2611
2612 evlist__for_each_entry(evsel_list, counter)
2613 perf_stat_process_counter(&stat_config, counter);
2614
2615 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2616 update_stats(&walltime_nsecs_stats, stat_round->time);
2617
2618 if (stat_config.interval && stat_round->time) {
2619 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2620 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2621 ts = &tsh;
2622 }
2623
2624 print_counters(ts, argc, argv);
2625 return 0;
2626 }
2627
2628 static
2629 int process_stat_config_event(struct perf_tool *tool,
2630 union perf_event *event,
2631 struct perf_session *session __maybe_unused)
2632 {
2633 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2634
2635 perf_event__read_stat_config(&stat_config, &event->stat_config);
2636
2637 if (cpu_map__empty(st->cpus)) {
2638 if (st->aggr_mode != AGGR_UNSET)
2639 pr_warning("warning: processing task data, aggregation mode not set\n");
2640 return 0;
2641 }
2642
2643 if (st->aggr_mode != AGGR_UNSET)
2644 stat_config.aggr_mode = st->aggr_mode;
2645
2646 if (perf_stat.data.is_pipe)
2647 perf_stat_init_aggr_mode();
2648 else
2649 perf_stat_init_aggr_mode_file(st);
2650
2651 return 0;
2652 }
2653
2654 static int set_maps(struct perf_stat *st)
2655 {
2656 if (!st->cpus || !st->threads)
2657 return 0;
2658
2659 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2660 return -EINVAL;
2661
2662 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2663
2664 if (perf_evlist__alloc_stats(evsel_list, true))
2665 return -ENOMEM;
2666
2667 st->maps_allocated = true;
2668 return 0;
2669 }
2670
2671 static
2672 int process_thread_map_event(struct perf_tool *tool,
2673 union perf_event *event,
2674 struct perf_session *session __maybe_unused)
2675 {
2676 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2677
2678 if (st->threads) {
2679 pr_warning("Extra thread map event, ignoring.\n");
2680 return 0;
2681 }
2682
2683 st->threads = thread_map__new_event(&event->thread_map);
2684 if (!st->threads)
2685 return -ENOMEM;
2686
2687 return set_maps(st);
2688 }
2689
2690 static
2691 int process_cpu_map_event(struct perf_tool *tool,
2692 union perf_event *event,
2693 struct perf_session *session __maybe_unused)
2694 {
2695 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2696 struct cpu_map *cpus;
2697
2698 if (st->cpus) {
2699 pr_warning("Extra cpu map event, ignoring.\n");
2700 return 0;
2701 }
2702
2703 cpus = cpu_map__new_data(&event->cpu_map.data);
2704 if (!cpus)
2705 return -ENOMEM;
2706
2707 st->cpus = cpus;
2708 return set_maps(st);
2709 }
2710
2711 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
2712 {
2713 int i;
2714
2715 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
2716 if (!config->stats)
2717 return -1;
2718
2719 config->stats_num = nthreads;
2720
2721 for (i = 0; i < nthreads; i++)
2722 runtime_stat__init(&config->stats[i]);
2723
2724 return 0;
2725 }
2726
2727 static void runtime_stat_delete(struct perf_stat_config *config)
2728 {
2729 int i;
2730
2731 if (!config->stats)
2732 return;
2733
2734 for (i = 0; i < config->stats_num; i++)
2735 runtime_stat__exit(&config->stats[i]);
2736
2737 free(config->stats);
2738 }
2739
2740 static const char * const stat_report_usage[] = {
2741 "perf stat report [<options>]",
2742 NULL,
2743 };
2744
2745 static struct perf_stat perf_stat = {
2746 .tool = {
2747 .attr = perf_event__process_attr,
2748 .event_update = perf_event__process_event_update,
2749 .thread_map = process_thread_map_event,
2750 .cpu_map = process_cpu_map_event,
2751 .stat_config = process_stat_config_event,
2752 .stat = perf_event__process_stat_event,
2753 .stat_round = process_stat_round_event,
2754 },
2755 .aggr_mode = AGGR_UNSET,
2756 };
2757
2758 static int __cmd_report(int argc, const char **argv)
2759 {
2760 struct perf_session *session;
2761 const struct option options[] = {
2762 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2763 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2764 "aggregate counts per processor socket", AGGR_SOCKET),
2765 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2766 "aggregate counts per physical processor core", AGGR_CORE),
2767 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2768 "disable CPU count aggregation", AGGR_NONE),
2769 OPT_END()
2770 };
2771 struct stat st;
2772 int ret;
2773
2774 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2775
2776 if (!input_name || !strlen(input_name)) {
2777 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2778 input_name = "-";
2779 else
2780 input_name = "perf.data";
2781 }
2782
2783 perf_stat.data.file.path = input_name;
2784 perf_stat.data.mode = PERF_DATA_MODE_READ;
2785
2786 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2787 if (session == NULL)
2788 return -1;
2789
2790 perf_stat.session = session;
2791 stat_config.output = stderr;
2792 evsel_list = session->evlist;
2793
2794 ret = perf_session__process_events(session);
2795 if (ret)
2796 return ret;
2797
2798 perf_session__delete(session);
2799 return 0;
2800 }
2801
2802 static void setup_system_wide(int forks)
2803 {
2804 /*
2805 * Make system wide (-a) the default target if
2806 * no target was specified and one of following
2807 * conditions is met:
2808 *
2809 * - there's no workload specified
2810 * - there is workload specified but all requested
2811 * events are system wide events
2812 */
2813 if (!target__none(&target))
2814 return;
2815
2816 if (!forks)
2817 target.system_wide = true;
2818 else {
2819 struct perf_evsel *counter;
2820
2821 evlist__for_each_entry(evsel_list, counter) {
2822 if (!counter->system_wide)
2823 return;
2824 }
2825
2826 if (evsel_list->nr_entries)
2827 target.system_wide = true;
2828 }
2829 }
2830
2831 int cmd_stat(int argc, const char **argv)
2832 {
2833 const char * const stat_usage[] = {
2834 "perf stat [<options>] [<command>]",
2835 NULL
2836 };
2837 int status = -EINVAL, run_idx;
2838 const char *mode;
2839 FILE *output = stderr;
2840 unsigned int interval, timeout;
2841 const char * const stat_subcommands[] = { "record", "report" };
2842
2843 setlocale(LC_ALL, "");
2844
2845 evsel_list = perf_evlist__new();
2846 if (evsel_list == NULL)
2847 return -ENOMEM;
2848
2849 parse_events__shrink_config_terms();
2850 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2851 (const char **) stat_usage,
2852 PARSE_OPT_STOP_AT_NON_OPTION);
2853 perf_stat__collect_metric_expr(evsel_list);
2854 perf_stat__init_shadow_stats();
2855
2856 if (csv_sep) {
2857 csv_output = true;
2858 if (!strcmp(csv_sep, "\\t"))
2859 csv_sep = "\t";
2860 } else
2861 csv_sep = DEFAULT_SEPARATOR;
2862
2863 if (argc && !strncmp(argv[0], "rec", 3)) {
2864 argc = __cmd_record(argc, argv);
2865 if (argc < 0)
2866 return -1;
2867 } else if (argc && !strncmp(argv[0], "rep", 3))
2868 return __cmd_report(argc, argv);
2869
2870 interval = stat_config.interval;
2871 timeout = stat_config.timeout;
2872
2873 /*
2874 * For record command the -o is already taken care of.
2875 */
2876 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2877 output = NULL;
2878
2879 if (output_name && output_fd) {
2880 fprintf(stderr, "cannot use both --output and --log-fd\n");
2881 parse_options_usage(stat_usage, stat_options, "o", 1);
2882 parse_options_usage(NULL, stat_options, "log-fd", 0);
2883 goto out;
2884 }
2885
2886 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2887 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2888 goto out;
2889 }
2890
2891 if (metric_only && run_count > 1) {
2892 fprintf(stderr, "--metric-only is not supported with -r\n");
2893 goto out;
2894 }
2895
2896 if (walltime_run_table && run_count <= 1) {
2897 fprintf(stderr, "--table is only supported with -r\n");
2898 parse_options_usage(stat_usage, stat_options, "r", 1);
2899 parse_options_usage(NULL, stat_options, "table", 0);
2900 goto out;
2901 }
2902
2903 if (output_fd < 0) {
2904 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2905 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2906 goto out;
2907 }
2908
2909 if (!output) {
2910 struct timespec tm;
2911 mode = append_file ? "a" : "w";
2912
2913 output = fopen(output_name, mode);
2914 if (!output) {
2915 perror("failed to create output file");
2916 return -1;
2917 }
2918 clock_gettime(CLOCK_REALTIME, &tm);
2919 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2920 } else if (output_fd > 0) {
2921 mode = append_file ? "a" : "w";
2922 output = fdopen(output_fd, mode);
2923 if (!output) {
2924 perror("Failed opening logfd");
2925 return -errno;
2926 }
2927 }
2928
2929 stat_config.output = output;
2930
2931 /*
2932 * let the spreadsheet do the pretty-printing
2933 */
2934 if (csv_output) {
2935 /* User explicitly passed -B? */
2936 if (big_num_opt == 1) {
2937 fprintf(stderr, "-B option not supported with -x\n");
2938 parse_options_usage(stat_usage, stat_options, "B", 1);
2939 parse_options_usage(NULL, stat_options, "x", 1);
2940 goto out;
2941 } else /* Nope, so disable big number formatting */
2942 big_num = false;
2943 } else if (big_num_opt == 0) /* User passed --no-big-num */
2944 big_num = false;
2945
2946 setup_system_wide(argc);
2947
2948 /*
2949 * Display user/system times only for single
2950 * run and when there's specified tracee.
2951 */
2952 if ((run_count == 1) && target__none(&target))
2953 ru_display = true;
2954
2955 if (run_count < 0) {
2956 pr_err("Run count must be a positive number\n");
2957 parse_options_usage(stat_usage, stat_options, "r", 1);
2958 goto out;
2959 } else if (run_count == 0) {
2960 forever = true;
2961 run_count = 1;
2962 }
2963
2964 if (walltime_run_table) {
2965 walltime_run = zalloc(run_count * sizeof(walltime_run[0]));
2966 if (!walltime_run) {
2967 pr_err("failed to setup -r option");
2968 goto out;
2969 }
2970 }
2971
2972 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2973 !target__has_task(&target)) {
2974 if (!target.system_wide || target.cpu_list) {
2975 fprintf(stderr, "The --per-thread option is only "
2976 "available when monitoring via -p -t -a "
2977 "options or only --per-thread.\n");
2978 parse_options_usage(NULL, stat_options, "p", 1);
2979 parse_options_usage(NULL, stat_options, "t", 1);
2980 goto out;
2981 }
2982 }
2983
2984 /*
2985 * no_aggr, cgroup are for system-wide only
2986 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2987 */
2988 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2989 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2990 !target__has_cpu(&target)) {
2991 fprintf(stderr, "both cgroup and no-aggregation "
2992 "modes only available in system-wide mode\n");
2993
2994 parse_options_usage(stat_usage, stat_options, "G", 1);
2995 parse_options_usage(NULL, stat_options, "A", 1);
2996 parse_options_usage(NULL, stat_options, "a", 1);
2997 goto out;
2998 }
2999
3000 if (add_default_attributes())
3001 goto out;
3002
3003 target__validate(&target);
3004
3005 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
3006 target.per_thread = true;
3007
3008 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
3009 if (target__has_task(&target)) {
3010 pr_err("Problems finding threads of monitor\n");
3011 parse_options_usage(stat_usage, stat_options, "p", 1);
3012 parse_options_usage(NULL, stat_options, "t", 1);
3013 } else if (target__has_cpu(&target)) {
3014 perror("failed to parse CPUs map");
3015 parse_options_usage(stat_usage, stat_options, "C", 1);
3016 parse_options_usage(NULL, stat_options, "a", 1);
3017 }
3018 goto out;
3019 }
3020
3021 /*
3022 * Initialize thread_map with comm names,
3023 * so we could print it out on output.
3024 */
3025 if (stat_config.aggr_mode == AGGR_THREAD) {
3026 thread_map__read_comms(evsel_list->threads);
3027 if (target.system_wide) {
3028 if (runtime_stat_new(&stat_config,
3029 thread_map__nr(evsel_list->threads))) {
3030 goto out;
3031 }
3032 }
3033 }
3034
3035 if (stat_config.times && interval)
3036 interval_count = true;
3037 else if (stat_config.times && !interval) {
3038 pr_err("interval-count option should be used together with "
3039 "interval-print.\n");
3040 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
3041 parse_options_usage(stat_usage, stat_options, "I", 1);
3042 goto out;
3043 }
3044
3045 if (timeout && timeout < 100) {
3046 if (timeout < 10) {
3047 pr_err("timeout must be >= 10ms.\n");
3048 parse_options_usage(stat_usage, stat_options, "timeout", 0);
3049 goto out;
3050 } else
3051 pr_warning("timeout < 100ms. "
3052 "The overhead percentage could be high in some cases. "
3053 "Please proceed with caution.\n");
3054 }
3055 if (timeout && interval) {
3056 pr_err("timeout option is not supported with interval-print.\n");
3057 parse_options_usage(stat_usage, stat_options, "timeout", 0);
3058 parse_options_usage(stat_usage, stat_options, "I", 1);
3059 goto out;
3060 }
3061
3062 if (perf_evlist__alloc_stats(evsel_list, interval))
3063 goto out;
3064
3065 if (perf_stat_init_aggr_mode())
3066 goto out;
3067
3068 /*
3069 * We dont want to block the signals - that would cause
3070 * child tasks to inherit that and Ctrl-C would not work.
3071 * What we want is for Ctrl-C to work in the exec()-ed
3072 * task, but being ignored by perf stat itself:
3073 */
3074 atexit(sig_atexit);
3075 if (!forever)
3076 signal(SIGINT, skip_signal);
3077 signal(SIGCHLD, skip_signal);
3078 signal(SIGALRM, skip_signal);
3079 signal(SIGABRT, skip_signal);
3080
3081 status = 0;
3082 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
3083 if (run_count != 1 && verbose > 0)
3084 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
3085 run_idx + 1);
3086
3087 status = run_perf_stat(argc, argv, run_idx);
3088 if (forever && status != -1) {
3089 print_counters(NULL, argc, argv);
3090 perf_stat__reset_stats();
3091 }
3092 }
3093
3094 if (!forever && status != -1 && !interval)
3095 print_counters(NULL, argc, argv);
3096
3097 if (STAT_RECORD) {
3098 /*
3099 * We synthesize the kernel mmap record just so that older tools
3100 * don't emit warnings about not being able to resolve symbols
3101 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
3102 * a saner message about no samples being in the perf.data file.
3103 *
3104 * This also serves to suppress a warning about f_header.data.size == 0
3105 * in header.c at the moment 'perf stat record' gets introduced, which
3106 * is not really needed once we start adding the stat specific PERF_RECORD_
3107 * records, but the need to suppress the kptr_restrict messages in older
3108 * tools remain -acme
3109 */
3110 int fd = perf_data__fd(&perf_stat.data);
3111 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
3112 process_synthesized_event,
3113 &perf_stat.session->machines.host);
3114 if (err) {
3115 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
3116 "older tools may produce warnings about this file\n.");
3117 }
3118
3119 if (!interval) {
3120 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
3121 pr_err("failed to write stat round event\n");
3122 }
3123
3124 if (!perf_stat.data.is_pipe) {
3125 perf_stat.session->header.data_size += perf_stat.bytes_written;
3126 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
3127 }
3128
3129 perf_session__delete(perf_stat.session);
3130 }
3131
3132 perf_stat__exit_aggr_mode();
3133 perf_evlist__free_stats(evsel_list);
3134 out:
3135 free(walltime_run);
3136
3137 if (smi_cost && smi_reset)
3138 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
3139
3140 perf_evlist__delete(evsel_list);
3141
3142 runtime_stat_delete(&stat_config);
3143
3144 return status;
3145 }
(deprecated) Btrfs devel tree