projects / deliverable / linux.git / blob
? search:


fabcadba1f19ff1889e954328fb356cf31df5142
[deliverable/linux.git] / tools / perf / builtin-stat.c
1 /*
2 * builtin-stat.c
3 *
4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
6 *
7 * Sample output:
8
9 $ perf stat ./hackbench 10
10
11 Time: 0.118
12
13 Performance counter stats for './hackbench 10':
14
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
26
27 0.154822978 seconds time elapsed
28
29 *
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
31 *
32 * Improvements and fixes by:
33 *
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
40 *
41 * Released under the GPL v2. (and only v2, not any later version)
42 */
43
44 #include "perf.h"
45 #include "builtin.h"
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include <subcmd/parse-options.h>
49 #include "util/parse-events.h"
50 #include "util/pmu.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/color.h"
56 #include "util/stat.h"
57 #include "util/header.h"
58 #include "util/cpumap.h"
59 #include "util/thread.h"
60 #include "util/thread_map.h"
61 #include "util/counts.h"
62 #include "util/session.h"
63 #include "util/tool.h"
64 #include "asm/bug.h"
65
66 #include <stdlib.h>
67 #include <sys/prctl.h>
68 #include <locale.h>
69
70 #define DEFAULT_SEPARATOR " "
71 #define CNTR_NOT_SUPPORTED "<not supported>"
72 #define CNTR_NOT_COUNTED "<not counted>"
73
74 static void print_counters(struct timespec *ts, int argc, const char **argv);
75
76 /* Default events used for perf stat -T */
77 static const char *transaction_attrs = {
78 "task-clock,"
79 "{"
80 "instructions,"
81 "cycles,"
82 "cpu/cycles-t/,"
83 "cpu/tx-start/,"
84 "cpu/el-start/,"
85 "cpu/cycles-ct/"
86 "}"
87 };
88
89 /* More limited version when the CPU does not have all events. */
90 static const char * transaction_limited_attrs = {
91 "task-clock,"
92 "{"
93 "instructions,"
94 "cycles,"
95 "cpu/cycles-t/,"
96 "cpu/tx-start/"
97 "}"
98 };
99
100 static struct perf_evlist *evsel_list;
101
102 static struct target target = {
103 .uid = UINT_MAX,
104 };
105
106 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
107
108 static int run_count = 1;
109 static bool no_inherit = false;
110 static volatile pid_t child_pid = -1;
111 static bool null_run = false;
112 static int detailed_run = 0;
113 static bool transaction_run;
114 static bool big_num = true;
115 static int big_num_opt = -1;
116 static const char *csv_sep = NULL;
117 static bool csv_output = false;
118 static bool group = false;
119 static const char *pre_cmd = NULL;
120 static const char *post_cmd = NULL;
121 static bool sync_run = false;
122 static unsigned int initial_delay = 0;
123 static unsigned int unit_width = 4; /* strlen("unit") */
124 static bool forever = false;
125 static struct timespec ref_time;
126 static struct cpu_map *aggr_map;
127 static aggr_get_id_t aggr_get_id;
128 static bool append_file;
129 static const char *output_name;
130 static int output_fd;
131
132 struct perf_stat {
133 bool record;
134 struct perf_data_file file;
135 struct perf_session *session;
136 u64 bytes_written;
137 struct perf_tool tool;
138 bool maps_allocated;
139 struct cpu_map *cpus;
140 struct thread_map *threads;
141 enum aggr_mode aggr_mode;
142 };
143
144 static struct perf_stat perf_stat;
145 #define STAT_RECORD perf_stat.record
146
147 static volatile int done = 0;
148
149 static struct perf_stat_config stat_config = {
150 .aggr_mode = AGGR_GLOBAL,
151 .scale = true,
152 };
153
154 static inline void diff_timespec(struct timespec *r, struct timespec *a,
155 struct timespec *b)
156 {
157 r->tv_sec = a->tv_sec - b->tv_sec;
158 if (a->tv_nsec < b->tv_nsec) {
159 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
160 r->tv_sec--;
161 } else {
162 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
163 }
164 }
165
166 static void perf_stat__reset_stats(void)
167 {
168 perf_evlist__reset_stats(evsel_list);
169 perf_stat__reset_shadow_stats();
170 }
171
172 static int create_perf_stat_counter(struct perf_evsel *evsel)
173 {
174 struct perf_event_attr *attr = &evsel->attr;
175
176 if (stat_config.scale)
177 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
178 PERF_FORMAT_TOTAL_TIME_RUNNING;
179
180 attr->inherit = !no_inherit;
181
182 /*
183 * Some events get initialized with sample_(period/type) set,
184 * like tracepoints. Clear it up for counting.
185 */
186 attr->sample_period = 0;
187
188 /*
189 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
190 * while avoiding that older tools show confusing messages.
191 *
192 * However for pipe sessions we need to keep it zero,
193 * because script's perf_evsel__check_attr is triggered
194 * by attr->sample_type != 0, and we can't run it on
195 * stat sessions.
196 */
197 if (!(STAT_RECORD && perf_stat.file.is_pipe))
198 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
199
200 /*
201 * Disabling all counters initially, they will be enabled
202 * either manually by us or by kernel via enable_on_exec
203 * set later.
204 */
205 if (perf_evsel__is_group_leader(evsel)) {
206 attr->disabled = 1;
207
208 /*
209 * In case of initial_delay we enable tracee
210 * events manually.
211 */
212 if (target__none(&target) && !initial_delay)
213 attr->enable_on_exec = 1;
214 }
215
216 if (target__has_cpu(&target))
217 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
218
219 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
220 }
221
222 /*
223 * Does the counter have nsecs as a unit?
224 */
225 static inline int nsec_counter(struct perf_evsel *evsel)
226 {
227 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
228 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
229 return 1;
230
231 return 0;
232 }
233
234 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
235 union perf_event *event,
236 struct perf_sample *sample __maybe_unused,
237 struct machine *machine __maybe_unused)
238 {
239 if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
240 pr_err("failed to write perf data, error: %m\n");
241 return -1;
242 }
243
244 perf_stat.bytes_written += event->header.size;
245 return 0;
246 }
247
248 static int write_stat_round_event(u64 tm, u64 type)
249 {
250 return perf_event__synthesize_stat_round(NULL, tm, type,
251 process_synthesized_event,
252 NULL);
253 }
254
255 #define WRITE_STAT_ROUND_EVENT(time, interval) \
256 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
257
258 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
259
260 static int
261 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
262 struct perf_counts_values *count)
263 {
264 struct perf_sample_id *sid = SID(counter, cpu, thread);
265
266 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
267 process_synthesized_event, NULL);
268 }
269
270 /*
271 * Read out the results of a single counter:
272 * do not aggregate counts across CPUs in system-wide mode
273 */
274 static int read_counter(struct perf_evsel *counter)
275 {
276 int nthreads = thread_map__nr(evsel_list->threads);
277 int ncpus = perf_evsel__nr_cpus(counter);
278 int cpu, thread;
279
280 if (!counter->supported)
281 return -ENOENT;
282
283 if (counter->system_wide)
284 nthreads = 1;
285
286 for (thread = 0; thread < nthreads; thread++) {
287 for (cpu = 0; cpu < ncpus; cpu++) {
288 struct perf_counts_values *count;
289
290 count = perf_counts(counter->counts, cpu, thread);
291 if (perf_evsel__read(counter, cpu, thread, count))
292 return -1;
293
294 if (STAT_RECORD) {
295 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
296 pr_err("failed to write stat event\n");
297 return -1;
298 }
299 }
300 }
301 }
302
303 return 0;
304 }
305
306 static void read_counters(bool close_counters)
307 {
308 struct perf_evsel *counter;
309
310 evlist__for_each(evsel_list, counter) {
311 if (read_counter(counter))
312 pr_debug("failed to read counter %s\n", counter->name);
313
314 if (perf_stat_process_counter(&stat_config, counter))
315 pr_warning("failed to process counter %s\n", counter->name);
316
317 if (close_counters) {
318 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
319 thread_map__nr(evsel_list->threads));
320 }
321 }
322 }
323
324 static void process_interval(void)
325 {
326 struct timespec ts, rs;
327
328 read_counters(false);
329
330 clock_gettime(CLOCK_MONOTONIC, &ts);
331 diff_timespec(&rs, &ts, &ref_time);
332
333 if (STAT_RECORD) {
334 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSECS_PER_SEC + rs.tv_nsec, INTERVAL))
335 pr_err("failed to write stat round event\n");
336 }
337
338 print_counters(&rs, 0, NULL);
339 }
340
341 static void enable_counters(void)
342 {
343 if (initial_delay)
344 usleep(initial_delay * 1000);
345
346 /*
347 * We need to enable counters only if:
348 * - we don't have tracee (attaching to task or cpu)
349 * - we have initial delay configured
350 */
351 if (!target__none(&target) || initial_delay)
352 perf_evlist__enable(evsel_list);
353 }
354
355 static volatile int workload_exec_errno;
356
357 /*
358 * perf_evlist__prepare_workload will send a SIGUSR1
359 * if the fork fails, since we asked by setting its
360 * want_signal to true.
361 */
362 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
363 void *ucontext __maybe_unused)
364 {
365 workload_exec_errno = info->si_value.sival_int;
366 }
367
368 static bool has_unit(struct perf_evsel *counter)
369 {
370 return counter->unit && *counter->unit;
371 }
372
373 static bool has_scale(struct perf_evsel *counter)
374 {
375 return counter->scale != 1;
376 }
377
378 static int perf_stat_synthesize_config(bool is_pipe)
379 {
380 struct perf_evsel *counter;
381 int err;
382
383 if (is_pipe) {
384 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
385 process_synthesized_event);
386 if (err < 0) {
387 pr_err("Couldn't synthesize attrs.\n");
388 return err;
389 }
390 }
391
392 /*
393 * Synthesize other events stuff not carried within
394 * attr event - unit, scale, name
395 */
396 evlist__for_each(evsel_list, counter) {
397 if (!counter->supported)
398 continue;
399
400 /*
401 * Synthesize unit and scale only if it's defined.
402 */
403 if (has_unit(counter)) {
404 err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
405 if (err < 0) {
406 pr_err("Couldn't synthesize evsel unit.\n");
407 return err;
408 }
409 }
410
411 if (has_scale(counter)) {
412 err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
413 if (err < 0) {
414 pr_err("Couldn't synthesize evsel scale.\n");
415 return err;
416 }
417 }
418
419 if (counter->own_cpus) {
420 err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
421 if (err < 0) {
422 pr_err("Couldn't synthesize evsel scale.\n");
423 return err;
424 }
425 }
426
427 /*
428 * Name is needed only for pipe output,
429 * perf.data carries event names.
430 */
431 if (is_pipe) {
432 err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
433 if (err < 0) {
434 pr_err("Couldn't synthesize evsel name.\n");
435 return err;
436 }
437 }
438 }
439
440 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
441 process_synthesized_event,
442 NULL);
443 if (err < 0) {
444 pr_err("Couldn't synthesize thread map.\n");
445 return err;
446 }
447
448 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
449 process_synthesized_event, NULL);
450 if (err < 0) {
451 pr_err("Couldn't synthesize thread map.\n");
452 return err;
453 }
454
455 err = perf_event__synthesize_stat_config(NULL, &stat_config,
456 process_synthesized_event, NULL);
457 if (err < 0) {
458 pr_err("Couldn't synthesize config.\n");
459 return err;
460 }
461
462 return 0;
463 }
464
465 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
466
467 static int __store_counter_ids(struct perf_evsel *counter,
468 struct cpu_map *cpus,
469 struct thread_map *threads)
470 {
471 int cpu, thread;
472
473 for (cpu = 0; cpu < cpus->nr; cpu++) {
474 for (thread = 0; thread < threads->nr; thread++) {
475 int fd = FD(counter, cpu, thread);
476
477 if (perf_evlist__id_add_fd(evsel_list, counter,
478 cpu, thread, fd) < 0)
479 return -1;
480 }
481 }
482
483 return 0;
484 }
485
486 static int store_counter_ids(struct perf_evsel *counter)
487 {
488 struct cpu_map *cpus = counter->cpus;
489 struct thread_map *threads = counter->threads;
490
491 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
492 return -ENOMEM;
493
494 return __store_counter_ids(counter, cpus, threads);
495 }
496
497 static int __run_perf_stat(int argc, const char **argv)
498 {
499 int interval = stat_config.interval;
500 char msg[512];
501 unsigned long long t0, t1;
502 struct perf_evsel *counter;
503 struct timespec ts;
504 size_t l;
505 int status = 0;
506 const bool forks = (argc > 0);
507 bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
508
509 if (interval) {
510 ts.tv_sec = interval / 1000;
511 ts.tv_nsec = (interval % 1000) * 1000000;
512 } else {
513 ts.tv_sec = 1;
514 ts.tv_nsec = 0;
515 }
516
517 if (forks) {
518 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
519 workload_exec_failed_signal) < 0) {
520 perror("failed to prepare workload");
521 return -1;
522 }
523 child_pid = evsel_list->workload.pid;
524 }
525
526 if (group)
527 perf_evlist__set_leader(evsel_list);
528
529 evlist__for_each(evsel_list, counter) {
530 if (create_perf_stat_counter(counter) < 0) {
531 /*
532 * PPC returns ENXIO for HW counters until 2.6.37
533 * (behavior changed with commit b0a873e).
534 */
535 if (errno == EINVAL || errno == ENOSYS ||
536 errno == ENOENT || errno == EOPNOTSUPP ||
537 errno == ENXIO) {
538 if (verbose)
539 ui__warning("%s event is not supported by the kernel.\n",
540 perf_evsel__name(counter));
541 counter->supported = false;
542
543 if ((counter->leader != counter) ||
544 !(counter->leader->nr_members > 1))
545 continue;
546 }
547
548 perf_evsel__open_strerror(counter, &target,
549 errno, msg, sizeof(msg));
550 ui__error("%s\n", msg);
551
552 if (child_pid != -1)
553 kill(child_pid, SIGTERM);
554
555 return -1;
556 }
557 counter->supported = true;
558
559 l = strlen(counter->unit);
560 if (l > unit_width)
561 unit_width = l;
562
563 if (STAT_RECORD && store_counter_ids(counter))
564 return -1;
565 }
566
567 if (perf_evlist__apply_filters(evsel_list, &counter)) {
568 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
569 counter->filter, perf_evsel__name(counter), errno,
570 strerror_r(errno, msg, sizeof(msg)));
571 return -1;
572 }
573
574 if (STAT_RECORD) {
575 int err, fd = perf_data_file__fd(&perf_stat.file);
576
577 if (is_pipe) {
578 err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
579 } else {
580 err = perf_session__write_header(perf_stat.session, evsel_list,
581 fd, false);
582 }
583
584 if (err < 0)
585 return err;
586
587 err = perf_stat_synthesize_config(is_pipe);
588 if (err < 0)
589 return err;
590 }
591
592 /*
593 * Enable counters and exec the command:
594 */
595 t0 = rdclock();
596 clock_gettime(CLOCK_MONOTONIC, &ref_time);
597
598 if (forks) {
599 perf_evlist__start_workload(evsel_list);
600 enable_counters();
601
602 if (interval) {
603 while (!waitpid(child_pid, &status, WNOHANG)) {
604 nanosleep(&ts, NULL);
605 process_interval();
606 }
607 }
608 wait(&status);
609
610 if (workload_exec_errno) {
611 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
612 pr_err("Workload failed: %s\n", emsg);
613 return -1;
614 }
615
616 if (WIFSIGNALED(status))
617 psignal(WTERMSIG(status), argv[0]);
618 } else {
619 enable_counters();
620 while (!done) {
621 nanosleep(&ts, NULL);
622 if (interval)
623 process_interval();
624 }
625 }
626
627 t1 = rdclock();
628
629 update_stats(&walltime_nsecs_stats, t1 - t0);
630
631 read_counters(true);
632
633 return WEXITSTATUS(status);
634 }
635
636 static int run_perf_stat(int argc, const char **argv)
637 {
638 int ret;
639
640 if (pre_cmd) {
641 ret = system(pre_cmd);
642 if (ret)
643 return ret;
644 }
645
646 if (sync_run)
647 sync();
648
649 ret = __run_perf_stat(argc, argv);
650 if (ret)
651 return ret;
652
653 if (post_cmd) {
654 ret = system(post_cmd);
655 if (ret)
656 return ret;
657 }
658
659 return ret;
660 }
661
662 static void print_running(u64 run, u64 ena)
663 {
664 if (csv_output) {
665 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
666 csv_sep,
667 run,
668 csv_sep,
669 ena ? 100.0 * run / ena : 100.0);
670 } else if (run != ena) {
671 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
672 }
673 }
674
675 static void print_noise_pct(double total, double avg)
676 {
677 double pct = rel_stddev_stats(total, avg);
678
679 if (csv_output)
680 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
681 else if (pct)
682 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
683 }
684
685 static void print_noise(struct perf_evsel *evsel, double avg)
686 {
687 struct perf_stat_evsel *ps;
688
689 if (run_count == 1)
690 return;
691
692 ps = evsel->priv;
693 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
694 }
695
696 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
697 {
698 switch (stat_config.aggr_mode) {
699 case AGGR_CORE:
700 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
701 cpu_map__id_to_socket(id),
702 csv_output ? 0 : -8,
703 cpu_map__id_to_cpu(id),
704 csv_sep,
705 csv_output ? 0 : 4,
706 nr,
707 csv_sep);
708 break;
709 case AGGR_SOCKET:
710 fprintf(stat_config.output, "S%*d%s%*d%s",
711 csv_output ? 0 : -5,
712 id,
713 csv_sep,
714 csv_output ? 0 : 4,
715 nr,
716 csv_sep);
717 break;
718 case AGGR_NONE:
719 fprintf(stat_config.output, "CPU%*d%s",
720 csv_output ? 0 : -4,
721 perf_evsel__cpus(evsel)->map[id], csv_sep);
722 break;
723 case AGGR_THREAD:
724 fprintf(stat_config.output, "%*s-%*d%s",
725 csv_output ? 0 : 16,
726 thread_map__comm(evsel->threads, id),
727 csv_output ? 0 : -8,
728 thread_map__pid(evsel->threads, id),
729 csv_sep);
730 break;
731 case AGGR_GLOBAL:
732 case AGGR_UNSET:
733 default:
734 break;
735 }
736 }
737
738 struct outstate {
739 FILE *fh;
740 bool newline;
741 };
742
743 #define METRIC_LEN 35
744
745 static void new_line_std(void *ctx)
746 {
747 struct outstate *os = ctx;
748
749 os->newline = true;
750 }
751
752 static void do_new_line_std(struct outstate *os)
753 {
754 fputc('\n', os->fh);
755 if (stat_config.aggr_mode == AGGR_NONE)
756 fprintf(os->fh, " ");
757 if (stat_config.aggr_mode == AGGR_CORE)
758 fprintf(os->fh, " ");
759 if (stat_config.aggr_mode == AGGR_SOCKET)
760 fprintf(os->fh, " ");
761 fprintf(os->fh, " ");
762 }
763
764 static void print_metric_std(void *ctx, const char *color, const char *fmt,
765 const char *unit, double val)
766 {
767 struct outstate *os = ctx;
768 FILE *out = os->fh;
769 int n;
770 bool newline = os->newline;
771
772 os->newline = false;
773
774 if (unit == NULL || fmt == NULL) {
775 fprintf(out, "%-*s", METRIC_LEN, "");
776 return;
777 }
778
779 if (newline)
780 do_new_line_std(os);
781
782 n = fprintf(out, " # ");
783 if (color)
784 n += color_fprintf(out, color, fmt, val);
785 else
786 n += fprintf(out, fmt, val);
787 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
788 }
789
790 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
791 {
792 FILE *output = stat_config.output;
793 double msecs = avg / 1e6;
794 const char *fmt_v, *fmt_n;
795 char name[25];
796
797 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
798 fmt_n = csv_output ? "%s" : "%-25s";
799
800 aggr_printout(evsel, id, nr);
801
802 scnprintf(name, sizeof(name), "%s%s",
803 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
804
805 fprintf(output, fmt_v, msecs, csv_sep);
806
807 if (csv_output)
808 fprintf(output, "%s%s", evsel->unit, csv_sep);
809 else
810 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
811
812 fprintf(output, fmt_n, name);
813
814 if (evsel->cgrp)
815 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
816 }
817
818 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
819 {
820 FILE *output = stat_config.output;
821 double sc = evsel->scale;
822 const char *fmt;
823
824 if (csv_output) {
825 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
826 } else {
827 if (big_num)
828 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
829 else
830 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
831 }
832
833 aggr_printout(evsel, id, nr);
834
835 fprintf(output, fmt, avg, csv_sep);
836
837 if (evsel->unit)
838 fprintf(output, "%-*s%s",
839 csv_output ? 0 : unit_width,
840 evsel->unit, csv_sep);
841
842 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
843
844 if (evsel->cgrp)
845 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
846 }
847
848 static void printout(int id, int nr, struct perf_evsel *counter, double uval)
849 {
850 struct outstate os = { .fh = stat_config.output };
851 struct perf_stat_output_ctx out;
852 print_metric_t pm = print_metric_std;
853 void (*nl)(void *);
854
855 nl = new_line_std;
856
857 if (nsec_counter(counter))
858 nsec_printout(id, nr, counter, uval);
859 else
860 abs_printout(id, nr, counter, uval);
861
862 out.print_metric = pm;
863 out.new_line = nl;
864 out.ctx = &os;
865
866 if (!csv_output && !stat_config.interval)
867 perf_stat__print_shadow_stats(counter, uval,
868 stat_config.aggr_mode == AGGR_GLOBAL ? 0 :
869 cpu_map__id_to_cpu(id),
870 &out);
871 }
872
873 static void print_aggr(char *prefix)
874 {
875 FILE *output = stat_config.output;
876 struct perf_evsel *counter;
877 int cpu, s, s2, id, nr;
878 double uval;
879 u64 ena, run, val;
880
881 if (!(aggr_map || aggr_get_id))
882 return;
883
884 for (s = 0; s < aggr_map->nr; s++) {
885 id = aggr_map->map[s];
886 evlist__for_each(evsel_list, counter) {
887 val = ena = run = 0;
888 nr = 0;
889 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
890 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
891 if (s2 != id)
892 continue;
893 val += perf_counts(counter->counts, cpu, 0)->val;
894 ena += perf_counts(counter->counts, cpu, 0)->ena;
895 run += perf_counts(counter->counts, cpu, 0)->run;
896 nr++;
897 }
898 if (prefix)
899 fprintf(output, "%s", prefix);
900
901 if (run == 0 || ena == 0) {
902 aggr_printout(counter, id, nr);
903
904 fprintf(output, "%*s%s",
905 csv_output ? 0 : 18,
906 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
907 csv_sep);
908
909 fprintf(output, "%-*s%s",
910 csv_output ? 0 : unit_width,
911 counter->unit, csv_sep);
912
913 fprintf(output, "%*s",
914 csv_output ? 0 : -25,
915 perf_evsel__name(counter));
916
917 if (counter->cgrp)
918 fprintf(output, "%s%s",
919 csv_sep, counter->cgrp->name);
920
921 print_running(run, ena);
922 fputc('\n', output);
923 continue;
924 }
925 uval = val * counter->scale;
926 printout(id, nr, counter, uval);
927 if (!csv_output)
928 print_noise(counter, 1.0);
929
930 print_running(run, ena);
931 fputc('\n', output);
932 }
933 }
934 }
935
936 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
937 {
938 FILE *output = stat_config.output;
939 int nthreads = thread_map__nr(counter->threads);
940 int ncpus = cpu_map__nr(counter->cpus);
941 int cpu, thread;
942 double uval;
943
944 for (thread = 0; thread < nthreads; thread++) {
945 u64 ena = 0, run = 0, val = 0;
946
947 for (cpu = 0; cpu < ncpus; cpu++) {
948 val += perf_counts(counter->counts, cpu, thread)->val;
949 ena += perf_counts(counter->counts, cpu, thread)->ena;
950 run += perf_counts(counter->counts, cpu, thread)->run;
951 }
952
953 if (prefix)
954 fprintf(output, "%s", prefix);
955
956 uval = val * counter->scale;
957 printout(thread, 0, counter, uval);
958
959 if (!csv_output)
960 print_noise(counter, 1.0);
961
962 print_running(run, ena);
963 fputc('\n', output);
964 }
965 }
966
967 /*
968 * Print out the results of a single counter:
969 * aggregated counts in system-wide mode
970 */
971 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
972 {
973 FILE *output = stat_config.output;
974 struct perf_stat_evsel *ps = counter->priv;
975 double avg = avg_stats(&ps->res_stats[0]);
976 int scaled = counter->counts->scaled;
977 double uval;
978 double avg_enabled, avg_running;
979
980 avg_enabled = avg_stats(&ps->res_stats[1]);
981 avg_running = avg_stats(&ps->res_stats[2]);
982
983 if (prefix)
984 fprintf(output, "%s", prefix);
985
986 if (scaled == -1 || !counter->supported) {
987 fprintf(output, "%*s%s",
988 csv_output ? 0 : 18,
989 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
990 csv_sep);
991 fprintf(output, "%-*s%s",
992 csv_output ? 0 : unit_width,
993 counter->unit, csv_sep);
994 fprintf(output, "%*s",
995 csv_output ? 0 : -25,
996 perf_evsel__name(counter));
997
998 if (counter->cgrp)
999 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1000
1001 print_running(avg_running, avg_enabled);
1002 fputc('\n', output);
1003 return;
1004 }
1005
1006 uval = avg * counter->scale;
1007 printout(-1, 0, counter, uval);
1008
1009 print_noise(counter, avg);
1010
1011 print_running(avg_running, avg_enabled);
1012 fprintf(output, "\n");
1013 }
1014
1015 /*
1016 * Print out the results of a single counter:
1017 * does not use aggregated count in system-wide
1018 */
1019 static void print_counter(struct perf_evsel *counter, char *prefix)
1020 {
1021 FILE *output = stat_config.output;
1022 u64 ena, run, val;
1023 double uval;
1024 int cpu;
1025
1026 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1027 val = perf_counts(counter->counts, cpu, 0)->val;
1028 ena = perf_counts(counter->counts, cpu, 0)->ena;
1029 run = perf_counts(counter->counts, cpu, 0)->run;
1030
1031 if (prefix)
1032 fprintf(output, "%s", prefix);
1033
1034 if (run == 0 || ena == 0) {
1035 fprintf(output, "CPU%*d%s%*s%s",
1036 csv_output ? 0 : -4,
1037 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1038 csv_output ? 0 : 18,
1039 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1040 csv_sep);
1041
1042 fprintf(output, "%-*s%s",
1043 csv_output ? 0 : unit_width,
1044 counter->unit, csv_sep);
1045
1046 fprintf(output, "%*s",
1047 csv_output ? 0 : -25,
1048 perf_evsel__name(counter));
1049
1050 if (counter->cgrp)
1051 fprintf(output, "%s%s",
1052 csv_sep, counter->cgrp->name);
1053
1054 print_running(run, ena);
1055 fputc('\n', output);
1056 continue;
1057 }
1058
1059 uval = val * counter->scale;
1060 printout(cpu, 0, counter, uval);
1061 if (!csv_output)
1062 print_noise(counter, 1.0);
1063 print_running(run, ena);
1064
1065 fputc('\n', output);
1066 }
1067 }
1068
1069 static void print_interval(char *prefix, struct timespec *ts)
1070 {
1071 FILE *output = stat_config.output;
1072 static int num_print_interval;
1073
1074 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1075
1076 if (num_print_interval == 0 && !csv_output) {
1077 switch (stat_config.aggr_mode) {
1078 case AGGR_SOCKET:
1079 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
1080 break;
1081 case AGGR_CORE:
1082 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
1083 break;
1084 case AGGR_NONE:
1085 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
1086 break;
1087 case AGGR_THREAD:
1088 fprintf(output, "# time comm-pid counts %*s events\n", unit_width, "unit");
1089 break;
1090 case AGGR_GLOBAL:
1091 default:
1092 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
1093 case AGGR_UNSET:
1094 break;
1095 }
1096 }
1097
1098 if (++num_print_interval == 25)
1099 num_print_interval = 0;
1100 }
1101
1102 static void print_header(int argc, const char **argv)
1103 {
1104 FILE *output = stat_config.output;
1105 int i;
1106
1107 fflush(stdout);
1108
1109 if (!csv_output) {
1110 fprintf(output, "\n");
1111 fprintf(output, " Performance counter stats for ");
1112 if (target.system_wide)
1113 fprintf(output, "\'system wide");
1114 else if (target.cpu_list)
1115 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1116 else if (!target__has_task(&target)) {
1117 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1118 for (i = 1; argv && (i < argc); i++)
1119 fprintf(output, " %s", argv[i]);
1120 } else if (target.pid)
1121 fprintf(output, "process id \'%s", target.pid);
1122 else
1123 fprintf(output, "thread id \'%s", target.tid);
1124
1125 fprintf(output, "\'");
1126 if (run_count > 1)
1127 fprintf(output, " (%d runs)", run_count);
1128 fprintf(output, ":\n\n");
1129 }
1130 }
1131
1132 static void print_footer(void)
1133 {
1134 FILE *output = stat_config.output;
1135
1136 if (!null_run)
1137 fprintf(output, "\n");
1138 fprintf(output, " %17.9f seconds time elapsed",
1139 avg_stats(&walltime_nsecs_stats)/1e9);
1140 if (run_count > 1) {
1141 fprintf(output, " ");
1142 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1143 avg_stats(&walltime_nsecs_stats));
1144 }
1145 fprintf(output, "\n\n");
1146 }
1147
1148 static void print_counters(struct timespec *ts, int argc, const char **argv)
1149 {
1150 int interval = stat_config.interval;
1151 struct perf_evsel *counter;
1152 char buf[64], *prefix = NULL;
1153
1154 /* Do not print anything if we record to the pipe. */
1155 if (STAT_RECORD && perf_stat.file.is_pipe)
1156 return;
1157
1158 if (interval)
1159 print_interval(prefix = buf, ts);
1160 else
1161 print_header(argc, argv);
1162
1163 switch (stat_config.aggr_mode) {
1164 case AGGR_CORE:
1165 case AGGR_SOCKET:
1166 print_aggr(prefix);
1167 break;
1168 case AGGR_THREAD:
1169 evlist__for_each(evsel_list, counter)
1170 print_aggr_thread(counter, prefix);
1171 break;
1172 case AGGR_GLOBAL:
1173 evlist__for_each(evsel_list, counter)
1174 print_counter_aggr(counter, prefix);
1175 break;
1176 case AGGR_NONE:
1177 evlist__for_each(evsel_list, counter)
1178 print_counter(counter, prefix);
1179 break;
1180 case AGGR_UNSET:
1181 default:
1182 break;
1183 }
1184
1185 if (!interval && !csv_output)
1186 print_footer();
1187
1188 fflush(stat_config.output);
1189 }
1190
1191 static volatile int signr = -1;
1192
1193 static void skip_signal(int signo)
1194 {
1195 if ((child_pid == -1) || stat_config.interval)
1196 done = 1;
1197
1198 signr = signo;
1199 /*
1200 * render child_pid harmless
1201 * won't send SIGTERM to a random
1202 * process in case of race condition
1203 * and fast PID recycling
1204 */
1205 child_pid = -1;
1206 }
1207
1208 static void sig_atexit(void)
1209 {
1210 sigset_t set, oset;
1211
1212 /*
1213 * avoid race condition with SIGCHLD handler
1214 * in skip_signal() which is modifying child_pid
1215 * goal is to avoid send SIGTERM to a random
1216 * process
1217 */
1218 sigemptyset(&set);
1219 sigaddset(&set, SIGCHLD);
1220 sigprocmask(SIG_BLOCK, &set, &oset);
1221
1222 if (child_pid != -1)
1223 kill(child_pid, SIGTERM);
1224
1225 sigprocmask(SIG_SETMASK, &oset, NULL);
1226
1227 if (signr == -1)
1228 return;
1229
1230 signal(signr, SIG_DFL);
1231 kill(getpid(), signr);
1232 }
1233
1234 static int stat__set_big_num(const struct option *opt __maybe_unused,
1235 const char *s __maybe_unused, int unset)
1236 {
1237 big_num_opt = unset ? 0 : 1;
1238 return 0;
1239 }
1240
1241 static const struct option stat_options[] = {
1242 OPT_BOOLEAN('T', "transaction", &transaction_run,
1243 "hardware transaction statistics"),
1244 OPT_CALLBACK('e', "event", &evsel_list, "event",
1245 "event selector. use 'perf list' to list available events",
1246 parse_events_option),
1247 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1248 "event filter", parse_filter),
1249 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1250 "child tasks do not inherit counters"),
1251 OPT_STRING('p', "pid", &target.pid, "pid",
1252 "stat events on existing process id"),
1253 OPT_STRING('t', "tid", &target.tid, "tid",
1254 "stat events on existing thread id"),
1255 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1256 "system-wide collection from all CPUs"),
1257 OPT_BOOLEAN('g', "group", &group,
1258 "put the counters into a counter group"),
1259 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1260 OPT_INCR('v', "verbose", &verbose,
1261 "be more verbose (show counter open errors, etc)"),
1262 OPT_INTEGER('r', "repeat", &run_count,
1263 "repeat command and print average + stddev (max: 100, forever: 0)"),
1264 OPT_BOOLEAN('n', "null", &null_run,
1265 "null run - dont start any counters"),
1266 OPT_INCR('d', "detailed", &detailed_run,
1267 "detailed run - start a lot of events"),
1268 OPT_BOOLEAN('S', "sync", &sync_run,
1269 "call sync() before starting a run"),
1270 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1271 "print large numbers with thousands\' separators",
1272 stat__set_big_num),
1273 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1274 "list of cpus to monitor in system-wide"),
1275 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1276 "disable CPU count aggregation", AGGR_NONE),
1277 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1278 "print counts with custom separator"),
1279 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1280 "monitor event in cgroup name only", parse_cgroups),
1281 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1282 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1283 OPT_INTEGER(0, "log-fd", &output_fd,
1284 "log output to fd, instead of stderr"),
1285 OPT_STRING(0, "pre", &pre_cmd, "command",
1286 "command to run prior to the measured command"),
1287 OPT_STRING(0, "post", &post_cmd, "command",
1288 "command to run after to the measured command"),
1289 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1290 "print counts at regular interval in ms (>= 10)"),
1291 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1292 "aggregate counts per processor socket", AGGR_SOCKET),
1293 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1294 "aggregate counts per physical processor core", AGGR_CORE),
1295 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1296 "aggregate counts per thread", AGGR_THREAD),
1297 OPT_UINTEGER('D', "delay", &initial_delay,
1298 "ms to wait before starting measurement after program start"),
1299 OPT_END()
1300 };
1301
1302 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1303 {
1304 return cpu_map__get_socket(map, cpu, NULL);
1305 }
1306
1307 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1308 {
1309 return cpu_map__get_core(map, cpu, NULL);
1310 }
1311
1312 static int cpu_map__get_max(struct cpu_map *map)
1313 {
1314 int i, max = -1;
1315
1316 for (i = 0; i < map->nr; i++) {
1317 if (map->map[i] > max)
1318 max = map->map[i];
1319 }
1320
1321 return max;
1322 }
1323
1324 static struct cpu_map *cpus_aggr_map;
1325
1326 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1327 {
1328 int cpu;
1329
1330 if (idx >= map->nr)
1331 return -1;
1332
1333 cpu = map->map[idx];
1334
1335 if (cpus_aggr_map->map[cpu] == -1)
1336 cpus_aggr_map->map[cpu] = get_id(map, idx);
1337
1338 return cpus_aggr_map->map[cpu];
1339 }
1340
1341 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1342 {
1343 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1344 }
1345
1346 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1347 {
1348 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1349 }
1350
1351 static int perf_stat_init_aggr_mode(void)
1352 {
1353 int nr;
1354
1355 switch (stat_config.aggr_mode) {
1356 case AGGR_SOCKET:
1357 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1358 perror("cannot build socket map");
1359 return -1;
1360 }
1361 aggr_get_id = perf_stat__get_socket_cached;
1362 break;
1363 case AGGR_CORE:
1364 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1365 perror("cannot build core map");
1366 return -1;
1367 }
1368 aggr_get_id = perf_stat__get_core_cached;
1369 break;
1370 case AGGR_NONE:
1371 case AGGR_GLOBAL:
1372 case AGGR_THREAD:
1373 case AGGR_UNSET:
1374 default:
1375 break;
1376 }
1377
1378 /*
1379 * The evsel_list->cpus is the base we operate on,
1380 * taking the highest cpu number to be the size of
1381 * the aggregation translate cpumap.
1382 */
1383 nr = cpu_map__get_max(evsel_list->cpus);
1384 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1385 return cpus_aggr_map ? 0 : -ENOMEM;
1386 }
1387
1388 static void perf_stat__exit_aggr_mode(void)
1389 {
1390 cpu_map__put(aggr_map);
1391 cpu_map__put(cpus_aggr_map);
1392 aggr_map = NULL;
1393 cpus_aggr_map = NULL;
1394 }
1395
1396 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1397 {
1398 int cpu;
1399
1400 if (idx > map->nr)
1401 return -1;
1402
1403 cpu = map->map[idx];
1404
1405 if (cpu >= env->nr_cpus_online)
1406 return -1;
1407
1408 return cpu;
1409 }
1410
1411 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1412 {
1413 struct perf_env *env = data;
1414 int cpu = perf_env__get_cpu(env, map, idx);
1415
1416 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1417 }
1418
1419 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1420 {
1421 struct perf_env *env = data;
1422 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1423
1424 if (cpu != -1) {
1425 int socket_id = env->cpu[cpu].socket_id;
1426
1427 /*
1428 * Encode socket in upper 16 bits
1429 * core_id is relative to socket, and
1430 * we need a global id. So we combine
1431 * socket + core id.
1432 */
1433 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1434 }
1435
1436 return core;
1437 }
1438
1439 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1440 struct cpu_map **sockp)
1441 {
1442 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1443 }
1444
1445 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1446 struct cpu_map **corep)
1447 {
1448 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1449 }
1450
1451 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1452 {
1453 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1454 }
1455
1456 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1457 {
1458 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1459 }
1460
1461 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1462 {
1463 struct perf_env *env = &st->session->header.env;
1464
1465 switch (stat_config.aggr_mode) {
1466 case AGGR_SOCKET:
1467 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1468 perror("cannot build socket map");
1469 return -1;
1470 }
1471 aggr_get_id = perf_stat__get_socket_file;
1472 break;
1473 case AGGR_CORE:
1474 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1475 perror("cannot build core map");
1476 return -1;
1477 }
1478 aggr_get_id = perf_stat__get_core_file;
1479 break;
1480 case AGGR_NONE:
1481 case AGGR_GLOBAL:
1482 case AGGR_THREAD:
1483 case AGGR_UNSET:
1484 default:
1485 break;
1486 }
1487
1488 return 0;
1489 }
1490
1491 /*
1492 * Add default attributes, if there were no attributes specified or
1493 * if -d/--detailed, -d -d or -d -d -d is used:
1494 */
1495 static int add_default_attributes(void)
1496 {
1497 struct perf_event_attr default_attrs[] = {
1498
1499 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1500 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1501 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1502 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1503
1504 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1505 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1506 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1507 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1508 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1509 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1510
1511 };
1512
1513 /*
1514 * Detailed stats (-d), covering the L1 and last level data caches:
1515 */
1516 struct perf_event_attr detailed_attrs[] = {
1517
1518 { .type = PERF_TYPE_HW_CACHE,
1519 .config =
1520 PERF_COUNT_HW_CACHE_L1D << 0 |
1521 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1522 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1523
1524 { .type = PERF_TYPE_HW_CACHE,
1525 .config =
1526 PERF_COUNT_HW_CACHE_L1D << 0 |
1527 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1528 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1529
1530 { .type = PERF_TYPE_HW_CACHE,
1531 .config =
1532 PERF_COUNT_HW_CACHE_LL << 0 |
1533 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1534 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1535
1536 { .type = PERF_TYPE_HW_CACHE,
1537 .config =
1538 PERF_COUNT_HW_CACHE_LL << 0 |
1539 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1540 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1541 };
1542
1543 /*
1544 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1545 */
1546 struct perf_event_attr very_detailed_attrs[] = {
1547
1548 { .type = PERF_TYPE_HW_CACHE,
1549 .config =
1550 PERF_COUNT_HW_CACHE_L1I << 0 |
1551 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1552 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1553
1554 { .type = PERF_TYPE_HW_CACHE,
1555 .config =
1556 PERF_COUNT_HW_CACHE_L1I << 0 |
1557 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1558 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1559
1560 { .type = PERF_TYPE_HW_CACHE,
1561 .config =
1562 PERF_COUNT_HW_CACHE_DTLB << 0 |
1563 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1564 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1565
1566 { .type = PERF_TYPE_HW_CACHE,
1567 .config =
1568 PERF_COUNT_HW_CACHE_DTLB << 0 |
1569 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1570 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1571
1572 { .type = PERF_TYPE_HW_CACHE,
1573 .config =
1574 PERF_COUNT_HW_CACHE_ITLB << 0 |
1575 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1576 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1577
1578 { .type = PERF_TYPE_HW_CACHE,
1579 .config =
1580 PERF_COUNT_HW_CACHE_ITLB << 0 |
1581 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1582 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1583
1584 };
1585
1586 /*
1587 * Very, very detailed stats (-d -d -d), adding prefetch events:
1588 */
1589 struct perf_event_attr very_very_detailed_attrs[] = {
1590
1591 { .type = PERF_TYPE_HW_CACHE,
1592 .config =
1593 PERF_COUNT_HW_CACHE_L1D << 0 |
1594 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1595 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1596
1597 { .type = PERF_TYPE_HW_CACHE,
1598 .config =
1599 PERF_COUNT_HW_CACHE_L1D << 0 |
1600 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1601 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1602 };
1603
1604 /* Set attrs if no event is selected and !null_run: */
1605 if (null_run)
1606 return 0;
1607
1608 if (transaction_run) {
1609 int err;
1610 if (pmu_have_event("cpu", "cycles-ct") &&
1611 pmu_have_event("cpu", "el-start"))
1612 err = parse_events(evsel_list, transaction_attrs, NULL);
1613 else
1614 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1615 if (err) {
1616 fprintf(stderr, "Cannot set up transaction events\n");
1617 return -1;
1618 }
1619 return 0;
1620 }
1621
1622 if (!evsel_list->nr_entries) {
1623 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1624 return -1;
1625 }
1626
1627 /* Detailed events get appended to the event list: */
1628
1629 if (detailed_run < 1)
1630 return 0;
1631
1632 /* Append detailed run extra attributes: */
1633 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1634 return -1;
1635
1636 if (detailed_run < 2)
1637 return 0;
1638
1639 /* Append very detailed run extra attributes: */
1640 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1641 return -1;
1642
1643 if (detailed_run < 3)
1644 return 0;
1645
1646 /* Append very, very detailed run extra attributes: */
1647 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1648 }
1649
1650 static const char * const stat_record_usage[] = {
1651 "perf stat record [<options>]",
1652 NULL,
1653 };
1654
1655 static void init_features(struct perf_session *session)
1656 {
1657 int feat;
1658
1659 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1660 perf_header__set_feat(&session->header, feat);
1661
1662 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1663 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1664 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1665 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1666 }
1667
1668 static int __cmd_record(int argc, const char **argv)
1669 {
1670 struct perf_session *session;
1671 struct perf_data_file *file = &perf_stat.file;
1672
1673 argc = parse_options(argc, argv, stat_options, stat_record_usage,
1674 PARSE_OPT_STOP_AT_NON_OPTION);
1675
1676 if (output_name)
1677 file->path = output_name;
1678
1679 if (run_count != 1 || forever) {
1680 pr_err("Cannot use -r option with perf stat record.\n");
1681 return -1;
1682 }
1683
1684 session = perf_session__new(file, false, NULL);
1685 if (session == NULL) {
1686 pr_err("Perf session creation failed.\n");
1687 return -1;
1688 }
1689
1690 init_features(session);
1691
1692 session->evlist = evsel_list;
1693 perf_stat.session = session;
1694 perf_stat.record = true;
1695 return argc;
1696 }
1697
1698 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
1699 union perf_event *event,
1700 struct perf_session *session)
1701 {
1702 struct stat_round_event *round = &event->stat_round;
1703 struct perf_evsel *counter;
1704 struct timespec tsh, *ts = NULL;
1705 const char **argv = session->header.env.cmdline_argv;
1706 int argc = session->header.env.nr_cmdline;
1707
1708 evlist__for_each(evsel_list, counter)
1709 perf_stat_process_counter(&stat_config, counter);
1710
1711 if (round->type == PERF_STAT_ROUND_TYPE__FINAL)
1712 update_stats(&walltime_nsecs_stats, round->time);
1713
1714 if (stat_config.interval && round->time) {
1715 tsh.tv_sec = round->time / NSECS_PER_SEC;
1716 tsh.tv_nsec = round->time % NSECS_PER_SEC;
1717 ts = &tsh;
1718 }
1719
1720 print_counters(ts, argc, argv);
1721 return 0;
1722 }
1723
1724 static
1725 int process_stat_config_event(struct perf_tool *tool __maybe_unused,
1726 union perf_event *event,
1727 struct perf_session *session __maybe_unused)
1728 {
1729 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1730
1731 perf_event__read_stat_config(&stat_config, &event->stat_config);
1732
1733 if (cpu_map__empty(st->cpus)) {
1734 if (st->aggr_mode != AGGR_UNSET)
1735 pr_warning("warning: processing task data, aggregation mode not set\n");
1736 return 0;
1737 }
1738
1739 if (st->aggr_mode != AGGR_UNSET)
1740 stat_config.aggr_mode = st->aggr_mode;
1741
1742 if (perf_stat.file.is_pipe)
1743 perf_stat_init_aggr_mode();
1744 else
1745 perf_stat_init_aggr_mode_file(st);
1746
1747 return 0;
1748 }
1749
1750 static int set_maps(struct perf_stat *st)
1751 {
1752 if (!st->cpus || !st->threads)
1753 return 0;
1754
1755 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1756 return -EINVAL;
1757
1758 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
1759
1760 if (perf_evlist__alloc_stats(evsel_list, true))
1761 return -ENOMEM;
1762
1763 st->maps_allocated = true;
1764 return 0;
1765 }
1766
1767 static
1768 int process_thread_map_event(struct perf_tool *tool __maybe_unused,
1769 union perf_event *event,
1770 struct perf_session *session __maybe_unused)
1771 {
1772 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1773
1774 if (st->threads) {
1775 pr_warning("Extra thread map event, ignoring.\n");
1776 return 0;
1777 }
1778
1779 st->threads = thread_map__new_event(&event->thread_map);
1780 if (!st->threads)
1781 return -ENOMEM;
1782
1783 return set_maps(st);
1784 }
1785
1786 static
1787 int process_cpu_map_event(struct perf_tool *tool __maybe_unused,
1788 union perf_event *event,
1789 struct perf_session *session __maybe_unused)
1790 {
1791 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1792 struct cpu_map *cpus;
1793
1794 if (st->cpus) {
1795 pr_warning("Extra cpu map event, ignoring.\n");
1796 return 0;
1797 }
1798
1799 cpus = cpu_map__new_data(&event->cpu_map.data);
1800 if (!cpus)
1801 return -ENOMEM;
1802
1803 st->cpus = cpus;
1804 return set_maps(st);
1805 }
1806
1807 static const char * const stat_report_usage[] = {
1808 "perf stat report [<options>]",
1809 NULL,
1810 };
1811
1812 static struct perf_stat perf_stat = {
1813 .tool = {
1814 .attr = perf_event__process_attr,
1815 .event_update = perf_event__process_event_update,
1816 .thread_map = process_thread_map_event,
1817 .cpu_map = process_cpu_map_event,
1818 .stat_config = process_stat_config_event,
1819 .stat = perf_event__process_stat_event,
1820 .stat_round = process_stat_round_event,
1821 },
1822 .aggr_mode = AGGR_UNSET,
1823 };
1824
1825 static int __cmd_report(int argc, const char **argv)
1826 {
1827 struct perf_session *session;
1828 const struct option options[] = {
1829 OPT_STRING('i', "input", &input_name, "file", "input file name"),
1830 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1831 "aggregate counts per processor socket", AGGR_SOCKET),
1832 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1833 "aggregate counts per physical processor core", AGGR_CORE),
1834 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1835 "disable CPU count aggregation", AGGR_NONE),
1836 OPT_END()
1837 };
1838 struct stat st;
1839 int ret;
1840
1841 argc = parse_options(argc, argv, options, stat_report_usage, 0);
1842
1843 if (!input_name || !strlen(input_name)) {
1844 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1845 input_name = "-";
1846 else
1847 input_name = "perf.data";
1848 }
1849
1850 perf_stat.file.path = input_name;
1851 perf_stat.file.mode = PERF_DATA_MODE_READ;
1852
1853 session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
1854 if (session == NULL)
1855 return -1;
1856
1857 perf_stat.session = session;
1858 stat_config.output = stderr;
1859 evsel_list = session->evlist;
1860
1861 ret = perf_session__process_events(session);
1862 if (ret)
1863 return ret;
1864
1865 perf_session__delete(session);
1866 return 0;
1867 }
1868
1869 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1870 {
1871 const char * const stat_usage[] = {
1872 "perf stat [<options>] [<command>]",
1873 NULL
1874 };
1875 int status = -EINVAL, run_idx;
1876 const char *mode;
1877 FILE *output = stderr;
1878 unsigned int interval;
1879 const char * const stat_subcommands[] = { "record", "report" };
1880
1881 setlocale(LC_ALL, "");
1882
1883 evsel_list = perf_evlist__new();
1884 if (evsel_list == NULL)
1885 return -ENOMEM;
1886
1887 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1888 (const char **) stat_usage,
1889 PARSE_OPT_STOP_AT_NON_OPTION);
1890
1891 if (csv_sep) {
1892 csv_output = true;
1893 if (!strcmp(csv_sep, "\\t"))
1894 csv_sep = "\t";
1895 } else
1896 csv_sep = DEFAULT_SEPARATOR;
1897
1898 if (argc && !strncmp(argv[0], "rec", 3)) {
1899 argc = __cmd_record(argc, argv);
1900 if (argc < 0)
1901 return -1;
1902 } else if (argc && !strncmp(argv[0], "rep", 3))
1903 return __cmd_report(argc, argv);
1904
1905 interval = stat_config.interval;
1906
1907 /*
1908 * For record command the -o is already taken care of.
1909 */
1910 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1911 output = NULL;
1912
1913 if (output_name && output_fd) {
1914 fprintf(stderr, "cannot use both --output and --log-fd\n");
1915 parse_options_usage(stat_usage, stat_options, "o", 1);
1916 parse_options_usage(NULL, stat_options, "log-fd", 0);
1917 goto out;
1918 }
1919
1920 if (output_fd < 0) {
1921 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1922 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1923 goto out;
1924 }
1925
1926 if (!output) {
1927 struct timespec tm;
1928 mode = append_file ? "a" : "w";
1929
1930 output = fopen(output_name, mode);
1931 if (!output) {
1932 perror("failed to create output file");
1933 return -1;
1934 }
1935 clock_gettime(CLOCK_REALTIME, &tm);
1936 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1937 } else if (output_fd > 0) {
1938 mode = append_file ? "a" : "w";
1939 output = fdopen(output_fd, mode);
1940 if (!output) {
1941 perror("Failed opening logfd");
1942 return -errno;
1943 }
1944 }
1945
1946 stat_config.output = output;
1947
1948 /*
1949 * let the spreadsheet do the pretty-printing
1950 */
1951 if (csv_output) {
1952 /* User explicitly passed -B? */
1953 if (big_num_opt == 1) {
1954 fprintf(stderr, "-B option not supported with -x\n");
1955 parse_options_usage(stat_usage, stat_options, "B", 1);
1956 parse_options_usage(NULL, stat_options, "x", 1);
1957 goto out;
1958 } else /* Nope, so disable big number formatting */
1959 big_num = false;
1960 } else if (big_num_opt == 0) /* User passed --no-big-num */
1961 big_num = false;
1962
1963 if (!argc && target__none(&target))
1964 usage_with_options(stat_usage, stat_options);
1965
1966 if (run_count < 0) {
1967 pr_err("Run count must be a positive number\n");
1968 parse_options_usage(stat_usage, stat_options, "r", 1);
1969 goto out;
1970 } else if (run_count == 0) {
1971 forever = true;
1972 run_count = 1;
1973 }
1974
1975 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
1976 fprintf(stderr, "The --per-thread option is only available "
1977 "when monitoring via -p -t options.\n");
1978 parse_options_usage(NULL, stat_options, "p", 1);
1979 parse_options_usage(NULL, stat_options, "t", 1);
1980 goto out;
1981 }
1982
1983 /*
1984 * no_aggr, cgroup are for system-wide only
1985 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1986 */
1987 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1988 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1989 !target__has_cpu(&target)) {
1990 fprintf(stderr, "both cgroup and no-aggregation "
1991 "modes only available in system-wide mode\n");
1992
1993 parse_options_usage(stat_usage, stat_options, "G", 1);
1994 parse_options_usage(NULL, stat_options, "A", 1);
1995 parse_options_usage(NULL, stat_options, "a", 1);
1996 goto out;
1997 }
1998
1999 if (add_default_attributes())
2000 goto out;
2001
2002 target__validate(&target);
2003
2004 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2005 if (target__has_task(&target)) {
2006 pr_err("Problems finding threads of monitor\n");
2007 parse_options_usage(stat_usage, stat_options, "p", 1);
2008 parse_options_usage(NULL, stat_options, "t", 1);
2009 } else if (target__has_cpu(&target)) {
2010 perror("failed to parse CPUs map");
2011 parse_options_usage(stat_usage, stat_options, "C", 1);
2012 parse_options_usage(NULL, stat_options, "a", 1);
2013 }
2014 goto out;
2015 }
2016
2017 /*
2018 * Initialize thread_map with comm names,
2019 * so we could print it out on output.
2020 */
2021 if (stat_config.aggr_mode == AGGR_THREAD)
2022 thread_map__read_comms(evsel_list->threads);
2023
2024 if (interval && interval < 100) {
2025 if (interval < 10) {
2026 pr_err("print interval must be >= 10ms\n");
2027 parse_options_usage(stat_usage, stat_options, "I", 1);
2028 goto out;
2029 } else
2030 pr_warning("print interval < 100ms. "
2031 "The overhead percentage could be high in some cases. "
2032 "Please proceed with caution.\n");
2033 }
2034
2035 if (perf_evlist__alloc_stats(evsel_list, interval))
2036 goto out;
2037
2038 if (perf_stat_init_aggr_mode())
2039 goto out;
2040
2041 /*
2042 * We dont want to block the signals - that would cause
2043 * child tasks to inherit that and Ctrl-C would not work.
2044 * What we want is for Ctrl-C to work in the exec()-ed
2045 * task, but being ignored by perf stat itself:
2046 */
2047 atexit(sig_atexit);
2048 if (!forever)
2049 signal(SIGINT, skip_signal);
2050 signal(SIGCHLD, skip_signal);
2051 signal(SIGALRM, skip_signal);
2052 signal(SIGABRT, skip_signal);
2053
2054 status = 0;
2055 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2056 if (run_count != 1 && verbose)
2057 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2058 run_idx + 1);
2059
2060 status = run_perf_stat(argc, argv);
2061 if (forever && status != -1) {
2062 print_counters(NULL, argc, argv);
2063 perf_stat__reset_stats();
2064 }
2065 }
2066
2067 if (!forever && status != -1 && !interval)
2068 print_counters(NULL, argc, argv);
2069
2070 if (STAT_RECORD) {
2071 /*
2072 * We synthesize the kernel mmap record just so that older tools
2073 * don't emit warnings about not being able to resolve symbols
2074 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2075 * a saner message about no samples being in the perf.data file.
2076 *
2077 * This also serves to suppress a warning about f_header.data.size == 0
2078 * in header.c at the moment 'perf stat record' gets introduced, which
2079 * is not really needed once we start adding the stat specific PERF_RECORD_
2080 * records, but the need to suppress the kptr_restrict messages in older
2081 * tools remain -acme
2082 */
2083 int fd = perf_data_file__fd(&perf_stat.file);
2084 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2085 process_synthesized_event,
2086 &perf_stat.session->machines.host);
2087 if (err) {
2088 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2089 "older tools may produce warnings about this file\n.");
2090 }
2091
2092 if (!interval) {
2093 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2094 pr_err("failed to write stat round event\n");
2095 }
2096
2097 if (!perf_stat.file.is_pipe) {
2098 perf_stat.session->header.data_size += perf_stat.bytes_written;
2099 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2100 }
2101
2102 perf_session__delete(perf_stat.session);
2103 }
2104
2105 perf_stat__exit_aggr_mode();
2106 perf_evlist__free_stats(evsel_list);
2107 out:
2108 perf_evlist__delete(evsel_list);
2109 return status;
2110 }
Linux kernel - Deliverables
This page took 0.074693 seconds and 5 git commands to generate.