]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - tools/perf/builtin-stat.c
projects / mirror_ubuntu-artful-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge branch 'mm-readonly-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[mirror_ubuntu-artful-kernel.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 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
739 {
740 FILE *output = stat_config.output;
741 double msecs = avg / 1e6;
742 const char *fmt_v, *fmt_n;
743 char name[25];
744
745 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
746 fmt_n = csv_output ? "%s" : "%-25s";
747
748 aggr_printout(evsel, id, nr);
749
750 scnprintf(name, sizeof(name), "%s%s",
751 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
752
753 fprintf(output, fmt_v, msecs, csv_sep);
754
755 if (csv_output)
756 fprintf(output, "%s%s", evsel->unit, csv_sep);
757 else
758 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
759
760 fprintf(output, fmt_n, name);
761
762 if (evsel->cgrp)
763 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
764 }
765
766 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
767 {
768 FILE *output = stat_config.output;
769 double sc = evsel->scale;
770 const char *fmt;
771
772 if (csv_output) {
773 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
774 } else {
775 if (big_num)
776 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
777 else
778 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
779 }
780
781 aggr_printout(evsel, id, nr);
782
783 fprintf(output, fmt, avg, csv_sep);
784
785 if (evsel->unit)
786 fprintf(output, "%-*s%s",
787 csv_output ? 0 : unit_width,
788 evsel->unit, csv_sep);
789
790 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
791
792 if (evsel->cgrp)
793 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
794 }
795
796 static void printout(int id, int nr, struct perf_evsel *counter, double uval)
797 {
798 int cpu = cpu_map__id_to_cpu(id);
799
800 if (stat_config.aggr_mode == AGGR_GLOBAL)
801 cpu = 0;
802
803 if (nsec_counter(counter))
804 nsec_printout(id, nr, counter, uval);
805 else
806 abs_printout(id, nr, counter, uval);
807
808 if (!csv_output && !stat_config.interval)
809 perf_stat__print_shadow_stats(stat_config.output, counter,
810 uval, cpu,
811 stat_config.aggr_mode);
812 }
813
814 static void print_aggr(char *prefix)
815 {
816 FILE *output = stat_config.output;
817 struct perf_evsel *counter;
818 int cpu, s, s2, id, nr;
819 double uval;
820 u64 ena, run, val;
821
822 if (!(aggr_map || aggr_get_id))
823 return;
824
825 for (s = 0; s < aggr_map->nr; s++) {
826 id = aggr_map->map[s];
827 evlist__for_each(evsel_list, counter) {
828 val = ena = run = 0;
829 nr = 0;
830 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
831 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
832 if (s2 != id)
833 continue;
834 val += perf_counts(counter->counts, cpu, 0)->val;
835 ena += perf_counts(counter->counts, cpu, 0)->ena;
836 run += perf_counts(counter->counts, cpu, 0)->run;
837 nr++;
838 }
839 if (prefix)
840 fprintf(output, "%s", prefix);
841
842 if (run == 0 || ena == 0) {
843 aggr_printout(counter, id, nr);
844
845 fprintf(output, "%*s%s",
846 csv_output ? 0 : 18,
847 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
848 csv_sep);
849
850 fprintf(output, "%-*s%s",
851 csv_output ? 0 : unit_width,
852 counter->unit, csv_sep);
853
854 fprintf(output, "%*s",
855 csv_output ? 0 : -25,
856 perf_evsel__name(counter));
857
858 if (counter->cgrp)
859 fprintf(output, "%s%s",
860 csv_sep, counter->cgrp->name);
861
862 print_running(run, ena);
863 fputc('\n', output);
864 continue;
865 }
866 uval = val * counter->scale;
867 printout(id, nr, counter, uval);
868 if (!csv_output)
869 print_noise(counter, 1.0);
870
871 print_running(run, ena);
872 fputc('\n', output);
873 }
874 }
875 }
876
877 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
878 {
879 FILE *output = stat_config.output;
880 int nthreads = thread_map__nr(counter->threads);
881 int ncpus = cpu_map__nr(counter->cpus);
882 int cpu, thread;
883 double uval;
884
885 for (thread = 0; thread < nthreads; thread++) {
886 u64 ena = 0, run = 0, val = 0;
887
888 for (cpu = 0; cpu < ncpus; cpu++) {
889 val += perf_counts(counter->counts, cpu, thread)->val;
890 ena += perf_counts(counter->counts, cpu, thread)->ena;
891 run += perf_counts(counter->counts, cpu, thread)->run;
892 }
893
894 if (prefix)
895 fprintf(output, "%s", prefix);
896
897 uval = val * counter->scale;
898 printout(thread, 0, counter, uval);
899
900 if (!csv_output)
901 print_noise(counter, 1.0);
902
903 print_running(run, ena);
904 fputc('\n', output);
905 }
906 }
907
908 /*
909 * Print out the results of a single counter:
910 * aggregated counts in system-wide mode
911 */
912 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
913 {
914 FILE *output = stat_config.output;
915 struct perf_stat_evsel *ps = counter->priv;
916 double avg = avg_stats(&ps->res_stats[0]);
917 int scaled = counter->counts->scaled;
918 double uval;
919 double avg_enabled, avg_running;
920
921 avg_enabled = avg_stats(&ps->res_stats[1]);
922 avg_running = avg_stats(&ps->res_stats[2]);
923
924 if (prefix)
925 fprintf(output, "%s", prefix);
926
927 if (scaled == -1 || !counter->supported) {
928 fprintf(output, "%*s%s",
929 csv_output ? 0 : 18,
930 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
931 csv_sep);
932 fprintf(output, "%-*s%s",
933 csv_output ? 0 : unit_width,
934 counter->unit, csv_sep);
935 fprintf(output, "%*s",
936 csv_output ? 0 : -25,
937 perf_evsel__name(counter));
938
939 if (counter->cgrp)
940 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
941
942 print_running(avg_running, avg_enabled);
943 fputc('\n', output);
944 return;
945 }
946
947 uval = avg * counter->scale;
948 printout(-1, 0, counter, uval);
949
950 print_noise(counter, avg);
951
952 print_running(avg_running, avg_enabled);
953 fprintf(output, "\n");
954 }
955
956 /*
957 * Print out the results of a single counter:
958 * does not use aggregated count in system-wide
959 */
960 static void print_counter(struct perf_evsel *counter, char *prefix)
961 {
962 FILE *output = stat_config.output;
963 u64 ena, run, val;
964 double uval;
965 int cpu;
966
967 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
968 val = perf_counts(counter->counts, cpu, 0)->val;
969 ena = perf_counts(counter->counts, cpu, 0)->ena;
970 run = perf_counts(counter->counts, cpu, 0)->run;
971
972 if (prefix)
973 fprintf(output, "%s", prefix);
974
975 if (run == 0 || ena == 0) {
976 fprintf(output, "CPU%*d%s%*s%s",
977 csv_output ? 0 : -4,
978 perf_evsel__cpus(counter)->map[cpu], csv_sep,
979 csv_output ? 0 : 18,
980 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
981 csv_sep);
982
983 fprintf(output, "%-*s%s",
984 csv_output ? 0 : unit_width,
985 counter->unit, csv_sep);
986
987 fprintf(output, "%*s",
988 csv_output ? 0 : -25,
989 perf_evsel__name(counter));
990
991 if (counter->cgrp)
992 fprintf(output, "%s%s",
993 csv_sep, counter->cgrp->name);
994
995 print_running(run, ena);
996 fputc('\n', output);
997 continue;
998 }
999
1000 uval = val * counter->scale;
1001 printout(cpu, 0, counter, uval);
1002 if (!csv_output)
1003 print_noise(counter, 1.0);
1004 print_running(run, ena);
1005
1006 fputc('\n', output);
1007 }
1008 }
1009
1010 static void print_interval(char *prefix, struct timespec *ts)
1011 {
1012 FILE *output = stat_config.output;
1013 static int num_print_interval;
1014
1015 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1016
1017 if (num_print_interval == 0 && !csv_output) {
1018 switch (stat_config.aggr_mode) {
1019 case AGGR_SOCKET:
1020 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
1021 break;
1022 case AGGR_CORE:
1023 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
1024 break;
1025 case AGGR_NONE:
1026 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
1027 break;
1028 case AGGR_THREAD:
1029 fprintf(output, "# time comm-pid counts %*s events\n", unit_width, "unit");
1030 break;
1031 case AGGR_GLOBAL:
1032 default:
1033 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
1034 case AGGR_UNSET:
1035 break;
1036 }
1037 }
1038
1039 if (++num_print_interval == 25)
1040 num_print_interval = 0;
1041 }
1042
1043 static void print_header(int argc, const char **argv)
1044 {
1045 FILE *output = stat_config.output;
1046 int i;
1047
1048 fflush(stdout);
1049
1050 if (!csv_output) {
1051 fprintf(output, "\n");
1052 fprintf(output, " Performance counter stats for ");
1053 if (target.system_wide)
1054 fprintf(output, "\'system wide");
1055 else if (target.cpu_list)
1056 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1057 else if (!target__has_task(&target)) {
1058 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1059 for (i = 1; argv && (i < argc); i++)
1060 fprintf(output, " %s", argv[i]);
1061 } else if (target.pid)
1062 fprintf(output, "process id \'%s", target.pid);
1063 else
1064 fprintf(output, "thread id \'%s", target.tid);
1065
1066 fprintf(output, "\'");
1067 if (run_count > 1)
1068 fprintf(output, " (%d runs)", run_count);
1069 fprintf(output, ":\n\n");
1070 }
1071 }
1072
1073 static void print_footer(void)
1074 {
1075 FILE *output = stat_config.output;
1076
1077 if (!null_run)
1078 fprintf(output, "\n");
1079 fprintf(output, " %17.9f seconds time elapsed",
1080 avg_stats(&walltime_nsecs_stats)/1e9);
1081 if (run_count > 1) {
1082 fprintf(output, " ");
1083 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1084 avg_stats(&walltime_nsecs_stats));
1085 }
1086 fprintf(output, "\n\n");
1087 }
1088
1089 static void print_counters(struct timespec *ts, int argc, const char **argv)
1090 {
1091 int interval = stat_config.interval;
1092 struct perf_evsel *counter;
1093 char buf[64], *prefix = NULL;
1094
1095 /* Do not print anything if we record to the pipe. */
1096 if (STAT_RECORD && perf_stat.file.is_pipe)
1097 return;
1098
1099 if (interval)
1100 print_interval(prefix = buf, ts);
1101 else
1102 print_header(argc, argv);
1103
1104 switch (stat_config.aggr_mode) {
1105 case AGGR_CORE:
1106 case AGGR_SOCKET:
1107 print_aggr(prefix);
1108 break;
1109 case AGGR_THREAD:
1110 evlist__for_each(evsel_list, counter)
1111 print_aggr_thread(counter, prefix);
1112 break;
1113 case AGGR_GLOBAL:
1114 evlist__for_each(evsel_list, counter)
1115 print_counter_aggr(counter, prefix);
1116 break;
1117 case AGGR_NONE:
1118 evlist__for_each(evsel_list, counter)
1119 print_counter(counter, prefix);
1120 break;
1121 case AGGR_UNSET:
1122 default:
1123 break;
1124 }
1125
1126 if (!interval && !csv_output)
1127 print_footer();
1128
1129 fflush(stat_config.output);
1130 }
1131
1132 static volatile int signr = -1;
1133
1134 static void skip_signal(int signo)
1135 {
1136 if ((child_pid == -1) || stat_config.interval)
1137 done = 1;
1138
1139 signr = signo;
1140 /*
1141 * render child_pid harmless
1142 * won't send SIGTERM to a random
1143 * process in case of race condition
1144 * and fast PID recycling
1145 */
1146 child_pid = -1;
1147 }
1148
1149 static void sig_atexit(void)
1150 {
1151 sigset_t set, oset;
1152
1153 /*
1154 * avoid race condition with SIGCHLD handler
1155 * in skip_signal() which is modifying child_pid
1156 * goal is to avoid send SIGTERM to a random
1157 * process
1158 */
1159 sigemptyset(&set);
1160 sigaddset(&set, SIGCHLD);
1161 sigprocmask(SIG_BLOCK, &set, &oset);
1162
1163 if (child_pid != -1)
1164 kill(child_pid, SIGTERM);
1165
1166 sigprocmask(SIG_SETMASK, &oset, NULL);
1167
1168 if (signr == -1)
1169 return;
1170
1171 signal(signr, SIG_DFL);
1172 kill(getpid(), signr);
1173 }
1174
1175 static int stat__set_big_num(const struct option *opt __maybe_unused,
1176 const char *s __maybe_unused, int unset)
1177 {
1178 big_num_opt = unset ? 0 : 1;
1179 return 0;
1180 }
1181
1182 static const struct option stat_options[] = {
1183 OPT_BOOLEAN('T', "transaction", &transaction_run,
1184 "hardware transaction statistics"),
1185 OPT_CALLBACK('e', "event", &evsel_list, "event",
1186 "event selector. use 'perf list' to list available events",
1187 parse_events_option),
1188 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1189 "event filter", parse_filter),
1190 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1191 "child tasks do not inherit counters"),
1192 OPT_STRING('p', "pid", &target.pid, "pid",
1193 "stat events on existing process id"),
1194 OPT_STRING('t', "tid", &target.tid, "tid",
1195 "stat events on existing thread id"),
1196 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1197 "system-wide collection from all CPUs"),
1198 OPT_BOOLEAN('g', "group", &group,
1199 "put the counters into a counter group"),
1200 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1201 OPT_INCR('v', "verbose", &verbose,
1202 "be more verbose (show counter open errors, etc)"),
1203 OPT_INTEGER('r', "repeat", &run_count,
1204 "repeat command and print average + stddev (max: 100, forever: 0)"),
1205 OPT_BOOLEAN('n', "null", &null_run,
1206 "null run - dont start any counters"),
1207 OPT_INCR('d', "detailed", &detailed_run,
1208 "detailed run - start a lot of events"),
1209 OPT_BOOLEAN('S', "sync", &sync_run,
1210 "call sync() before starting a run"),
1211 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1212 "print large numbers with thousands\' separators",
1213 stat__set_big_num),
1214 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1215 "list of cpus to monitor in system-wide"),
1216 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1217 "disable CPU count aggregation", AGGR_NONE),
1218 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1219 "print counts with custom separator"),
1220 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1221 "monitor event in cgroup name only", parse_cgroups),
1222 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1223 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1224 OPT_INTEGER(0, "log-fd", &output_fd,
1225 "log output to fd, instead of stderr"),
1226 OPT_STRING(0, "pre", &pre_cmd, "command",
1227 "command to run prior to the measured command"),
1228 OPT_STRING(0, "post", &post_cmd, "command",
1229 "command to run after to the measured command"),
1230 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1231 "print counts at regular interval in ms (>= 10)"),
1232 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1233 "aggregate counts per processor socket", AGGR_SOCKET),
1234 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1235 "aggregate counts per physical processor core", AGGR_CORE),
1236 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1237 "aggregate counts per thread", AGGR_THREAD),
1238 OPT_UINTEGER('D', "delay", &initial_delay,
1239 "ms to wait before starting measurement after program start"),
1240 OPT_END()
1241 };
1242
1243 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1244 {
1245 return cpu_map__get_socket(map, cpu, NULL);
1246 }
1247
1248 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1249 {
1250 return cpu_map__get_core(map, cpu, NULL);
1251 }
1252
1253 static int cpu_map__get_max(struct cpu_map *map)
1254 {
1255 int i, max = -1;
1256
1257 for (i = 0; i < map->nr; i++) {
1258 if (map->map[i] > max)
1259 max = map->map[i];
1260 }
1261
1262 return max;
1263 }
1264
1265 static struct cpu_map *cpus_aggr_map;
1266
1267 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1268 {
1269 int cpu;
1270
1271 if (idx >= map->nr)
1272 return -1;
1273
1274 cpu = map->map[idx];
1275
1276 if (cpus_aggr_map->map[cpu] == -1)
1277 cpus_aggr_map->map[cpu] = get_id(map, idx);
1278
1279 return cpus_aggr_map->map[cpu];
1280 }
1281
1282 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1283 {
1284 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1285 }
1286
1287 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1288 {
1289 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1290 }
1291
1292 static int perf_stat_init_aggr_mode(void)
1293 {
1294 int nr;
1295
1296 switch (stat_config.aggr_mode) {
1297 case AGGR_SOCKET:
1298 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1299 perror("cannot build socket map");
1300 return -1;
1301 }
1302 aggr_get_id = perf_stat__get_socket_cached;
1303 break;
1304 case AGGR_CORE:
1305 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1306 perror("cannot build core map");
1307 return -1;
1308 }
1309 aggr_get_id = perf_stat__get_core_cached;
1310 break;
1311 case AGGR_NONE:
1312 case AGGR_GLOBAL:
1313 case AGGR_THREAD:
1314 case AGGR_UNSET:
1315 default:
1316 break;
1317 }
1318
1319 /*
1320 * The evsel_list->cpus is the base we operate on,
1321 * taking the highest cpu number to be the size of
1322 * the aggregation translate cpumap.
1323 */
1324 nr = cpu_map__get_max(evsel_list->cpus);
1325 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1326 return cpus_aggr_map ? 0 : -ENOMEM;
1327 }
1328
1329 static void perf_stat__exit_aggr_mode(void)
1330 {
1331 cpu_map__put(aggr_map);
1332 cpu_map__put(cpus_aggr_map);
1333 aggr_map = NULL;
1334 cpus_aggr_map = NULL;
1335 }
1336
1337 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1338 {
1339 int cpu;
1340
1341 if (idx > map->nr)
1342 return -1;
1343
1344 cpu = map->map[idx];
1345
1346 if (cpu >= env->nr_cpus_online)
1347 return -1;
1348
1349 return cpu;
1350 }
1351
1352 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1353 {
1354 struct perf_env *env = data;
1355 int cpu = perf_env__get_cpu(env, map, idx);
1356
1357 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1358 }
1359
1360 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1361 {
1362 struct perf_env *env = data;
1363 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1364
1365 if (cpu != -1) {
1366 int socket_id = env->cpu[cpu].socket_id;
1367
1368 /*
1369 * Encode socket in upper 16 bits
1370 * core_id is relative to socket, and
1371 * we need a global id. So we combine
1372 * socket + core id.
1373 */
1374 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1375 }
1376
1377 return core;
1378 }
1379
1380 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1381 struct cpu_map **sockp)
1382 {
1383 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1384 }
1385
1386 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1387 struct cpu_map **corep)
1388 {
1389 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1390 }
1391
1392 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1393 {
1394 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1395 }
1396
1397 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1398 {
1399 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1400 }
1401
1402 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1403 {
1404 struct perf_env *env = &st->session->header.env;
1405
1406 switch (stat_config.aggr_mode) {
1407 case AGGR_SOCKET:
1408 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1409 perror("cannot build socket map");
1410 return -1;
1411 }
1412 aggr_get_id = perf_stat__get_socket_file;
1413 break;
1414 case AGGR_CORE:
1415 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1416 perror("cannot build core map");
1417 return -1;
1418 }
1419 aggr_get_id = perf_stat__get_core_file;
1420 break;
1421 case AGGR_NONE:
1422 case AGGR_GLOBAL:
1423 case AGGR_THREAD:
1424 case AGGR_UNSET:
1425 default:
1426 break;
1427 }
1428
1429 return 0;
1430 }
1431
1432 /*
1433 * Add default attributes, if there were no attributes specified or
1434 * if -d/--detailed, -d -d or -d -d -d is used:
1435 */
1436 static int add_default_attributes(void)
1437 {
1438 struct perf_event_attr default_attrs[] = {
1439
1440 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1441 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1442 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1443 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1444
1445 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1446 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1447 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1448 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1449 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1450 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1451
1452 };
1453
1454 /*
1455 * Detailed stats (-d), covering the L1 and last level data caches:
1456 */
1457 struct perf_event_attr detailed_attrs[] = {
1458
1459 { .type = PERF_TYPE_HW_CACHE,
1460 .config =
1461 PERF_COUNT_HW_CACHE_L1D << 0 |
1462 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1463 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1464
1465 { .type = PERF_TYPE_HW_CACHE,
1466 .config =
1467 PERF_COUNT_HW_CACHE_L1D << 0 |
1468 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1469 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1470
1471 { .type = PERF_TYPE_HW_CACHE,
1472 .config =
1473 PERF_COUNT_HW_CACHE_LL << 0 |
1474 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1475 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1476
1477 { .type = PERF_TYPE_HW_CACHE,
1478 .config =
1479 PERF_COUNT_HW_CACHE_LL << 0 |
1480 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1481 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1482 };
1483
1484 /*
1485 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1486 */
1487 struct perf_event_attr very_detailed_attrs[] = {
1488
1489 { .type = PERF_TYPE_HW_CACHE,
1490 .config =
1491 PERF_COUNT_HW_CACHE_L1I << 0 |
1492 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1493 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1494
1495 { .type = PERF_TYPE_HW_CACHE,
1496 .config =
1497 PERF_COUNT_HW_CACHE_L1I << 0 |
1498 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1499 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1500
1501 { .type = PERF_TYPE_HW_CACHE,
1502 .config =
1503 PERF_COUNT_HW_CACHE_DTLB << 0 |
1504 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1505 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1506
1507 { .type = PERF_TYPE_HW_CACHE,
1508 .config =
1509 PERF_COUNT_HW_CACHE_DTLB << 0 |
1510 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1511 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1512
1513 { .type = PERF_TYPE_HW_CACHE,
1514 .config =
1515 PERF_COUNT_HW_CACHE_ITLB << 0 |
1516 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1517 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1518
1519 { .type = PERF_TYPE_HW_CACHE,
1520 .config =
1521 PERF_COUNT_HW_CACHE_ITLB << 0 |
1522 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1523 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1524
1525 };
1526
1527 /*
1528 * Very, very detailed stats (-d -d -d), adding prefetch events:
1529 */
1530 struct perf_event_attr very_very_detailed_attrs[] = {
1531
1532 { .type = PERF_TYPE_HW_CACHE,
1533 .config =
1534 PERF_COUNT_HW_CACHE_L1D << 0 |
1535 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1536 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1537
1538 { .type = PERF_TYPE_HW_CACHE,
1539 .config =
1540 PERF_COUNT_HW_CACHE_L1D << 0 |
1541 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1542 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1543 };
1544
1545 /* Set attrs if no event is selected and !null_run: */
1546 if (null_run)
1547 return 0;
1548
1549 if (transaction_run) {
1550 int err;
1551 if (pmu_have_event("cpu", "cycles-ct") &&
1552 pmu_have_event("cpu", "el-start"))
1553 err = parse_events(evsel_list, transaction_attrs, NULL);
1554 else
1555 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1556 if (err) {
1557 fprintf(stderr, "Cannot set up transaction events\n");
1558 return -1;
1559 }
1560 return 0;
1561 }
1562
1563 if (!evsel_list->nr_entries) {
1564 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1565 return -1;
1566 }
1567
1568 /* Detailed events get appended to the event list: */
1569
1570 if (detailed_run < 1)
1571 return 0;
1572
1573 /* Append detailed run extra attributes: */
1574 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1575 return -1;
1576
1577 if (detailed_run < 2)
1578 return 0;
1579
1580 /* Append very detailed run extra attributes: */
1581 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1582 return -1;
1583
1584 if (detailed_run < 3)
1585 return 0;
1586
1587 /* Append very, very detailed run extra attributes: */
1588 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1589 }
1590
1591 static const char * const stat_record_usage[] = {
1592 "perf stat record [<options>]",
1593 NULL,
1594 };
1595
1596 static void init_features(struct perf_session *session)
1597 {
1598 int feat;
1599
1600 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1601 perf_header__set_feat(&session->header, feat);
1602
1603 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1604 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1605 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1606 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1607 }
1608
1609 static int __cmd_record(int argc, const char **argv)
1610 {
1611 struct perf_session *session;
1612 struct perf_data_file *file = &perf_stat.file;
1613
1614 argc = parse_options(argc, argv, stat_options, stat_record_usage,
1615 PARSE_OPT_STOP_AT_NON_OPTION);
1616
1617 if (output_name)
1618 file->path = output_name;
1619
1620 if (run_count != 1 || forever) {
1621 pr_err("Cannot use -r option with perf stat record.\n");
1622 return -1;
1623 }
1624
1625 session = perf_session__new(file, false, NULL);
1626 if (session == NULL) {
1627 pr_err("Perf session creation failed.\n");
1628 return -1;
1629 }
1630
1631 init_features(session);
1632
1633 session->evlist = evsel_list;
1634 perf_stat.session = session;
1635 perf_stat.record = true;
1636 return argc;
1637 }
1638
1639 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
1640 union perf_event *event,
1641 struct perf_session *session)
1642 {
1643 struct stat_round_event *round = &event->stat_round;
1644 struct perf_evsel *counter;
1645 struct timespec tsh, *ts = NULL;
1646 const char **argv = session->header.env.cmdline_argv;
1647 int argc = session->header.env.nr_cmdline;
1648
1649 evlist__for_each(evsel_list, counter)
1650 perf_stat_process_counter(&stat_config, counter);
1651
1652 if (round->type == PERF_STAT_ROUND_TYPE__FINAL)
1653 update_stats(&walltime_nsecs_stats, round->time);
1654
1655 if (stat_config.interval && round->time) {
1656 tsh.tv_sec = round->time / NSECS_PER_SEC;
1657 tsh.tv_nsec = round->time % NSECS_PER_SEC;
1658 ts = &tsh;
1659 }
1660
1661 print_counters(ts, argc, argv);
1662 return 0;
1663 }
1664
1665 static
1666 int process_stat_config_event(struct perf_tool *tool __maybe_unused,
1667 union perf_event *event,
1668 struct perf_session *session __maybe_unused)
1669 {
1670 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1671
1672 perf_event__read_stat_config(&stat_config, &event->stat_config);
1673
1674 if (cpu_map__empty(st->cpus)) {
1675 if (st->aggr_mode != AGGR_UNSET)
1676 pr_warning("warning: processing task data, aggregation mode not set\n");
1677 return 0;
1678 }
1679
1680 if (st->aggr_mode != AGGR_UNSET)
1681 stat_config.aggr_mode = st->aggr_mode;
1682
1683 if (perf_stat.file.is_pipe)
1684 perf_stat_init_aggr_mode();
1685 else
1686 perf_stat_init_aggr_mode_file(st);
1687
1688 return 0;
1689 }
1690
1691 static int set_maps(struct perf_stat *st)
1692 {
1693 if (!st->cpus || !st->threads)
1694 return 0;
1695
1696 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1697 return -EINVAL;
1698
1699 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
1700
1701 if (perf_evlist__alloc_stats(evsel_list, true))
1702 return -ENOMEM;
1703
1704 st->maps_allocated = true;
1705 return 0;
1706 }
1707
1708 static
1709 int process_thread_map_event(struct perf_tool *tool __maybe_unused,
1710 union perf_event *event,
1711 struct perf_session *session __maybe_unused)
1712 {
1713 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1714
1715 if (st->threads) {
1716 pr_warning("Extra thread map event, ignoring.\n");
1717 return 0;
1718 }
1719
1720 st->threads = thread_map__new_event(&event->thread_map);
1721 if (!st->threads)
1722 return -ENOMEM;
1723
1724 return set_maps(st);
1725 }
1726
1727 static
1728 int process_cpu_map_event(struct perf_tool *tool __maybe_unused,
1729 union perf_event *event,
1730 struct perf_session *session __maybe_unused)
1731 {
1732 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1733 struct cpu_map *cpus;
1734
1735 if (st->cpus) {
1736 pr_warning("Extra cpu map event, ignoring.\n");
1737 return 0;
1738 }
1739
1740 cpus = cpu_map__new_data(&event->cpu_map.data);
1741 if (!cpus)
1742 return -ENOMEM;
1743
1744 st->cpus = cpus;
1745 return set_maps(st);
1746 }
1747
1748 static const char * const stat_report_usage[] = {
1749 "perf stat report [<options>]",
1750 NULL,
1751 };
1752
1753 static struct perf_stat perf_stat = {
1754 .tool = {
1755 .attr = perf_event__process_attr,
1756 .event_update = perf_event__process_event_update,
1757 .thread_map = process_thread_map_event,
1758 .cpu_map = process_cpu_map_event,
1759 .stat_config = process_stat_config_event,
1760 .stat = perf_event__process_stat_event,
1761 .stat_round = process_stat_round_event,
1762 },
1763 .aggr_mode = AGGR_UNSET,
1764 };
1765
1766 static int __cmd_report(int argc, const char **argv)
1767 {
1768 struct perf_session *session;
1769 const struct option options[] = {
1770 OPT_STRING('i', "input", &input_name, "file", "input file name"),
1771 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1772 "aggregate counts per processor socket", AGGR_SOCKET),
1773 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1774 "aggregate counts per physical processor core", AGGR_CORE),
1775 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1776 "disable CPU count aggregation", AGGR_NONE),
1777 OPT_END()
1778 };
1779 struct stat st;
1780 int ret;
1781
1782 argc = parse_options(argc, argv, options, stat_report_usage, 0);
1783
1784 if (!input_name || !strlen(input_name)) {
1785 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1786 input_name = "-";
1787 else
1788 input_name = "perf.data";
1789 }
1790
1791 perf_stat.file.path = input_name;
1792 perf_stat.file.mode = PERF_DATA_MODE_READ;
1793
1794 session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
1795 if (session == NULL)
1796 return -1;
1797
1798 perf_stat.session = session;
1799 stat_config.output = stderr;
1800 evsel_list = session->evlist;
1801
1802 ret = perf_session__process_events(session);
1803 if (ret)
1804 return ret;
1805
1806 perf_session__delete(session);
1807 return 0;
1808 }
1809
1810 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1811 {
1812 const char * const stat_usage[] = {
1813 "perf stat [<options>] [<command>]",
1814 NULL
1815 };
1816 int status = -EINVAL, run_idx;
1817 const char *mode;
1818 FILE *output = stderr;
1819 unsigned int interval;
1820 const char * const stat_subcommands[] = { "record", "report" };
1821
1822 setlocale(LC_ALL, "");
1823
1824 evsel_list = perf_evlist__new();
1825 if (evsel_list == NULL)
1826 return -ENOMEM;
1827
1828 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1829 (const char **) stat_usage,
1830 PARSE_OPT_STOP_AT_NON_OPTION);
1831
1832 if (csv_sep) {
1833 csv_output = true;
1834 if (!strcmp(csv_sep, "\\t"))
1835 csv_sep = "\t";
1836 } else
1837 csv_sep = DEFAULT_SEPARATOR;
1838
1839 if (argc && !strncmp(argv[0], "rec", 3)) {
1840 argc = __cmd_record(argc, argv);
1841 if (argc < 0)
1842 return -1;
1843 } else if (argc && !strncmp(argv[0], "rep", 3))
1844 return __cmd_report(argc, argv);
1845
1846 interval = stat_config.interval;
1847
1848 /*
1849 * For record command the -o is already taken care of.
1850 */
1851 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1852 output = NULL;
1853
1854 if (output_name && output_fd) {
1855 fprintf(stderr, "cannot use both --output and --log-fd\n");
1856 parse_options_usage(stat_usage, stat_options, "o", 1);
1857 parse_options_usage(NULL, stat_options, "log-fd", 0);
1858 goto out;
1859 }
1860
1861 if (output_fd < 0) {
1862 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1863 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1864 goto out;
1865 }
1866
1867 if (!output) {
1868 struct timespec tm;
1869 mode = append_file ? "a" : "w";
1870
1871 output = fopen(output_name, mode);
1872 if (!output) {
1873 perror("failed to create output file");
1874 return -1;
1875 }
1876 clock_gettime(CLOCK_REALTIME, &tm);
1877 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1878 } else if (output_fd > 0) {
1879 mode = append_file ? "a" : "w";
1880 output = fdopen(output_fd, mode);
1881 if (!output) {
1882 perror("Failed opening logfd");
1883 return -errno;
1884 }
1885 }
1886
1887 stat_config.output = output;
1888
1889 /*
1890 * let the spreadsheet do the pretty-printing
1891 */
1892 if (csv_output) {
1893 /* User explicitly passed -B? */
1894 if (big_num_opt == 1) {
1895 fprintf(stderr, "-B option not supported with -x\n");
1896 parse_options_usage(stat_usage, stat_options, "B", 1);
1897 parse_options_usage(NULL, stat_options, "x", 1);
1898 goto out;
1899 } else /* Nope, so disable big number formatting */
1900 big_num = false;
1901 } else if (big_num_opt == 0) /* User passed --no-big-num */
1902 big_num = false;
1903
1904 if (!argc && target__none(&target))
1905 usage_with_options(stat_usage, stat_options);
1906
1907 if (run_count < 0) {
1908 pr_err("Run count must be a positive number\n");
1909 parse_options_usage(stat_usage, stat_options, "r", 1);
1910 goto out;
1911 } else if (run_count == 0) {
1912 forever = true;
1913 run_count = 1;
1914 }
1915
1916 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
1917 fprintf(stderr, "The --per-thread option is only available "
1918 "when monitoring via -p -t options.\n");
1919 parse_options_usage(NULL, stat_options, "p", 1);
1920 parse_options_usage(NULL, stat_options, "t", 1);
1921 goto out;
1922 }
1923
1924 /*
1925 * no_aggr, cgroup are for system-wide only
1926 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1927 */
1928 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1929 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1930 !target__has_cpu(&target)) {
1931 fprintf(stderr, "both cgroup and no-aggregation "
1932 "modes only available in system-wide mode\n");
1933
1934 parse_options_usage(stat_usage, stat_options, "G", 1);
1935 parse_options_usage(NULL, stat_options, "A", 1);
1936 parse_options_usage(NULL, stat_options, "a", 1);
1937 goto out;
1938 }
1939
1940 if (add_default_attributes())
1941 goto out;
1942
1943 target__validate(&target);
1944
1945 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1946 if (target__has_task(&target)) {
1947 pr_err("Problems finding threads of monitor\n");
1948 parse_options_usage(stat_usage, stat_options, "p", 1);
1949 parse_options_usage(NULL, stat_options, "t", 1);
1950 } else if (target__has_cpu(&target)) {
1951 perror("failed to parse CPUs map");
1952 parse_options_usage(stat_usage, stat_options, "C", 1);
1953 parse_options_usage(NULL, stat_options, "a", 1);
1954 }
1955 goto out;
1956 }
1957
1958 /*
1959 * Initialize thread_map with comm names,
1960 * so we could print it out on output.
1961 */
1962 if (stat_config.aggr_mode == AGGR_THREAD)
1963 thread_map__read_comms(evsel_list->threads);
1964
1965 if (interval && interval < 100) {
1966 if (interval < 10) {
1967 pr_err("print interval must be >= 10ms\n");
1968 parse_options_usage(stat_usage, stat_options, "I", 1);
1969 goto out;
1970 } else
1971 pr_warning("print interval < 100ms. "
1972 "The overhead percentage could be high in some cases. "
1973 "Please proceed with caution.\n");
1974 }
1975
1976 if (perf_evlist__alloc_stats(evsel_list, interval))
1977 goto out;
1978
1979 if (perf_stat_init_aggr_mode())
1980 goto out;
1981
1982 /*
1983 * We dont want to block the signals - that would cause
1984 * child tasks to inherit that and Ctrl-C would not work.
1985 * What we want is for Ctrl-C to work in the exec()-ed
1986 * task, but being ignored by perf stat itself:
1987 */
1988 atexit(sig_atexit);
1989 if (!forever)
1990 signal(SIGINT, skip_signal);
1991 signal(SIGCHLD, skip_signal);
1992 signal(SIGALRM, skip_signal);
1993 signal(SIGABRT, skip_signal);
1994
1995 status = 0;
1996 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1997 if (run_count != 1 && verbose)
1998 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1999 run_idx + 1);
2000
2001 status = run_perf_stat(argc, argv);
2002 if (forever && status != -1) {
2003 print_counters(NULL, argc, argv);
2004 perf_stat__reset_stats();
2005 }
2006 }
2007
2008 if (!forever && status != -1 && !interval)
2009 print_counters(NULL, argc, argv);
2010
2011 if (STAT_RECORD) {
2012 /*
2013 * We synthesize the kernel mmap record just so that older tools
2014 * don't emit warnings about not being able to resolve symbols
2015 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2016 * a saner message about no samples being in the perf.data file.
2017 *
2018 * This also serves to suppress a warning about f_header.data.size == 0
2019 * in header.c at the moment 'perf stat record' gets introduced, which
2020 * is not really needed once we start adding the stat specific PERF_RECORD_
2021 * records, but the need to suppress the kptr_restrict messages in older
2022 * tools remain -acme
2023 */
2024 int fd = perf_data_file__fd(&perf_stat.file);
2025 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2026 process_synthesized_event,
2027 &perf_stat.session->machines.host);
2028 if (err) {
2029 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2030 "older tools may produce warnings about this file\n.");
2031 }
2032
2033 if (!interval) {
2034 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2035 pr_err("failed to write stat round event\n");
2036 }
2037
2038 if (!perf_stat.file.is_pipe) {
2039 perf_stat.session->header.data_size += perf_stat.bytes_written;
2040 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2041 }
2042
2043 perf_session__delete(perf_stat.session);
2044 }
2045
2046 perf_stat__exit_aggr_mode();
2047 perf_evlist__free_stats(evsel_list);
2048 out:
2049 perf_evlist__delete(evsel_list);
2050 return status;
2051 }
Ubuntu Artful kernel mirror