Loading...
1/*
2 * builtin-stat.c
3 *
4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
6 *
7 * Sample output:
8
9 $ perf stat ./hackbench 10
10
11 Time: 0.118
12
13 Performance counter stats for './hackbench 10':
14
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
26
27 0.154822978 seconds time elapsed
28
29 *
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
31 *
32 * Improvements and fixes by:
33 *
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
40 *
41 * Released under the GPL v2. (and only v2, not any later version)
42 */
43
44#include "perf.h"
45#include "builtin.h"
46#include "util/cgroup.h"
47#include "util/util.h"
48#include <subcmd/parse-options.h>
49#include "util/parse-events.h"
50#include "util/pmu.h"
51#include "util/event.h"
52#include "util/evlist.h"
53#include "util/evsel.h"
54#include "util/debug.h"
55#include "util/drv_configs.h"
56#include "util/color.h"
57#include "util/stat.h"
58#include "util/header.h"
59#include "util/cpumap.h"
60#include "util/thread.h"
61#include "util/thread_map.h"
62#include "util/counts.h"
63#include "util/group.h"
64#include "util/session.h"
65#include "util/tool.h"
66#include "util/string2.h"
67#include "util/metricgroup.h"
68#include "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
91static void print_counters(struct timespec *ts, int argc, const char **argv);
92
93/* Default events used for perf stat -T */
94static 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. */
107static const char * transaction_limited_attrs = {
108 "task-clock,"
109 "{"
110 "instructions,"
111 "cycles,"
112 "cpu/cycles-t/,"
113 "cpu/tx-start/"
114 "}"
115};
116
117static 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
126static const char *smi_cost_attrs = {
127 "{"
128 "msr/aperf/,"
129 "msr/smi/,"
130 "cycles"
131 "}"
132};
133
134static struct perf_evlist *evsel_list;
135
136static struct rblist metric_events;
137
138static struct target target = {
139 .uid = UINT_MAX,
140};
141
142typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
143
144static int run_count = 1;
145static bool no_inherit = false;
146static volatile pid_t child_pid = -1;
147static bool null_run = false;
148static int detailed_run = 0;
149static bool transaction_run;
150static bool topdown_run = false;
151static bool smi_cost = false;
152static bool smi_reset = false;
153static bool big_num = true;
154static int big_num_opt = -1;
155static const char *csv_sep = NULL;
156static bool csv_output = false;
157static bool group = false;
158static const char *pre_cmd = NULL;
159static const char *post_cmd = NULL;
160static bool sync_run = false;
161static unsigned int initial_delay = 0;
162static unsigned int unit_width = 4; /* strlen("unit") */
163static bool forever = false;
164static bool metric_only = false;
165static bool force_metric_only = false;
166static bool no_merge = false;
167static struct timespec ref_time;
168static struct cpu_map *aggr_map;
169static aggr_get_id_t aggr_get_id;
170static bool append_file;
171static bool interval_count;
172static const char *output_name;
173static int output_fd;
174static int print_free_counters_hint;
175static int print_mixed_hw_group_error;
176
177struct perf_stat {
178 bool record;
179 struct perf_data data;
180 struct perf_session *session;
181 u64 bytes_written;
182 struct perf_tool tool;
183 bool maps_allocated;
184 struct cpu_map *cpus;
185 struct thread_map *threads;
186 enum aggr_mode aggr_mode;
187};
188
189static struct perf_stat perf_stat;
190#define STAT_RECORD perf_stat.record
191
192static volatile int done = 0;
193
194static struct perf_stat_config stat_config = {
195 .aggr_mode = AGGR_GLOBAL,
196 .scale = true,
197};
198
199static bool is_duration_time(struct perf_evsel *evsel)
200{
201 return !strcmp(evsel->name, "duration_time");
202}
203
204static inline void diff_timespec(struct timespec *r, struct timespec *a,
205 struct timespec *b)
206{
207 r->tv_sec = a->tv_sec - b->tv_sec;
208 if (a->tv_nsec < b->tv_nsec) {
209 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
210 r->tv_sec--;
211 } else {
212 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
213 }
214}
215
216static void perf_stat__reset_stats(void)
217{
218 int i;
219
220 perf_evlist__reset_stats(evsel_list);
221 perf_stat__reset_shadow_stats();
222
223 for (i = 0; i < stat_config.stats_num; i++)
224 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
225}
226
227static int create_perf_stat_counter(struct perf_evsel *evsel)
228{
229 struct perf_event_attr *attr = &evsel->attr;
230 struct perf_evsel *leader = evsel->leader;
231
232 if (stat_config.scale) {
233 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
234 PERF_FORMAT_TOTAL_TIME_RUNNING;
235 }
236
237 /*
238 * The event is part of non trivial group, let's enable
239 * the group read (for leader) and ID retrieval for all
240 * members.
241 */
242 if (leader->nr_members > 1)
243 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
244
245 attr->inherit = !no_inherit;
246
247 /*
248 * Some events get initialized with sample_(period/type) set,
249 * like tracepoints. Clear it up for counting.
250 */
251 attr->sample_period = 0;
252
253 /*
254 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
255 * while avoiding that older tools show confusing messages.
256 *
257 * However for pipe sessions we need to keep it zero,
258 * because script's perf_evsel__check_attr is triggered
259 * by attr->sample_type != 0, and we can't run it on
260 * stat sessions.
261 */
262 if (!(STAT_RECORD && perf_stat.data.is_pipe))
263 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
264
265 /*
266 * Disabling all counters initially, they will be enabled
267 * either manually by us or by kernel via enable_on_exec
268 * set later.
269 */
270 if (perf_evsel__is_group_leader(evsel)) {
271 attr->disabled = 1;
272
273 /*
274 * In case of initial_delay we enable tracee
275 * events manually.
276 */
277 if (target__none(&target) && !initial_delay)
278 attr->enable_on_exec = 1;
279 }
280
281 if (target__has_cpu(&target) && !target__has_per_thread(&target))
282 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
283
284 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
285}
286
287/*
288 * Does the counter have nsecs as a unit?
289 */
290static inline int nsec_counter(struct perf_evsel *evsel)
291{
292 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
293 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
294 return 1;
295
296 return 0;
297}
298
299static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
300 union perf_event *event,
301 struct perf_sample *sample __maybe_unused,
302 struct machine *machine __maybe_unused)
303{
304 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
305 pr_err("failed to write perf data, error: %m\n");
306 return -1;
307 }
308
309 perf_stat.bytes_written += event->header.size;
310 return 0;
311}
312
313static int write_stat_round_event(u64 tm, u64 type)
314{
315 return perf_event__synthesize_stat_round(NULL, tm, type,
316 process_synthesized_event,
317 NULL);
318}
319
320#define WRITE_STAT_ROUND_EVENT(time, interval) \
321 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
322
323#define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
324
325static int
326perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
327 struct perf_counts_values *count)
328{
329 struct perf_sample_id *sid = SID(counter, cpu, thread);
330
331 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
332 process_synthesized_event, NULL);
333}
334
335/*
336 * Read out the results of a single counter:
337 * do not aggregate counts across CPUs in system-wide mode
338 */
339static int read_counter(struct perf_evsel *counter)
340{
341 int nthreads = thread_map__nr(evsel_list->threads);
342 int ncpus, cpu, thread;
343
344 if (target__has_cpu(&target) && !target__has_per_thread(&target))
345 ncpus = perf_evsel__nr_cpus(counter);
346 else
347 ncpus = 1;
348
349 if (!counter->supported)
350 return -ENOENT;
351
352 if (counter->system_wide)
353 nthreads = 1;
354
355 for (thread = 0; thread < nthreads; thread++) {
356 for (cpu = 0; cpu < ncpus; cpu++) {
357 struct perf_counts_values *count;
358
359 count = perf_counts(counter->counts, cpu, thread);
360
361 /*
362 * The leader's group read loads data into its group members
363 * (via perf_evsel__read_counter) and sets threir count->loaded.
364 */
365 if (!count->loaded &&
366 perf_evsel__read_counter(counter, cpu, thread)) {
367 counter->counts->scaled = -1;
368 perf_counts(counter->counts, cpu, thread)->ena = 0;
369 perf_counts(counter->counts, cpu, thread)->run = 0;
370 return -1;
371 }
372
373 count->loaded = false;
374
375 if (STAT_RECORD) {
376 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
377 pr_err("failed to write stat event\n");
378 return -1;
379 }
380 }
381
382 if (verbose > 1) {
383 fprintf(stat_config.output,
384 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
385 perf_evsel__name(counter),
386 cpu,
387 count->val, count->ena, count->run);
388 }
389 }
390 }
391
392 return 0;
393}
394
395static void read_counters(void)
396{
397 struct perf_evsel *counter;
398 int ret;
399
400 evlist__for_each_entry(evsel_list, counter) {
401 ret = read_counter(counter);
402 if (ret)
403 pr_debug("failed to read counter %s\n", counter->name);
404
405 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
406 pr_warning("failed to process counter %s\n", counter->name);
407 }
408}
409
410static void process_interval(void)
411{
412 struct timespec ts, rs;
413
414 read_counters();
415
416 clock_gettime(CLOCK_MONOTONIC, &ts);
417 diff_timespec(&rs, &ts, &ref_time);
418
419 if (STAT_RECORD) {
420 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
421 pr_err("failed to write stat round event\n");
422 }
423
424 init_stats(&walltime_nsecs_stats);
425 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
426 print_counters(&rs, 0, NULL);
427}
428
429static void enable_counters(void)
430{
431 if (initial_delay)
432 usleep(initial_delay * USEC_PER_MSEC);
433
434 /*
435 * We need to enable counters only if:
436 * - we don't have tracee (attaching to task or cpu)
437 * - we have initial delay configured
438 */
439 if (!target__none(&target) || initial_delay)
440 perf_evlist__enable(evsel_list);
441}
442
443static void disable_counters(void)
444{
445 /*
446 * If we don't have tracee (attaching to task or cpu), counters may
447 * still be running. To get accurate group ratios, we must stop groups
448 * from counting before reading their constituent counters.
449 */
450 if (!target__none(&target))
451 perf_evlist__disable(evsel_list);
452}
453
454static volatile int workload_exec_errno;
455
456/*
457 * perf_evlist__prepare_workload will send a SIGUSR1
458 * if the fork fails, since we asked by setting its
459 * want_signal to true.
460 */
461static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
462 void *ucontext __maybe_unused)
463{
464 workload_exec_errno = info->si_value.sival_int;
465}
466
467static int perf_stat_synthesize_config(bool is_pipe)
468{
469 int err;
470
471 if (is_pipe) {
472 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
473 process_synthesized_event);
474 if (err < 0) {
475 pr_err("Couldn't synthesize attrs.\n");
476 return err;
477 }
478 }
479
480 err = perf_event__synthesize_extra_attr(NULL,
481 evsel_list,
482 process_synthesized_event,
483 is_pipe);
484
485 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
486 process_synthesized_event,
487 NULL);
488 if (err < 0) {
489 pr_err("Couldn't synthesize thread map.\n");
490 return err;
491 }
492
493 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
494 process_synthesized_event, NULL);
495 if (err < 0) {
496 pr_err("Couldn't synthesize thread map.\n");
497 return err;
498 }
499
500 err = perf_event__synthesize_stat_config(NULL, &stat_config,
501 process_synthesized_event, NULL);
502 if (err < 0) {
503 pr_err("Couldn't synthesize config.\n");
504 return err;
505 }
506
507 return 0;
508}
509
510#define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
511
512static int __store_counter_ids(struct perf_evsel *counter)
513{
514 int cpu, thread;
515
516 for (cpu = 0; cpu < xyarray__max_x(counter->fd); cpu++) {
517 for (thread = 0; thread < xyarray__max_y(counter->fd);
518 thread++) {
519 int fd = FD(counter, cpu, thread);
520
521 if (perf_evlist__id_add_fd(evsel_list, counter,
522 cpu, thread, fd) < 0)
523 return -1;
524 }
525 }
526
527 return 0;
528}
529
530static int store_counter_ids(struct perf_evsel *counter)
531{
532 struct cpu_map *cpus = counter->cpus;
533 struct thread_map *threads = counter->threads;
534
535 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
536 return -ENOMEM;
537
538 return __store_counter_ids(counter);
539}
540
541static bool perf_evsel__should_store_id(struct perf_evsel *counter)
542{
543 return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
544}
545
546static struct perf_evsel *perf_evsel__reset_weak_group(struct perf_evsel *evsel)
547{
548 struct perf_evsel *c2, *leader;
549 bool is_open = true;
550
551 leader = evsel->leader;
552 pr_debug("Weak group for %s/%d failed\n",
553 leader->name, leader->nr_members);
554
555 /*
556 * for_each_group_member doesn't work here because it doesn't
557 * include the first entry.
558 */
559 evlist__for_each_entry(evsel_list, c2) {
560 if (c2 == evsel)
561 is_open = false;
562 if (c2->leader == leader) {
563 if (is_open)
564 perf_evsel__close(c2);
565 c2->leader = c2;
566 c2->nr_members = 0;
567 }
568 }
569 return leader;
570}
571
572static int __run_perf_stat(int argc, const char **argv)
573{
574 int interval = stat_config.interval;
575 int times = stat_config.times;
576 int timeout = stat_config.timeout;
577 char msg[BUFSIZ];
578 unsigned long long t0, t1;
579 struct perf_evsel *counter;
580 struct timespec ts;
581 size_t l;
582 int status = 0;
583 const bool forks = (argc > 0);
584 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
585 struct perf_evsel_config_term *err_term;
586
587 if (interval) {
588 ts.tv_sec = interval / USEC_PER_MSEC;
589 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
590 } else if (timeout) {
591 ts.tv_sec = timeout / USEC_PER_MSEC;
592 ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
593 } else {
594 ts.tv_sec = 1;
595 ts.tv_nsec = 0;
596 }
597
598 if (forks) {
599 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
600 workload_exec_failed_signal) < 0) {
601 perror("failed to prepare workload");
602 return -1;
603 }
604 child_pid = evsel_list->workload.pid;
605 }
606
607 if (group)
608 perf_evlist__set_leader(evsel_list);
609
610 evlist__for_each_entry(evsel_list, counter) {
611try_again:
612 if (create_perf_stat_counter(counter) < 0) {
613
614 /* Weak group failed. Reset the group. */
615 if ((errno == EINVAL || errno == EBADF) &&
616 counter->leader != counter &&
617 counter->weak_group) {
618 counter = perf_evsel__reset_weak_group(counter);
619 goto try_again;
620 }
621
622 /*
623 * PPC returns ENXIO for HW counters until 2.6.37
624 * (behavior changed with commit b0a873e).
625 */
626 if (errno == EINVAL || errno == ENOSYS ||
627 errno == ENOENT || errno == EOPNOTSUPP ||
628 errno == ENXIO) {
629 if (verbose > 0)
630 ui__warning("%s event is not supported by the kernel.\n",
631 perf_evsel__name(counter));
632 counter->supported = false;
633
634 if ((counter->leader != counter) ||
635 !(counter->leader->nr_members > 1))
636 continue;
637 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
638 if (verbose > 0)
639 ui__warning("%s\n", msg);
640 goto try_again;
641 } else if (target__has_per_thread(&target) &&
642 evsel_list->threads &&
643 evsel_list->threads->err_thread != -1) {
644 /*
645 * For global --per-thread case, skip current
646 * error thread.
647 */
648 if (!thread_map__remove(evsel_list->threads,
649 evsel_list->threads->err_thread)) {
650 evsel_list->threads->err_thread = -1;
651 goto try_again;
652 }
653 }
654
655 perf_evsel__open_strerror(counter, &target,
656 errno, msg, sizeof(msg));
657 ui__error("%s\n", msg);
658
659 if (child_pid != -1)
660 kill(child_pid, SIGTERM);
661
662 return -1;
663 }
664 counter->supported = true;
665
666 l = strlen(counter->unit);
667 if (l > unit_width)
668 unit_width = l;
669
670 if (perf_evsel__should_store_id(counter) &&
671 store_counter_ids(counter))
672 return -1;
673 }
674
675 if (perf_evlist__apply_filters(evsel_list, &counter)) {
676 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
677 counter->filter, perf_evsel__name(counter), errno,
678 str_error_r(errno, msg, sizeof(msg)));
679 return -1;
680 }
681
682 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
683 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
684 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
685 str_error_r(errno, msg, sizeof(msg)));
686 return -1;
687 }
688
689 if (STAT_RECORD) {
690 int err, fd = perf_data__fd(&perf_stat.data);
691
692 if (is_pipe) {
693 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
694 } else {
695 err = perf_session__write_header(perf_stat.session, evsel_list,
696 fd, false);
697 }
698
699 if (err < 0)
700 return err;
701
702 err = perf_stat_synthesize_config(is_pipe);
703 if (err < 0)
704 return err;
705 }
706
707 /*
708 * Enable counters and exec the command:
709 */
710 t0 = rdclock();
711 clock_gettime(CLOCK_MONOTONIC, &ref_time);
712
713 if (forks) {
714 perf_evlist__start_workload(evsel_list);
715 enable_counters();
716
717 if (interval || timeout) {
718 while (!waitpid(child_pid, &status, WNOHANG)) {
719 nanosleep(&ts, NULL);
720 if (timeout)
721 break;
722 process_interval();
723 if (interval_count && !(--times))
724 break;
725 }
726 }
727 waitpid(child_pid, &status, 0);
728
729 if (workload_exec_errno) {
730 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
731 pr_err("Workload failed: %s\n", emsg);
732 return -1;
733 }
734
735 if (WIFSIGNALED(status))
736 psignal(WTERMSIG(status), argv[0]);
737 } else {
738 enable_counters();
739 while (!done) {
740 nanosleep(&ts, NULL);
741 if (timeout)
742 break;
743 if (interval) {
744 process_interval();
745 if (interval_count && !(--times))
746 break;
747 }
748 }
749 }
750
751 disable_counters();
752
753 t1 = rdclock();
754
755 update_stats(&walltime_nsecs_stats, t1 - t0);
756
757 /*
758 * Closing a group leader splits the group, and as we only disable
759 * group leaders, results in remaining events becoming enabled. To
760 * avoid arbitrary skew, we must read all counters before closing any
761 * group leaders.
762 */
763 read_counters();
764 perf_evlist__close(evsel_list);
765
766 return WEXITSTATUS(status);
767}
768
769static int run_perf_stat(int argc, const char **argv)
770{
771 int ret;
772
773 if (pre_cmd) {
774 ret = system(pre_cmd);
775 if (ret)
776 return ret;
777 }
778
779 if (sync_run)
780 sync();
781
782 ret = __run_perf_stat(argc, argv);
783 if (ret)
784 return ret;
785
786 if (post_cmd) {
787 ret = system(post_cmd);
788 if (ret)
789 return ret;
790 }
791
792 return ret;
793}
794
795static void print_running(u64 run, u64 ena)
796{
797 if (csv_output) {
798 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
799 csv_sep,
800 run,
801 csv_sep,
802 ena ? 100.0 * run / ena : 100.0);
803 } else if (run != ena) {
804 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
805 }
806}
807
808static void print_noise_pct(double total, double avg)
809{
810 double pct = rel_stddev_stats(total, avg);
811
812 if (csv_output)
813 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
814 else if (pct)
815 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
816}
817
818static void print_noise(struct perf_evsel *evsel, double avg)
819{
820 struct perf_stat_evsel *ps;
821
822 if (run_count == 1)
823 return;
824
825 ps = evsel->stats;
826 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
827}
828
829static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
830{
831 switch (stat_config.aggr_mode) {
832 case AGGR_CORE:
833 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
834 cpu_map__id_to_socket(id),
835 csv_output ? 0 : -8,
836 cpu_map__id_to_cpu(id),
837 csv_sep,
838 csv_output ? 0 : 4,
839 nr,
840 csv_sep);
841 break;
842 case AGGR_SOCKET:
843 fprintf(stat_config.output, "S%*d%s%*d%s",
844 csv_output ? 0 : -5,
845 id,
846 csv_sep,
847 csv_output ? 0 : 4,
848 nr,
849 csv_sep);
850 break;
851 case AGGR_NONE:
852 fprintf(stat_config.output, "CPU%*d%s",
853 csv_output ? 0 : -4,
854 perf_evsel__cpus(evsel)->map[id], csv_sep);
855 break;
856 case AGGR_THREAD:
857 fprintf(stat_config.output, "%*s-%*d%s",
858 csv_output ? 0 : 16,
859 thread_map__comm(evsel->threads, id),
860 csv_output ? 0 : -8,
861 thread_map__pid(evsel->threads, id),
862 csv_sep);
863 break;
864 case AGGR_GLOBAL:
865 case AGGR_UNSET:
866 default:
867 break;
868 }
869}
870
871struct outstate {
872 FILE *fh;
873 bool newline;
874 const char *prefix;
875 int nfields;
876 int id, nr;
877 struct perf_evsel *evsel;
878};
879
880#define METRIC_LEN 35
881
882static void new_line_std(void *ctx)
883{
884 struct outstate *os = ctx;
885
886 os->newline = true;
887}
888
889static void do_new_line_std(struct outstate *os)
890{
891 fputc('\n', os->fh);
892 fputs(os->prefix, os->fh);
893 aggr_printout(os->evsel, os->id, os->nr);
894 if (stat_config.aggr_mode == AGGR_NONE)
895 fprintf(os->fh, " ");
896 fprintf(os->fh, " ");
897}
898
899static void print_metric_std(void *ctx, const char *color, const char *fmt,
900 const char *unit, double val)
901{
902 struct outstate *os = ctx;
903 FILE *out = os->fh;
904 int n;
905 bool newline = os->newline;
906
907 os->newline = false;
908
909 if (unit == NULL || fmt == NULL) {
910 fprintf(out, "%-*s", METRIC_LEN, "");
911 return;
912 }
913
914 if (newline)
915 do_new_line_std(os);
916
917 n = fprintf(out, " # ");
918 if (color)
919 n += color_fprintf(out, color, fmt, val);
920 else
921 n += fprintf(out, fmt, val);
922 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
923}
924
925static void new_line_csv(void *ctx)
926{
927 struct outstate *os = ctx;
928 int i;
929
930 fputc('\n', os->fh);
931 if (os->prefix)
932 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
933 aggr_printout(os->evsel, os->id, os->nr);
934 for (i = 0; i < os->nfields; i++)
935 fputs(csv_sep, os->fh);
936}
937
938static void print_metric_csv(void *ctx,
939 const char *color __maybe_unused,
940 const char *fmt, const char *unit, double val)
941{
942 struct outstate *os = ctx;
943 FILE *out = os->fh;
944 char buf[64], *vals, *ends;
945
946 if (unit == NULL || fmt == NULL) {
947 fprintf(out, "%s%s", csv_sep, csv_sep);
948 return;
949 }
950 snprintf(buf, sizeof(buf), fmt, val);
951 ends = vals = ltrim(buf);
952 while (isdigit(*ends) || *ends == '.')
953 ends++;
954 *ends = 0;
955 while (isspace(*unit))
956 unit++;
957 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
958}
959
960#define METRIC_ONLY_LEN 20
961
962/* Filter out some columns that don't work well in metrics only mode */
963
964static bool valid_only_metric(const char *unit)
965{
966 if (!unit)
967 return false;
968 if (strstr(unit, "/sec") ||
969 strstr(unit, "hz") ||
970 strstr(unit, "Hz") ||
971 strstr(unit, "CPUs utilized"))
972 return false;
973 return true;
974}
975
976static const char *fixunit(char *buf, struct perf_evsel *evsel,
977 const char *unit)
978{
979 if (!strncmp(unit, "of all", 6)) {
980 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
981 unit);
982 return buf;
983 }
984 return unit;
985}
986
987static void print_metric_only(void *ctx, const char *color, const char *fmt,
988 const char *unit, double val)
989{
990 struct outstate *os = ctx;
991 FILE *out = os->fh;
992 int n;
993 char buf[1024];
994 unsigned mlen = METRIC_ONLY_LEN;
995
996 if (!valid_only_metric(unit))
997 return;
998 unit = fixunit(buf, os->evsel, unit);
999 if (color)
1000 n = color_fprintf(out, color, fmt, val);
1001 else
1002 n = fprintf(out, fmt, val);
1003 if (n > METRIC_ONLY_LEN)
1004 n = METRIC_ONLY_LEN;
1005 if (mlen < strlen(unit))
1006 mlen = strlen(unit) + 1;
1007 fprintf(out, "%*s", mlen - n, "");
1008}
1009
1010static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
1011 const char *fmt,
1012 const char *unit, double val)
1013{
1014 struct outstate *os = ctx;
1015 FILE *out = os->fh;
1016 char buf[64], *vals, *ends;
1017 char tbuf[1024];
1018
1019 if (!valid_only_metric(unit))
1020 return;
1021 unit = fixunit(tbuf, os->evsel, unit);
1022 snprintf(buf, sizeof buf, fmt, val);
1023 ends = vals = ltrim(buf);
1024 while (isdigit(*ends) || *ends == '.')
1025 ends++;
1026 *ends = 0;
1027 fprintf(out, "%s%s", vals, csv_sep);
1028}
1029
1030static void new_line_metric(void *ctx __maybe_unused)
1031{
1032}
1033
1034static void print_metric_header(void *ctx, const char *color __maybe_unused,
1035 const char *fmt __maybe_unused,
1036 const char *unit, double val __maybe_unused)
1037{
1038 struct outstate *os = ctx;
1039 char tbuf[1024];
1040
1041 if (!valid_only_metric(unit))
1042 return;
1043 unit = fixunit(tbuf, os->evsel, unit);
1044 if (csv_output)
1045 fprintf(os->fh, "%s%s", unit, csv_sep);
1046 else
1047 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
1048}
1049
1050static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1051{
1052 FILE *output = stat_config.output;
1053 double msecs = avg / NSEC_PER_MSEC;
1054 const char *fmt_v, *fmt_n;
1055 char name[25];
1056
1057 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
1058 fmt_n = csv_output ? "%s" : "%-25s";
1059
1060 aggr_printout(evsel, id, nr);
1061
1062 scnprintf(name, sizeof(name), "%s%s",
1063 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
1064
1065 fprintf(output, fmt_v, msecs, csv_sep);
1066
1067 if (csv_output)
1068 fprintf(output, "%s%s", evsel->unit, csv_sep);
1069 else
1070 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
1071
1072 fprintf(output, fmt_n, name);
1073
1074 if (evsel->cgrp)
1075 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1076}
1077
1078static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1079{
1080 int i;
1081
1082 if (!aggr_get_id)
1083 return 0;
1084
1085 if (stat_config.aggr_mode == AGGR_NONE)
1086 return id;
1087
1088 if (stat_config.aggr_mode == AGGR_GLOBAL)
1089 return 0;
1090
1091 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1092 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1093
1094 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1095 return cpu2;
1096 }
1097 return 0;
1098}
1099
1100static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1101{
1102 FILE *output = stat_config.output;
1103 double sc = evsel->scale;
1104 const char *fmt;
1105
1106 if (csv_output) {
1107 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
1108 } else {
1109 if (big_num)
1110 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1111 else
1112 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1113 }
1114
1115 aggr_printout(evsel, id, nr);
1116
1117 fprintf(output, fmt, avg, csv_sep);
1118
1119 if (evsel->unit)
1120 fprintf(output, "%-*s%s",
1121 csv_output ? 0 : unit_width,
1122 evsel->unit, csv_sep);
1123
1124 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1125
1126 if (evsel->cgrp)
1127 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1128}
1129
1130static bool is_mixed_hw_group(struct perf_evsel *counter)
1131{
1132 struct perf_evlist *evlist = counter->evlist;
1133 u32 pmu_type = counter->attr.type;
1134 struct perf_evsel *pos;
1135
1136 if (counter->nr_members < 2)
1137 return false;
1138
1139 evlist__for_each_entry(evlist, pos) {
1140 /* software events can be part of any hardware group */
1141 if (pos->attr.type == PERF_TYPE_SOFTWARE)
1142 continue;
1143 if (pmu_type == PERF_TYPE_SOFTWARE) {
1144 pmu_type = pos->attr.type;
1145 continue;
1146 }
1147 if (pmu_type != pos->attr.type)
1148 return true;
1149 }
1150
1151 return false;
1152}
1153
1154static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1155 char *prefix, u64 run, u64 ena, double noise,
1156 struct runtime_stat *st)
1157{
1158 struct perf_stat_output_ctx out;
1159 struct outstate os = {
1160 .fh = stat_config.output,
1161 .prefix = prefix ? prefix : "",
1162 .id = id,
1163 .nr = nr,
1164 .evsel = counter,
1165 };
1166 print_metric_t pm = print_metric_std;
1167 void (*nl)(void *);
1168
1169 if (metric_only) {
1170 nl = new_line_metric;
1171 if (csv_output)
1172 pm = print_metric_only_csv;
1173 else
1174 pm = print_metric_only;
1175 } else
1176 nl = new_line_std;
1177
1178 if (csv_output && !metric_only) {
1179 static int aggr_fields[] = {
1180 [AGGR_GLOBAL] = 0,
1181 [AGGR_THREAD] = 1,
1182 [AGGR_NONE] = 1,
1183 [AGGR_SOCKET] = 2,
1184 [AGGR_CORE] = 2,
1185 };
1186
1187 pm = print_metric_csv;
1188 nl = new_line_csv;
1189 os.nfields = 3;
1190 os.nfields += aggr_fields[stat_config.aggr_mode];
1191 if (counter->cgrp)
1192 os.nfields++;
1193 }
1194 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1195 if (metric_only) {
1196 pm(&os, NULL, "", "", 0);
1197 return;
1198 }
1199 aggr_printout(counter, id, nr);
1200
1201 fprintf(stat_config.output, "%*s%s",
1202 csv_output ? 0 : 18,
1203 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1204 csv_sep);
1205
1206 if (counter->supported) {
1207 print_free_counters_hint = 1;
1208 if (is_mixed_hw_group(counter))
1209 print_mixed_hw_group_error = 1;
1210 }
1211
1212 fprintf(stat_config.output, "%-*s%s",
1213 csv_output ? 0 : unit_width,
1214 counter->unit, csv_sep);
1215
1216 fprintf(stat_config.output, "%*s",
1217 csv_output ? 0 : -25,
1218 perf_evsel__name(counter));
1219
1220 if (counter->cgrp)
1221 fprintf(stat_config.output, "%s%s",
1222 csv_sep, counter->cgrp->name);
1223
1224 if (!csv_output)
1225 pm(&os, NULL, NULL, "", 0);
1226 print_noise(counter, noise);
1227 print_running(run, ena);
1228 if (csv_output)
1229 pm(&os, NULL, NULL, "", 0);
1230 return;
1231 }
1232
1233 if (metric_only)
1234 /* nothing */;
1235 else if (nsec_counter(counter))
1236 nsec_printout(id, nr, counter, uval);
1237 else
1238 abs_printout(id, nr, counter, uval);
1239
1240 out.print_metric = pm;
1241 out.new_line = nl;
1242 out.ctx = &os;
1243 out.force_header = false;
1244
1245 if (csv_output && !metric_only) {
1246 print_noise(counter, noise);
1247 print_running(run, ena);
1248 }
1249
1250 perf_stat__print_shadow_stats(counter, uval,
1251 first_shadow_cpu(counter, id),
1252 &out, &metric_events, st);
1253 if (!csv_output && !metric_only) {
1254 print_noise(counter, noise);
1255 print_running(run, ena);
1256 }
1257}
1258
1259static void aggr_update_shadow(void)
1260{
1261 int cpu, s2, id, s;
1262 u64 val;
1263 struct perf_evsel *counter;
1264
1265 for (s = 0; s < aggr_map->nr; s++) {
1266 id = aggr_map->map[s];
1267 evlist__for_each_entry(evsel_list, counter) {
1268 val = 0;
1269 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1270 s2 = aggr_get_id(evsel_list->cpus, cpu);
1271 if (s2 != id)
1272 continue;
1273 val += perf_counts(counter->counts, cpu, 0)->val;
1274 }
1275 perf_stat__update_shadow_stats(counter, val,
1276 first_shadow_cpu(counter, id),
1277 &rt_stat);
1278 }
1279 }
1280}
1281
1282static void uniquify_event_name(struct perf_evsel *counter)
1283{
1284 char *new_name;
1285 char *config;
1286
1287 if (counter->uniquified_name ||
1288 !counter->pmu_name || !strncmp(counter->name, counter->pmu_name,
1289 strlen(counter->pmu_name)))
1290 return;
1291
1292 config = strchr(counter->name, '/');
1293 if (config) {
1294 if (asprintf(&new_name,
1295 "%s%s", counter->pmu_name, config) > 0) {
1296 free(counter->name);
1297 counter->name = new_name;
1298 }
1299 } else {
1300 if (asprintf(&new_name,
1301 "%s [%s]", counter->name, counter->pmu_name) > 0) {
1302 free(counter->name);
1303 counter->name = new_name;
1304 }
1305 }
1306
1307 counter->uniquified_name = true;
1308}
1309
1310static void collect_all_aliases(struct perf_evsel *counter,
1311 void (*cb)(struct perf_evsel *counter, void *data,
1312 bool first),
1313 void *data)
1314{
1315 struct perf_evsel *alias;
1316
1317 alias = list_prepare_entry(counter, &(evsel_list->entries), node);
1318 list_for_each_entry_continue (alias, &evsel_list->entries, node) {
1319 if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) ||
1320 alias->scale != counter->scale ||
1321 alias->cgrp != counter->cgrp ||
1322 strcmp(alias->unit, counter->unit) ||
1323 nsec_counter(alias) != nsec_counter(counter))
1324 break;
1325 alias->merged_stat = true;
1326 cb(alias, data, false);
1327 }
1328}
1329
1330static bool collect_data(struct perf_evsel *counter,
1331 void (*cb)(struct perf_evsel *counter, void *data,
1332 bool first),
1333 void *data)
1334{
1335 if (counter->merged_stat)
1336 return false;
1337 cb(counter, data, true);
1338 if (no_merge)
1339 uniquify_event_name(counter);
1340 else if (counter->auto_merge_stats)
1341 collect_all_aliases(counter, cb, data);
1342 return true;
1343}
1344
1345struct aggr_data {
1346 u64 ena, run, val;
1347 int id;
1348 int nr;
1349 int cpu;
1350};
1351
1352static void aggr_cb(struct perf_evsel *counter, void *data, bool first)
1353{
1354 struct aggr_data *ad = data;
1355 int cpu, s2;
1356
1357 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1358 struct perf_counts_values *counts;
1359
1360 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1361 if (s2 != ad->id)
1362 continue;
1363 if (first)
1364 ad->nr++;
1365 counts = perf_counts(counter->counts, cpu, 0);
1366 /*
1367 * When any result is bad, make them all to give
1368 * consistent output in interval mode.
1369 */
1370 if (counts->ena == 0 || counts->run == 0 ||
1371 counter->counts->scaled == -1) {
1372 ad->ena = 0;
1373 ad->run = 0;
1374 break;
1375 }
1376 ad->val += counts->val;
1377 ad->ena += counts->ena;
1378 ad->run += counts->run;
1379 }
1380}
1381
1382static void print_aggr(char *prefix)
1383{
1384 FILE *output = stat_config.output;
1385 struct perf_evsel *counter;
1386 int s, id, nr;
1387 double uval;
1388 u64 ena, run, val;
1389 bool first;
1390
1391 if (!(aggr_map || aggr_get_id))
1392 return;
1393
1394 aggr_update_shadow();
1395
1396 /*
1397 * With metric_only everything is on a single line.
1398 * Without each counter has its own line.
1399 */
1400 for (s = 0; s < aggr_map->nr; s++) {
1401 struct aggr_data ad;
1402 if (prefix && metric_only)
1403 fprintf(output, "%s", prefix);
1404
1405 ad.id = id = aggr_map->map[s];
1406 first = true;
1407 evlist__for_each_entry(evsel_list, counter) {
1408 if (is_duration_time(counter))
1409 continue;
1410
1411 ad.val = ad.ena = ad.run = 0;
1412 ad.nr = 0;
1413 if (!collect_data(counter, aggr_cb, &ad))
1414 continue;
1415 nr = ad.nr;
1416 ena = ad.ena;
1417 run = ad.run;
1418 val = ad.val;
1419 if (first && metric_only) {
1420 first = false;
1421 aggr_printout(counter, id, nr);
1422 }
1423 if (prefix && !metric_only)
1424 fprintf(output, "%s", prefix);
1425
1426 uval = val * counter->scale;
1427 printout(id, nr, counter, uval, prefix, run, ena, 1.0,
1428 &rt_stat);
1429 if (!metric_only)
1430 fputc('\n', output);
1431 }
1432 if (metric_only)
1433 fputc('\n', output);
1434 }
1435}
1436
1437static int cmp_val(const void *a, const void *b)
1438{
1439 return ((struct perf_aggr_thread_value *)b)->val -
1440 ((struct perf_aggr_thread_value *)a)->val;
1441}
1442
1443static struct perf_aggr_thread_value *sort_aggr_thread(
1444 struct perf_evsel *counter,
1445 int nthreads, int ncpus,
1446 int *ret)
1447{
1448 int cpu, thread, i = 0;
1449 double uval;
1450 struct perf_aggr_thread_value *buf;
1451
1452 buf = calloc(nthreads, sizeof(struct perf_aggr_thread_value));
1453 if (!buf)
1454 return NULL;
1455
1456 for (thread = 0; thread < nthreads; thread++) {
1457 u64 ena = 0, run = 0, val = 0;
1458
1459 for (cpu = 0; cpu < ncpus; cpu++) {
1460 val += perf_counts(counter->counts, cpu, thread)->val;
1461 ena += perf_counts(counter->counts, cpu, thread)->ena;
1462 run += perf_counts(counter->counts, cpu, thread)->run;
1463 }
1464
1465 uval = val * counter->scale;
1466
1467 /*
1468 * Skip value 0 when enabling --per-thread globally,
1469 * otherwise too many 0 output.
1470 */
1471 if (uval == 0.0 && target__has_per_thread(&target))
1472 continue;
1473
1474 buf[i].counter = counter;
1475 buf[i].id = thread;
1476 buf[i].uval = uval;
1477 buf[i].val = val;
1478 buf[i].run = run;
1479 buf[i].ena = ena;
1480 i++;
1481 }
1482
1483 qsort(buf, i, sizeof(struct perf_aggr_thread_value), cmp_val);
1484
1485 if (ret)
1486 *ret = i;
1487
1488 return buf;
1489}
1490
1491static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1492{
1493 FILE *output = stat_config.output;
1494 int nthreads = thread_map__nr(counter->threads);
1495 int ncpus = cpu_map__nr(counter->cpus);
1496 int thread, sorted_threads, id;
1497 struct perf_aggr_thread_value *buf;
1498
1499 buf = sort_aggr_thread(counter, nthreads, ncpus, &sorted_threads);
1500 if (!buf) {
1501 perror("cannot sort aggr thread");
1502 return;
1503 }
1504
1505 for (thread = 0; thread < sorted_threads; thread++) {
1506 if (prefix)
1507 fprintf(output, "%s", prefix);
1508
1509 id = buf[thread].id;
1510 if (stat_config.stats)
1511 printout(id, 0, buf[thread].counter, buf[thread].uval,
1512 prefix, buf[thread].run, buf[thread].ena, 1.0,
1513 &stat_config.stats[id]);
1514 else
1515 printout(id, 0, buf[thread].counter, buf[thread].uval,
1516 prefix, buf[thread].run, buf[thread].ena, 1.0,
1517 &rt_stat);
1518 fputc('\n', output);
1519 }
1520
1521 free(buf);
1522}
1523
1524struct caggr_data {
1525 double avg, avg_enabled, avg_running;
1526};
1527
1528static void counter_aggr_cb(struct perf_evsel *counter, void *data,
1529 bool first __maybe_unused)
1530{
1531 struct caggr_data *cd = data;
1532 struct perf_stat_evsel *ps = counter->stats;
1533
1534 cd->avg += avg_stats(&ps->res_stats[0]);
1535 cd->avg_enabled += avg_stats(&ps->res_stats[1]);
1536 cd->avg_running += avg_stats(&ps->res_stats[2]);
1537}
1538
1539/*
1540 * Print out the results of a single counter:
1541 * aggregated counts in system-wide mode
1542 */
1543static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1544{
1545 FILE *output = stat_config.output;
1546 double uval;
1547 struct caggr_data cd = { .avg = 0.0 };
1548
1549 if (!collect_data(counter, counter_aggr_cb, &cd))
1550 return;
1551
1552 if (prefix && !metric_only)
1553 fprintf(output, "%s", prefix);
1554
1555 uval = cd.avg * counter->scale;
1556 printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled,
1557 cd.avg, &rt_stat);
1558 if (!metric_only)
1559 fprintf(output, "\n");
1560}
1561
1562static void counter_cb(struct perf_evsel *counter, void *data,
1563 bool first __maybe_unused)
1564{
1565 struct aggr_data *ad = data;
1566
1567 ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
1568 ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
1569 ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
1570}
1571
1572/*
1573 * Print out the results of a single counter:
1574 * does not use aggregated count in system-wide
1575 */
1576static void print_counter(struct perf_evsel *counter, char *prefix)
1577{
1578 FILE *output = stat_config.output;
1579 u64 ena, run, val;
1580 double uval;
1581 int cpu;
1582
1583 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1584 struct aggr_data ad = { .cpu = cpu };
1585
1586 if (!collect_data(counter, counter_cb, &ad))
1587 return;
1588 val = ad.val;
1589 ena = ad.ena;
1590 run = ad.run;
1591
1592 if (prefix)
1593 fprintf(output, "%s", prefix);
1594
1595 uval = val * counter->scale;
1596 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0,
1597 &rt_stat);
1598
1599 fputc('\n', output);
1600 }
1601}
1602
1603static void print_no_aggr_metric(char *prefix)
1604{
1605 int cpu;
1606 int nrcpus = 0;
1607 struct perf_evsel *counter;
1608 u64 ena, run, val;
1609 double uval;
1610
1611 nrcpus = evsel_list->cpus->nr;
1612 for (cpu = 0; cpu < nrcpus; cpu++) {
1613 bool first = true;
1614
1615 if (prefix)
1616 fputs(prefix, stat_config.output);
1617 evlist__for_each_entry(evsel_list, counter) {
1618 if (is_duration_time(counter))
1619 continue;
1620 if (first) {
1621 aggr_printout(counter, cpu, 0);
1622 first = false;
1623 }
1624 val = perf_counts(counter->counts, cpu, 0)->val;
1625 ena = perf_counts(counter->counts, cpu, 0)->ena;
1626 run = perf_counts(counter->counts, cpu, 0)->run;
1627
1628 uval = val * counter->scale;
1629 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0,
1630 &rt_stat);
1631 }
1632 fputc('\n', stat_config.output);
1633 }
1634}
1635
1636static int aggr_header_lens[] = {
1637 [AGGR_CORE] = 18,
1638 [AGGR_SOCKET] = 12,
1639 [AGGR_NONE] = 6,
1640 [AGGR_THREAD] = 24,
1641 [AGGR_GLOBAL] = 0,
1642};
1643
1644static const char *aggr_header_csv[] = {
1645 [AGGR_CORE] = "core,cpus,",
1646 [AGGR_SOCKET] = "socket,cpus",
1647 [AGGR_NONE] = "cpu,",
1648 [AGGR_THREAD] = "comm-pid,",
1649 [AGGR_GLOBAL] = ""
1650};
1651
1652static void print_metric_headers(const char *prefix, bool no_indent)
1653{
1654 struct perf_stat_output_ctx out;
1655 struct perf_evsel *counter;
1656 struct outstate os = {
1657 .fh = stat_config.output
1658 };
1659
1660 if (prefix)
1661 fprintf(stat_config.output, "%s", prefix);
1662
1663 if (!csv_output && !no_indent)
1664 fprintf(stat_config.output, "%*s",
1665 aggr_header_lens[stat_config.aggr_mode], "");
1666 if (csv_output) {
1667 if (stat_config.interval)
1668 fputs("time,", stat_config.output);
1669 fputs(aggr_header_csv[stat_config.aggr_mode],
1670 stat_config.output);
1671 }
1672
1673 /* Print metrics headers only */
1674 evlist__for_each_entry(evsel_list, counter) {
1675 if (is_duration_time(counter))
1676 continue;
1677 os.evsel = counter;
1678 out.ctx = &os;
1679 out.print_metric = print_metric_header;
1680 out.new_line = new_line_metric;
1681 out.force_header = true;
1682 os.evsel = counter;
1683 perf_stat__print_shadow_stats(counter, 0,
1684 0,
1685 &out,
1686 &metric_events,
1687 &rt_stat);
1688 }
1689 fputc('\n', stat_config.output);
1690}
1691
1692static void print_interval(char *prefix, struct timespec *ts)
1693{
1694 FILE *output = stat_config.output;
1695 static int num_print_interval;
1696
1697 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1698
1699 if (num_print_interval == 0 && !csv_output) {
1700 switch (stat_config.aggr_mode) {
1701 case AGGR_SOCKET:
1702 fprintf(output, "# time socket cpus");
1703 if (!metric_only)
1704 fprintf(output, " counts %*s events\n", unit_width, "unit");
1705 break;
1706 case AGGR_CORE:
1707 fprintf(output, "# time core cpus");
1708 if (!metric_only)
1709 fprintf(output, " counts %*s events\n", unit_width, "unit");
1710 break;
1711 case AGGR_NONE:
1712 fprintf(output, "# time CPU");
1713 if (!metric_only)
1714 fprintf(output, " counts %*s events\n", unit_width, "unit");
1715 break;
1716 case AGGR_THREAD:
1717 fprintf(output, "# time comm-pid");
1718 if (!metric_only)
1719 fprintf(output, " counts %*s events\n", unit_width, "unit");
1720 break;
1721 case AGGR_GLOBAL:
1722 default:
1723 fprintf(output, "# time");
1724 if (!metric_only)
1725 fprintf(output, " counts %*s events\n", unit_width, "unit");
1726 case AGGR_UNSET:
1727 break;
1728 }
1729 }
1730
1731 if (num_print_interval == 0 && metric_only)
1732 print_metric_headers(" ", true);
1733 if (++num_print_interval == 25)
1734 num_print_interval = 0;
1735}
1736
1737static void print_header(int argc, const char **argv)
1738{
1739 FILE *output = stat_config.output;
1740 int i;
1741
1742 fflush(stdout);
1743
1744 if (!csv_output) {
1745 fprintf(output, "\n");
1746 fprintf(output, " Performance counter stats for ");
1747 if (target.system_wide)
1748 fprintf(output, "\'system wide");
1749 else if (target.cpu_list)
1750 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1751 else if (!target__has_task(&target)) {
1752 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1753 for (i = 1; argv && (i < argc); i++)
1754 fprintf(output, " %s", argv[i]);
1755 } else if (target.pid)
1756 fprintf(output, "process id \'%s", target.pid);
1757 else
1758 fprintf(output, "thread id \'%s", target.tid);
1759
1760 fprintf(output, "\'");
1761 if (run_count > 1)
1762 fprintf(output, " (%d runs)", run_count);
1763 fprintf(output, ":\n\n");
1764 }
1765}
1766
1767static void print_footer(void)
1768{
1769 FILE *output = stat_config.output;
1770 int n;
1771
1772 if (!null_run)
1773 fprintf(output, "\n");
1774 fprintf(output, " %17.9f seconds time elapsed",
1775 avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC);
1776 if (run_count > 1) {
1777 fprintf(output, " ");
1778 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1779 avg_stats(&walltime_nsecs_stats));
1780 }
1781 fprintf(output, "\n\n");
1782
1783 if (print_free_counters_hint &&
1784 sysctl__read_int("kernel/nmi_watchdog", &n) >= 0 &&
1785 n > 0)
1786 fprintf(output,
1787"Some events weren't counted. Try disabling the NMI watchdog:\n"
1788" echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1789" perf stat ...\n"
1790" echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1791
1792 if (print_mixed_hw_group_error)
1793 fprintf(output,
1794 "The events in group usually have to be from "
1795 "the same PMU. Try reorganizing the group.\n");
1796}
1797
1798static void print_counters(struct timespec *ts, int argc, const char **argv)
1799{
1800 int interval = stat_config.interval;
1801 struct perf_evsel *counter;
1802 char buf[64], *prefix = NULL;
1803
1804 /* Do not print anything if we record to the pipe. */
1805 if (STAT_RECORD && perf_stat.data.is_pipe)
1806 return;
1807
1808 if (interval)
1809 print_interval(prefix = buf, ts);
1810 else
1811 print_header(argc, argv);
1812
1813 if (metric_only) {
1814 static int num_print_iv;
1815
1816 if (num_print_iv == 0 && !interval)
1817 print_metric_headers(prefix, false);
1818 if (num_print_iv++ == 25)
1819 num_print_iv = 0;
1820 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1821 fprintf(stat_config.output, "%s", prefix);
1822 }
1823
1824 switch (stat_config.aggr_mode) {
1825 case AGGR_CORE:
1826 case AGGR_SOCKET:
1827 print_aggr(prefix);
1828 break;
1829 case AGGR_THREAD:
1830 evlist__for_each_entry(evsel_list, counter) {
1831 if (is_duration_time(counter))
1832 continue;
1833 print_aggr_thread(counter, prefix);
1834 }
1835 break;
1836 case AGGR_GLOBAL:
1837 evlist__for_each_entry(evsel_list, counter) {
1838 if (is_duration_time(counter))
1839 continue;
1840 print_counter_aggr(counter, prefix);
1841 }
1842 if (metric_only)
1843 fputc('\n', stat_config.output);
1844 break;
1845 case AGGR_NONE:
1846 if (metric_only)
1847 print_no_aggr_metric(prefix);
1848 else {
1849 evlist__for_each_entry(evsel_list, counter) {
1850 if (is_duration_time(counter))
1851 continue;
1852 print_counter(counter, prefix);
1853 }
1854 }
1855 break;
1856 case AGGR_UNSET:
1857 default:
1858 break;
1859 }
1860
1861 if (!interval && !csv_output)
1862 print_footer();
1863
1864 fflush(stat_config.output);
1865}
1866
1867static volatile int signr = -1;
1868
1869static void skip_signal(int signo)
1870{
1871 if ((child_pid == -1) || stat_config.interval)
1872 done = 1;
1873
1874 signr = signo;
1875 /*
1876 * render child_pid harmless
1877 * won't send SIGTERM to a random
1878 * process in case of race condition
1879 * and fast PID recycling
1880 */
1881 child_pid = -1;
1882}
1883
1884static void sig_atexit(void)
1885{
1886 sigset_t set, oset;
1887
1888 /*
1889 * avoid race condition with SIGCHLD handler
1890 * in skip_signal() which is modifying child_pid
1891 * goal is to avoid send SIGTERM to a random
1892 * process
1893 */
1894 sigemptyset(&set);
1895 sigaddset(&set, SIGCHLD);
1896 sigprocmask(SIG_BLOCK, &set, &oset);
1897
1898 if (child_pid != -1)
1899 kill(child_pid, SIGTERM);
1900
1901 sigprocmask(SIG_SETMASK, &oset, NULL);
1902
1903 if (signr == -1)
1904 return;
1905
1906 signal(signr, SIG_DFL);
1907 kill(getpid(), signr);
1908}
1909
1910static int stat__set_big_num(const struct option *opt __maybe_unused,
1911 const char *s __maybe_unused, int unset)
1912{
1913 big_num_opt = unset ? 0 : 1;
1914 return 0;
1915}
1916
1917static int enable_metric_only(const struct option *opt __maybe_unused,
1918 const char *s __maybe_unused, int unset)
1919{
1920 force_metric_only = true;
1921 metric_only = !unset;
1922 return 0;
1923}
1924
1925static int parse_metric_groups(const struct option *opt,
1926 const char *str,
1927 int unset __maybe_unused)
1928{
1929 return metricgroup__parse_groups(opt, str, &metric_events);
1930}
1931
1932static const struct option stat_options[] = {
1933 OPT_BOOLEAN('T', "transaction", &transaction_run,
1934 "hardware transaction statistics"),
1935 OPT_CALLBACK('e', "event", &evsel_list, "event",
1936 "event selector. use 'perf list' to list available events",
1937 parse_events_option),
1938 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1939 "event filter", parse_filter),
1940 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1941 "child tasks do not inherit counters"),
1942 OPT_STRING('p', "pid", &target.pid, "pid",
1943 "stat events on existing process id"),
1944 OPT_STRING('t', "tid", &target.tid, "tid",
1945 "stat events on existing thread id"),
1946 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1947 "system-wide collection from all CPUs"),
1948 OPT_BOOLEAN('g', "group", &group,
1949 "put the counters into a counter group"),
1950 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1951 OPT_INCR('v', "verbose", &verbose,
1952 "be more verbose (show counter open errors, etc)"),
1953 OPT_INTEGER('r', "repeat", &run_count,
1954 "repeat command and print average + stddev (max: 100, forever: 0)"),
1955 OPT_BOOLEAN('n', "null", &null_run,
1956 "null run - dont start any counters"),
1957 OPT_INCR('d', "detailed", &detailed_run,
1958 "detailed run - start a lot of events"),
1959 OPT_BOOLEAN('S', "sync", &sync_run,
1960 "call sync() before starting a run"),
1961 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1962 "print large numbers with thousands\' separators",
1963 stat__set_big_num),
1964 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1965 "list of cpus to monitor in system-wide"),
1966 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1967 "disable CPU count aggregation", AGGR_NONE),
1968 OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"),
1969 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1970 "print counts with custom separator"),
1971 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1972 "monitor event in cgroup name only", parse_cgroups),
1973 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1974 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1975 OPT_INTEGER(0, "log-fd", &output_fd,
1976 "log output to fd, instead of stderr"),
1977 OPT_STRING(0, "pre", &pre_cmd, "command",
1978 "command to run prior to the measured command"),
1979 OPT_STRING(0, "post", &post_cmd, "command",
1980 "command to run after to the measured command"),
1981 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1982 "print counts at regular interval in ms "
1983 "(overhead is possible for values <= 100ms)"),
1984 OPT_INTEGER(0, "interval-count", &stat_config.times,
1985 "print counts for fixed number of times"),
1986 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1987 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1988 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1989 "aggregate counts per processor socket", AGGR_SOCKET),
1990 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1991 "aggregate counts per physical processor core", AGGR_CORE),
1992 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1993 "aggregate counts per thread", AGGR_THREAD),
1994 OPT_UINTEGER('D', "delay", &initial_delay,
1995 "ms to wait before starting measurement after program start"),
1996 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1997 "Only print computed metrics. No raw values", enable_metric_only),
1998 OPT_BOOLEAN(0, "topdown", &topdown_run,
1999 "measure topdown level 1 statistics"),
2000 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
2001 "measure SMI cost"),
2002 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
2003 "monitor specified metrics or metric groups (separated by ,)",
2004 parse_metric_groups),
2005 OPT_END()
2006};
2007
2008static int perf_stat__get_socket(struct cpu_map *map, int cpu)
2009{
2010 return cpu_map__get_socket(map, cpu, NULL);
2011}
2012
2013static int perf_stat__get_core(struct cpu_map *map, int cpu)
2014{
2015 return cpu_map__get_core(map, cpu, NULL);
2016}
2017
2018static int cpu_map__get_max(struct cpu_map *map)
2019{
2020 int i, max = -1;
2021
2022 for (i = 0; i < map->nr; i++) {
2023 if (map->map[i] > max)
2024 max = map->map[i];
2025 }
2026
2027 return max;
2028}
2029
2030static struct cpu_map *cpus_aggr_map;
2031
2032static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
2033{
2034 int cpu;
2035
2036 if (idx >= map->nr)
2037 return -1;
2038
2039 cpu = map->map[idx];
2040
2041 if (cpus_aggr_map->map[cpu] == -1)
2042 cpus_aggr_map->map[cpu] = get_id(map, idx);
2043
2044 return cpus_aggr_map->map[cpu];
2045}
2046
2047static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
2048{
2049 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
2050}
2051
2052static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
2053{
2054 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
2055}
2056
2057static int perf_stat_init_aggr_mode(void)
2058{
2059 int nr;
2060
2061 switch (stat_config.aggr_mode) {
2062 case AGGR_SOCKET:
2063 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
2064 perror("cannot build socket map");
2065 return -1;
2066 }
2067 aggr_get_id = perf_stat__get_socket_cached;
2068 break;
2069 case AGGR_CORE:
2070 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
2071 perror("cannot build core map");
2072 return -1;
2073 }
2074 aggr_get_id = perf_stat__get_core_cached;
2075 break;
2076 case AGGR_NONE:
2077 case AGGR_GLOBAL:
2078 case AGGR_THREAD:
2079 case AGGR_UNSET:
2080 default:
2081 break;
2082 }
2083
2084 /*
2085 * The evsel_list->cpus is the base we operate on,
2086 * taking the highest cpu number to be the size of
2087 * the aggregation translate cpumap.
2088 */
2089 nr = cpu_map__get_max(evsel_list->cpus);
2090 cpus_aggr_map = cpu_map__empty_new(nr + 1);
2091 return cpus_aggr_map ? 0 : -ENOMEM;
2092}
2093
2094static void perf_stat__exit_aggr_mode(void)
2095{
2096 cpu_map__put(aggr_map);
2097 cpu_map__put(cpus_aggr_map);
2098 aggr_map = NULL;
2099 cpus_aggr_map = NULL;
2100}
2101
2102static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
2103{
2104 int cpu;
2105
2106 if (idx > map->nr)
2107 return -1;
2108
2109 cpu = map->map[idx];
2110
2111 if (cpu >= env->nr_cpus_avail)
2112 return -1;
2113
2114 return cpu;
2115}
2116
2117static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
2118{
2119 struct perf_env *env = data;
2120 int cpu = perf_env__get_cpu(env, map, idx);
2121
2122 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
2123}
2124
2125static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
2126{
2127 struct perf_env *env = data;
2128 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
2129
2130 if (cpu != -1) {
2131 int socket_id = env->cpu[cpu].socket_id;
2132
2133 /*
2134 * Encode socket in upper 16 bits
2135 * core_id is relative to socket, and
2136 * we need a global id. So we combine
2137 * socket + core id.
2138 */
2139 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
2140 }
2141
2142 return core;
2143}
2144
2145static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
2146 struct cpu_map **sockp)
2147{
2148 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
2149}
2150
2151static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
2152 struct cpu_map **corep)
2153{
2154 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
2155}
2156
2157static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
2158{
2159 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
2160}
2161
2162static int perf_stat__get_core_file(struct cpu_map *map, int idx)
2163{
2164 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
2165}
2166
2167static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
2168{
2169 struct perf_env *env = &st->session->header.env;
2170
2171 switch (stat_config.aggr_mode) {
2172 case AGGR_SOCKET:
2173 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
2174 perror("cannot build socket map");
2175 return -1;
2176 }
2177 aggr_get_id = perf_stat__get_socket_file;
2178 break;
2179 case AGGR_CORE:
2180 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
2181 perror("cannot build core map");
2182 return -1;
2183 }
2184 aggr_get_id = perf_stat__get_core_file;
2185 break;
2186 case AGGR_NONE:
2187 case AGGR_GLOBAL:
2188 case AGGR_THREAD:
2189 case AGGR_UNSET:
2190 default:
2191 break;
2192 }
2193
2194 return 0;
2195}
2196
2197static int topdown_filter_events(const char **attr, char **str, bool use_group)
2198{
2199 int off = 0;
2200 int i;
2201 int len = 0;
2202 char *s;
2203
2204 for (i = 0; attr[i]; i++) {
2205 if (pmu_have_event("cpu", attr[i])) {
2206 len += strlen(attr[i]) + 1;
2207 attr[i - off] = attr[i];
2208 } else
2209 off++;
2210 }
2211 attr[i - off] = NULL;
2212
2213 *str = malloc(len + 1 + 2);
2214 if (!*str)
2215 return -1;
2216 s = *str;
2217 if (i - off == 0) {
2218 *s = 0;
2219 return 0;
2220 }
2221 if (use_group)
2222 *s++ = '{';
2223 for (i = 0; attr[i]; i++) {
2224 strcpy(s, attr[i]);
2225 s += strlen(s);
2226 *s++ = ',';
2227 }
2228 if (use_group) {
2229 s[-1] = '}';
2230 *s = 0;
2231 } else
2232 s[-1] = 0;
2233 return 0;
2234}
2235
2236__weak bool arch_topdown_check_group(bool *warn)
2237{
2238 *warn = false;
2239 return false;
2240}
2241
2242__weak void arch_topdown_group_warn(void)
2243{
2244}
2245
2246/*
2247 * Add default attributes, if there were no attributes specified or
2248 * if -d/--detailed, -d -d or -d -d -d is used:
2249 */
2250static int add_default_attributes(void)
2251{
2252 int err;
2253 struct perf_event_attr default_attrs0[] = {
2254
2255 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
2256 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
2257 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
2258 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
2259
2260 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
2261};
2262 struct perf_event_attr frontend_attrs[] = {
2263 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
2264};
2265 struct perf_event_attr backend_attrs[] = {
2266 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
2267};
2268 struct perf_event_attr default_attrs1[] = {
2269 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
2270 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
2271 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
2272
2273};
2274
2275/*
2276 * Detailed stats (-d), covering the L1 and last level data caches:
2277 */
2278 struct perf_event_attr detailed_attrs[] = {
2279
2280 { .type = PERF_TYPE_HW_CACHE,
2281 .config =
2282 PERF_COUNT_HW_CACHE_L1D << 0 |
2283 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2284 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2285
2286 { .type = PERF_TYPE_HW_CACHE,
2287 .config =
2288 PERF_COUNT_HW_CACHE_L1D << 0 |
2289 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2290 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2291
2292 { .type = PERF_TYPE_HW_CACHE,
2293 .config =
2294 PERF_COUNT_HW_CACHE_LL << 0 |
2295 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2296 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2297
2298 { .type = PERF_TYPE_HW_CACHE,
2299 .config =
2300 PERF_COUNT_HW_CACHE_LL << 0 |
2301 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2302 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2303};
2304
2305/*
2306 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2307 */
2308 struct perf_event_attr very_detailed_attrs[] = {
2309
2310 { .type = PERF_TYPE_HW_CACHE,
2311 .config =
2312 PERF_COUNT_HW_CACHE_L1I << 0 |
2313 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2314 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2315
2316 { .type = PERF_TYPE_HW_CACHE,
2317 .config =
2318 PERF_COUNT_HW_CACHE_L1I << 0 |
2319 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2320 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2321
2322 { .type = PERF_TYPE_HW_CACHE,
2323 .config =
2324 PERF_COUNT_HW_CACHE_DTLB << 0 |
2325 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2326 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2327
2328 { .type = PERF_TYPE_HW_CACHE,
2329 .config =
2330 PERF_COUNT_HW_CACHE_DTLB << 0 |
2331 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2332 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2333
2334 { .type = PERF_TYPE_HW_CACHE,
2335 .config =
2336 PERF_COUNT_HW_CACHE_ITLB << 0 |
2337 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2338 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2339
2340 { .type = PERF_TYPE_HW_CACHE,
2341 .config =
2342 PERF_COUNT_HW_CACHE_ITLB << 0 |
2343 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2344 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2345
2346};
2347
2348/*
2349 * Very, very detailed stats (-d -d -d), adding prefetch events:
2350 */
2351 struct perf_event_attr very_very_detailed_attrs[] = {
2352
2353 { .type = PERF_TYPE_HW_CACHE,
2354 .config =
2355 PERF_COUNT_HW_CACHE_L1D << 0 |
2356 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2357 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2358
2359 { .type = PERF_TYPE_HW_CACHE,
2360 .config =
2361 PERF_COUNT_HW_CACHE_L1D << 0 |
2362 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2363 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2364};
2365
2366 /* Set attrs if no event is selected and !null_run: */
2367 if (null_run)
2368 return 0;
2369
2370 if (transaction_run) {
2371 struct parse_events_error errinfo;
2372
2373 if (pmu_have_event("cpu", "cycles-ct") &&
2374 pmu_have_event("cpu", "el-start"))
2375 err = parse_events(evsel_list, transaction_attrs,
2376 &errinfo);
2377 else
2378 err = parse_events(evsel_list,
2379 transaction_limited_attrs,
2380 &errinfo);
2381 if (err) {
2382 fprintf(stderr, "Cannot set up transaction events\n");
2383 return -1;
2384 }
2385 return 0;
2386 }
2387
2388 if (smi_cost) {
2389 int smi;
2390
2391 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2392 fprintf(stderr, "freeze_on_smi is not supported.\n");
2393 return -1;
2394 }
2395
2396 if (!smi) {
2397 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2398 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2399 return -1;
2400 }
2401 smi_reset = true;
2402 }
2403
2404 if (pmu_have_event("msr", "aperf") &&
2405 pmu_have_event("msr", "smi")) {
2406 if (!force_metric_only)
2407 metric_only = true;
2408 err = parse_events(evsel_list, smi_cost_attrs, NULL);
2409 } else {
2410 fprintf(stderr, "To measure SMI cost, it needs "
2411 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
2412 return -1;
2413 }
2414 if (err) {
2415 fprintf(stderr, "Cannot set up SMI cost events\n");
2416 return -1;
2417 }
2418 return 0;
2419 }
2420
2421 if (topdown_run) {
2422 char *str = NULL;
2423 bool warn = false;
2424
2425 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2426 stat_config.aggr_mode != AGGR_CORE) {
2427 pr_err("top down event configuration requires --per-core mode\n");
2428 return -1;
2429 }
2430 stat_config.aggr_mode = AGGR_CORE;
2431 if (nr_cgroups || !target__has_cpu(&target)) {
2432 pr_err("top down event configuration requires system-wide mode (-a)\n");
2433 return -1;
2434 }
2435
2436 if (!force_metric_only)
2437 metric_only = true;
2438 if (topdown_filter_events(topdown_attrs, &str,
2439 arch_topdown_check_group(&warn)) < 0) {
2440 pr_err("Out of memory\n");
2441 return -1;
2442 }
2443 if (topdown_attrs[0] && str) {
2444 if (warn)
2445 arch_topdown_group_warn();
2446 err = parse_events(evsel_list, str, NULL);
2447 if (err) {
2448 fprintf(stderr,
2449 "Cannot set up top down events %s: %d\n",
2450 str, err);
2451 free(str);
2452 return -1;
2453 }
2454 } else {
2455 fprintf(stderr, "System does not support topdown\n");
2456 return -1;
2457 }
2458 free(str);
2459 }
2460
2461 if (!evsel_list->nr_entries) {
2462 if (target__has_cpu(&target))
2463 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2464
2465 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2466 return -1;
2467 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2468 if (perf_evlist__add_default_attrs(evsel_list,
2469 frontend_attrs) < 0)
2470 return -1;
2471 }
2472 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2473 if (perf_evlist__add_default_attrs(evsel_list,
2474 backend_attrs) < 0)
2475 return -1;
2476 }
2477 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2478 return -1;
2479 }
2480
2481 /* Detailed events get appended to the event list: */
2482
2483 if (detailed_run < 1)
2484 return 0;
2485
2486 /* Append detailed run extra attributes: */
2487 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2488 return -1;
2489
2490 if (detailed_run < 2)
2491 return 0;
2492
2493 /* Append very detailed run extra attributes: */
2494 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2495 return -1;
2496
2497 if (detailed_run < 3)
2498 return 0;
2499
2500 /* Append very, very detailed run extra attributes: */
2501 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2502}
2503
2504static const char * const stat_record_usage[] = {
2505 "perf stat record [<options>]",
2506 NULL,
2507};
2508
2509static void init_features(struct perf_session *session)
2510{
2511 int feat;
2512
2513 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2514 perf_header__set_feat(&session->header, feat);
2515
2516 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2517 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2518 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2519 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2520}
2521
2522static int __cmd_record(int argc, const char **argv)
2523{
2524 struct perf_session *session;
2525 struct perf_data *data = &perf_stat.data;
2526
2527 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2528 PARSE_OPT_STOP_AT_NON_OPTION);
2529
2530 if (output_name)
2531 data->file.path = output_name;
2532
2533 if (run_count != 1 || forever) {
2534 pr_err("Cannot use -r option with perf stat record.\n");
2535 return -1;
2536 }
2537
2538 session = perf_session__new(data, false, NULL);
2539 if (session == NULL) {
2540 pr_err("Perf session creation failed.\n");
2541 return -1;
2542 }
2543
2544 init_features(session);
2545
2546 session->evlist = evsel_list;
2547 perf_stat.session = session;
2548 perf_stat.record = true;
2549 return argc;
2550}
2551
2552static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2553 union perf_event *event,
2554 struct perf_session *session)
2555{
2556 struct stat_round_event *stat_round = &event->stat_round;
2557 struct perf_evsel *counter;
2558 struct timespec tsh, *ts = NULL;
2559 const char **argv = session->header.env.cmdline_argv;
2560 int argc = session->header.env.nr_cmdline;
2561
2562 evlist__for_each_entry(evsel_list, counter)
2563 perf_stat_process_counter(&stat_config, counter);
2564
2565 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2566 update_stats(&walltime_nsecs_stats, stat_round->time);
2567
2568 if (stat_config.interval && stat_round->time) {
2569 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2570 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2571 ts = &tsh;
2572 }
2573
2574 print_counters(ts, argc, argv);
2575 return 0;
2576}
2577
2578static
2579int process_stat_config_event(struct perf_tool *tool,
2580 union perf_event *event,
2581 struct perf_session *session __maybe_unused)
2582{
2583 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2584
2585 perf_event__read_stat_config(&stat_config, &event->stat_config);
2586
2587 if (cpu_map__empty(st->cpus)) {
2588 if (st->aggr_mode != AGGR_UNSET)
2589 pr_warning("warning: processing task data, aggregation mode not set\n");
2590 return 0;
2591 }
2592
2593 if (st->aggr_mode != AGGR_UNSET)
2594 stat_config.aggr_mode = st->aggr_mode;
2595
2596 if (perf_stat.data.is_pipe)
2597 perf_stat_init_aggr_mode();
2598 else
2599 perf_stat_init_aggr_mode_file(st);
2600
2601 return 0;
2602}
2603
2604static int set_maps(struct perf_stat *st)
2605{
2606 if (!st->cpus || !st->threads)
2607 return 0;
2608
2609 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2610 return -EINVAL;
2611
2612 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2613
2614 if (perf_evlist__alloc_stats(evsel_list, true))
2615 return -ENOMEM;
2616
2617 st->maps_allocated = true;
2618 return 0;
2619}
2620
2621static
2622int process_thread_map_event(struct perf_tool *tool,
2623 union perf_event *event,
2624 struct perf_session *session __maybe_unused)
2625{
2626 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2627
2628 if (st->threads) {
2629 pr_warning("Extra thread map event, ignoring.\n");
2630 return 0;
2631 }
2632
2633 st->threads = thread_map__new_event(&event->thread_map);
2634 if (!st->threads)
2635 return -ENOMEM;
2636
2637 return set_maps(st);
2638}
2639
2640static
2641int process_cpu_map_event(struct perf_tool *tool,
2642 union perf_event *event,
2643 struct perf_session *session __maybe_unused)
2644{
2645 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2646 struct cpu_map *cpus;
2647
2648 if (st->cpus) {
2649 pr_warning("Extra cpu map event, ignoring.\n");
2650 return 0;
2651 }
2652
2653 cpus = cpu_map__new_data(&event->cpu_map.data);
2654 if (!cpus)
2655 return -ENOMEM;
2656
2657 st->cpus = cpus;
2658 return set_maps(st);
2659}
2660
2661static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
2662{
2663 int i;
2664
2665 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
2666 if (!config->stats)
2667 return -1;
2668
2669 config->stats_num = nthreads;
2670
2671 for (i = 0; i < nthreads; i++)
2672 runtime_stat__init(&config->stats[i]);
2673
2674 return 0;
2675}
2676
2677static void runtime_stat_delete(struct perf_stat_config *config)
2678{
2679 int i;
2680
2681 if (!config->stats)
2682 return;
2683
2684 for (i = 0; i < config->stats_num; i++)
2685 runtime_stat__exit(&config->stats[i]);
2686
2687 free(config->stats);
2688}
2689
2690static const char * const stat_report_usage[] = {
2691 "perf stat report [<options>]",
2692 NULL,
2693};
2694
2695static struct perf_stat perf_stat = {
2696 .tool = {
2697 .attr = perf_event__process_attr,
2698 .event_update = perf_event__process_event_update,
2699 .thread_map = process_thread_map_event,
2700 .cpu_map = process_cpu_map_event,
2701 .stat_config = process_stat_config_event,
2702 .stat = perf_event__process_stat_event,
2703 .stat_round = process_stat_round_event,
2704 },
2705 .aggr_mode = AGGR_UNSET,
2706};
2707
2708static int __cmd_report(int argc, const char **argv)
2709{
2710 struct perf_session *session;
2711 const struct option options[] = {
2712 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2713 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2714 "aggregate counts per processor socket", AGGR_SOCKET),
2715 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2716 "aggregate counts per physical processor core", AGGR_CORE),
2717 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2718 "disable CPU count aggregation", AGGR_NONE),
2719 OPT_END()
2720 };
2721 struct stat st;
2722 int ret;
2723
2724 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2725
2726 if (!input_name || !strlen(input_name)) {
2727 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2728 input_name = "-";
2729 else
2730 input_name = "perf.data";
2731 }
2732
2733 perf_stat.data.file.path = input_name;
2734 perf_stat.data.mode = PERF_DATA_MODE_READ;
2735
2736 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2737 if (session == NULL)
2738 return -1;
2739
2740 perf_stat.session = session;
2741 stat_config.output = stderr;
2742 evsel_list = session->evlist;
2743
2744 ret = perf_session__process_events(session);
2745 if (ret)
2746 return ret;
2747
2748 perf_session__delete(session);
2749 return 0;
2750}
2751
2752static void setup_system_wide(int forks)
2753{
2754 /*
2755 * Make system wide (-a) the default target if
2756 * no target was specified and one of following
2757 * conditions is met:
2758 *
2759 * - there's no workload specified
2760 * - there is workload specified but all requested
2761 * events are system wide events
2762 */
2763 if (!target__none(&target))
2764 return;
2765
2766 if (!forks)
2767 target.system_wide = true;
2768 else {
2769 struct perf_evsel *counter;
2770
2771 evlist__for_each_entry(evsel_list, counter) {
2772 if (!counter->system_wide)
2773 return;
2774 }
2775
2776 if (evsel_list->nr_entries)
2777 target.system_wide = true;
2778 }
2779}
2780
2781int cmd_stat(int argc, const char **argv)
2782{
2783 const char * const stat_usage[] = {
2784 "perf stat [<options>] [<command>]",
2785 NULL
2786 };
2787 int status = -EINVAL, run_idx;
2788 const char *mode;
2789 FILE *output = stderr;
2790 unsigned int interval, timeout;
2791 const char * const stat_subcommands[] = { "record", "report" };
2792
2793 setlocale(LC_ALL, "");
2794
2795 evsel_list = perf_evlist__new();
2796 if (evsel_list == NULL)
2797 return -ENOMEM;
2798
2799 parse_events__shrink_config_terms();
2800 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2801 (const char **) stat_usage,
2802 PARSE_OPT_STOP_AT_NON_OPTION);
2803 perf_stat__collect_metric_expr(evsel_list);
2804 perf_stat__init_shadow_stats();
2805
2806 if (csv_sep) {
2807 csv_output = true;
2808 if (!strcmp(csv_sep, "\\t"))
2809 csv_sep = "\t";
2810 } else
2811 csv_sep = DEFAULT_SEPARATOR;
2812
2813 if (argc && !strncmp(argv[0], "rec", 3)) {
2814 argc = __cmd_record(argc, argv);
2815 if (argc < 0)
2816 return -1;
2817 } else if (argc && !strncmp(argv[0], "rep", 3))
2818 return __cmd_report(argc, argv);
2819
2820 interval = stat_config.interval;
2821 timeout = stat_config.timeout;
2822
2823 /*
2824 * For record command the -o is already taken care of.
2825 */
2826 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2827 output = NULL;
2828
2829 if (output_name && output_fd) {
2830 fprintf(stderr, "cannot use both --output and --log-fd\n");
2831 parse_options_usage(stat_usage, stat_options, "o", 1);
2832 parse_options_usage(NULL, stat_options, "log-fd", 0);
2833 goto out;
2834 }
2835
2836 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2837 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2838 goto out;
2839 }
2840
2841 if (metric_only && run_count > 1) {
2842 fprintf(stderr, "--metric-only is not supported with -r\n");
2843 goto out;
2844 }
2845
2846 if (output_fd < 0) {
2847 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2848 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2849 goto out;
2850 }
2851
2852 if (!output) {
2853 struct timespec tm;
2854 mode = append_file ? "a" : "w";
2855
2856 output = fopen(output_name, mode);
2857 if (!output) {
2858 perror("failed to create output file");
2859 return -1;
2860 }
2861 clock_gettime(CLOCK_REALTIME, &tm);
2862 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2863 } else if (output_fd > 0) {
2864 mode = append_file ? "a" : "w";
2865 output = fdopen(output_fd, mode);
2866 if (!output) {
2867 perror("Failed opening logfd");
2868 return -errno;
2869 }
2870 }
2871
2872 stat_config.output = output;
2873
2874 /*
2875 * let the spreadsheet do the pretty-printing
2876 */
2877 if (csv_output) {
2878 /* User explicitly passed -B? */
2879 if (big_num_opt == 1) {
2880 fprintf(stderr, "-B option not supported with -x\n");
2881 parse_options_usage(stat_usage, stat_options, "B", 1);
2882 parse_options_usage(NULL, stat_options, "x", 1);
2883 goto out;
2884 } else /* Nope, so disable big number formatting */
2885 big_num = false;
2886 } else if (big_num_opt == 0) /* User passed --no-big-num */
2887 big_num = false;
2888
2889 setup_system_wide(argc);
2890
2891 if (run_count < 0) {
2892 pr_err("Run count must be a positive number\n");
2893 parse_options_usage(stat_usage, stat_options, "r", 1);
2894 goto out;
2895 } else if (run_count == 0) {
2896 forever = true;
2897 run_count = 1;
2898 }
2899
2900 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2901 !target__has_task(&target)) {
2902 if (!target.system_wide || target.cpu_list) {
2903 fprintf(stderr, "The --per-thread option is only "
2904 "available when monitoring via -p -t -a "
2905 "options or only --per-thread.\n");
2906 parse_options_usage(NULL, stat_options, "p", 1);
2907 parse_options_usage(NULL, stat_options, "t", 1);
2908 goto out;
2909 }
2910 }
2911
2912 /*
2913 * no_aggr, cgroup are for system-wide only
2914 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2915 */
2916 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2917 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2918 !target__has_cpu(&target)) {
2919 fprintf(stderr, "both cgroup and no-aggregation "
2920 "modes only available in system-wide mode\n");
2921
2922 parse_options_usage(stat_usage, stat_options, "G", 1);
2923 parse_options_usage(NULL, stat_options, "A", 1);
2924 parse_options_usage(NULL, stat_options, "a", 1);
2925 goto out;
2926 }
2927
2928 if (add_default_attributes())
2929 goto out;
2930
2931 target__validate(&target);
2932
2933 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2934 target.per_thread = true;
2935
2936 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2937 if (target__has_task(&target)) {
2938 pr_err("Problems finding threads of monitor\n");
2939 parse_options_usage(stat_usage, stat_options, "p", 1);
2940 parse_options_usage(NULL, stat_options, "t", 1);
2941 } else if (target__has_cpu(&target)) {
2942 perror("failed to parse CPUs map");
2943 parse_options_usage(stat_usage, stat_options, "C", 1);
2944 parse_options_usage(NULL, stat_options, "a", 1);
2945 }
2946 goto out;
2947 }
2948
2949 /*
2950 * Initialize thread_map with comm names,
2951 * so we could print it out on output.
2952 */
2953 if (stat_config.aggr_mode == AGGR_THREAD) {
2954 thread_map__read_comms(evsel_list->threads);
2955 if (target.system_wide) {
2956 if (runtime_stat_new(&stat_config,
2957 thread_map__nr(evsel_list->threads))) {
2958 goto out;
2959 }
2960 }
2961 }
2962
2963 if (stat_config.times && interval)
2964 interval_count = true;
2965 else if (stat_config.times && !interval) {
2966 pr_err("interval-count option should be used together with "
2967 "interval-print.\n");
2968 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2969 parse_options_usage(stat_usage, stat_options, "I", 1);
2970 goto out;
2971 }
2972
2973 if (timeout && timeout < 100) {
2974 if (timeout < 10) {
2975 pr_err("timeout must be >= 10ms.\n");
2976 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2977 goto out;
2978 } else
2979 pr_warning("timeout < 100ms. "
2980 "The overhead percentage could be high in some cases. "
2981 "Please proceed with caution.\n");
2982 }
2983 if (timeout && interval) {
2984 pr_err("timeout option is not supported with interval-print.\n");
2985 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2986 parse_options_usage(stat_usage, stat_options, "I", 1);
2987 goto out;
2988 }
2989
2990 if (perf_evlist__alloc_stats(evsel_list, interval))
2991 goto out;
2992
2993 if (perf_stat_init_aggr_mode())
2994 goto out;
2995
2996 /*
2997 * We dont want to block the signals - that would cause
2998 * child tasks to inherit that and Ctrl-C would not work.
2999 * What we want is for Ctrl-C to work in the exec()-ed
3000 * task, but being ignored by perf stat itself:
3001 */
3002 atexit(sig_atexit);
3003 if (!forever)
3004 signal(SIGINT, skip_signal);
3005 signal(SIGCHLD, skip_signal);
3006 signal(SIGALRM, skip_signal);
3007 signal(SIGABRT, skip_signal);
3008
3009 status = 0;
3010 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
3011 if (run_count != 1 && verbose > 0)
3012 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
3013 run_idx + 1);
3014
3015 status = run_perf_stat(argc, argv);
3016 if (forever && status != -1) {
3017 print_counters(NULL, argc, argv);
3018 perf_stat__reset_stats();
3019 }
3020 }
3021
3022 if (!forever && status != -1 && !interval)
3023 print_counters(NULL, argc, argv);
3024
3025 if (STAT_RECORD) {
3026 /*
3027 * We synthesize the kernel mmap record just so that older tools
3028 * don't emit warnings about not being able to resolve symbols
3029 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
3030 * a saner message about no samples being in the perf.data file.
3031 *
3032 * This also serves to suppress a warning about f_header.data.size == 0
3033 * in header.c at the moment 'perf stat record' gets introduced, which
3034 * is not really needed once we start adding the stat specific PERF_RECORD_
3035 * records, but the need to suppress the kptr_restrict messages in older
3036 * tools remain -acme
3037 */
3038 int fd = perf_data__fd(&perf_stat.data);
3039 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
3040 process_synthesized_event,
3041 &perf_stat.session->machines.host);
3042 if (err) {
3043 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
3044 "older tools may produce warnings about this file\n.");
3045 }
3046
3047 if (!interval) {
3048 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
3049 pr_err("failed to write stat round event\n");
3050 }
3051
3052 if (!perf_stat.data.is_pipe) {
3053 perf_stat.session->header.data_size += perf_stat.bytes_written;
3054 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
3055 }
3056
3057 perf_session__delete(perf_stat.session);
3058 }
3059
3060 perf_stat__exit_aggr_mode();
3061 perf_evlist__free_stats(evsel_list);
3062out:
3063 if (smi_cost && smi_reset)
3064 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
3065
3066 perf_evlist__delete(evsel_list);
3067
3068 runtime_stat_delete(&stat_config);
3069
3070 return status;
3071}
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
74static void print_counters(struct timespec *ts, int argc, const char **argv);
75
76/* Default events used for perf stat -T */
77static 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. */
90static const char * transaction_limited_attrs = {
91 "task-clock,"
92 "{"
93 "instructions,"
94 "cycles,"
95 "cpu/cycles-t/,"
96 "cpu/tx-start/"
97 "}"
98};
99
100static struct perf_evlist *evsel_list;
101
102static struct target target = {
103 .uid = UINT_MAX,
104};
105
106typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
107
108static int run_count = 1;
109static bool no_inherit = false;
110static volatile pid_t child_pid = -1;
111static bool null_run = false;
112static int detailed_run = 0;
113static bool transaction_run;
114static bool big_num = true;
115static int big_num_opt = -1;
116static const char *csv_sep = NULL;
117static bool csv_output = false;
118static bool group = false;
119static const char *pre_cmd = NULL;
120static const char *post_cmd = NULL;
121static bool sync_run = false;
122static unsigned int initial_delay = 0;
123static unsigned int unit_width = 4; /* strlen("unit") */
124static bool forever = false;
125static bool metric_only = false;
126static struct timespec ref_time;
127static struct cpu_map *aggr_map;
128static aggr_get_id_t aggr_get_id;
129static bool append_file;
130static const char *output_name;
131static int output_fd;
132
133struct perf_stat {
134 bool record;
135 struct perf_data_file file;
136 struct perf_session *session;
137 u64 bytes_written;
138 struct perf_tool tool;
139 bool maps_allocated;
140 struct cpu_map *cpus;
141 struct thread_map *threads;
142 enum aggr_mode aggr_mode;
143};
144
145static struct perf_stat perf_stat;
146#define STAT_RECORD perf_stat.record
147
148static volatile int done = 0;
149
150static struct perf_stat_config stat_config = {
151 .aggr_mode = AGGR_GLOBAL,
152 .scale = true,
153};
154
155static inline void diff_timespec(struct timespec *r, struct timespec *a,
156 struct timespec *b)
157{
158 r->tv_sec = a->tv_sec - b->tv_sec;
159 if (a->tv_nsec < b->tv_nsec) {
160 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
161 r->tv_sec--;
162 } else {
163 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
164 }
165}
166
167static void perf_stat__reset_stats(void)
168{
169 perf_evlist__reset_stats(evsel_list);
170 perf_stat__reset_shadow_stats();
171}
172
173static int create_perf_stat_counter(struct perf_evsel *evsel)
174{
175 struct perf_event_attr *attr = &evsel->attr;
176
177 if (stat_config.scale)
178 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
179 PERF_FORMAT_TOTAL_TIME_RUNNING;
180
181 attr->inherit = !no_inherit;
182
183 /*
184 * Some events get initialized with sample_(period/type) set,
185 * like tracepoints. Clear it up for counting.
186 */
187 attr->sample_period = 0;
188
189 /*
190 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
191 * while avoiding that older tools show confusing messages.
192 *
193 * However for pipe sessions we need to keep it zero,
194 * because script's perf_evsel__check_attr is triggered
195 * by attr->sample_type != 0, and we can't run it on
196 * stat sessions.
197 */
198 if (!(STAT_RECORD && perf_stat.file.is_pipe))
199 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
200
201 /*
202 * Disabling all counters initially, they will be enabled
203 * either manually by us or by kernel via enable_on_exec
204 * set later.
205 */
206 if (perf_evsel__is_group_leader(evsel)) {
207 attr->disabled = 1;
208
209 /*
210 * In case of initial_delay we enable tracee
211 * events manually.
212 */
213 if (target__none(&target) && !initial_delay)
214 attr->enable_on_exec = 1;
215 }
216
217 if (target__has_cpu(&target))
218 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
219
220 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
221}
222
223/*
224 * Does the counter have nsecs as a unit?
225 */
226static inline int nsec_counter(struct perf_evsel *evsel)
227{
228 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
229 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
230 return 1;
231
232 return 0;
233}
234
235static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
236 union perf_event *event,
237 struct perf_sample *sample __maybe_unused,
238 struct machine *machine __maybe_unused)
239{
240 if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
241 pr_err("failed to write perf data, error: %m\n");
242 return -1;
243 }
244
245 perf_stat.bytes_written += event->header.size;
246 return 0;
247}
248
249static int write_stat_round_event(u64 tm, u64 type)
250{
251 return perf_event__synthesize_stat_round(NULL, tm, type,
252 process_synthesized_event,
253 NULL);
254}
255
256#define WRITE_STAT_ROUND_EVENT(time, interval) \
257 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
258
259#define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
260
261static int
262perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
263 struct perf_counts_values *count)
264{
265 struct perf_sample_id *sid = SID(counter, cpu, thread);
266
267 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
268 process_synthesized_event, NULL);
269}
270
271/*
272 * Read out the results of a single counter:
273 * do not aggregate counts across CPUs in system-wide mode
274 */
275static int read_counter(struct perf_evsel *counter)
276{
277 int nthreads = thread_map__nr(evsel_list->threads);
278 int ncpus = perf_evsel__nr_cpus(counter);
279 int cpu, thread;
280
281 if (!counter->supported)
282 return -ENOENT;
283
284 if (counter->system_wide)
285 nthreads = 1;
286
287 for (thread = 0; thread < nthreads; thread++) {
288 for (cpu = 0; cpu < ncpus; cpu++) {
289 struct perf_counts_values *count;
290
291 count = perf_counts(counter->counts, cpu, thread);
292 if (perf_evsel__read(counter, cpu, thread, count))
293 return -1;
294
295 if (STAT_RECORD) {
296 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
297 pr_err("failed to write stat event\n");
298 return -1;
299 }
300 }
301 }
302 }
303
304 return 0;
305}
306
307static void read_counters(bool close_counters)
308{
309 struct perf_evsel *counter;
310
311 evlist__for_each(evsel_list, counter) {
312 if (read_counter(counter))
313 pr_debug("failed to read counter %s\n", counter->name);
314
315 if (perf_stat_process_counter(&stat_config, counter))
316 pr_warning("failed to process counter %s\n", counter->name);
317
318 if (close_counters) {
319 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
320 thread_map__nr(evsel_list->threads));
321 }
322 }
323}
324
325static void process_interval(void)
326{
327 struct timespec ts, rs;
328
329 read_counters(false);
330
331 clock_gettime(CLOCK_MONOTONIC, &ts);
332 diff_timespec(&rs, &ts, &ref_time);
333
334 if (STAT_RECORD) {
335 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSECS_PER_SEC + rs.tv_nsec, INTERVAL))
336 pr_err("failed to write stat round event\n");
337 }
338
339 print_counters(&rs, 0, NULL);
340}
341
342static void enable_counters(void)
343{
344 if (initial_delay)
345 usleep(initial_delay * 1000);
346
347 /*
348 * We need to enable counters only if:
349 * - we don't have tracee (attaching to task or cpu)
350 * - we have initial delay configured
351 */
352 if (!target__none(&target) || initial_delay)
353 perf_evlist__enable(evsel_list);
354}
355
356static volatile int workload_exec_errno;
357
358/*
359 * perf_evlist__prepare_workload will send a SIGUSR1
360 * if the fork fails, since we asked by setting its
361 * want_signal to true.
362 */
363static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
364 void *ucontext __maybe_unused)
365{
366 workload_exec_errno = info->si_value.sival_int;
367}
368
369static bool has_unit(struct perf_evsel *counter)
370{
371 return counter->unit && *counter->unit;
372}
373
374static bool has_scale(struct perf_evsel *counter)
375{
376 return counter->scale != 1;
377}
378
379static int perf_stat_synthesize_config(bool is_pipe)
380{
381 struct perf_evsel *counter;
382 int err;
383
384 if (is_pipe) {
385 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
386 process_synthesized_event);
387 if (err < 0) {
388 pr_err("Couldn't synthesize attrs.\n");
389 return err;
390 }
391 }
392
393 /*
394 * Synthesize other events stuff not carried within
395 * attr event - unit, scale, name
396 */
397 evlist__for_each(evsel_list, counter) {
398 if (!counter->supported)
399 continue;
400
401 /*
402 * Synthesize unit and scale only if it's defined.
403 */
404 if (has_unit(counter)) {
405 err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
406 if (err < 0) {
407 pr_err("Couldn't synthesize evsel unit.\n");
408 return err;
409 }
410 }
411
412 if (has_scale(counter)) {
413 err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
414 if (err < 0) {
415 pr_err("Couldn't synthesize evsel scale.\n");
416 return err;
417 }
418 }
419
420 if (counter->own_cpus) {
421 err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
422 if (err < 0) {
423 pr_err("Couldn't synthesize evsel scale.\n");
424 return err;
425 }
426 }
427
428 /*
429 * Name is needed only for pipe output,
430 * perf.data carries event names.
431 */
432 if (is_pipe) {
433 err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
434 if (err < 0) {
435 pr_err("Couldn't synthesize evsel name.\n");
436 return err;
437 }
438 }
439 }
440
441 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
442 process_synthesized_event,
443 NULL);
444 if (err < 0) {
445 pr_err("Couldn't synthesize thread map.\n");
446 return err;
447 }
448
449 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
450 process_synthesized_event, NULL);
451 if (err < 0) {
452 pr_err("Couldn't synthesize thread map.\n");
453 return err;
454 }
455
456 err = perf_event__synthesize_stat_config(NULL, &stat_config,
457 process_synthesized_event, NULL);
458 if (err < 0) {
459 pr_err("Couldn't synthesize config.\n");
460 return err;
461 }
462
463 return 0;
464}
465
466#define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
467
468static int __store_counter_ids(struct perf_evsel *counter,
469 struct cpu_map *cpus,
470 struct thread_map *threads)
471{
472 int cpu, thread;
473
474 for (cpu = 0; cpu < cpus->nr; cpu++) {
475 for (thread = 0; thread < threads->nr; thread++) {
476 int fd = FD(counter, cpu, thread);
477
478 if (perf_evlist__id_add_fd(evsel_list, counter,
479 cpu, thread, fd) < 0)
480 return -1;
481 }
482 }
483
484 return 0;
485}
486
487static int store_counter_ids(struct perf_evsel *counter)
488{
489 struct cpu_map *cpus = counter->cpus;
490 struct thread_map *threads = counter->threads;
491
492 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
493 return -ENOMEM;
494
495 return __store_counter_ids(counter, cpus, threads);
496}
497
498static int __run_perf_stat(int argc, const char **argv)
499{
500 int interval = stat_config.interval;
501 char msg[512];
502 unsigned long long t0, t1;
503 struct perf_evsel *counter;
504 struct timespec ts;
505 size_t l;
506 int status = 0;
507 const bool forks = (argc > 0);
508 bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
509
510 if (interval) {
511 ts.tv_sec = interval / 1000;
512 ts.tv_nsec = (interval % 1000) * 1000000;
513 } else {
514 ts.tv_sec = 1;
515 ts.tv_nsec = 0;
516 }
517
518 if (forks) {
519 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
520 workload_exec_failed_signal) < 0) {
521 perror("failed to prepare workload");
522 return -1;
523 }
524 child_pid = evsel_list->workload.pid;
525 }
526
527 if (group)
528 perf_evlist__set_leader(evsel_list);
529
530 evlist__for_each(evsel_list, counter) {
531try_again:
532 if (create_perf_stat_counter(counter) < 0) {
533 /*
534 * PPC returns ENXIO for HW counters until 2.6.37
535 * (behavior changed with commit b0a873e).
536 */
537 if (errno == EINVAL || errno == ENOSYS ||
538 errno == ENOENT || errno == EOPNOTSUPP ||
539 errno == ENXIO) {
540 if (verbose)
541 ui__warning("%s event is not supported by the kernel.\n",
542 perf_evsel__name(counter));
543 counter->supported = false;
544
545 if ((counter->leader != counter) ||
546 !(counter->leader->nr_members > 1))
547 continue;
548 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
549 if (verbose)
550 ui__warning("%s\n", msg);
551 goto try_again;
552 }
553
554 perf_evsel__open_strerror(counter, &target,
555 errno, msg, sizeof(msg));
556 ui__error("%s\n", msg);
557
558 if (child_pid != -1)
559 kill(child_pid, SIGTERM);
560
561 return -1;
562 }
563 counter->supported = true;
564
565 l = strlen(counter->unit);
566 if (l > unit_width)
567 unit_width = l;
568
569 if (STAT_RECORD && store_counter_ids(counter))
570 return -1;
571 }
572
573 if (perf_evlist__apply_filters(evsel_list, &counter)) {
574 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
575 counter->filter, perf_evsel__name(counter), errno,
576 strerror_r(errno, msg, sizeof(msg)));
577 return -1;
578 }
579
580 if (STAT_RECORD) {
581 int err, fd = perf_data_file__fd(&perf_stat.file);
582
583 if (is_pipe) {
584 err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
585 } else {
586 err = perf_session__write_header(perf_stat.session, evsel_list,
587 fd, false);
588 }
589
590 if (err < 0)
591 return err;
592
593 err = perf_stat_synthesize_config(is_pipe);
594 if (err < 0)
595 return err;
596 }
597
598 /*
599 * Enable counters and exec the command:
600 */
601 t0 = rdclock();
602 clock_gettime(CLOCK_MONOTONIC, &ref_time);
603
604 if (forks) {
605 perf_evlist__start_workload(evsel_list);
606 enable_counters();
607
608 if (interval) {
609 while (!waitpid(child_pid, &status, WNOHANG)) {
610 nanosleep(&ts, NULL);
611 process_interval();
612 }
613 }
614 wait(&status);
615
616 if (workload_exec_errno) {
617 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
618 pr_err("Workload failed: %s\n", emsg);
619 return -1;
620 }
621
622 if (WIFSIGNALED(status))
623 psignal(WTERMSIG(status), argv[0]);
624 } else {
625 enable_counters();
626 while (!done) {
627 nanosleep(&ts, NULL);
628 if (interval)
629 process_interval();
630 }
631 }
632
633 t1 = rdclock();
634
635 update_stats(&walltime_nsecs_stats, t1 - t0);
636
637 read_counters(true);
638
639 return WEXITSTATUS(status);
640}
641
642static int run_perf_stat(int argc, const char **argv)
643{
644 int ret;
645
646 if (pre_cmd) {
647 ret = system(pre_cmd);
648 if (ret)
649 return ret;
650 }
651
652 if (sync_run)
653 sync();
654
655 ret = __run_perf_stat(argc, argv);
656 if (ret)
657 return ret;
658
659 if (post_cmd) {
660 ret = system(post_cmd);
661 if (ret)
662 return ret;
663 }
664
665 return ret;
666}
667
668static void print_running(u64 run, u64 ena)
669{
670 if (csv_output) {
671 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
672 csv_sep,
673 run,
674 csv_sep,
675 ena ? 100.0 * run / ena : 100.0);
676 } else if (run != ena) {
677 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
678 }
679}
680
681static void print_noise_pct(double total, double avg)
682{
683 double pct = rel_stddev_stats(total, avg);
684
685 if (csv_output)
686 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
687 else if (pct)
688 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
689}
690
691static void print_noise(struct perf_evsel *evsel, double avg)
692{
693 struct perf_stat_evsel *ps;
694
695 if (run_count == 1)
696 return;
697
698 ps = evsel->priv;
699 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
700}
701
702static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
703{
704 switch (stat_config.aggr_mode) {
705 case AGGR_CORE:
706 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
707 cpu_map__id_to_socket(id),
708 csv_output ? 0 : -8,
709 cpu_map__id_to_cpu(id),
710 csv_sep,
711 csv_output ? 0 : 4,
712 nr,
713 csv_sep);
714 break;
715 case AGGR_SOCKET:
716 fprintf(stat_config.output, "S%*d%s%*d%s",
717 csv_output ? 0 : -5,
718 id,
719 csv_sep,
720 csv_output ? 0 : 4,
721 nr,
722 csv_sep);
723 break;
724 case AGGR_NONE:
725 fprintf(stat_config.output, "CPU%*d%s",
726 csv_output ? 0 : -4,
727 perf_evsel__cpus(evsel)->map[id], csv_sep);
728 break;
729 case AGGR_THREAD:
730 fprintf(stat_config.output, "%*s-%*d%s",
731 csv_output ? 0 : 16,
732 thread_map__comm(evsel->threads, id),
733 csv_output ? 0 : -8,
734 thread_map__pid(evsel->threads, id),
735 csv_sep);
736 break;
737 case AGGR_GLOBAL:
738 case AGGR_UNSET:
739 default:
740 break;
741 }
742}
743
744struct outstate {
745 FILE *fh;
746 bool newline;
747 const char *prefix;
748 int nfields;
749 int id, nr;
750 struct perf_evsel *evsel;
751};
752
753#define METRIC_LEN 35
754
755static void new_line_std(void *ctx)
756{
757 struct outstate *os = ctx;
758
759 os->newline = true;
760}
761
762static void do_new_line_std(struct outstate *os)
763{
764 fputc('\n', os->fh);
765 fputs(os->prefix, os->fh);
766 aggr_printout(os->evsel, os->id, os->nr);
767 if (stat_config.aggr_mode == AGGR_NONE)
768 fprintf(os->fh, " ");
769 fprintf(os->fh, " ");
770}
771
772static void print_metric_std(void *ctx, const char *color, const char *fmt,
773 const char *unit, double val)
774{
775 struct outstate *os = ctx;
776 FILE *out = os->fh;
777 int n;
778 bool newline = os->newline;
779
780 os->newline = false;
781
782 if (unit == NULL || fmt == NULL) {
783 fprintf(out, "%-*s", METRIC_LEN, "");
784 return;
785 }
786
787 if (newline)
788 do_new_line_std(os);
789
790 n = fprintf(out, " # ");
791 if (color)
792 n += color_fprintf(out, color, fmt, val);
793 else
794 n += fprintf(out, fmt, val);
795 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
796}
797
798static void new_line_csv(void *ctx)
799{
800 struct outstate *os = ctx;
801 int i;
802
803 fputc('\n', os->fh);
804 if (os->prefix)
805 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
806 aggr_printout(os->evsel, os->id, os->nr);
807 for (i = 0; i < os->nfields; i++)
808 fputs(csv_sep, os->fh);
809}
810
811static void print_metric_csv(void *ctx,
812 const char *color __maybe_unused,
813 const char *fmt, const char *unit, double val)
814{
815 struct outstate *os = ctx;
816 FILE *out = os->fh;
817 char buf[64], *vals, *ends;
818
819 if (unit == NULL || fmt == NULL) {
820 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
821 return;
822 }
823 snprintf(buf, sizeof(buf), fmt, val);
824 vals = buf;
825 while (isspace(*vals))
826 vals++;
827 ends = vals;
828 while (isdigit(*ends) || *ends == '.')
829 ends++;
830 *ends = 0;
831 while (isspace(*unit))
832 unit++;
833 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
834}
835
836#define METRIC_ONLY_LEN 20
837
838/* Filter out some columns that don't work well in metrics only mode */
839
840static bool valid_only_metric(const char *unit)
841{
842 if (!unit)
843 return false;
844 if (strstr(unit, "/sec") ||
845 strstr(unit, "hz") ||
846 strstr(unit, "Hz") ||
847 strstr(unit, "CPUs utilized"))
848 return false;
849 return true;
850}
851
852static const char *fixunit(char *buf, struct perf_evsel *evsel,
853 const char *unit)
854{
855 if (!strncmp(unit, "of all", 6)) {
856 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
857 unit);
858 return buf;
859 }
860 return unit;
861}
862
863static void print_metric_only(void *ctx, const char *color, const char *fmt,
864 const char *unit, double val)
865{
866 struct outstate *os = ctx;
867 FILE *out = os->fh;
868 int n;
869 char buf[1024];
870 unsigned mlen = METRIC_ONLY_LEN;
871
872 if (!valid_only_metric(unit))
873 return;
874 unit = fixunit(buf, os->evsel, unit);
875 if (color)
876 n = color_fprintf(out, color, fmt, val);
877 else
878 n = fprintf(out, fmt, val);
879 if (n > METRIC_ONLY_LEN)
880 n = METRIC_ONLY_LEN;
881 if (mlen < strlen(unit))
882 mlen = strlen(unit) + 1;
883 fprintf(out, "%*s", mlen - n, "");
884}
885
886static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
887 const char *fmt,
888 const char *unit, double val)
889{
890 struct outstate *os = ctx;
891 FILE *out = os->fh;
892 char buf[64], *vals, *ends;
893 char tbuf[1024];
894
895 if (!valid_only_metric(unit))
896 return;
897 unit = fixunit(tbuf, os->evsel, unit);
898 snprintf(buf, sizeof buf, fmt, val);
899 vals = buf;
900 while (isspace(*vals))
901 vals++;
902 ends = vals;
903 while (isdigit(*ends) || *ends == '.')
904 ends++;
905 *ends = 0;
906 fprintf(out, "%s%s", vals, csv_sep);
907}
908
909static void new_line_metric(void *ctx __maybe_unused)
910{
911}
912
913static void print_metric_header(void *ctx, const char *color __maybe_unused,
914 const char *fmt __maybe_unused,
915 const char *unit, double val __maybe_unused)
916{
917 struct outstate *os = ctx;
918 char tbuf[1024];
919
920 if (!valid_only_metric(unit))
921 return;
922 unit = fixunit(tbuf, os->evsel, unit);
923 if (csv_output)
924 fprintf(os->fh, "%s%s", unit, csv_sep);
925 else
926 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
927}
928
929static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
930{
931 FILE *output = stat_config.output;
932 double msecs = avg / 1e6;
933 const char *fmt_v, *fmt_n;
934 char name[25];
935
936 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
937 fmt_n = csv_output ? "%s" : "%-25s";
938
939 aggr_printout(evsel, id, nr);
940
941 scnprintf(name, sizeof(name), "%s%s",
942 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
943
944 fprintf(output, fmt_v, msecs, csv_sep);
945
946 if (csv_output)
947 fprintf(output, "%s%s", evsel->unit, csv_sep);
948 else
949 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
950
951 fprintf(output, fmt_n, name);
952
953 if (evsel->cgrp)
954 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
955}
956
957static int first_shadow_cpu(struct perf_evsel *evsel, int id)
958{
959 int i;
960
961 if (!aggr_get_id)
962 return 0;
963
964 if (stat_config.aggr_mode == AGGR_NONE)
965 return id;
966
967 if (stat_config.aggr_mode == AGGR_GLOBAL)
968 return 0;
969
970 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
971 int cpu2 = perf_evsel__cpus(evsel)->map[i];
972
973 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
974 return cpu2;
975 }
976 return 0;
977}
978
979static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
980{
981 FILE *output = stat_config.output;
982 double sc = evsel->scale;
983 const char *fmt;
984
985 if (csv_output) {
986 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
987 } else {
988 if (big_num)
989 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
990 else
991 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
992 }
993
994 aggr_printout(evsel, id, nr);
995
996 fprintf(output, fmt, avg, csv_sep);
997
998 if (evsel->unit)
999 fprintf(output, "%-*s%s",
1000 csv_output ? 0 : unit_width,
1001 evsel->unit, csv_sep);
1002
1003 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1004
1005 if (evsel->cgrp)
1006 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1007}
1008
1009static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1010 char *prefix, u64 run, u64 ena, double noise)
1011{
1012 struct perf_stat_output_ctx out;
1013 struct outstate os = {
1014 .fh = stat_config.output,
1015 .prefix = prefix ? prefix : "",
1016 .id = id,
1017 .nr = nr,
1018 .evsel = counter,
1019 };
1020 print_metric_t pm = print_metric_std;
1021 void (*nl)(void *);
1022
1023 if (metric_only) {
1024 nl = new_line_metric;
1025 if (csv_output)
1026 pm = print_metric_only_csv;
1027 else
1028 pm = print_metric_only;
1029 } else
1030 nl = new_line_std;
1031
1032 if (csv_output && !metric_only) {
1033 static int aggr_fields[] = {
1034 [AGGR_GLOBAL] = 0,
1035 [AGGR_THREAD] = 1,
1036 [AGGR_NONE] = 1,
1037 [AGGR_SOCKET] = 2,
1038 [AGGR_CORE] = 2,
1039 };
1040
1041 pm = print_metric_csv;
1042 nl = new_line_csv;
1043 os.nfields = 3;
1044 os.nfields += aggr_fields[stat_config.aggr_mode];
1045 if (counter->cgrp)
1046 os.nfields++;
1047 }
1048 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1049 if (metric_only) {
1050 pm(&os, NULL, "", "", 0);
1051 return;
1052 }
1053 aggr_printout(counter, id, nr);
1054
1055 fprintf(stat_config.output, "%*s%s",
1056 csv_output ? 0 : 18,
1057 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1058 csv_sep);
1059
1060 fprintf(stat_config.output, "%-*s%s",
1061 csv_output ? 0 : unit_width,
1062 counter->unit, csv_sep);
1063
1064 fprintf(stat_config.output, "%*s",
1065 csv_output ? 0 : -25,
1066 perf_evsel__name(counter));
1067
1068 if (counter->cgrp)
1069 fprintf(stat_config.output, "%s%s",
1070 csv_sep, counter->cgrp->name);
1071
1072 if (!csv_output)
1073 pm(&os, NULL, NULL, "", 0);
1074 print_noise(counter, noise);
1075 print_running(run, ena);
1076 if (csv_output)
1077 pm(&os, NULL, NULL, "", 0);
1078 return;
1079 }
1080
1081 if (metric_only)
1082 /* nothing */;
1083 else if (nsec_counter(counter))
1084 nsec_printout(id, nr, counter, uval);
1085 else
1086 abs_printout(id, nr, counter, uval);
1087
1088 out.print_metric = pm;
1089 out.new_line = nl;
1090 out.ctx = &os;
1091
1092 if (csv_output && !metric_only) {
1093 print_noise(counter, noise);
1094 print_running(run, ena);
1095 }
1096
1097 perf_stat__print_shadow_stats(counter, uval,
1098 first_shadow_cpu(counter, id),
1099 &out);
1100 if (!csv_output && !metric_only) {
1101 print_noise(counter, noise);
1102 print_running(run, ena);
1103 }
1104}
1105
1106static void aggr_update_shadow(void)
1107{
1108 int cpu, s2, id, s;
1109 u64 val;
1110 struct perf_evsel *counter;
1111
1112 for (s = 0; s < aggr_map->nr; s++) {
1113 id = aggr_map->map[s];
1114 evlist__for_each(evsel_list, counter) {
1115 val = 0;
1116 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1117 s2 = aggr_get_id(evsel_list->cpus, cpu);
1118 if (s2 != id)
1119 continue;
1120 val += perf_counts(counter->counts, cpu, 0)->val;
1121 }
1122 val = val * counter->scale;
1123 perf_stat__update_shadow_stats(counter, &val,
1124 first_shadow_cpu(counter, id));
1125 }
1126 }
1127}
1128
1129static void print_aggr(char *prefix)
1130{
1131 FILE *output = stat_config.output;
1132 struct perf_evsel *counter;
1133 int cpu, s, s2, id, nr;
1134 double uval;
1135 u64 ena, run, val;
1136 bool first;
1137
1138 if (!(aggr_map || aggr_get_id))
1139 return;
1140
1141 aggr_update_shadow();
1142
1143 /*
1144 * With metric_only everything is on a single line.
1145 * Without each counter has its own line.
1146 */
1147 for (s = 0; s < aggr_map->nr; s++) {
1148 if (prefix && metric_only)
1149 fprintf(output, "%s", prefix);
1150
1151 id = aggr_map->map[s];
1152 first = true;
1153 evlist__for_each(evsel_list, counter) {
1154 val = ena = run = 0;
1155 nr = 0;
1156 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1157 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1158 if (s2 != id)
1159 continue;
1160 val += perf_counts(counter->counts, cpu, 0)->val;
1161 ena += perf_counts(counter->counts, cpu, 0)->ena;
1162 run += perf_counts(counter->counts, cpu, 0)->run;
1163 nr++;
1164 }
1165 if (first && metric_only) {
1166 first = false;
1167 aggr_printout(counter, id, nr);
1168 }
1169 if (prefix && !metric_only)
1170 fprintf(output, "%s", prefix);
1171
1172 uval = val * counter->scale;
1173 printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1174 if (!metric_only)
1175 fputc('\n', output);
1176 }
1177 if (metric_only)
1178 fputc('\n', output);
1179 }
1180}
1181
1182static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1183{
1184 FILE *output = stat_config.output;
1185 int nthreads = thread_map__nr(counter->threads);
1186 int ncpus = cpu_map__nr(counter->cpus);
1187 int cpu, thread;
1188 double uval;
1189
1190 for (thread = 0; thread < nthreads; thread++) {
1191 u64 ena = 0, run = 0, val = 0;
1192
1193 for (cpu = 0; cpu < ncpus; cpu++) {
1194 val += perf_counts(counter->counts, cpu, thread)->val;
1195 ena += perf_counts(counter->counts, cpu, thread)->ena;
1196 run += perf_counts(counter->counts, cpu, thread)->run;
1197 }
1198
1199 if (prefix)
1200 fprintf(output, "%s", prefix);
1201
1202 uval = val * counter->scale;
1203 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1204 fputc('\n', output);
1205 }
1206}
1207
1208/*
1209 * Print out the results of a single counter:
1210 * aggregated counts in system-wide mode
1211 */
1212static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1213{
1214 FILE *output = stat_config.output;
1215 struct perf_stat_evsel *ps = counter->priv;
1216 double avg = avg_stats(&ps->res_stats[0]);
1217 double uval;
1218 double avg_enabled, avg_running;
1219
1220 avg_enabled = avg_stats(&ps->res_stats[1]);
1221 avg_running = avg_stats(&ps->res_stats[2]);
1222
1223 if (prefix && !metric_only)
1224 fprintf(output, "%s", prefix);
1225
1226 uval = avg * counter->scale;
1227 printout(-1, 0, counter, uval, prefix, avg_running, avg_enabled, avg);
1228 if (!metric_only)
1229 fprintf(output, "\n");
1230}
1231
1232/*
1233 * Print out the results of a single counter:
1234 * does not use aggregated count in system-wide
1235 */
1236static void print_counter(struct perf_evsel *counter, char *prefix)
1237{
1238 FILE *output = stat_config.output;
1239 u64 ena, run, val;
1240 double uval;
1241 int cpu;
1242
1243 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1244 val = perf_counts(counter->counts, cpu, 0)->val;
1245 ena = perf_counts(counter->counts, cpu, 0)->ena;
1246 run = perf_counts(counter->counts, cpu, 0)->run;
1247
1248 if (prefix)
1249 fprintf(output, "%s", prefix);
1250
1251 uval = val * counter->scale;
1252 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1253
1254 fputc('\n', output);
1255 }
1256}
1257
1258static void print_no_aggr_metric(char *prefix)
1259{
1260 int cpu;
1261 int nrcpus = 0;
1262 struct perf_evsel *counter;
1263 u64 ena, run, val;
1264 double uval;
1265
1266 nrcpus = evsel_list->cpus->nr;
1267 for (cpu = 0; cpu < nrcpus; cpu++) {
1268 bool first = true;
1269
1270 if (prefix)
1271 fputs(prefix, stat_config.output);
1272 evlist__for_each(evsel_list, counter) {
1273 if (first) {
1274 aggr_printout(counter, cpu, 0);
1275 first = false;
1276 }
1277 val = perf_counts(counter->counts, cpu, 0)->val;
1278 ena = perf_counts(counter->counts, cpu, 0)->ena;
1279 run = perf_counts(counter->counts, cpu, 0)->run;
1280
1281 uval = val * counter->scale;
1282 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1283 }
1284 fputc('\n', stat_config.output);
1285 }
1286}
1287
1288static int aggr_header_lens[] = {
1289 [AGGR_CORE] = 18,
1290 [AGGR_SOCKET] = 12,
1291 [AGGR_NONE] = 6,
1292 [AGGR_THREAD] = 24,
1293 [AGGR_GLOBAL] = 0,
1294};
1295
1296static void print_metric_headers(char *prefix)
1297{
1298 struct perf_stat_output_ctx out;
1299 struct perf_evsel *counter;
1300 struct outstate os = {
1301 .fh = stat_config.output
1302 };
1303
1304 if (prefix)
1305 fprintf(stat_config.output, "%s", prefix);
1306
1307 if (!csv_output)
1308 fprintf(stat_config.output, "%*s",
1309 aggr_header_lens[stat_config.aggr_mode], "");
1310
1311 /* Print metrics headers only */
1312 evlist__for_each(evsel_list, counter) {
1313 os.evsel = counter;
1314 out.ctx = &os;
1315 out.print_metric = print_metric_header;
1316 out.new_line = new_line_metric;
1317 os.evsel = counter;
1318 perf_stat__print_shadow_stats(counter, 0,
1319 0,
1320 &out);
1321 }
1322 fputc('\n', stat_config.output);
1323}
1324
1325static void print_interval(char *prefix, struct timespec *ts)
1326{
1327 FILE *output = stat_config.output;
1328 static int num_print_interval;
1329
1330 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1331
1332 if (num_print_interval == 0 && !csv_output && !metric_only) {
1333 switch (stat_config.aggr_mode) {
1334 case AGGR_SOCKET:
1335 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
1336 break;
1337 case AGGR_CORE:
1338 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
1339 break;
1340 case AGGR_NONE:
1341 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
1342 break;
1343 case AGGR_THREAD:
1344 fprintf(output, "# time comm-pid counts %*s events\n", unit_width, "unit");
1345 break;
1346 case AGGR_GLOBAL:
1347 default:
1348 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
1349 case AGGR_UNSET:
1350 break;
1351 }
1352 }
1353
1354 if (++num_print_interval == 25)
1355 num_print_interval = 0;
1356}
1357
1358static void print_header(int argc, const char **argv)
1359{
1360 FILE *output = stat_config.output;
1361 int i;
1362
1363 fflush(stdout);
1364
1365 if (!csv_output) {
1366 fprintf(output, "\n");
1367 fprintf(output, " Performance counter stats for ");
1368 if (target.system_wide)
1369 fprintf(output, "\'system wide");
1370 else if (target.cpu_list)
1371 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1372 else if (!target__has_task(&target)) {
1373 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1374 for (i = 1; argv && (i < argc); i++)
1375 fprintf(output, " %s", argv[i]);
1376 } else if (target.pid)
1377 fprintf(output, "process id \'%s", target.pid);
1378 else
1379 fprintf(output, "thread id \'%s", target.tid);
1380
1381 fprintf(output, "\'");
1382 if (run_count > 1)
1383 fprintf(output, " (%d runs)", run_count);
1384 fprintf(output, ":\n\n");
1385 }
1386}
1387
1388static void print_footer(void)
1389{
1390 FILE *output = stat_config.output;
1391
1392 if (!null_run)
1393 fprintf(output, "\n");
1394 fprintf(output, " %17.9f seconds time elapsed",
1395 avg_stats(&walltime_nsecs_stats)/1e9);
1396 if (run_count > 1) {
1397 fprintf(output, " ");
1398 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1399 avg_stats(&walltime_nsecs_stats));
1400 }
1401 fprintf(output, "\n\n");
1402}
1403
1404static void print_counters(struct timespec *ts, int argc, const char **argv)
1405{
1406 int interval = stat_config.interval;
1407 struct perf_evsel *counter;
1408 char buf[64], *prefix = NULL;
1409
1410 /* Do not print anything if we record to the pipe. */
1411 if (STAT_RECORD && perf_stat.file.is_pipe)
1412 return;
1413
1414 if (interval)
1415 print_interval(prefix = buf, ts);
1416 else
1417 print_header(argc, argv);
1418
1419 if (metric_only) {
1420 static int num_print_iv;
1421
1422 if (num_print_iv == 0)
1423 print_metric_headers(prefix);
1424 if (num_print_iv++ == 25)
1425 num_print_iv = 0;
1426 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1427 fprintf(stat_config.output, "%s", prefix);
1428 }
1429
1430 switch (stat_config.aggr_mode) {
1431 case AGGR_CORE:
1432 case AGGR_SOCKET:
1433 print_aggr(prefix);
1434 break;
1435 case AGGR_THREAD:
1436 evlist__for_each(evsel_list, counter)
1437 print_aggr_thread(counter, prefix);
1438 break;
1439 case AGGR_GLOBAL:
1440 evlist__for_each(evsel_list, counter)
1441 print_counter_aggr(counter, prefix);
1442 if (metric_only)
1443 fputc('\n', stat_config.output);
1444 break;
1445 case AGGR_NONE:
1446 if (metric_only)
1447 print_no_aggr_metric(prefix);
1448 else {
1449 evlist__for_each(evsel_list, counter)
1450 print_counter(counter, prefix);
1451 }
1452 break;
1453 case AGGR_UNSET:
1454 default:
1455 break;
1456 }
1457
1458 if (!interval && !csv_output)
1459 print_footer();
1460
1461 fflush(stat_config.output);
1462}
1463
1464static volatile int signr = -1;
1465
1466static void skip_signal(int signo)
1467{
1468 if ((child_pid == -1) || stat_config.interval)
1469 done = 1;
1470
1471 signr = signo;
1472 /*
1473 * render child_pid harmless
1474 * won't send SIGTERM to a random
1475 * process in case of race condition
1476 * and fast PID recycling
1477 */
1478 child_pid = -1;
1479}
1480
1481static void sig_atexit(void)
1482{
1483 sigset_t set, oset;
1484
1485 /*
1486 * avoid race condition with SIGCHLD handler
1487 * in skip_signal() which is modifying child_pid
1488 * goal is to avoid send SIGTERM to a random
1489 * process
1490 */
1491 sigemptyset(&set);
1492 sigaddset(&set, SIGCHLD);
1493 sigprocmask(SIG_BLOCK, &set, &oset);
1494
1495 if (child_pid != -1)
1496 kill(child_pid, SIGTERM);
1497
1498 sigprocmask(SIG_SETMASK, &oset, NULL);
1499
1500 if (signr == -1)
1501 return;
1502
1503 signal(signr, SIG_DFL);
1504 kill(getpid(), signr);
1505}
1506
1507static int stat__set_big_num(const struct option *opt __maybe_unused,
1508 const char *s __maybe_unused, int unset)
1509{
1510 big_num_opt = unset ? 0 : 1;
1511 return 0;
1512}
1513
1514static const struct option stat_options[] = {
1515 OPT_BOOLEAN('T', "transaction", &transaction_run,
1516 "hardware transaction statistics"),
1517 OPT_CALLBACK('e', "event", &evsel_list, "event",
1518 "event selector. use 'perf list' to list available events",
1519 parse_events_option),
1520 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1521 "event filter", parse_filter),
1522 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1523 "child tasks do not inherit counters"),
1524 OPT_STRING('p', "pid", &target.pid, "pid",
1525 "stat events on existing process id"),
1526 OPT_STRING('t', "tid", &target.tid, "tid",
1527 "stat events on existing thread id"),
1528 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1529 "system-wide collection from all CPUs"),
1530 OPT_BOOLEAN('g', "group", &group,
1531 "put the counters into a counter group"),
1532 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1533 OPT_INCR('v', "verbose", &verbose,
1534 "be more verbose (show counter open errors, etc)"),
1535 OPT_INTEGER('r', "repeat", &run_count,
1536 "repeat command and print average + stddev (max: 100, forever: 0)"),
1537 OPT_BOOLEAN('n', "null", &null_run,
1538 "null run - dont start any counters"),
1539 OPT_INCR('d', "detailed", &detailed_run,
1540 "detailed run - start a lot of events"),
1541 OPT_BOOLEAN('S', "sync", &sync_run,
1542 "call sync() before starting a run"),
1543 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1544 "print large numbers with thousands\' separators",
1545 stat__set_big_num),
1546 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1547 "list of cpus to monitor in system-wide"),
1548 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1549 "disable CPU count aggregation", AGGR_NONE),
1550 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1551 "print counts with custom separator"),
1552 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1553 "monitor event in cgroup name only", parse_cgroups),
1554 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1555 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1556 OPT_INTEGER(0, "log-fd", &output_fd,
1557 "log output to fd, instead of stderr"),
1558 OPT_STRING(0, "pre", &pre_cmd, "command",
1559 "command to run prior to the measured command"),
1560 OPT_STRING(0, "post", &post_cmd, "command",
1561 "command to run after to the measured command"),
1562 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1563 "print counts at regular interval in ms (>= 10)"),
1564 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1565 "aggregate counts per processor socket", AGGR_SOCKET),
1566 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1567 "aggregate counts per physical processor core", AGGR_CORE),
1568 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1569 "aggregate counts per thread", AGGR_THREAD),
1570 OPT_UINTEGER('D', "delay", &initial_delay,
1571 "ms to wait before starting measurement after program start"),
1572 OPT_BOOLEAN(0, "metric-only", &metric_only,
1573 "Only print computed metrics. No raw values"),
1574 OPT_END()
1575};
1576
1577static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1578{
1579 return cpu_map__get_socket(map, cpu, NULL);
1580}
1581
1582static int perf_stat__get_core(struct cpu_map *map, int cpu)
1583{
1584 return cpu_map__get_core(map, cpu, NULL);
1585}
1586
1587static int cpu_map__get_max(struct cpu_map *map)
1588{
1589 int i, max = -1;
1590
1591 for (i = 0; i < map->nr; i++) {
1592 if (map->map[i] > max)
1593 max = map->map[i];
1594 }
1595
1596 return max;
1597}
1598
1599static struct cpu_map *cpus_aggr_map;
1600
1601static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1602{
1603 int cpu;
1604
1605 if (idx >= map->nr)
1606 return -1;
1607
1608 cpu = map->map[idx];
1609
1610 if (cpus_aggr_map->map[cpu] == -1)
1611 cpus_aggr_map->map[cpu] = get_id(map, idx);
1612
1613 return cpus_aggr_map->map[cpu];
1614}
1615
1616static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1617{
1618 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1619}
1620
1621static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1622{
1623 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1624}
1625
1626static int perf_stat_init_aggr_mode(void)
1627{
1628 int nr;
1629
1630 switch (stat_config.aggr_mode) {
1631 case AGGR_SOCKET:
1632 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1633 perror("cannot build socket map");
1634 return -1;
1635 }
1636 aggr_get_id = perf_stat__get_socket_cached;
1637 break;
1638 case AGGR_CORE:
1639 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1640 perror("cannot build core map");
1641 return -1;
1642 }
1643 aggr_get_id = perf_stat__get_core_cached;
1644 break;
1645 case AGGR_NONE:
1646 case AGGR_GLOBAL:
1647 case AGGR_THREAD:
1648 case AGGR_UNSET:
1649 default:
1650 break;
1651 }
1652
1653 /*
1654 * The evsel_list->cpus is the base we operate on,
1655 * taking the highest cpu number to be the size of
1656 * the aggregation translate cpumap.
1657 */
1658 nr = cpu_map__get_max(evsel_list->cpus);
1659 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1660 return cpus_aggr_map ? 0 : -ENOMEM;
1661}
1662
1663static void perf_stat__exit_aggr_mode(void)
1664{
1665 cpu_map__put(aggr_map);
1666 cpu_map__put(cpus_aggr_map);
1667 aggr_map = NULL;
1668 cpus_aggr_map = NULL;
1669}
1670
1671static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1672{
1673 int cpu;
1674
1675 if (idx > map->nr)
1676 return -1;
1677
1678 cpu = map->map[idx];
1679
1680 if (cpu >= env->nr_cpus_online)
1681 return -1;
1682
1683 return cpu;
1684}
1685
1686static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1687{
1688 struct perf_env *env = data;
1689 int cpu = perf_env__get_cpu(env, map, idx);
1690
1691 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1692}
1693
1694static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1695{
1696 struct perf_env *env = data;
1697 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1698
1699 if (cpu != -1) {
1700 int socket_id = env->cpu[cpu].socket_id;
1701
1702 /*
1703 * Encode socket in upper 16 bits
1704 * core_id is relative to socket, and
1705 * we need a global id. So we combine
1706 * socket + core id.
1707 */
1708 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1709 }
1710
1711 return core;
1712}
1713
1714static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1715 struct cpu_map **sockp)
1716{
1717 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1718}
1719
1720static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1721 struct cpu_map **corep)
1722{
1723 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1724}
1725
1726static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1727{
1728 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1729}
1730
1731static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1732{
1733 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1734}
1735
1736static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1737{
1738 struct perf_env *env = &st->session->header.env;
1739
1740 switch (stat_config.aggr_mode) {
1741 case AGGR_SOCKET:
1742 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1743 perror("cannot build socket map");
1744 return -1;
1745 }
1746 aggr_get_id = perf_stat__get_socket_file;
1747 break;
1748 case AGGR_CORE:
1749 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1750 perror("cannot build core map");
1751 return -1;
1752 }
1753 aggr_get_id = perf_stat__get_core_file;
1754 break;
1755 case AGGR_NONE:
1756 case AGGR_GLOBAL:
1757 case AGGR_THREAD:
1758 case AGGR_UNSET:
1759 default:
1760 break;
1761 }
1762
1763 return 0;
1764}
1765
1766/*
1767 * Add default attributes, if there were no attributes specified or
1768 * if -d/--detailed, -d -d or -d -d -d is used:
1769 */
1770static int add_default_attributes(void)
1771{
1772 struct perf_event_attr default_attrs0[] = {
1773
1774 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1775 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1776 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1777 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1778
1779 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1780};
1781 struct perf_event_attr frontend_attrs[] = {
1782 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1783};
1784 struct perf_event_attr backend_attrs[] = {
1785 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1786};
1787 struct perf_event_attr default_attrs1[] = {
1788 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1789 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1790 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1791
1792};
1793
1794/*
1795 * Detailed stats (-d), covering the L1 and last level data caches:
1796 */
1797 struct perf_event_attr detailed_attrs[] = {
1798
1799 { .type = PERF_TYPE_HW_CACHE,
1800 .config =
1801 PERF_COUNT_HW_CACHE_L1D << 0 |
1802 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1803 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1804
1805 { .type = PERF_TYPE_HW_CACHE,
1806 .config =
1807 PERF_COUNT_HW_CACHE_L1D << 0 |
1808 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1809 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1810
1811 { .type = PERF_TYPE_HW_CACHE,
1812 .config =
1813 PERF_COUNT_HW_CACHE_LL << 0 |
1814 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1815 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1816
1817 { .type = PERF_TYPE_HW_CACHE,
1818 .config =
1819 PERF_COUNT_HW_CACHE_LL << 0 |
1820 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1821 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1822};
1823
1824/*
1825 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1826 */
1827 struct perf_event_attr very_detailed_attrs[] = {
1828
1829 { .type = PERF_TYPE_HW_CACHE,
1830 .config =
1831 PERF_COUNT_HW_CACHE_L1I << 0 |
1832 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1833 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1834
1835 { .type = PERF_TYPE_HW_CACHE,
1836 .config =
1837 PERF_COUNT_HW_CACHE_L1I << 0 |
1838 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1839 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1840
1841 { .type = PERF_TYPE_HW_CACHE,
1842 .config =
1843 PERF_COUNT_HW_CACHE_DTLB << 0 |
1844 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1845 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1846
1847 { .type = PERF_TYPE_HW_CACHE,
1848 .config =
1849 PERF_COUNT_HW_CACHE_DTLB << 0 |
1850 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1851 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1852
1853 { .type = PERF_TYPE_HW_CACHE,
1854 .config =
1855 PERF_COUNT_HW_CACHE_ITLB << 0 |
1856 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1857 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1858
1859 { .type = PERF_TYPE_HW_CACHE,
1860 .config =
1861 PERF_COUNT_HW_CACHE_ITLB << 0 |
1862 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1863 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1864
1865};
1866
1867/*
1868 * Very, very detailed stats (-d -d -d), adding prefetch events:
1869 */
1870 struct perf_event_attr very_very_detailed_attrs[] = {
1871
1872 { .type = PERF_TYPE_HW_CACHE,
1873 .config =
1874 PERF_COUNT_HW_CACHE_L1D << 0 |
1875 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1876 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1877
1878 { .type = PERF_TYPE_HW_CACHE,
1879 .config =
1880 PERF_COUNT_HW_CACHE_L1D << 0 |
1881 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1882 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1883};
1884
1885 /* Set attrs if no event is selected and !null_run: */
1886 if (null_run)
1887 return 0;
1888
1889 if (transaction_run) {
1890 int err;
1891 if (pmu_have_event("cpu", "cycles-ct") &&
1892 pmu_have_event("cpu", "el-start"))
1893 err = parse_events(evsel_list, transaction_attrs, NULL);
1894 else
1895 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1896 if (err) {
1897 fprintf(stderr, "Cannot set up transaction events\n");
1898 return -1;
1899 }
1900 return 0;
1901 }
1902
1903 if (!evsel_list->nr_entries) {
1904 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1905 return -1;
1906 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1907 if (perf_evlist__add_default_attrs(evsel_list,
1908 frontend_attrs) < 0)
1909 return -1;
1910 }
1911 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1912 if (perf_evlist__add_default_attrs(evsel_list,
1913 backend_attrs) < 0)
1914 return -1;
1915 }
1916 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1917 return -1;
1918 }
1919
1920 /* Detailed events get appended to the event list: */
1921
1922 if (detailed_run < 1)
1923 return 0;
1924
1925 /* Append detailed run extra attributes: */
1926 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1927 return -1;
1928
1929 if (detailed_run < 2)
1930 return 0;
1931
1932 /* Append very detailed run extra attributes: */
1933 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1934 return -1;
1935
1936 if (detailed_run < 3)
1937 return 0;
1938
1939 /* Append very, very detailed run extra attributes: */
1940 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1941}
1942
1943static const char * const stat_record_usage[] = {
1944 "perf stat record [<options>]",
1945 NULL,
1946};
1947
1948static void init_features(struct perf_session *session)
1949{
1950 int feat;
1951
1952 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1953 perf_header__set_feat(&session->header, feat);
1954
1955 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1956 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1957 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1958 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1959}
1960
1961static int __cmd_record(int argc, const char **argv)
1962{
1963 struct perf_session *session;
1964 struct perf_data_file *file = &perf_stat.file;
1965
1966 argc = parse_options(argc, argv, stat_options, stat_record_usage,
1967 PARSE_OPT_STOP_AT_NON_OPTION);
1968
1969 if (output_name)
1970 file->path = output_name;
1971
1972 if (run_count != 1 || forever) {
1973 pr_err("Cannot use -r option with perf stat record.\n");
1974 return -1;
1975 }
1976
1977 session = perf_session__new(file, false, NULL);
1978 if (session == NULL) {
1979 pr_err("Perf session creation failed.\n");
1980 return -1;
1981 }
1982
1983 init_features(session);
1984
1985 session->evlist = evsel_list;
1986 perf_stat.session = session;
1987 perf_stat.record = true;
1988 return argc;
1989}
1990
1991static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
1992 union perf_event *event,
1993 struct perf_session *session)
1994{
1995 struct stat_round_event *round = &event->stat_round;
1996 struct perf_evsel *counter;
1997 struct timespec tsh, *ts = NULL;
1998 const char **argv = session->header.env.cmdline_argv;
1999 int argc = session->header.env.nr_cmdline;
2000
2001 evlist__for_each(evsel_list, counter)
2002 perf_stat_process_counter(&stat_config, counter);
2003
2004 if (round->type == PERF_STAT_ROUND_TYPE__FINAL)
2005 update_stats(&walltime_nsecs_stats, round->time);
2006
2007 if (stat_config.interval && round->time) {
2008 tsh.tv_sec = round->time / NSECS_PER_SEC;
2009 tsh.tv_nsec = round->time % NSECS_PER_SEC;
2010 ts = &tsh;
2011 }
2012
2013 print_counters(ts, argc, argv);
2014 return 0;
2015}
2016
2017static
2018int process_stat_config_event(struct perf_tool *tool __maybe_unused,
2019 union perf_event *event,
2020 struct perf_session *session __maybe_unused)
2021{
2022 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2023
2024 perf_event__read_stat_config(&stat_config, &event->stat_config);
2025
2026 if (cpu_map__empty(st->cpus)) {
2027 if (st->aggr_mode != AGGR_UNSET)
2028 pr_warning("warning: processing task data, aggregation mode not set\n");
2029 return 0;
2030 }
2031
2032 if (st->aggr_mode != AGGR_UNSET)
2033 stat_config.aggr_mode = st->aggr_mode;
2034
2035 if (perf_stat.file.is_pipe)
2036 perf_stat_init_aggr_mode();
2037 else
2038 perf_stat_init_aggr_mode_file(st);
2039
2040 return 0;
2041}
2042
2043static int set_maps(struct perf_stat *st)
2044{
2045 if (!st->cpus || !st->threads)
2046 return 0;
2047
2048 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2049 return -EINVAL;
2050
2051 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2052
2053 if (perf_evlist__alloc_stats(evsel_list, true))
2054 return -ENOMEM;
2055
2056 st->maps_allocated = true;
2057 return 0;
2058}
2059
2060static
2061int process_thread_map_event(struct perf_tool *tool __maybe_unused,
2062 union perf_event *event,
2063 struct perf_session *session __maybe_unused)
2064{
2065 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2066
2067 if (st->threads) {
2068 pr_warning("Extra thread map event, ignoring.\n");
2069 return 0;
2070 }
2071
2072 st->threads = thread_map__new_event(&event->thread_map);
2073 if (!st->threads)
2074 return -ENOMEM;
2075
2076 return set_maps(st);
2077}
2078
2079static
2080int process_cpu_map_event(struct perf_tool *tool __maybe_unused,
2081 union perf_event *event,
2082 struct perf_session *session __maybe_unused)
2083{
2084 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2085 struct cpu_map *cpus;
2086
2087 if (st->cpus) {
2088 pr_warning("Extra cpu map event, ignoring.\n");
2089 return 0;
2090 }
2091
2092 cpus = cpu_map__new_data(&event->cpu_map.data);
2093 if (!cpus)
2094 return -ENOMEM;
2095
2096 st->cpus = cpus;
2097 return set_maps(st);
2098}
2099
2100static const char * const stat_report_usage[] = {
2101 "perf stat report [<options>]",
2102 NULL,
2103};
2104
2105static struct perf_stat perf_stat = {
2106 .tool = {
2107 .attr = perf_event__process_attr,
2108 .event_update = perf_event__process_event_update,
2109 .thread_map = process_thread_map_event,
2110 .cpu_map = process_cpu_map_event,
2111 .stat_config = process_stat_config_event,
2112 .stat = perf_event__process_stat_event,
2113 .stat_round = process_stat_round_event,
2114 },
2115 .aggr_mode = AGGR_UNSET,
2116};
2117
2118static int __cmd_report(int argc, const char **argv)
2119{
2120 struct perf_session *session;
2121 const struct option options[] = {
2122 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2123 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2124 "aggregate counts per processor socket", AGGR_SOCKET),
2125 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2126 "aggregate counts per physical processor core", AGGR_CORE),
2127 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2128 "disable CPU count aggregation", AGGR_NONE),
2129 OPT_END()
2130 };
2131 struct stat st;
2132 int ret;
2133
2134 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2135
2136 if (!input_name || !strlen(input_name)) {
2137 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2138 input_name = "-";
2139 else
2140 input_name = "perf.data";
2141 }
2142
2143 perf_stat.file.path = input_name;
2144 perf_stat.file.mode = PERF_DATA_MODE_READ;
2145
2146 session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
2147 if (session == NULL)
2148 return -1;
2149
2150 perf_stat.session = session;
2151 stat_config.output = stderr;
2152 evsel_list = session->evlist;
2153
2154 ret = perf_session__process_events(session);
2155 if (ret)
2156 return ret;
2157
2158 perf_session__delete(session);
2159 return 0;
2160}
2161
2162int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
2163{
2164 const char * const stat_usage[] = {
2165 "perf stat [<options>] [<command>]",
2166 NULL
2167 };
2168 int status = -EINVAL, run_idx;
2169 const char *mode;
2170 FILE *output = stderr;
2171 unsigned int interval;
2172 const char * const stat_subcommands[] = { "record", "report" };
2173
2174 setlocale(LC_ALL, "");
2175
2176 evsel_list = perf_evlist__new();
2177 if (evsel_list == NULL)
2178 return -ENOMEM;
2179
2180 parse_events__shrink_config_terms();
2181 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2182 (const char **) stat_usage,
2183 PARSE_OPT_STOP_AT_NON_OPTION);
2184 perf_stat__init_shadow_stats();
2185
2186 if (csv_sep) {
2187 csv_output = true;
2188 if (!strcmp(csv_sep, "\\t"))
2189 csv_sep = "\t";
2190 } else
2191 csv_sep = DEFAULT_SEPARATOR;
2192
2193 if (argc && !strncmp(argv[0], "rec", 3)) {
2194 argc = __cmd_record(argc, argv);
2195 if (argc < 0)
2196 return -1;
2197 } else if (argc && !strncmp(argv[0], "rep", 3))
2198 return __cmd_report(argc, argv);
2199
2200 interval = stat_config.interval;
2201
2202 /*
2203 * For record command the -o is already taken care of.
2204 */
2205 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2206 output = NULL;
2207
2208 if (output_name && output_fd) {
2209 fprintf(stderr, "cannot use both --output and --log-fd\n");
2210 parse_options_usage(stat_usage, stat_options, "o", 1);
2211 parse_options_usage(NULL, stat_options, "log-fd", 0);
2212 goto out;
2213 }
2214
2215 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2216 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2217 goto out;
2218 }
2219
2220 if (metric_only && run_count > 1) {
2221 fprintf(stderr, "--metric-only is not supported with -r\n");
2222 goto out;
2223 }
2224
2225 if (output_fd < 0) {
2226 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2227 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2228 goto out;
2229 }
2230
2231 if (!output) {
2232 struct timespec tm;
2233 mode = append_file ? "a" : "w";
2234
2235 output = fopen(output_name, mode);
2236 if (!output) {
2237 perror("failed to create output file");
2238 return -1;
2239 }
2240 clock_gettime(CLOCK_REALTIME, &tm);
2241 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2242 } else if (output_fd > 0) {
2243 mode = append_file ? "a" : "w";
2244 output = fdopen(output_fd, mode);
2245 if (!output) {
2246 perror("Failed opening logfd");
2247 return -errno;
2248 }
2249 }
2250
2251 stat_config.output = output;
2252
2253 /*
2254 * let the spreadsheet do the pretty-printing
2255 */
2256 if (csv_output) {
2257 /* User explicitly passed -B? */
2258 if (big_num_opt == 1) {
2259 fprintf(stderr, "-B option not supported with -x\n");
2260 parse_options_usage(stat_usage, stat_options, "B", 1);
2261 parse_options_usage(NULL, stat_options, "x", 1);
2262 goto out;
2263 } else /* Nope, so disable big number formatting */
2264 big_num = false;
2265 } else if (big_num_opt == 0) /* User passed --no-big-num */
2266 big_num = false;
2267
2268 if (!argc && target__none(&target))
2269 usage_with_options(stat_usage, stat_options);
2270
2271 if (run_count < 0) {
2272 pr_err("Run count must be a positive number\n");
2273 parse_options_usage(stat_usage, stat_options, "r", 1);
2274 goto out;
2275 } else if (run_count == 0) {
2276 forever = true;
2277 run_count = 1;
2278 }
2279
2280 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2281 fprintf(stderr, "The --per-thread option is only available "
2282 "when monitoring via -p -t options.\n");
2283 parse_options_usage(NULL, stat_options, "p", 1);
2284 parse_options_usage(NULL, stat_options, "t", 1);
2285 goto out;
2286 }
2287
2288 /*
2289 * no_aggr, cgroup are for system-wide only
2290 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2291 */
2292 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2293 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2294 !target__has_cpu(&target)) {
2295 fprintf(stderr, "both cgroup and no-aggregation "
2296 "modes only available in system-wide mode\n");
2297
2298 parse_options_usage(stat_usage, stat_options, "G", 1);
2299 parse_options_usage(NULL, stat_options, "A", 1);
2300 parse_options_usage(NULL, stat_options, "a", 1);
2301 goto out;
2302 }
2303
2304 if (add_default_attributes())
2305 goto out;
2306
2307 target__validate(&target);
2308
2309 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2310 if (target__has_task(&target)) {
2311 pr_err("Problems finding threads of monitor\n");
2312 parse_options_usage(stat_usage, stat_options, "p", 1);
2313 parse_options_usage(NULL, stat_options, "t", 1);
2314 } else if (target__has_cpu(&target)) {
2315 perror("failed to parse CPUs map");
2316 parse_options_usage(stat_usage, stat_options, "C", 1);
2317 parse_options_usage(NULL, stat_options, "a", 1);
2318 }
2319 goto out;
2320 }
2321
2322 /*
2323 * Initialize thread_map with comm names,
2324 * so we could print it out on output.
2325 */
2326 if (stat_config.aggr_mode == AGGR_THREAD)
2327 thread_map__read_comms(evsel_list->threads);
2328
2329 if (interval && interval < 100) {
2330 if (interval < 10) {
2331 pr_err("print interval must be >= 10ms\n");
2332 parse_options_usage(stat_usage, stat_options, "I", 1);
2333 goto out;
2334 } else
2335 pr_warning("print interval < 100ms. "
2336 "The overhead percentage could be high in some cases. "
2337 "Please proceed with caution.\n");
2338 }
2339
2340 if (perf_evlist__alloc_stats(evsel_list, interval))
2341 goto out;
2342
2343 if (perf_stat_init_aggr_mode())
2344 goto out;
2345
2346 /*
2347 * We dont want to block the signals - that would cause
2348 * child tasks to inherit that and Ctrl-C would not work.
2349 * What we want is for Ctrl-C to work in the exec()-ed
2350 * task, but being ignored by perf stat itself:
2351 */
2352 atexit(sig_atexit);
2353 if (!forever)
2354 signal(SIGINT, skip_signal);
2355 signal(SIGCHLD, skip_signal);
2356 signal(SIGALRM, skip_signal);
2357 signal(SIGABRT, skip_signal);
2358
2359 status = 0;
2360 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2361 if (run_count != 1 && verbose)
2362 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2363 run_idx + 1);
2364
2365 status = run_perf_stat(argc, argv);
2366 if (forever && status != -1) {
2367 print_counters(NULL, argc, argv);
2368 perf_stat__reset_stats();
2369 }
2370 }
2371
2372 if (!forever && status != -1 && !interval)
2373 print_counters(NULL, argc, argv);
2374
2375 if (STAT_RECORD) {
2376 /*
2377 * We synthesize the kernel mmap record just so that older tools
2378 * don't emit warnings about not being able to resolve symbols
2379 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2380 * a saner message about no samples being in the perf.data file.
2381 *
2382 * This also serves to suppress a warning about f_header.data.size == 0
2383 * in header.c at the moment 'perf stat record' gets introduced, which
2384 * is not really needed once we start adding the stat specific PERF_RECORD_
2385 * records, but the need to suppress the kptr_restrict messages in older
2386 * tools remain -acme
2387 */
2388 int fd = perf_data_file__fd(&perf_stat.file);
2389 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2390 process_synthesized_event,
2391 &perf_stat.session->machines.host);
2392 if (err) {
2393 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2394 "older tools may produce warnings about this file\n.");
2395 }
2396
2397 if (!interval) {
2398 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2399 pr_err("failed to write stat round event\n");
2400 }
2401
2402 if (!perf_stat.file.is_pipe) {
2403 perf_stat.session->header.data_size += perf_stat.bytes_written;
2404 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2405 }
2406
2407 perf_session__delete(perf_stat.session);
2408 }
2409
2410 perf_stat__exit_aggr_mode();
2411 perf_evlist__free_stats(evsel_list);
2412out:
2413 perf_evlist__delete(evsel_list);
2414 return status;
2415}