?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
kvm,x86: Fix apf_task_wake_one() wq serialization
[cris-mirror.git] / tools / perf / builtin-stat.c
blob866da7aa54bf0356af42fdd484c03401cf2e6f15
1 /*
2 * builtin-stat.c
3 *
4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
6 *
7 * Sample output:
8
9 $ perf stat ./hackbench 10
10
11 Time: 0.118
12
13 Performance counter stats for './hackbench 10':
14
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
26
27 0.154822978 seconds time elapsed
28
29 *
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
31 *
32 * Improvements and fixes by:
33 *
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
40 *
41 * Released under the GPL v2. (and only v2, not any later version)
42 */
43
44 #include "perf.h"
45 #include "builtin.h"
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include <subcmd/parse-options.h>
49 #include "util/parse-events.h"
50 #include "util/pmu.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/drv_configs.h"
56 #include "util/color.h"
57 #include "util/stat.h"
58 #include "util/header.h"
59 #include "util/cpumap.h"
60 #include "util/thread.h"
61 #include "util/thread_map.h"
62 #include "util/counts.h"
63 #include "util/group.h"
64 #include "util/session.h"
65 #include "util/tool.h"
66 #include "util/group.h"
67 #include "util/string2.h"
68 #include "asm/bug.h"
69
70 #include <linux/time64.h>
71 #include <api/fs/fs.h>
72 #include <errno.h>
73 #include <signal.h>
74 #include <stdlib.h>
75 #include <sys/prctl.h>
76 #include <inttypes.h>
77 #include <locale.h>
78 #include <math.h>
79 #include <sys/types.h>
80 #include <sys/stat.h>
81 #include <sys/wait.h>
82 #include <unistd.h>
83
84 #include "sane_ctype.h"
85
86 #define DEFAULT_SEPARATOR " "
87 #define CNTR_NOT_SUPPORTED "<not supported>"
88 #define CNTR_NOT_COUNTED "<not counted>"
89 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
90
91 static void print_counters(struct timespec *ts, int argc, const char **argv);
92
93 /* Default events used for perf stat -T */
94 static const char *transaction_attrs = {
95 "task-clock,"
96 "{"
97 "instructions,"
98 "cycles,"
99 "cpu/cycles-t/,"
100 "cpu/tx-start/,"
101 "cpu/el-start/,"
102 "cpu/cycles-ct/"
103 "}"
104 };
105
106 /* More limited version when the CPU does not have all events. */
107 static const char * transaction_limited_attrs = {
108 "task-clock,"
109 "{"
110 "instructions,"
111 "cycles,"
112 "cpu/cycles-t/,"
113 "cpu/tx-start/"
114 "}"
115 };
116
117 static const char * topdown_attrs[] = {
118 "topdown-total-slots",
119 "topdown-slots-retired",
120 "topdown-recovery-bubbles",
121 "topdown-fetch-bubbles",
122 "topdown-slots-issued",
123 NULL,
124 };
125
126 static const char *smi_cost_attrs = {
127 "{"
128 "msr/aperf/,"
129 "msr/smi/,"
130 "cycles"
131 "}"
132 };
133
134 static struct perf_evlist *evsel_list;
135
136 static struct target target = {
137 .uid = UINT_MAX,
138 };
139
140 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
141
142 static int run_count = 1;
143 static bool no_inherit = false;
144 static volatile pid_t child_pid = -1;
145 static bool null_run = false;
146 static int detailed_run = 0;
147 static bool transaction_run;
148 static bool topdown_run = false;
149 static bool smi_cost = false;
150 static bool smi_reset = false;
151 static bool big_num = true;
152 static int big_num_opt = -1;
153 static const char *csv_sep = NULL;
154 static bool csv_output = false;
155 static bool group = false;
156 static const char *pre_cmd = NULL;
157 static const char *post_cmd = NULL;
158 static bool sync_run = false;
159 static unsigned int initial_delay = 0;
160 static unsigned int unit_width = 4; /* strlen("unit") */
161 static bool forever = false;
162 static bool metric_only = false;
163 static bool force_metric_only = false;
164 static bool no_merge = false;
165 static struct timespec ref_time;
166 static struct cpu_map *aggr_map;
167 static aggr_get_id_t aggr_get_id;
168 static bool append_file;
169 static const char *output_name;
170 static int output_fd;
171 static int print_free_counters_hint;
172
173 struct perf_stat {
174 bool record;
175 struct perf_data_file file;
176 struct perf_session *session;
177 u64 bytes_written;
178 struct perf_tool tool;
179 bool maps_allocated;
180 struct cpu_map *cpus;
181 struct thread_map *threads;
182 enum aggr_mode aggr_mode;
183 };
184
185 static struct perf_stat perf_stat;
186 #define STAT_RECORD perf_stat.record
187
188 static volatile int done = 0;
189
190 static struct perf_stat_config stat_config = {
191 .aggr_mode = AGGR_GLOBAL,
192 .scale = true,
193 };
194
195 static inline void diff_timespec(struct timespec *r, struct timespec *a,
196 struct timespec *b)
197 {
198 r->tv_sec = a->tv_sec - b->tv_sec;
199 if (a->tv_nsec < b->tv_nsec) {
200 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
201 r->tv_sec--;
202 } else {
203 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
204 }
205 }
206
207 static void perf_stat__reset_stats(void)
208 {
209 perf_evlist__reset_stats(evsel_list);
210 perf_stat__reset_shadow_stats();
211 }
212
213 static int create_perf_stat_counter(struct perf_evsel *evsel)
214 {
215 struct perf_event_attr *attr = &evsel->attr;
216 struct perf_evsel *leader = evsel->leader;
217
218 if (stat_config.scale) {
219 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
220 PERF_FORMAT_TOTAL_TIME_RUNNING;
221 }
222
223 /*
224 * The event is part of non trivial group, let's enable
225 * the group read (for leader) and ID retrieval for all
226 * members.
227 */
228 if (leader->nr_members > 1)
229 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
230
231 attr->inherit = !no_inherit;
232
233 /*
234 * Some events get initialized with sample_(period/type) set,
235 * like tracepoints. Clear it up for counting.
236 */
237 attr->sample_period = 0;
238
239 /*
240 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
241 * while avoiding that older tools show confusing messages.
242 *
243 * However for pipe sessions we need to keep it zero,
244 * because script's perf_evsel__check_attr is triggered
245 * by attr->sample_type != 0, and we can't run it on
246 * stat sessions.
247 */
248 if (!(STAT_RECORD && perf_stat.file.is_pipe))
249 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
250
251 /*
252 * Disabling all counters initially, they will be enabled
253 * either manually by us or by kernel via enable_on_exec
254 * set later.
255 */
256 if (perf_evsel__is_group_leader(evsel)) {
257 attr->disabled = 1;
258
259 /*
260 * In case of initial_delay we enable tracee
261 * events manually.
262 */
263 if (target__none(&target) && !initial_delay)
264 attr->enable_on_exec = 1;
265 }
266
267 if (target__has_cpu(&target))
268 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
269
270 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
271 }
272
273 /*
274 * Does the counter have nsecs as a unit?
275 */
276 static inline int nsec_counter(struct perf_evsel *evsel)
277 {
278 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
279 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
280 return 1;
281
282 return 0;
283 }
284
285 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
286 union perf_event *event,
287 struct perf_sample *sample __maybe_unused,
288 struct machine *machine __maybe_unused)
289 {
290 if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
291 pr_err("failed to write perf data, error: %m\n");
292 return -1;
293 }
294
295 perf_stat.bytes_written += event->header.size;
296 return 0;
297 }
298
299 static int write_stat_round_event(u64 tm, u64 type)
300 {
301 return perf_event__synthesize_stat_round(NULL, tm, type,
302 process_synthesized_event,
303 NULL);
304 }
305
306 #define WRITE_STAT_ROUND_EVENT(time, interval) \
307 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
308
309 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
310
311 static int
312 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
313 struct perf_counts_values *count)
314 {
315 struct perf_sample_id *sid = SID(counter, cpu, thread);
316
317 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
318 process_synthesized_event, NULL);
319 }
320
321 /*
322 * Read out the results of a single counter:
323 * do not aggregate counts across CPUs in system-wide mode
324 */
325 static int read_counter(struct perf_evsel *counter)
326 {
327 int nthreads = thread_map__nr(evsel_list->threads);
328 int ncpus, cpu, thread;
329
330 if (target__has_cpu(&target))
331 ncpus = perf_evsel__nr_cpus(counter);
332 else
333 ncpus = 1;
334
335 if (!counter->supported)
336 return -ENOENT;
337
338 if (counter->system_wide)
339 nthreads = 1;
340
341 for (thread = 0; thread < nthreads; thread++) {
342 for (cpu = 0; cpu < ncpus; cpu++) {
343 struct perf_counts_values *count;
344
345 count = perf_counts(counter->counts, cpu, thread);
346
347 /*
348 * The leader's group read loads data into its group members
349 * (via perf_evsel__read_counter) and sets threir count->loaded.
350 */
351 if (!count->loaded &&
352 perf_evsel__read_counter(counter, cpu, thread)) {
353 counter->counts->scaled = -1;
354 perf_counts(counter->counts, cpu, thread)->ena = 0;
355 perf_counts(counter->counts, cpu, thread)->run = 0;
356 return -1;
357 }
358
359 count->loaded = false;
360
361 if (STAT_RECORD) {
362 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
363 pr_err("failed to write stat event\n");
364 return -1;
365 }
366 }
367
368 if (verbose > 1) {
369 fprintf(stat_config.output,
370 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
371 perf_evsel__name(counter),
372 cpu,
373 count->val, count->ena, count->run);
374 }
375 }
376 }
377
378 return 0;
379 }
380
381 static void read_counters(void)
382 {
383 struct perf_evsel *counter;
384 int ret;
385
386 evlist__for_each_entry(evsel_list, counter) {
387 ret = read_counter(counter);
388 if (ret)
389 pr_debug("failed to read counter %s\n", counter->name);
390
391 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
392 pr_warning("failed to process counter %s\n", counter->name);
393 }
394 }
395
396 static void process_interval(void)
397 {
398 struct timespec ts, rs;
399
400 read_counters();
401
402 clock_gettime(CLOCK_MONOTONIC, &ts);
403 diff_timespec(&rs, &ts, &ref_time);
404
405 if (STAT_RECORD) {
406 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
407 pr_err("failed to write stat round event\n");
408 }
409
410 print_counters(&rs, 0, NULL);
411 }
412
413 static void enable_counters(void)
414 {
415 if (initial_delay)
416 usleep(initial_delay * USEC_PER_MSEC);
417
418 /*
419 * We need to enable counters only if:
420 * - we don't have tracee (attaching to task or cpu)
421 * - we have initial delay configured
422 */
423 if (!target__none(&target) || initial_delay)
424 perf_evlist__enable(evsel_list);
425 }
426
427 static void disable_counters(void)
428 {
429 /*
430 * If we don't have tracee (attaching to task or cpu), counters may
431 * still be running. To get accurate group ratios, we must stop groups
432 * from counting before reading their constituent counters.
433 */
434 if (!target__none(&target))
435 perf_evlist__disable(evsel_list);
436 }
437
438 static volatile int workload_exec_errno;
439
440 /*
441 * perf_evlist__prepare_workload will send a SIGUSR1
442 * if the fork fails, since we asked by setting its
443 * want_signal to true.
444 */
445 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
446 void *ucontext __maybe_unused)
447 {
448 workload_exec_errno = info->si_value.sival_int;
449 }
450
451 static bool has_unit(struct perf_evsel *counter)
452 {
453 return counter->unit && *counter->unit;
454 }
455
456 static bool has_scale(struct perf_evsel *counter)
457 {
458 return counter->scale != 1;
459 }
460
461 static int perf_stat_synthesize_config(bool is_pipe)
462 {
463 struct perf_evsel *counter;
464 int err;
465
466 if (is_pipe) {
467 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
468 process_synthesized_event);
469 if (err < 0) {
470 pr_err("Couldn't synthesize attrs.\n");
471 return err;
472 }
473 }
474
475 /*
476 * Synthesize other events stuff not carried within
477 * attr event - unit, scale, name
478 */
479 evlist__for_each_entry(evsel_list, counter) {
480 if (!counter->supported)
481 continue;
482
483 /*
484 * Synthesize unit and scale only if it's defined.
485 */
486 if (has_unit(counter)) {
487 err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
488 if (err < 0) {
489 pr_err("Couldn't synthesize evsel unit.\n");
490 return err;
491 }
492 }
493
494 if (has_scale(counter)) {
495 err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
496 if (err < 0) {
497 pr_err("Couldn't synthesize evsel scale.\n");
498 return err;
499 }
500 }
501
502 if (counter->own_cpus) {
503 err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
504 if (err < 0) {
505 pr_err("Couldn't synthesize evsel scale.\n");
506 return err;
507 }
508 }
509
510 /*
511 * Name is needed only for pipe output,
512 * perf.data carries event names.
513 */
514 if (is_pipe) {
515 err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
516 if (err < 0) {
517 pr_err("Couldn't synthesize evsel name.\n");
518 return err;
519 }
520 }
521 }
522
523 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
524 process_synthesized_event,
525 NULL);
526 if (err < 0) {
527 pr_err("Couldn't synthesize thread map.\n");
528 return err;
529 }
530
531 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
532 process_synthesized_event, NULL);
533 if (err < 0) {
534 pr_err("Couldn't synthesize thread map.\n");
535 return err;
536 }
537
538 err = perf_event__synthesize_stat_config(NULL, &stat_config,
539 process_synthesized_event, NULL);
540 if (err < 0) {
541 pr_err("Couldn't synthesize config.\n");
542 return err;
543 }
544
545 return 0;
546 }
547
548 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
549
550 static int __store_counter_ids(struct perf_evsel *counter,
551 struct cpu_map *cpus,
552 struct thread_map *threads)
553 {
554 int cpu, thread;
555
556 for (cpu = 0; cpu < cpus->nr; cpu++) {
557 for (thread = 0; thread < threads->nr; thread++) {
558 int fd = FD(counter, cpu, thread);
559
560 if (perf_evlist__id_add_fd(evsel_list, counter,
561 cpu, thread, fd) < 0)
562 return -1;
563 }
564 }
565
566 return 0;
567 }
568
569 static int store_counter_ids(struct perf_evsel *counter)
570 {
571 struct cpu_map *cpus = counter->cpus;
572 struct thread_map *threads = counter->threads;
573
574 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
575 return -ENOMEM;
576
577 return __store_counter_ids(counter, cpus, threads);
578 }
579
580 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
581 {
582 return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
583 }
584
585 static int __run_perf_stat(int argc, const char **argv)
586 {
587 int interval = stat_config.interval;
588 char msg[BUFSIZ];
589 unsigned long long t0, t1;
590 struct perf_evsel *counter;
591 struct timespec ts;
592 size_t l;
593 int status = 0;
594 const bool forks = (argc > 0);
595 bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
596 struct perf_evsel_config_term *err_term;
597
598 if (interval) {
599 ts.tv_sec = interval / USEC_PER_MSEC;
600 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
601 } else {
602 ts.tv_sec = 1;
603 ts.tv_nsec = 0;
604 }
605
606 if (forks) {
607 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
608 workload_exec_failed_signal) < 0) {
609 perror("failed to prepare workload");
610 return -1;
611 }
612 child_pid = evsel_list->workload.pid;
613 }
614
615 if (group)
616 perf_evlist__set_leader(evsel_list);
617
618 evlist__for_each_entry(evsel_list, counter) {
619 try_again:
620 if (create_perf_stat_counter(counter) < 0) {
621 /*
622 * PPC returns ENXIO for HW counters until 2.6.37
623 * (behavior changed with commit b0a873e).
624 */
625 if (errno == EINVAL || errno == ENOSYS ||
626 errno == ENOENT || errno == EOPNOTSUPP ||
627 errno == ENXIO) {
628 if (verbose > 0)
629 ui__warning("%s event is not supported by the kernel.\n",
630 perf_evsel__name(counter));
631 counter->supported = false;
632
633 if ((counter->leader != counter) ||
634 !(counter->leader->nr_members > 1))
635 continue;
636 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
637 if (verbose > 0)
638 ui__warning("%s\n", msg);
639 goto try_again;
640 }
641
642 perf_evsel__open_strerror(counter, &target,
643 errno, msg, sizeof(msg));
644 ui__error("%s\n", msg);
645
646 if (child_pid != -1)
647 kill(child_pid, SIGTERM);
648
649 return -1;
650 }
651 counter->supported = true;
652
653 l = strlen(counter->unit);
654 if (l > unit_width)
655 unit_width = l;
656
657 if (perf_evsel__should_store_id(counter) &&
658 store_counter_ids(counter))
659 return -1;
660 }
661
662 if (perf_evlist__apply_filters(evsel_list, &counter)) {
663 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
664 counter->filter, perf_evsel__name(counter), errno,
665 str_error_r(errno, msg, sizeof(msg)));
666 return -1;
667 }
668
669 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
670 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
671 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
672 str_error_r(errno, msg, sizeof(msg)));
673 return -1;
674 }
675
676 if (STAT_RECORD) {
677 int err, fd = perf_data_file__fd(&perf_stat.file);
678
679 if (is_pipe) {
680 err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
681 } else {
682 err = perf_session__write_header(perf_stat.session, evsel_list,
683 fd, false);
684 }
685
686 if (err < 0)
687 return err;
688
689 err = perf_stat_synthesize_config(is_pipe);
690 if (err < 0)
691 return err;
692 }
693
694 /*
695 * Enable counters and exec the command:
696 */
697 t0 = rdclock();
698 clock_gettime(CLOCK_MONOTONIC, &ref_time);
699
700 if (forks) {
701 perf_evlist__start_workload(evsel_list);
702 enable_counters();
703
704 if (interval) {
705 while (!waitpid(child_pid, &status, WNOHANG)) {
706 nanosleep(&ts, NULL);
707 process_interval();
708 }
709 }
710 wait(&status);
711
712 if (workload_exec_errno) {
713 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
714 pr_err("Workload failed: %s\n", emsg);
715 return -1;
716 }
717
718 if (WIFSIGNALED(status))
719 psignal(WTERMSIG(status), argv[0]);
720 } else {
721 enable_counters();
722 while (!done) {
723 nanosleep(&ts, NULL);
724 if (interval)
725 process_interval();
726 }
727 }
728
729 disable_counters();
730
731 t1 = rdclock();
732
733 update_stats(&walltime_nsecs_stats, t1 - t0);
734
735 /*
736 * Closing a group leader splits the group, and as we only disable
737 * group leaders, results in remaining events becoming enabled. To
738 * avoid arbitrary skew, we must read all counters before closing any
739 * group leaders.
740 */
741 read_counters();
742 perf_evlist__close(evsel_list);
743
744 return WEXITSTATUS(status);
745 }
746
747 static int run_perf_stat(int argc, const char **argv)
748 {
749 int ret;
750
751 if (pre_cmd) {
752 ret = system(pre_cmd);
753 if (ret)
754 return ret;
755 }
756
757 if (sync_run)
758 sync();
759
760 ret = __run_perf_stat(argc, argv);
761 if (ret)
762 return ret;
763
764 if (post_cmd) {
765 ret = system(post_cmd);
766 if (ret)
767 return ret;
768 }
769
770 return ret;
771 }
772
773 static void print_running(u64 run, u64 ena)
774 {
775 if (csv_output) {
776 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
777 csv_sep,
778 run,
779 csv_sep,
780 ena ? 100.0 * run / ena : 100.0);
781 } else if (run != ena) {
782 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
783 }
784 }
785
786 static void print_noise_pct(double total, double avg)
787 {
788 double pct = rel_stddev_stats(total, avg);
789
790 if (csv_output)
791 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
792 else if (pct)
793 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
794 }
795
796 static void print_noise(struct perf_evsel *evsel, double avg)
797 {
798 struct perf_stat_evsel *ps;
799
800 if (run_count == 1)
801 return;
802
803 ps = evsel->priv;
804 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
805 }
806
807 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
808 {
809 switch (stat_config.aggr_mode) {
810 case AGGR_CORE:
811 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
812 cpu_map__id_to_socket(id),
813 csv_output ? 0 : -8,
814 cpu_map__id_to_cpu(id),
815 csv_sep,
816 csv_output ? 0 : 4,
817 nr,
818 csv_sep);
819 break;
820 case AGGR_SOCKET:
821 fprintf(stat_config.output, "S%*d%s%*d%s",
822 csv_output ? 0 : -5,
823 id,
824 csv_sep,
825 csv_output ? 0 : 4,
826 nr,
827 csv_sep);
828 break;
829 case AGGR_NONE:
830 fprintf(stat_config.output, "CPU%*d%s",
831 csv_output ? 0 : -4,
832 perf_evsel__cpus(evsel)->map[id], csv_sep);
833 break;
834 case AGGR_THREAD:
835 fprintf(stat_config.output, "%*s-%*d%s",
836 csv_output ? 0 : 16,
837 thread_map__comm(evsel->threads, id),
838 csv_output ? 0 : -8,
839 thread_map__pid(evsel->threads, id),
840 csv_sep);
841 break;
842 case AGGR_GLOBAL:
843 case AGGR_UNSET:
844 default:
845 break;
846 }
847 }
848
849 struct outstate {
850 FILE *fh;
851 bool newline;
852 const char *prefix;
853 int nfields;
854 int id, nr;
855 struct perf_evsel *evsel;
856 };
857
858 #define METRIC_LEN 35
859
860 static void new_line_std(void *ctx)
861 {
862 struct outstate *os = ctx;
863
864 os->newline = true;
865 }
866
867 static void do_new_line_std(struct outstate *os)
868 {
869 fputc('\n', os->fh);
870 fputs(os->prefix, os->fh);
871 aggr_printout(os->evsel, os->id, os->nr);
872 if (stat_config.aggr_mode == AGGR_NONE)
873 fprintf(os->fh, " ");
874 fprintf(os->fh, " ");
875 }
876
877 static void print_metric_std(void *ctx, const char *color, const char *fmt,
878 const char *unit, double val)
879 {
880 struct outstate *os = ctx;
881 FILE *out = os->fh;
882 int n;
883 bool newline = os->newline;
884
885 os->newline = false;
886
887 if (unit == NULL || fmt == NULL) {
888 fprintf(out, "%-*s", METRIC_LEN, "");
889 return;
890 }
891
892 if (newline)
893 do_new_line_std(os);
894
895 n = fprintf(out, " # ");
896 if (color)
897 n += color_fprintf(out, color, fmt, val);
898 else
899 n += fprintf(out, fmt, val);
900 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
901 }
902
903 static void new_line_csv(void *ctx)
904 {
905 struct outstate *os = ctx;
906 int i;
907
908 fputc('\n', os->fh);
909 if (os->prefix)
910 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
911 aggr_printout(os->evsel, os->id, os->nr);
912 for (i = 0; i < os->nfields; i++)
913 fputs(csv_sep, os->fh);
914 }
915
916 static void print_metric_csv(void *ctx,
917 const char *color __maybe_unused,
918 const char *fmt, const char *unit, double val)
919 {
920 struct outstate *os = ctx;
921 FILE *out = os->fh;
922 char buf[64], *vals, *ends;
923
924 if (unit == NULL || fmt == NULL) {
925 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
926 return;
927 }
928 snprintf(buf, sizeof(buf), fmt, val);
929 ends = vals = ltrim(buf);
930 while (isdigit(*ends) || *ends == '.')
931 ends++;
932 *ends = 0;
933 while (isspace(*unit))
934 unit++;
935 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
936 }
937
938 #define METRIC_ONLY_LEN 20
939
940 /* Filter out some columns that don't work well in metrics only mode */
941
942 static bool valid_only_metric(const char *unit)
943 {
944 if (!unit)
945 return false;
946 if (strstr(unit, "/sec") ||
947 strstr(unit, "hz") ||
948 strstr(unit, "Hz") ||
949 strstr(unit, "CPUs utilized"))
950 return false;
951 return true;
952 }
953
954 static const char *fixunit(char *buf, struct perf_evsel *evsel,
955 const char *unit)
956 {
957 if (!strncmp(unit, "of all", 6)) {
958 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
959 unit);
960 return buf;
961 }
962 return unit;
963 }
964
965 static void print_metric_only(void *ctx, const char *color, const char *fmt,
966 const char *unit, double val)
967 {
968 struct outstate *os = ctx;
969 FILE *out = os->fh;
970 int n;
971 char buf[1024];
972 unsigned mlen = METRIC_ONLY_LEN;
973
974 if (!valid_only_metric(unit))
975 return;
976 unit = fixunit(buf, os->evsel, unit);
977 if (color)
978 n = color_fprintf(out, color, fmt, val);
979 else
980 n = fprintf(out, fmt, val);
981 if (n > METRIC_ONLY_LEN)
982 n = METRIC_ONLY_LEN;
983 if (mlen < strlen(unit))
984 mlen = strlen(unit) + 1;
985 fprintf(out, "%*s", mlen - n, "");
986 }
987
988 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
989 const char *fmt,
990 const char *unit, double val)
991 {
992 struct outstate *os = ctx;
993 FILE *out = os->fh;
994 char buf[64], *vals, *ends;
995 char tbuf[1024];
996
997 if (!valid_only_metric(unit))
998 return;
999 unit = fixunit(tbuf, os->evsel, unit);
1000 snprintf(buf, sizeof buf, fmt, val);
1001 ends = vals = ltrim(buf);
1002 while (isdigit(*ends) || *ends == '.')
1003 ends++;
1004 *ends = 0;
1005 fprintf(out, "%s%s", vals, csv_sep);
1006 }
1007
1008 static void new_line_metric(void *ctx __maybe_unused)
1009 {
1010 }
1011
1012 static void print_metric_header(void *ctx, const char *color __maybe_unused,
1013 const char *fmt __maybe_unused,
1014 const char *unit, double val __maybe_unused)
1015 {
1016 struct outstate *os = ctx;
1017 char tbuf[1024];
1018
1019 if (!valid_only_metric(unit))
1020 return;
1021 unit = fixunit(tbuf, os->evsel, unit);
1022 if (csv_output)
1023 fprintf(os->fh, "%s%s", unit, csv_sep);
1024 else
1025 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
1026 }
1027
1028 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1029 {
1030 FILE *output = stat_config.output;
1031 double msecs = avg / NSEC_PER_MSEC;
1032 const char *fmt_v, *fmt_n;
1033 char name[25];
1034
1035 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
1036 fmt_n = csv_output ? "%s" : "%-25s";
1037
1038 aggr_printout(evsel, id, nr);
1039
1040 scnprintf(name, sizeof(name), "%s%s",
1041 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
1042
1043 fprintf(output, fmt_v, msecs, csv_sep);
1044
1045 if (csv_output)
1046 fprintf(output, "%s%s", evsel->unit, csv_sep);
1047 else
1048 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
1049
1050 fprintf(output, fmt_n, name);
1051
1052 if (evsel->cgrp)
1053 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1054 }
1055
1056 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1057 {
1058 int i;
1059
1060 if (!aggr_get_id)
1061 return 0;
1062
1063 if (stat_config.aggr_mode == AGGR_NONE)
1064 return id;
1065
1066 if (stat_config.aggr_mode == AGGR_GLOBAL)
1067 return 0;
1068
1069 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1070 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1071
1072 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1073 return cpu2;
1074 }
1075 return 0;
1076 }
1077
1078 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1079 {
1080 FILE *output = stat_config.output;
1081 double sc = evsel->scale;
1082 const char *fmt;
1083
1084 if (csv_output) {
1085 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
1086 } else {
1087 if (big_num)
1088 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1089 else
1090 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1091 }
1092
1093 aggr_printout(evsel, id, nr);
1094
1095 fprintf(output, fmt, avg, csv_sep);
1096
1097 if (evsel->unit)
1098 fprintf(output, "%-*s%s",
1099 csv_output ? 0 : unit_width,
1100 evsel->unit, csv_sep);
1101
1102 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1103
1104 if (evsel->cgrp)
1105 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1106 }
1107
1108 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1109 char *prefix, u64 run, u64 ena, double noise)
1110 {
1111 struct perf_stat_output_ctx out;
1112 struct outstate os = {
1113 .fh = stat_config.output,
1114 .prefix = prefix ? prefix : "",
1115 .id = id,
1116 .nr = nr,
1117 .evsel = counter,
1118 };
1119 print_metric_t pm = print_metric_std;
1120 void (*nl)(void *);
1121
1122 if (metric_only) {
1123 nl = new_line_metric;
1124 if (csv_output)
1125 pm = print_metric_only_csv;
1126 else
1127 pm = print_metric_only;
1128 } else
1129 nl = new_line_std;
1130
1131 if (csv_output && !metric_only) {
1132 static int aggr_fields[] = {
1133 [AGGR_GLOBAL] = 0,
1134 [AGGR_THREAD] = 1,
1135 [AGGR_NONE] = 1,
1136 [AGGR_SOCKET] = 2,
1137 [AGGR_CORE] = 2,
1138 };
1139
1140 pm = print_metric_csv;
1141 nl = new_line_csv;
1142 os.nfields = 3;
1143 os.nfields += aggr_fields[stat_config.aggr_mode];
1144 if (counter->cgrp)
1145 os.nfields++;
1146 }
1147 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1148 if (metric_only) {
1149 pm(&os, NULL, "", "", 0);
1150 return;
1151 }
1152 aggr_printout(counter, id, nr);
1153
1154 fprintf(stat_config.output, "%*s%s",
1155 csv_output ? 0 : 18,
1156 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1157 csv_sep);
1158
1159 if (counter->supported)
1160 print_free_counters_hint = 1;
1161
1162 fprintf(stat_config.output, "%-*s%s",
1163 csv_output ? 0 : unit_width,
1164 counter->unit, csv_sep);
1165
1166 fprintf(stat_config.output, "%*s",
1167 csv_output ? 0 : -25,
1168 perf_evsel__name(counter));
1169
1170 if (counter->cgrp)
1171 fprintf(stat_config.output, "%s%s",
1172 csv_sep, counter->cgrp->name);
1173
1174 if (!csv_output)
1175 pm(&os, NULL, NULL, "", 0);
1176 print_noise(counter, noise);
1177 print_running(run, ena);
1178 if (csv_output)
1179 pm(&os, NULL, NULL, "", 0);
1180 return;
1181 }
1182
1183 if (metric_only)
1184 /* nothing */;
1185 else if (nsec_counter(counter))
1186 nsec_printout(id, nr, counter, uval);
1187 else
1188 abs_printout(id, nr, counter, uval);
1189
1190 out.print_metric = pm;
1191 out.new_line = nl;
1192 out.ctx = &os;
1193 out.force_header = false;
1194
1195 if (csv_output && !metric_only) {
1196 print_noise(counter, noise);
1197 print_running(run, ena);
1198 }
1199
1200 perf_stat__print_shadow_stats(counter, uval,
1201 first_shadow_cpu(counter, id),
1202 &out);
1203 if (!csv_output && !metric_only) {
1204 print_noise(counter, noise);
1205 print_running(run, ena);
1206 }
1207 }
1208
1209 static void aggr_update_shadow(void)
1210 {
1211 int cpu, s2, id, s;
1212 u64 val;
1213 struct perf_evsel *counter;
1214
1215 for (s = 0; s < aggr_map->nr; s++) {
1216 id = aggr_map->map[s];
1217 evlist__for_each_entry(evsel_list, counter) {
1218 val = 0;
1219 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1220 s2 = aggr_get_id(evsel_list->cpus, cpu);
1221 if (s2 != id)
1222 continue;
1223 val += perf_counts(counter->counts, cpu, 0)->val;
1224 }
1225 val = val * counter->scale;
1226 perf_stat__update_shadow_stats(counter, &val,
1227 first_shadow_cpu(counter, id));
1228 }
1229 }
1230 }
1231
1232 static void collect_all_aliases(struct perf_evsel *counter,
1233 void (*cb)(struct perf_evsel *counter, void *data,
1234 bool first),
1235 void *data)
1236 {
1237 struct perf_evsel *alias;
1238
1239 alias = list_prepare_entry(counter, &(evsel_list->entries), node);
1240 list_for_each_entry_continue (alias, &evsel_list->entries, node) {
1241 if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) ||
1242 alias->scale != counter->scale ||
1243 alias->cgrp != counter->cgrp ||
1244 strcmp(alias->unit, counter->unit) ||
1245 nsec_counter(alias) != nsec_counter(counter))
1246 break;
1247 alias->merged_stat = true;
1248 cb(alias, data, false);
1249 }
1250 }
1251
1252 static bool collect_data(struct perf_evsel *counter,
1253 void (*cb)(struct perf_evsel *counter, void *data,
1254 bool first),
1255 void *data)
1256 {
1257 if (counter->merged_stat)
1258 return false;
1259 cb(counter, data, true);
1260 if (!no_merge)
1261 collect_all_aliases(counter, cb, data);
1262 return true;
1263 }
1264
1265 struct aggr_data {
1266 u64 ena, run, val;
1267 int id;
1268 int nr;
1269 int cpu;
1270 };
1271
1272 static void aggr_cb(struct perf_evsel *counter, void *data, bool first)
1273 {
1274 struct aggr_data *ad = data;
1275 int cpu, s2;
1276
1277 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1278 struct perf_counts_values *counts;
1279
1280 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1281 if (s2 != ad->id)
1282 continue;
1283 if (first)
1284 ad->nr++;
1285 counts = perf_counts(counter->counts, cpu, 0);
1286 /*
1287 * When any result is bad, make them all to give
1288 * consistent output in interval mode.
1289 */
1290 if (counts->ena == 0 || counts->run == 0 ||
1291 counter->counts->scaled == -1) {
1292 ad->ena = 0;
1293 ad->run = 0;
1294 break;
1295 }
1296 ad->val += counts->val;
1297 ad->ena += counts->ena;
1298 ad->run += counts->run;
1299 }
1300 }
1301
1302 static void print_aggr(char *prefix)
1303 {
1304 FILE *output = stat_config.output;
1305 struct perf_evsel *counter;
1306 int s, id, nr;
1307 double uval;
1308 u64 ena, run, val;
1309 bool first;
1310
1311 if (!(aggr_map || aggr_get_id))
1312 return;
1313
1314 aggr_update_shadow();
1315
1316 /*
1317 * With metric_only everything is on a single line.
1318 * Without each counter has its own line.
1319 */
1320 for (s = 0; s < aggr_map->nr; s++) {
1321 struct aggr_data ad;
1322 if (prefix && metric_only)
1323 fprintf(output, "%s", prefix);
1324
1325 ad.id = id = aggr_map->map[s];
1326 first = true;
1327 evlist__for_each_entry(evsel_list, counter) {
1328 ad.val = ad.ena = ad.run = 0;
1329 ad.nr = 0;
1330 if (!collect_data(counter, aggr_cb, &ad))
1331 continue;
1332 nr = ad.nr;
1333 ena = ad.ena;
1334 run = ad.run;
1335 val = ad.val;
1336 if (first && metric_only) {
1337 first = false;
1338 aggr_printout(counter, id, nr);
1339 }
1340 if (prefix && !metric_only)
1341 fprintf(output, "%s", prefix);
1342
1343 uval = val * counter->scale;
1344 printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1345 if (!metric_only)
1346 fputc('\n', output);
1347 }
1348 if (metric_only)
1349 fputc('\n', output);
1350 }
1351 }
1352
1353 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1354 {
1355 FILE *output = stat_config.output;
1356 int nthreads = thread_map__nr(counter->threads);
1357 int ncpus = cpu_map__nr(counter->cpus);
1358 int cpu, thread;
1359 double uval;
1360
1361 for (thread = 0; thread < nthreads; thread++) {
1362 u64 ena = 0, run = 0, val = 0;
1363
1364 for (cpu = 0; cpu < ncpus; cpu++) {
1365 val += perf_counts(counter->counts, cpu, thread)->val;
1366 ena += perf_counts(counter->counts, cpu, thread)->ena;
1367 run += perf_counts(counter->counts, cpu, thread)->run;
1368 }
1369
1370 if (prefix)
1371 fprintf(output, "%s", prefix);
1372
1373 uval = val * counter->scale;
1374 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1375 fputc('\n', output);
1376 }
1377 }
1378
1379 struct caggr_data {
1380 double avg, avg_enabled, avg_running;
1381 };
1382
1383 static void counter_aggr_cb(struct perf_evsel *counter, void *data,
1384 bool first __maybe_unused)
1385 {
1386 struct caggr_data *cd = data;
1387 struct perf_stat_evsel *ps = counter->priv;
1388
1389 cd->avg += avg_stats(&ps->res_stats[0]);
1390 cd->avg_enabled += avg_stats(&ps->res_stats[1]);
1391 cd->avg_running += avg_stats(&ps->res_stats[2]);
1392 }
1393
1394 /*
1395 * Print out the results of a single counter:
1396 * aggregated counts in system-wide mode
1397 */
1398 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1399 {
1400 FILE *output = stat_config.output;
1401 double uval;
1402 struct caggr_data cd = { .avg = 0.0 };
1403
1404 if (!collect_data(counter, counter_aggr_cb, &cd))
1405 return;
1406
1407 if (prefix && !metric_only)
1408 fprintf(output, "%s", prefix);
1409
1410 uval = cd.avg * counter->scale;
1411 printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled, cd.avg);
1412 if (!metric_only)
1413 fprintf(output, "\n");
1414 }
1415
1416 static void counter_cb(struct perf_evsel *counter, void *data,
1417 bool first __maybe_unused)
1418 {
1419 struct aggr_data *ad = data;
1420
1421 ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
1422 ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
1423 ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
1424 }
1425
1426 /*
1427 * Print out the results of a single counter:
1428 * does not use aggregated count in system-wide
1429 */
1430 static void print_counter(struct perf_evsel *counter, char *prefix)
1431 {
1432 FILE *output = stat_config.output;
1433 u64 ena, run, val;
1434 double uval;
1435 int cpu;
1436
1437 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1438 struct aggr_data ad = { .cpu = cpu };
1439
1440 if (!collect_data(counter, counter_cb, &ad))
1441 return;
1442 val = ad.val;
1443 ena = ad.ena;
1444 run = ad.run;
1445
1446 if (prefix)
1447 fprintf(output, "%s", prefix);
1448
1449 uval = val * counter->scale;
1450 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1451
1452 fputc('\n', output);
1453 }
1454 }
1455
1456 static void print_no_aggr_metric(char *prefix)
1457 {
1458 int cpu;
1459 int nrcpus = 0;
1460 struct perf_evsel *counter;
1461 u64 ena, run, val;
1462 double uval;
1463
1464 nrcpus = evsel_list->cpus->nr;
1465 for (cpu = 0; cpu < nrcpus; cpu++) {
1466 bool first = true;
1467
1468 if (prefix)
1469 fputs(prefix, stat_config.output);
1470 evlist__for_each_entry(evsel_list, counter) {
1471 if (first) {
1472 aggr_printout(counter, cpu, 0);
1473 first = false;
1474 }
1475 val = perf_counts(counter->counts, cpu, 0)->val;
1476 ena = perf_counts(counter->counts, cpu, 0)->ena;
1477 run = perf_counts(counter->counts, cpu, 0)->run;
1478
1479 uval = val * counter->scale;
1480 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1481 }
1482 fputc('\n', stat_config.output);
1483 }
1484 }
1485
1486 static int aggr_header_lens[] = {
1487 [AGGR_CORE] = 18,
1488 [AGGR_SOCKET] = 12,
1489 [AGGR_NONE] = 6,
1490 [AGGR_THREAD] = 24,
1491 [AGGR_GLOBAL] = 0,
1492 };
1493
1494 static const char *aggr_header_csv[] = {
1495 [AGGR_CORE] = "core,cpus,",
1496 [AGGR_SOCKET] = "socket,cpus",
1497 [AGGR_NONE] = "cpu,",
1498 [AGGR_THREAD] = "comm-pid,",
1499 [AGGR_GLOBAL] = ""
1500 };
1501
1502 static void print_metric_headers(const char *prefix, bool no_indent)
1503 {
1504 struct perf_stat_output_ctx out;
1505 struct perf_evsel *counter;
1506 struct outstate os = {
1507 .fh = stat_config.output
1508 };
1509
1510 if (prefix)
1511 fprintf(stat_config.output, "%s", prefix);
1512
1513 if (!csv_output && !no_indent)
1514 fprintf(stat_config.output, "%*s",
1515 aggr_header_lens[stat_config.aggr_mode], "");
1516 if (csv_output) {
1517 if (stat_config.interval)
1518 fputs("time,", stat_config.output);
1519 fputs(aggr_header_csv[stat_config.aggr_mode],
1520 stat_config.output);
1521 }
1522
1523 /* Print metrics headers only */
1524 evlist__for_each_entry(evsel_list, counter) {
1525 os.evsel = counter;
1526 out.ctx = &os;
1527 out.print_metric = print_metric_header;
1528 out.new_line = new_line_metric;
1529 out.force_header = true;
1530 os.evsel = counter;
1531 perf_stat__print_shadow_stats(counter, 0,
1532 0,
1533 &out);
1534 }
1535 fputc('\n', stat_config.output);
1536 }
1537
1538 static void print_interval(char *prefix, struct timespec *ts)
1539 {
1540 FILE *output = stat_config.output;
1541 static int num_print_interval;
1542
1543 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1544
1545 if (num_print_interval == 0 && !csv_output) {
1546 switch (stat_config.aggr_mode) {
1547 case AGGR_SOCKET:
1548 fprintf(output, "# time socket cpus");
1549 if (!metric_only)
1550 fprintf(output, " counts %*s events\n", unit_width, "unit");
1551 break;
1552 case AGGR_CORE:
1553 fprintf(output, "# time core cpus");
1554 if (!metric_only)
1555 fprintf(output, " counts %*s events\n", unit_width, "unit");
1556 break;
1557 case AGGR_NONE:
1558 fprintf(output, "# time CPU");
1559 if (!metric_only)
1560 fprintf(output, " counts %*s events\n", unit_width, "unit");
1561 break;
1562 case AGGR_THREAD:
1563 fprintf(output, "# time comm-pid");
1564 if (!metric_only)
1565 fprintf(output, " counts %*s events\n", unit_width, "unit");
1566 break;
1567 case AGGR_GLOBAL:
1568 default:
1569 fprintf(output, "# time");
1570 if (!metric_only)
1571 fprintf(output, " counts %*s events\n", unit_width, "unit");
1572 case AGGR_UNSET:
1573 break;
1574 }
1575 }
1576
1577 if (num_print_interval == 0 && metric_only)
1578 print_metric_headers(" ", true);
1579 if (++num_print_interval == 25)
1580 num_print_interval = 0;
1581 }
1582
1583 static void print_header(int argc, const char **argv)
1584 {
1585 FILE *output = stat_config.output;
1586 int i;
1587
1588 fflush(stdout);
1589
1590 if (!csv_output) {
1591 fprintf(output, "\n");
1592 fprintf(output, " Performance counter stats for ");
1593 if (target.system_wide)
1594 fprintf(output, "\'system wide");
1595 else if (target.cpu_list)
1596 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1597 else if (!target__has_task(&target)) {
1598 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1599 for (i = 1; argv && (i < argc); i++)
1600 fprintf(output, " %s", argv[i]);
1601 } else if (target.pid)
1602 fprintf(output, "process id \'%s", target.pid);
1603 else
1604 fprintf(output, "thread id \'%s", target.tid);
1605
1606 fprintf(output, "\'");
1607 if (run_count > 1)
1608 fprintf(output, " (%d runs)", run_count);
1609 fprintf(output, ":\n\n");
1610 }
1611 }
1612
1613 static void print_footer(void)
1614 {
1615 FILE *output = stat_config.output;
1616 int n;
1617
1618 if (!null_run)
1619 fprintf(output, "\n");
1620 fprintf(output, " %17.9f seconds time elapsed",
1621 avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC);
1622 if (run_count > 1) {
1623 fprintf(output, " ");
1624 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1625 avg_stats(&walltime_nsecs_stats));
1626 }
1627 fprintf(output, "\n\n");
1628
1629 if (print_free_counters_hint &&
1630 sysctl__read_int("kernel/nmi_watchdog", &n) >= 0 &&
1631 n > 0)
1632 fprintf(output,
1633 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1634 " echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1635 " perf stat ...\n"
1636 " echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1637 }
1638
1639 static void print_counters(struct timespec *ts, int argc, const char **argv)
1640 {
1641 int interval = stat_config.interval;
1642 struct perf_evsel *counter;
1643 char buf[64], *prefix = NULL;
1644
1645 /* Do not print anything if we record to the pipe. */
1646 if (STAT_RECORD && perf_stat.file.is_pipe)
1647 return;
1648
1649 if (interval)
1650 print_interval(prefix = buf, ts);
1651 else
1652 print_header(argc, argv);
1653
1654 if (metric_only) {
1655 static int num_print_iv;
1656
1657 if (num_print_iv == 0 && !interval)
1658 print_metric_headers(prefix, false);
1659 if (num_print_iv++ == 25)
1660 num_print_iv = 0;
1661 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1662 fprintf(stat_config.output, "%s", prefix);
1663 }
1664
1665 switch (stat_config.aggr_mode) {
1666 case AGGR_CORE:
1667 case AGGR_SOCKET:
1668 print_aggr(prefix);
1669 break;
1670 case AGGR_THREAD:
1671 evlist__for_each_entry(evsel_list, counter)
1672 print_aggr_thread(counter, prefix);
1673 break;
1674 case AGGR_GLOBAL:
1675 evlist__for_each_entry(evsel_list, counter)
1676 print_counter_aggr(counter, prefix);
1677 if (metric_only)
1678 fputc('\n', stat_config.output);
1679 break;
1680 case AGGR_NONE:
1681 if (metric_only)
1682 print_no_aggr_metric(prefix);
1683 else {
1684 evlist__for_each_entry(evsel_list, counter)
1685 print_counter(counter, prefix);
1686 }
1687 break;
1688 case AGGR_UNSET:
1689 default:
1690 break;
1691 }
1692
1693 if (!interval && !csv_output)
1694 print_footer();
1695
1696 fflush(stat_config.output);
1697 }
1698
1699 static volatile int signr = -1;
1700
1701 static void skip_signal(int signo)
1702 {
1703 if ((child_pid == -1) || stat_config.interval)
1704 done = 1;
1705
1706 signr = signo;
1707 /*
1708 * render child_pid harmless
1709 * won't send SIGTERM to a random
1710 * process in case of race condition
1711 * and fast PID recycling
1712 */
1713 child_pid = -1;
1714 }
1715
1716 static void sig_atexit(void)
1717 {
1718 sigset_t set, oset;
1719
1720 /*
1721 * avoid race condition with SIGCHLD handler
1722 * in skip_signal() which is modifying child_pid
1723 * goal is to avoid send SIGTERM to a random
1724 * process
1725 */
1726 sigemptyset(&set);
1727 sigaddset(&set, SIGCHLD);
1728 sigprocmask(SIG_BLOCK, &set, &oset);
1729
1730 if (child_pid != -1)
1731 kill(child_pid, SIGTERM);
1732
1733 sigprocmask(SIG_SETMASK, &oset, NULL);
1734
1735 if (signr == -1)
1736 return;
1737
1738 signal(signr, SIG_DFL);
1739 kill(getpid(), signr);
1740 }
1741
1742 static int stat__set_big_num(const struct option *opt __maybe_unused,
1743 const char *s __maybe_unused, int unset)
1744 {
1745 big_num_opt = unset ? 0 : 1;
1746 return 0;
1747 }
1748
1749 static int enable_metric_only(const struct option *opt __maybe_unused,
1750 const char *s __maybe_unused, int unset)
1751 {
1752 force_metric_only = true;
1753 metric_only = !unset;
1754 return 0;
1755 }
1756
1757 static const struct option stat_options[] = {
1758 OPT_BOOLEAN('T', "transaction", &transaction_run,
1759 "hardware transaction statistics"),
1760 OPT_CALLBACK('e', "event", &evsel_list, "event",
1761 "event selector. use 'perf list' to list available events",
1762 parse_events_option),
1763 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1764 "event filter", parse_filter),
1765 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1766 "child tasks do not inherit counters"),
1767 OPT_STRING('p', "pid", &target.pid, "pid",
1768 "stat events on existing process id"),
1769 OPT_STRING('t', "tid", &target.tid, "tid",
1770 "stat events on existing thread id"),
1771 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1772 "system-wide collection from all CPUs"),
1773 OPT_BOOLEAN('g', "group", &group,
1774 "put the counters into a counter group"),
1775 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1776 OPT_INCR('v', "verbose", &verbose,
1777 "be more verbose (show counter open errors, etc)"),
1778 OPT_INTEGER('r', "repeat", &run_count,
1779 "repeat command and print average + stddev (max: 100, forever: 0)"),
1780 OPT_BOOLEAN('n', "null", &null_run,
1781 "null run - dont start any counters"),
1782 OPT_INCR('d', "detailed", &detailed_run,
1783 "detailed run - start a lot of events"),
1784 OPT_BOOLEAN('S', "sync", &sync_run,
1785 "call sync() before starting a run"),
1786 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1787 "print large numbers with thousands\' separators",
1788 stat__set_big_num),
1789 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1790 "list of cpus to monitor in system-wide"),
1791 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1792 "disable CPU count aggregation", AGGR_NONE),
1793 OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"),
1794 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1795 "print counts with custom separator"),
1796 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1797 "monitor event in cgroup name only", parse_cgroups),
1798 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1799 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1800 OPT_INTEGER(0, "log-fd", &output_fd,
1801 "log output to fd, instead of stderr"),
1802 OPT_STRING(0, "pre", &pre_cmd, "command",
1803 "command to run prior to the measured command"),
1804 OPT_STRING(0, "post", &post_cmd, "command",
1805 "command to run after to the measured command"),
1806 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1807 "print counts at regular interval in ms (>= 10)"),
1808 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1809 "aggregate counts per processor socket", AGGR_SOCKET),
1810 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1811 "aggregate counts per physical processor core", AGGR_CORE),
1812 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1813 "aggregate counts per thread", AGGR_THREAD),
1814 OPT_UINTEGER('D', "delay", &initial_delay,
1815 "ms to wait before starting measurement after program start"),
1816 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1817 "Only print computed metrics. No raw values", enable_metric_only),
1818 OPT_BOOLEAN(0, "topdown", &topdown_run,
1819 "measure topdown level 1 statistics"),
1820 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1821 "measure SMI cost"),
1822 OPT_END()
1823 };
1824
1825 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1826 {
1827 return cpu_map__get_socket(map, cpu, NULL);
1828 }
1829
1830 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1831 {
1832 return cpu_map__get_core(map, cpu, NULL);
1833 }
1834
1835 static int cpu_map__get_max(struct cpu_map *map)
1836 {
1837 int i, max = -1;
1838
1839 for (i = 0; i < map->nr; i++) {
1840 if (map->map[i] > max)
1841 max = map->map[i];
1842 }
1843
1844 return max;
1845 }
1846
1847 static struct cpu_map *cpus_aggr_map;
1848
1849 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1850 {
1851 int cpu;
1852
1853 if (idx >= map->nr)
1854 return -1;
1855
1856 cpu = map->map[idx];
1857
1858 if (cpus_aggr_map->map[cpu] == -1)
1859 cpus_aggr_map->map[cpu] = get_id(map, idx);
1860
1861 return cpus_aggr_map->map[cpu];
1862 }
1863
1864 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1865 {
1866 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1867 }
1868
1869 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1870 {
1871 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1872 }
1873
1874 static int perf_stat_init_aggr_mode(void)
1875 {
1876 int nr;
1877
1878 switch (stat_config.aggr_mode) {
1879 case AGGR_SOCKET:
1880 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1881 perror("cannot build socket map");
1882 return -1;
1883 }
1884 aggr_get_id = perf_stat__get_socket_cached;
1885 break;
1886 case AGGR_CORE:
1887 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1888 perror("cannot build core map");
1889 return -1;
1890 }
1891 aggr_get_id = perf_stat__get_core_cached;
1892 break;
1893 case AGGR_NONE:
1894 case AGGR_GLOBAL:
1895 case AGGR_THREAD:
1896 case AGGR_UNSET:
1897 default:
1898 break;
1899 }
1900
1901 /*
1902 * The evsel_list->cpus is the base we operate on,
1903 * taking the highest cpu number to be the size of
1904 * the aggregation translate cpumap.
1905 */
1906 nr = cpu_map__get_max(evsel_list->cpus);
1907 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1908 return cpus_aggr_map ? 0 : -ENOMEM;
1909 }
1910
1911 static void perf_stat__exit_aggr_mode(void)
1912 {
1913 cpu_map__put(aggr_map);
1914 cpu_map__put(cpus_aggr_map);
1915 aggr_map = NULL;
1916 cpus_aggr_map = NULL;
1917 }
1918
1919 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1920 {
1921 int cpu;
1922
1923 if (idx > map->nr)
1924 return -1;
1925
1926 cpu = map->map[idx];
1927
1928 if (cpu >= env->nr_cpus_avail)
1929 return -1;
1930
1931 return cpu;
1932 }
1933
1934 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1935 {
1936 struct perf_env *env = data;
1937 int cpu = perf_env__get_cpu(env, map, idx);
1938
1939 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1940 }
1941
1942 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1943 {
1944 struct perf_env *env = data;
1945 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1946
1947 if (cpu != -1) {
1948 int socket_id = env->cpu[cpu].socket_id;
1949
1950 /*
1951 * Encode socket in upper 16 bits
1952 * core_id is relative to socket, and
1953 * we need a global id. So we combine
1954 * socket + core id.
1955 */
1956 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1957 }
1958
1959 return core;
1960 }
1961
1962 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1963 struct cpu_map **sockp)
1964 {
1965 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1966 }
1967
1968 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1969 struct cpu_map **corep)
1970 {
1971 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1972 }
1973
1974 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1975 {
1976 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1977 }
1978
1979 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1980 {
1981 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1982 }
1983
1984 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1985 {
1986 struct perf_env *env = &st->session->header.env;
1987
1988 switch (stat_config.aggr_mode) {
1989 case AGGR_SOCKET:
1990 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1991 perror("cannot build socket map");
1992 return -1;
1993 }
1994 aggr_get_id = perf_stat__get_socket_file;
1995 break;
1996 case AGGR_CORE:
1997 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1998 perror("cannot build core map");
1999 return -1;
2000 }
2001 aggr_get_id = perf_stat__get_core_file;
2002 break;
2003 case AGGR_NONE:
2004 case AGGR_GLOBAL:
2005 case AGGR_THREAD:
2006 case AGGR_UNSET:
2007 default:
2008 break;
2009 }
2010
2011 return 0;
2012 }
2013
2014 static int topdown_filter_events(const char **attr, char **str, bool use_group)
2015 {
2016 int off = 0;
2017 int i;
2018 int len = 0;
2019 char *s;
2020
2021 for (i = 0; attr[i]; i++) {
2022 if (pmu_have_event("cpu", attr[i])) {
2023 len += strlen(attr[i]) + 1;
2024 attr[i - off] = attr[i];
2025 } else
2026 off++;
2027 }
2028 attr[i - off] = NULL;
2029
2030 *str = malloc(len + 1 + 2);
2031 if (!*str)
2032 return -1;
2033 s = *str;
2034 if (i - off == 0) {
2035 *s = 0;
2036 return 0;
2037 }
2038 if (use_group)
2039 *s++ = '{';
2040 for (i = 0; attr[i]; i++) {
2041 strcpy(s, attr[i]);
2042 s += strlen(s);
2043 *s++ = ',';
2044 }
2045 if (use_group) {
2046 s[-1] = '}';
2047 *s = 0;
2048 } else
2049 s[-1] = 0;
2050 return 0;
2051 }
2052
2053 __weak bool arch_topdown_check_group(bool *warn)
2054 {
2055 *warn = false;
2056 return false;
2057 }
2058
2059 __weak void arch_topdown_group_warn(void)
2060 {
2061 }
2062
2063 /*
2064 * Add default attributes, if there were no attributes specified or
2065 * if -d/--detailed, -d -d or -d -d -d is used:
2066 */
2067 static int add_default_attributes(void)
2068 {
2069 int err;
2070 struct perf_event_attr default_attrs0[] = {
2071
2072 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
2073 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
2074 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
2075 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
2076
2077 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
2078 };
2079 struct perf_event_attr frontend_attrs[] = {
2080 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
2081 };
2082 struct perf_event_attr backend_attrs[] = {
2083 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
2084 };
2085 struct perf_event_attr default_attrs1[] = {
2086 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
2087 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
2088 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
2089
2090 };
2091
2092 /*
2093 * Detailed stats (-d), covering the L1 and last level data caches:
2094 */
2095 struct perf_event_attr detailed_attrs[] = {
2096
2097 { .type = PERF_TYPE_HW_CACHE,
2098 .config =
2099 PERF_COUNT_HW_CACHE_L1D << 0 |
2100 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2101 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2102
2103 { .type = PERF_TYPE_HW_CACHE,
2104 .config =
2105 PERF_COUNT_HW_CACHE_L1D << 0 |
2106 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2107 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2108
2109 { .type = PERF_TYPE_HW_CACHE,
2110 .config =
2111 PERF_COUNT_HW_CACHE_LL << 0 |
2112 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2113 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2114
2115 { .type = PERF_TYPE_HW_CACHE,
2116 .config =
2117 PERF_COUNT_HW_CACHE_LL << 0 |
2118 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2119 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2120 };
2121
2122 /*
2123 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2124 */
2125 struct perf_event_attr very_detailed_attrs[] = {
2126
2127 { .type = PERF_TYPE_HW_CACHE,
2128 .config =
2129 PERF_COUNT_HW_CACHE_L1I << 0 |
2130 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2131 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2132
2133 { .type = PERF_TYPE_HW_CACHE,
2134 .config =
2135 PERF_COUNT_HW_CACHE_L1I << 0 |
2136 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2137 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2138
2139 { .type = PERF_TYPE_HW_CACHE,
2140 .config =
2141 PERF_COUNT_HW_CACHE_DTLB << 0 |
2142 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2143 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2144
2145 { .type = PERF_TYPE_HW_CACHE,
2146 .config =
2147 PERF_COUNT_HW_CACHE_DTLB << 0 |
2148 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2149 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2150
2151 { .type = PERF_TYPE_HW_CACHE,
2152 .config =
2153 PERF_COUNT_HW_CACHE_ITLB << 0 |
2154 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2155 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2156
2157 { .type = PERF_TYPE_HW_CACHE,
2158 .config =
2159 PERF_COUNT_HW_CACHE_ITLB << 0 |
2160 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2161 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2162
2163 };
2164
2165 /*
2166 * Very, very detailed stats (-d -d -d), adding prefetch events:
2167 */
2168 struct perf_event_attr very_very_detailed_attrs[] = {
2169
2170 { .type = PERF_TYPE_HW_CACHE,
2171 .config =
2172 PERF_COUNT_HW_CACHE_L1D << 0 |
2173 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2174 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2175
2176 { .type = PERF_TYPE_HW_CACHE,
2177 .config =
2178 PERF_COUNT_HW_CACHE_L1D << 0 |
2179 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2180 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2181 };
2182
2183 /* Set attrs if no event is selected and !null_run: */
2184 if (null_run)
2185 return 0;
2186
2187 if (transaction_run) {
2188 if (pmu_have_event("cpu", "cycles-ct") &&
2189 pmu_have_event("cpu", "el-start"))
2190 err = parse_events(evsel_list, transaction_attrs, NULL);
2191 else
2192 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
2193 if (err) {
2194 fprintf(stderr, "Cannot set up transaction events\n");
2195 return -1;
2196 }
2197 return 0;
2198 }
2199
2200 if (smi_cost) {
2201 int smi;
2202
2203 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2204 fprintf(stderr, "freeze_on_smi is not supported.\n");
2205 return -1;
2206 }
2207
2208 if (!smi) {
2209 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2210 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2211 return -1;
2212 }
2213 smi_reset = true;
2214 }
2215
2216 if (pmu_have_event("msr", "aperf") &&
2217 pmu_have_event("msr", "smi")) {
2218 if (!force_metric_only)
2219 metric_only = true;
2220 err = parse_events(evsel_list, smi_cost_attrs, NULL);
2221 } else {
2222 fprintf(stderr, "To measure SMI cost, it needs "
2223 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
2224 return -1;
2225 }
2226 if (err) {
2227 fprintf(stderr, "Cannot set up SMI cost events\n");
2228 return -1;
2229 }
2230 return 0;
2231 }
2232
2233 if (topdown_run) {
2234 char *str = NULL;
2235 bool warn = false;
2236
2237 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2238 stat_config.aggr_mode != AGGR_CORE) {
2239 pr_err("top down event configuration requires --per-core mode\n");
2240 return -1;
2241 }
2242 stat_config.aggr_mode = AGGR_CORE;
2243 if (nr_cgroups || !target__has_cpu(&target)) {
2244 pr_err("top down event configuration requires system-wide mode (-a)\n");
2245 return -1;
2246 }
2247
2248 if (!force_metric_only)
2249 metric_only = true;
2250 if (topdown_filter_events(topdown_attrs, &str,
2251 arch_topdown_check_group(&warn)) < 0) {
2252 pr_err("Out of memory\n");
2253 return -1;
2254 }
2255 if (topdown_attrs[0] && str) {
2256 if (warn)
2257 arch_topdown_group_warn();
2258 err = parse_events(evsel_list, str, NULL);
2259 if (err) {
2260 fprintf(stderr,
2261 "Cannot set up top down events %s: %d\n",
2262 str, err);
2263 free(str);
2264 return -1;
2265 }
2266 } else {
2267 fprintf(stderr, "System does not support topdown\n");
2268 return -1;
2269 }
2270 free(str);
2271 }
2272
2273 if (!evsel_list->nr_entries) {
2274 if (target__has_cpu(&target))
2275 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2276
2277 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2278 return -1;
2279 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2280 if (perf_evlist__add_default_attrs(evsel_list,
2281 frontend_attrs) < 0)
2282 return -1;
2283 }
2284 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2285 if (perf_evlist__add_default_attrs(evsel_list,
2286 backend_attrs) < 0)
2287 return -1;
2288 }
2289 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2290 return -1;
2291 }
2292
2293 /* Detailed events get appended to the event list: */
2294
2295 if (detailed_run < 1)
2296 return 0;
2297
2298 /* Append detailed run extra attributes: */
2299 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2300 return -1;
2301
2302 if (detailed_run < 2)
2303 return 0;
2304
2305 /* Append very detailed run extra attributes: */
2306 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2307 return -1;
2308
2309 if (detailed_run < 3)
2310 return 0;
2311
2312 /* Append very, very detailed run extra attributes: */
2313 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2314 }
2315
2316 static const char * const stat_record_usage[] = {
2317 "perf stat record [<options>]",
2318 NULL,
2319 };
2320
2321 static void init_features(struct perf_session *session)
2322 {
2323 int feat;
2324
2325 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2326 perf_header__set_feat(&session->header, feat);
2327
2328 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2329 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2330 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2331 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2332 }
2333
2334 static int __cmd_record(int argc, const char **argv)
2335 {
2336 struct perf_session *session;
2337 struct perf_data_file *file = &perf_stat.file;
2338
2339 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2340 PARSE_OPT_STOP_AT_NON_OPTION);
2341
2342 if (output_name)
2343 file->path = output_name;
2344
2345 if (run_count != 1 || forever) {
2346 pr_err("Cannot use -r option with perf stat record.\n");
2347 return -1;
2348 }
2349
2350 session = perf_session__new(file, false, NULL);
2351 if (session == NULL) {
2352 pr_err("Perf session creation failed.\n");
2353 return -1;
2354 }
2355
2356 init_features(session);
2357
2358 session->evlist = evsel_list;
2359 perf_stat.session = session;
2360 perf_stat.record = true;
2361 return argc;
2362 }
2363
2364 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2365 union perf_event *event,
2366 struct perf_session *session)
2367 {
2368 struct stat_round_event *stat_round = &event->stat_round;
2369 struct perf_evsel *counter;
2370 struct timespec tsh, *ts = NULL;
2371 const char **argv = session->header.env.cmdline_argv;
2372 int argc = session->header.env.nr_cmdline;
2373
2374 evlist__for_each_entry(evsel_list, counter)
2375 perf_stat_process_counter(&stat_config, counter);
2376
2377 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2378 update_stats(&walltime_nsecs_stats, stat_round->time);
2379
2380 if (stat_config.interval && stat_round->time) {
2381 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2382 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2383 ts = &tsh;
2384 }
2385
2386 print_counters(ts, argc, argv);
2387 return 0;
2388 }
2389
2390 static
2391 int process_stat_config_event(struct perf_tool *tool,
2392 union perf_event *event,
2393 struct perf_session *session __maybe_unused)
2394 {
2395 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2396
2397 perf_event__read_stat_config(&stat_config, &event->stat_config);
2398
2399 if (cpu_map__empty(st->cpus)) {
2400 if (st->aggr_mode != AGGR_UNSET)
2401 pr_warning("warning: processing task data, aggregation mode not set\n");
2402 return 0;
2403 }
2404
2405 if (st->aggr_mode != AGGR_UNSET)
2406 stat_config.aggr_mode = st->aggr_mode;
2407
2408 if (perf_stat.file.is_pipe)
2409 perf_stat_init_aggr_mode();
2410 else
2411 perf_stat_init_aggr_mode_file(st);
2412
2413 return 0;
2414 }
2415
2416 static int set_maps(struct perf_stat *st)
2417 {
2418 if (!st->cpus || !st->threads)
2419 return 0;
2420
2421 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2422 return -EINVAL;
2423
2424 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2425
2426 if (perf_evlist__alloc_stats(evsel_list, true))
2427 return -ENOMEM;
2428
2429 st->maps_allocated = true;
2430 return 0;
2431 }
2432
2433 static
2434 int process_thread_map_event(struct perf_tool *tool,
2435 union perf_event *event,
2436 struct perf_session *session __maybe_unused)
2437 {
2438 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2439
2440 if (st->threads) {
2441 pr_warning("Extra thread map event, ignoring.\n");
2442 return 0;
2443 }
2444
2445 st->threads = thread_map__new_event(&event->thread_map);
2446 if (!st->threads)
2447 return -ENOMEM;
2448
2449 return set_maps(st);
2450 }
2451
2452 static
2453 int process_cpu_map_event(struct perf_tool *tool,
2454 union perf_event *event,
2455 struct perf_session *session __maybe_unused)
2456 {
2457 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2458 struct cpu_map *cpus;
2459
2460 if (st->cpus) {
2461 pr_warning("Extra cpu map event, ignoring.\n");
2462 return 0;
2463 }
2464
2465 cpus = cpu_map__new_data(&event->cpu_map.data);
2466 if (!cpus)
2467 return -ENOMEM;
2468
2469 st->cpus = cpus;
2470 return set_maps(st);
2471 }
2472
2473 static const char * const stat_report_usage[] = {
2474 "perf stat report [<options>]",
2475 NULL,
2476 };
2477
2478 static struct perf_stat perf_stat = {
2479 .tool = {
2480 .attr = perf_event__process_attr,
2481 .event_update = perf_event__process_event_update,
2482 .thread_map = process_thread_map_event,
2483 .cpu_map = process_cpu_map_event,
2484 .stat_config = process_stat_config_event,
2485 .stat = perf_event__process_stat_event,
2486 .stat_round = process_stat_round_event,
2487 },
2488 .aggr_mode = AGGR_UNSET,
2489 };
2490
2491 static int __cmd_report(int argc, const char **argv)
2492 {
2493 struct perf_session *session;
2494 const struct option options[] = {
2495 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2496 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2497 "aggregate counts per processor socket", AGGR_SOCKET),
2498 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2499 "aggregate counts per physical processor core", AGGR_CORE),
2500 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2501 "disable CPU count aggregation", AGGR_NONE),
2502 OPT_END()
2503 };
2504 struct stat st;
2505 int ret;
2506
2507 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2508
2509 if (!input_name || !strlen(input_name)) {
2510 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2511 input_name = "-";
2512 else
2513 input_name = "perf.data";
2514 }
2515
2516 perf_stat.file.path = input_name;
2517 perf_stat.file.mode = PERF_DATA_MODE_READ;
2518
2519 session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
2520 if (session == NULL)
2521 return -1;
2522
2523 perf_stat.session = session;
2524 stat_config.output = stderr;
2525 evsel_list = session->evlist;
2526
2527 ret = perf_session__process_events(session);
2528 if (ret)
2529 return ret;
2530
2531 perf_session__delete(session);
2532 return 0;
2533 }
2534
2535 static void setup_system_wide(int forks)
2536 {
2537 /*
2538 * Make system wide (-a) the default target if
2539 * no target was specified and one of following
2540 * conditions is met:
2541 *
2542 * - there's no workload specified
2543 * - there is workload specified but all requested
2544 * events are system wide events
2545 */
2546 if (!target__none(&target))
2547 return;
2548
2549 if (!forks)
2550 target.system_wide = true;
2551 else {
2552 struct perf_evsel *counter;
2553
2554 evlist__for_each_entry(evsel_list, counter) {
2555 if (!counter->system_wide)
2556 return;
2557 }
2558
2559 if (evsel_list->nr_entries)
2560 target.system_wide = true;
2561 }
2562 }
2563
2564 int cmd_stat(int argc, const char **argv)
2565 {
2566 const char * const stat_usage[] = {
2567 "perf stat [<options>] [<command>]",
2568 NULL
2569 };
2570 int status = -EINVAL, run_idx;
2571 const char *mode;
2572 FILE *output = stderr;
2573 unsigned int interval;
2574 const char * const stat_subcommands[] = { "record", "report" };
2575
2576 setlocale(LC_ALL, "");
2577
2578 evsel_list = perf_evlist__new();
2579 if (evsel_list == NULL)
2580 return -ENOMEM;
2581
2582 parse_events__shrink_config_terms();
2583 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2584 (const char **) stat_usage,
2585 PARSE_OPT_STOP_AT_NON_OPTION);
2586 perf_stat__collect_metric_expr(evsel_list);
2587 perf_stat__init_shadow_stats();
2588
2589 if (csv_sep) {
2590 csv_output = true;
2591 if (!strcmp(csv_sep, "\\t"))
2592 csv_sep = "\t";
2593 } else
2594 csv_sep = DEFAULT_SEPARATOR;
2595
2596 if (argc && !strncmp(argv[0], "rec", 3)) {
2597 argc = __cmd_record(argc, argv);
2598 if (argc < 0)
2599 return -1;
2600 } else if (argc && !strncmp(argv[0], "rep", 3))
2601 return __cmd_report(argc, argv);
2602
2603 interval = stat_config.interval;
2604
2605 /*
2606 * For record command the -o is already taken care of.
2607 */
2608 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2609 output = NULL;
2610
2611 if (output_name && output_fd) {
2612 fprintf(stderr, "cannot use both --output and --log-fd\n");
2613 parse_options_usage(stat_usage, stat_options, "o", 1);
2614 parse_options_usage(NULL, stat_options, "log-fd", 0);
2615 goto out;
2616 }
2617
2618 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2619 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2620 goto out;
2621 }
2622
2623 if (metric_only && run_count > 1) {
2624 fprintf(stderr, "--metric-only is not supported with -r\n");
2625 goto out;
2626 }
2627
2628 if (output_fd < 0) {
2629 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2630 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2631 goto out;
2632 }
2633
2634 if (!output) {
2635 struct timespec tm;
2636 mode = append_file ? "a" : "w";
2637
2638 output = fopen(output_name, mode);
2639 if (!output) {
2640 perror("failed to create output file");
2641 return -1;
2642 }
2643 clock_gettime(CLOCK_REALTIME, &tm);
2644 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2645 } else if (output_fd > 0) {
2646 mode = append_file ? "a" : "w";
2647 output = fdopen(output_fd, mode);
2648 if (!output) {
2649 perror("Failed opening logfd");
2650 return -errno;
2651 }
2652 }
2653
2654 stat_config.output = output;
2655
2656 /*
2657 * let the spreadsheet do the pretty-printing
2658 */
2659 if (csv_output) {
2660 /* User explicitly passed -B? */
2661 if (big_num_opt == 1) {
2662 fprintf(stderr, "-B option not supported with -x\n");
2663 parse_options_usage(stat_usage, stat_options, "B", 1);
2664 parse_options_usage(NULL, stat_options, "x", 1);
2665 goto out;
2666 } else /* Nope, so disable big number formatting */
2667 big_num = false;
2668 } else if (big_num_opt == 0) /* User passed --no-big-num */
2669 big_num = false;
2670
2671 setup_system_wide(argc);
2672
2673 if (run_count < 0) {
2674 pr_err("Run count must be a positive number\n");
2675 parse_options_usage(stat_usage, stat_options, "r", 1);
2676 goto out;
2677 } else if (run_count == 0) {
2678 forever = true;
2679 run_count = 1;
2680 }
2681
2682 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2683 fprintf(stderr, "The --per-thread option is only available "
2684 "when monitoring via -p -t options.\n");
2685 parse_options_usage(NULL, stat_options, "p", 1);
2686 parse_options_usage(NULL, stat_options, "t", 1);
2687 goto out;
2688 }
2689
2690 /*
2691 * no_aggr, cgroup are for system-wide only
2692 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2693 */
2694 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2695 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2696 !target__has_cpu(&target)) {
2697 fprintf(stderr, "both cgroup and no-aggregation "
2698 "modes only available in system-wide mode\n");
2699
2700 parse_options_usage(stat_usage, stat_options, "G", 1);
2701 parse_options_usage(NULL, stat_options, "A", 1);
2702 parse_options_usage(NULL, stat_options, "a", 1);
2703 goto out;
2704 }
2705
2706 if (add_default_attributes())
2707 goto out;
2708
2709 target__validate(&target);
2710
2711 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2712 if (target__has_task(&target)) {
2713 pr_err("Problems finding threads of monitor\n");
2714 parse_options_usage(stat_usage, stat_options, "p", 1);
2715 parse_options_usage(NULL, stat_options, "t", 1);
2716 } else if (target__has_cpu(&target)) {
2717 perror("failed to parse CPUs map");
2718 parse_options_usage(stat_usage, stat_options, "C", 1);
2719 parse_options_usage(NULL, stat_options, "a", 1);
2720 }
2721 goto out;
2722 }
2723
2724 /*
2725 * Initialize thread_map with comm names,
2726 * so we could print it out on output.
2727 */
2728 if (stat_config.aggr_mode == AGGR_THREAD)
2729 thread_map__read_comms(evsel_list->threads);
2730
2731 if (interval && interval < 100) {
2732 if (interval < 10) {
2733 pr_err("print interval must be >= 10ms\n");
2734 parse_options_usage(stat_usage, stat_options, "I", 1);
2735 goto out;
2736 } else
2737 pr_warning("print interval < 100ms. "
2738 "The overhead percentage could be high in some cases. "
2739 "Please proceed with caution.\n");
2740 }
2741
2742 if (perf_evlist__alloc_stats(evsel_list, interval))
2743 goto out;
2744
2745 if (perf_stat_init_aggr_mode())
2746 goto out;
2747
2748 /*
2749 * We dont want to block the signals - that would cause
2750 * child tasks to inherit that and Ctrl-C would not work.
2751 * What we want is for Ctrl-C to work in the exec()-ed
2752 * task, but being ignored by perf stat itself:
2753 */
2754 atexit(sig_atexit);
2755 if (!forever)
2756 signal(SIGINT, skip_signal);
2757 signal(SIGCHLD, skip_signal);
2758 signal(SIGALRM, skip_signal);
2759 signal(SIGABRT, skip_signal);
2760
2761 status = 0;
2762 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2763 if (run_count != 1 && verbose > 0)
2764 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2765 run_idx + 1);
2766
2767 status = run_perf_stat(argc, argv);
2768 if (forever && status != -1) {
2769 print_counters(NULL, argc, argv);
2770 perf_stat__reset_stats();
2771 }
2772 }
2773
2774 if (!forever && status != -1 && !interval)
2775 print_counters(NULL, argc, argv);
2776
2777 if (STAT_RECORD) {
2778 /*
2779 * We synthesize the kernel mmap record just so that older tools
2780 * don't emit warnings about not being able to resolve symbols
2781 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2782 * a saner message about no samples being in the perf.data file.
2783 *
2784 * This also serves to suppress a warning about f_header.data.size == 0
2785 * in header.c at the moment 'perf stat record' gets introduced, which
2786 * is not really needed once we start adding the stat specific PERF_RECORD_
2787 * records, but the need to suppress the kptr_restrict messages in older
2788 * tools remain -acme
2789 */
2790 int fd = perf_data_file__fd(&perf_stat.file);
2791 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2792 process_synthesized_event,
2793 &perf_stat.session->machines.host);
2794 if (err) {
2795 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2796 "older tools may produce warnings about this file\n.");
2797 }
2798
2799 if (!interval) {
2800 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2801 pr_err("failed to write stat round event\n");
2802 }
2803
2804 if (!perf_stat.file.is_pipe) {
2805 perf_stat.session->header.data_size += perf_stat.bytes_written;
2806 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2807 }
2808
2809 perf_session__delete(perf_stat.session);
2810 }
2811
2812 perf_stat__exit_aggr_mode();
2813 perf_evlist__free_stats(evsel_list);
2814 out:
2815 if (smi_cost && smi_reset)
2816 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2817
2818 perf_evlist__delete(evsel_list);
2819 return status;
2820 }
CRIS linux kernel tree