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/util.h"
47#include "util/parse-options.h"
48#include "util/parse-events.h"
49#include "util/pmu.h"
50#include "util/event.h"
51#include "util/evlist.h"
52#include "util/evsel.h"
53#include "util/debug.h"
54#include "util/color.h"
55#include "util/stat.h"
56#include "util/header.h"
57#include "util/cpumap.h"
58#include "util/thread.h"
59#include "util/thread_map.h"
60
61#include <stdlib.h>
62#include <sys/prctl.h>
63#include <locale.h>
64
65#define DEFAULT_SEPARATOR " "
66#define CNTR_NOT_SUPPORTED "<not supported>"
67#define CNTR_NOT_COUNTED "<not counted>"
68
69static void print_stat(int argc, const char **argv);
70static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
71static void print_counter(struct perf_evsel *counter, char *prefix);
72static void print_aggr(char *prefix);
73
74/* Default events used for perf stat -T */
75static const char * const transaction_attrs[] = {
76 "task-clock",
77 "{"
78 "instructions,"
79 "cycles,"
80 "cpu/cycles-t/,"
81 "cpu/tx-start/,"
82 "cpu/el-start/,"
83 "cpu/cycles-ct/"
84 "}"
85};
86
87/* More limited version when the CPU does not have all events. */
88static const char * const transaction_limited_attrs[] = {
89 "task-clock",
90 "{"
91 "instructions,"
92 "cycles,"
93 "cpu/cycles-t/,"
94 "cpu/tx-start/"
95 "}"
96};
97
98/* must match transaction_attrs and the beginning limited_attrs */
99enum {
100 T_TASK_CLOCK,
101 T_INSTRUCTIONS,
102 T_CYCLES,
103 T_CYCLES_IN_TX,
104 T_TRANSACTION_START,
105 T_ELISION_START,
106 T_CYCLES_IN_TX_CP,
107};
108
109static struct perf_evlist *evsel_list;
110
111static struct target target = {
112 .uid = UINT_MAX,
113};
114
115enum aggr_mode {
116 AGGR_NONE,
117 AGGR_GLOBAL,
118 AGGR_SOCKET,
119 AGGR_CORE,
120};
121
122static int run_count = 1;
123static bool no_inherit = false;
124static bool scale = true;
125static enum aggr_mode aggr_mode = AGGR_GLOBAL;
126static volatile pid_t child_pid = -1;
127static bool null_run = false;
128static int detailed_run = 0;
129static bool transaction_run;
130static bool big_num = true;
131static int big_num_opt = -1;
132static const char *csv_sep = NULL;
133static bool csv_output = false;
134static bool group = false;
135static FILE *output = NULL;
136static const char *pre_cmd = NULL;
137static const char *post_cmd = NULL;
138static bool sync_run = false;
139static unsigned int interval = 0;
140static unsigned int initial_delay = 0;
141static unsigned int unit_width = 4; /* strlen("unit") */
142static bool forever = false;
143static struct timespec ref_time;
144static struct cpu_map *aggr_map;
145static int (*aggr_get_id)(struct cpu_map *m, int cpu);
146
147static volatile int done = 0;
148
149struct perf_stat {
150 struct stats res_stats[3];
151};
152
153static inline void diff_timespec(struct timespec *r, struct timespec *a,
154 struct timespec *b)
155{
156 r->tv_sec = a->tv_sec - b->tv_sec;
157 if (a->tv_nsec < b->tv_nsec) {
158 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
159 r->tv_sec--;
160 } else {
161 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
162 }
163}
164
165static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
166{
167 return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
168}
169
170static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
171{
172 return perf_evsel__cpus(evsel)->nr;
173}
174
175static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
176{
177 int i;
178 struct perf_stat *ps = evsel->priv;
179
180 for (i = 0; i < 3; i++)
181 init_stats(&ps->res_stats[i]);
182}
183
184static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
185{
186 evsel->priv = zalloc(sizeof(struct perf_stat));
187 if (evsel == NULL)
188 return -ENOMEM;
189 perf_evsel__reset_stat_priv(evsel);
190 return 0;
191}
192
193static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
194{
195 zfree(&evsel->priv);
196}
197
198static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
199{
200 void *addr;
201 size_t sz;
202
203 sz = sizeof(*evsel->counts) +
204 (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
205
206 addr = zalloc(sz);
207 if (!addr)
208 return -ENOMEM;
209
210 evsel->prev_raw_counts = addr;
211
212 return 0;
213}
214
215static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
216{
217 zfree(&evsel->prev_raw_counts);
218}
219
220static void perf_evlist__free_stats(struct perf_evlist *evlist)
221{
222 struct perf_evsel *evsel;
223
224 evlist__for_each(evlist, evsel) {
225 perf_evsel__free_stat_priv(evsel);
226 perf_evsel__free_counts(evsel);
227 perf_evsel__free_prev_raw_counts(evsel);
228 }
229}
230
231static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
232{
233 struct perf_evsel *evsel;
234
235 evlist__for_each(evlist, evsel) {
236 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
237 perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
238 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
239 goto out_free;
240 }
241
242 return 0;
243
244out_free:
245 perf_evlist__free_stats(evlist);
246 return -1;
247}
248
249static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
250static struct stats runtime_cycles_stats[MAX_NR_CPUS];
251static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
252static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
253static struct stats runtime_branches_stats[MAX_NR_CPUS];
254static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
255static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
256static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
257static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
258static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
259static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
260static struct stats runtime_cycles_in_tx_stats[MAX_NR_CPUS];
261static struct stats walltime_nsecs_stats;
262static struct stats runtime_transaction_stats[MAX_NR_CPUS];
263static struct stats runtime_elision_stats[MAX_NR_CPUS];
264
265static void perf_stat__reset_stats(struct perf_evlist *evlist)
266{
267 struct perf_evsel *evsel;
268
269 evlist__for_each(evlist, evsel) {
270 perf_evsel__reset_stat_priv(evsel);
271 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
272 }
273
274 memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
275 memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
276 memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
277 memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
278 memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
279 memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
280 memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
281 memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
282 memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
283 memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
284 memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
285 memset(runtime_cycles_in_tx_stats, 0,
286 sizeof(runtime_cycles_in_tx_stats));
287 memset(runtime_transaction_stats, 0,
288 sizeof(runtime_transaction_stats));
289 memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
290 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
291}
292
293static int create_perf_stat_counter(struct perf_evsel *evsel)
294{
295 struct perf_event_attr *attr = &evsel->attr;
296
297 if (scale)
298 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
299 PERF_FORMAT_TOTAL_TIME_RUNNING;
300
301 attr->inherit = !no_inherit;
302
303 if (target__has_cpu(&target))
304 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
305
306 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
307 attr->disabled = 1;
308 if (!initial_delay)
309 attr->enable_on_exec = 1;
310 }
311
312 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
313}
314
315/*
316 * Does the counter have nsecs as a unit?
317 */
318static inline int nsec_counter(struct perf_evsel *evsel)
319{
320 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
321 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
322 return 1;
323
324 return 0;
325}
326
327static struct perf_evsel *nth_evsel(int n)
328{
329 static struct perf_evsel **array;
330 static int array_len;
331 struct perf_evsel *ev;
332 int j;
333
334 /* Assumes this only called when evsel_list does not change anymore. */
335 if (!array) {
336 evlist__for_each(evsel_list, ev)
337 array_len++;
338 array = malloc(array_len * sizeof(void *));
339 if (!array)
340 exit(ENOMEM);
341 j = 0;
342 evlist__for_each(evsel_list, ev)
343 array[j++] = ev;
344 }
345 if (n < array_len)
346 return array[n];
347 return NULL;
348}
349
350/*
351 * Update various tracking values we maintain to print
352 * more semantic information such as miss/hit ratios,
353 * instruction rates, etc:
354 */
355static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
356{
357 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
358 update_stats(&runtime_nsecs_stats[0], count[0]);
359 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
360 update_stats(&runtime_cycles_stats[0], count[0]);
361 else if (transaction_run &&
362 perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX)))
363 update_stats(&runtime_cycles_in_tx_stats[0], count[0]);
364 else if (transaction_run &&
365 perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START)))
366 update_stats(&runtime_transaction_stats[0], count[0]);
367 else if (transaction_run &&
368 perf_evsel__cmp(counter, nth_evsel(T_ELISION_START)))
369 update_stats(&runtime_elision_stats[0], count[0]);
370 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
371 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
372 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
373 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
374 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
375 update_stats(&runtime_branches_stats[0], count[0]);
376 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
377 update_stats(&runtime_cacherefs_stats[0], count[0]);
378 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
379 update_stats(&runtime_l1_dcache_stats[0], count[0]);
380 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
381 update_stats(&runtime_l1_icache_stats[0], count[0]);
382 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
383 update_stats(&runtime_ll_cache_stats[0], count[0]);
384 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
385 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
386 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
387 update_stats(&runtime_itlb_cache_stats[0], count[0]);
388}
389
390/*
391 * Read out the results of a single counter:
392 * aggregate counts across CPUs in system-wide mode
393 */
394static int read_counter_aggr(struct perf_evsel *counter)
395{
396 struct perf_stat *ps = counter->priv;
397 u64 *count = counter->counts->aggr.values;
398 int i;
399
400 if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
401 thread_map__nr(evsel_list->threads), scale) < 0)
402 return -1;
403
404 for (i = 0; i < 3; i++)
405 update_stats(&ps->res_stats[i], count[i]);
406
407 if (verbose) {
408 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
409 perf_evsel__name(counter), count[0], count[1], count[2]);
410 }
411
412 /*
413 * Save the full runtime - to allow normalization during printout:
414 */
415 update_shadow_stats(counter, count);
416
417 return 0;
418}
419
420/*
421 * Read out the results of a single counter:
422 * do not aggregate counts across CPUs in system-wide mode
423 */
424static int read_counter(struct perf_evsel *counter)
425{
426 u64 *count;
427 int cpu;
428
429 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
430 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
431 return -1;
432
433 count = counter->counts->cpu[cpu].values;
434
435 update_shadow_stats(counter, count);
436 }
437
438 return 0;
439}
440
441static void print_interval(void)
442{
443 static int num_print_interval;
444 struct perf_evsel *counter;
445 struct perf_stat *ps;
446 struct timespec ts, rs;
447 char prefix[64];
448
449 if (aggr_mode == AGGR_GLOBAL) {
450 evlist__for_each(evsel_list, counter) {
451 ps = counter->priv;
452 memset(ps->res_stats, 0, sizeof(ps->res_stats));
453 read_counter_aggr(counter);
454 }
455 } else {
456 evlist__for_each(evsel_list, counter) {
457 ps = counter->priv;
458 memset(ps->res_stats, 0, sizeof(ps->res_stats));
459 read_counter(counter);
460 }
461 }
462
463 clock_gettime(CLOCK_MONOTONIC, &ts);
464 diff_timespec(&rs, &ts, &ref_time);
465 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
466
467 if (num_print_interval == 0 && !csv_output) {
468 switch (aggr_mode) {
469 case AGGR_SOCKET:
470 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
471 break;
472 case AGGR_CORE:
473 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
474 break;
475 case AGGR_NONE:
476 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
477 break;
478 case AGGR_GLOBAL:
479 default:
480 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
481 }
482 }
483
484 if (++num_print_interval == 25)
485 num_print_interval = 0;
486
487 switch (aggr_mode) {
488 case AGGR_CORE:
489 case AGGR_SOCKET:
490 print_aggr(prefix);
491 break;
492 case AGGR_NONE:
493 evlist__for_each(evsel_list, counter)
494 print_counter(counter, prefix);
495 break;
496 case AGGR_GLOBAL:
497 default:
498 evlist__for_each(evsel_list, counter)
499 print_counter_aggr(counter, prefix);
500 }
501
502 fflush(output);
503}
504
505static void handle_initial_delay(void)
506{
507 struct perf_evsel *counter;
508
509 if (initial_delay) {
510 const int ncpus = cpu_map__nr(evsel_list->cpus),
511 nthreads = thread_map__nr(evsel_list->threads);
512
513 usleep(initial_delay * 1000);
514 evlist__for_each(evsel_list, counter)
515 perf_evsel__enable(counter, ncpus, nthreads);
516 }
517}
518
519static volatile int workload_exec_errno;
520
521/*
522 * perf_evlist__prepare_workload will send a SIGUSR1
523 * if the fork fails, since we asked by setting its
524 * want_signal to true.
525 */
526static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
527 void *ucontext __maybe_unused)
528{
529 workload_exec_errno = info->si_value.sival_int;
530}
531
532static int __run_perf_stat(int argc, const char **argv)
533{
534 char msg[512];
535 unsigned long long t0, t1;
536 struct perf_evsel *counter;
537 struct timespec ts;
538 size_t l;
539 int status = 0;
540 const bool forks = (argc > 0);
541
542 if (interval) {
543 ts.tv_sec = interval / 1000;
544 ts.tv_nsec = (interval % 1000) * 1000000;
545 } else {
546 ts.tv_sec = 1;
547 ts.tv_nsec = 0;
548 }
549
550 if (forks) {
551 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
552 workload_exec_failed_signal) < 0) {
553 perror("failed to prepare workload");
554 return -1;
555 }
556 child_pid = evsel_list->workload.pid;
557 }
558
559 if (group)
560 perf_evlist__set_leader(evsel_list);
561
562 evlist__for_each(evsel_list, counter) {
563 if (create_perf_stat_counter(counter) < 0) {
564 /*
565 * PPC returns ENXIO for HW counters until 2.6.37
566 * (behavior changed with commit b0a873e).
567 */
568 if (errno == EINVAL || errno == ENOSYS ||
569 errno == ENOENT || errno == EOPNOTSUPP ||
570 errno == ENXIO) {
571 if (verbose)
572 ui__warning("%s event is not supported by the kernel.\n",
573 perf_evsel__name(counter));
574 counter->supported = false;
575 continue;
576 }
577
578 perf_evsel__open_strerror(counter, &target,
579 errno, msg, sizeof(msg));
580 ui__error("%s\n", msg);
581
582 if (child_pid != -1)
583 kill(child_pid, SIGTERM);
584
585 return -1;
586 }
587 counter->supported = true;
588
589 l = strlen(counter->unit);
590 if (l > unit_width)
591 unit_width = l;
592 }
593
594 if (perf_evlist__apply_filters(evsel_list)) {
595 error("failed to set filter with %d (%s)\n", errno,
596 strerror(errno));
597 return -1;
598 }
599
600 /*
601 * Enable counters and exec the command:
602 */
603 t0 = rdclock();
604 clock_gettime(CLOCK_MONOTONIC, &ref_time);
605
606 if (forks) {
607 perf_evlist__start_workload(evsel_list);
608 handle_initial_delay();
609
610 if (interval) {
611 while (!waitpid(child_pid, &status, WNOHANG)) {
612 nanosleep(&ts, NULL);
613 print_interval();
614 }
615 }
616 wait(&status);
617
618 if (workload_exec_errno) {
619 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
620 pr_err("Workload failed: %s\n", emsg);
621 return -1;
622 }
623
624 if (WIFSIGNALED(status))
625 psignal(WTERMSIG(status), argv[0]);
626 } else {
627 handle_initial_delay();
628 while (!done) {
629 nanosleep(&ts, NULL);
630 if (interval)
631 print_interval();
632 }
633 }
634
635 t1 = rdclock();
636
637 update_stats(&walltime_nsecs_stats, t1 - t0);
638
639 if (aggr_mode == AGGR_GLOBAL) {
640 evlist__for_each(evsel_list, counter) {
641 read_counter_aggr(counter);
642 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
643 thread_map__nr(evsel_list->threads));
644 }
645 } else {
646 evlist__for_each(evsel_list, counter) {
647 read_counter(counter);
648 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
649 }
650 }
651
652 return WEXITSTATUS(status);
653}
654
655static int run_perf_stat(int argc, const char **argv)
656{
657 int ret;
658
659 if (pre_cmd) {
660 ret = system(pre_cmd);
661 if (ret)
662 return ret;
663 }
664
665 if (sync_run)
666 sync();
667
668 ret = __run_perf_stat(argc, argv);
669 if (ret)
670 return ret;
671
672 if (post_cmd) {
673 ret = system(post_cmd);
674 if (ret)
675 return ret;
676 }
677
678 return ret;
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(output, "%s%.2f%%", csv_sep, pct);
687 else if (pct)
688 fprintf(output, " ( +-%6.2f%% )", pct);
689}
690
691static void print_noise(struct perf_evsel *evsel, double avg)
692{
693 struct perf_stat *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 (aggr_mode) {
705 case AGGR_CORE:
706 fprintf(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(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(output, "CPU%*d%s",
726 csv_output ? 0 : -4,
727 perf_evsel__cpus(evsel)->map[id], csv_sep);
728 break;
729 case AGGR_GLOBAL:
730 default:
731 break;
732 }
733}
734
735static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
736{
737 double msecs = avg / 1e6;
738 const char *fmt_v, *fmt_n;
739 char name[25];
740
741 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
742 fmt_n = csv_output ? "%s" : "%-25s";
743
744 aggr_printout(evsel, cpu, nr);
745
746 scnprintf(name, sizeof(name), "%s%s",
747 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
748
749 fprintf(output, fmt_v, msecs, csv_sep);
750
751 if (csv_output)
752 fprintf(output, "%s%s", evsel->unit, csv_sep);
753 else
754 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
755
756 fprintf(output, fmt_n, name);
757
758 if (evsel->cgrp)
759 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
760
761 if (csv_output || interval)
762 return;
763
764 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
765 fprintf(output, " # %8.3f CPUs utilized ",
766 avg / avg_stats(&walltime_nsecs_stats));
767 else
768 fprintf(output, " ");
769}
770
771/* used for get_ratio_color() */
772enum grc_type {
773 GRC_STALLED_CYCLES_FE,
774 GRC_STALLED_CYCLES_BE,
775 GRC_CACHE_MISSES,
776 GRC_MAX_NR
777};
778
779static const char *get_ratio_color(enum grc_type type, double ratio)
780{
781 static const double grc_table[GRC_MAX_NR][3] = {
782 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
783 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
784 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
785 };
786 const char *color = PERF_COLOR_NORMAL;
787
788 if (ratio > grc_table[type][0])
789 color = PERF_COLOR_RED;
790 else if (ratio > grc_table[type][1])
791 color = PERF_COLOR_MAGENTA;
792 else if (ratio > grc_table[type][2])
793 color = PERF_COLOR_YELLOW;
794
795 return color;
796}
797
798static void print_stalled_cycles_frontend(int cpu,
799 struct perf_evsel *evsel
800 __maybe_unused, double avg)
801{
802 double total, ratio = 0.0;
803 const char *color;
804
805 total = avg_stats(&runtime_cycles_stats[cpu]);
806
807 if (total)
808 ratio = avg / total * 100.0;
809
810 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
811
812 fprintf(output, " # ");
813 color_fprintf(output, color, "%6.2f%%", ratio);
814 fprintf(output, " frontend cycles idle ");
815}
816
817static void print_stalled_cycles_backend(int cpu,
818 struct perf_evsel *evsel
819 __maybe_unused, double avg)
820{
821 double total, ratio = 0.0;
822 const char *color;
823
824 total = avg_stats(&runtime_cycles_stats[cpu]);
825
826 if (total)
827 ratio = avg / total * 100.0;
828
829 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
830
831 fprintf(output, " # ");
832 color_fprintf(output, color, "%6.2f%%", ratio);
833 fprintf(output, " backend cycles idle ");
834}
835
836static void print_branch_misses(int cpu,
837 struct perf_evsel *evsel __maybe_unused,
838 double avg)
839{
840 double total, ratio = 0.0;
841 const char *color;
842
843 total = avg_stats(&runtime_branches_stats[cpu]);
844
845 if (total)
846 ratio = avg / total * 100.0;
847
848 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
849
850 fprintf(output, " # ");
851 color_fprintf(output, color, "%6.2f%%", ratio);
852 fprintf(output, " of all branches ");
853}
854
855static void print_l1_dcache_misses(int cpu,
856 struct perf_evsel *evsel __maybe_unused,
857 double avg)
858{
859 double total, ratio = 0.0;
860 const char *color;
861
862 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
863
864 if (total)
865 ratio = avg / total * 100.0;
866
867 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
868
869 fprintf(output, " # ");
870 color_fprintf(output, color, "%6.2f%%", ratio);
871 fprintf(output, " of all L1-dcache hits ");
872}
873
874static void print_l1_icache_misses(int cpu,
875 struct perf_evsel *evsel __maybe_unused,
876 double avg)
877{
878 double total, ratio = 0.0;
879 const char *color;
880
881 total = avg_stats(&runtime_l1_icache_stats[cpu]);
882
883 if (total)
884 ratio = avg / total * 100.0;
885
886 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
887
888 fprintf(output, " # ");
889 color_fprintf(output, color, "%6.2f%%", ratio);
890 fprintf(output, " of all L1-icache hits ");
891}
892
893static void print_dtlb_cache_misses(int cpu,
894 struct perf_evsel *evsel __maybe_unused,
895 double avg)
896{
897 double total, ratio = 0.0;
898 const char *color;
899
900 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
901
902 if (total)
903 ratio = avg / total * 100.0;
904
905 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
906
907 fprintf(output, " # ");
908 color_fprintf(output, color, "%6.2f%%", ratio);
909 fprintf(output, " of all dTLB cache hits ");
910}
911
912static void print_itlb_cache_misses(int cpu,
913 struct perf_evsel *evsel __maybe_unused,
914 double avg)
915{
916 double total, ratio = 0.0;
917 const char *color;
918
919 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
920
921 if (total)
922 ratio = avg / total * 100.0;
923
924 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
925
926 fprintf(output, " # ");
927 color_fprintf(output, color, "%6.2f%%", ratio);
928 fprintf(output, " of all iTLB cache hits ");
929}
930
931static void print_ll_cache_misses(int cpu,
932 struct perf_evsel *evsel __maybe_unused,
933 double avg)
934{
935 double total, ratio = 0.0;
936 const char *color;
937
938 total = avg_stats(&runtime_ll_cache_stats[cpu]);
939
940 if (total)
941 ratio = avg / total * 100.0;
942
943 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
944
945 fprintf(output, " # ");
946 color_fprintf(output, color, "%6.2f%%", ratio);
947 fprintf(output, " of all LL-cache hits ");
948}
949
950static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
951{
952 double total, ratio = 0.0, total2;
953 double sc = evsel->scale;
954 const char *fmt;
955
956 if (csv_output) {
957 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
958 } else {
959 if (big_num)
960 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
961 else
962 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
963 }
964
965 aggr_printout(evsel, cpu, nr);
966
967 if (aggr_mode == AGGR_GLOBAL)
968 cpu = 0;
969
970 fprintf(output, fmt, avg, csv_sep);
971
972 if (evsel->unit)
973 fprintf(output, "%-*s%s",
974 csv_output ? 0 : unit_width,
975 evsel->unit, csv_sep);
976
977 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
978
979 if (evsel->cgrp)
980 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
981
982 if (csv_output || interval)
983 return;
984
985 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
986 total = avg_stats(&runtime_cycles_stats[cpu]);
987 if (total) {
988 ratio = avg / total;
989 fprintf(output, " # %5.2f insns per cycle ", ratio);
990 }
991 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
992 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
993
994 if (total && avg) {
995 ratio = total / avg;
996 fprintf(output, "\n");
997 if (aggr_mode == AGGR_NONE)
998 fprintf(output, " ");
999 fprintf(output, " # %5.2f stalled cycles per insn", ratio);
1000 }
1001
1002 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
1003 runtime_branches_stats[cpu].n != 0) {
1004 print_branch_misses(cpu, evsel, avg);
1005 } else if (
1006 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1007 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
1008 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1009 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1010 runtime_l1_dcache_stats[cpu].n != 0) {
1011 print_l1_dcache_misses(cpu, evsel, avg);
1012 } else if (
1013 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1014 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
1015 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1016 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1017 runtime_l1_icache_stats[cpu].n != 0) {
1018 print_l1_icache_misses(cpu, evsel, avg);
1019 } else if (
1020 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1021 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
1022 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1023 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1024 runtime_dtlb_cache_stats[cpu].n != 0) {
1025 print_dtlb_cache_misses(cpu, evsel, avg);
1026 } else if (
1027 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1028 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
1029 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1030 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1031 runtime_itlb_cache_stats[cpu].n != 0) {
1032 print_itlb_cache_misses(cpu, evsel, avg);
1033 } else if (
1034 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1035 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
1036 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1037 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1038 runtime_ll_cache_stats[cpu].n != 0) {
1039 print_ll_cache_misses(cpu, evsel, avg);
1040 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
1041 runtime_cacherefs_stats[cpu].n != 0) {
1042 total = avg_stats(&runtime_cacherefs_stats[cpu]);
1043
1044 if (total)
1045 ratio = avg * 100 / total;
1046
1047 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
1048
1049 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1050 print_stalled_cycles_frontend(cpu, evsel, avg);
1051 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1052 print_stalled_cycles_backend(cpu, evsel, avg);
1053 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1054 total = avg_stats(&runtime_nsecs_stats[cpu]);
1055
1056 if (total) {
1057 ratio = avg / total;
1058 fprintf(output, " # %8.3f GHz ", ratio);
1059 }
1060 } else if (transaction_run &&
1061 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1062 total = avg_stats(&runtime_cycles_stats[cpu]);
1063 if (total)
1064 fprintf(output,
1065 " # %5.2f%% transactional cycles ",
1066 100.0 * (avg / total));
1067 } else if (transaction_run &&
1068 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1069 total = avg_stats(&runtime_cycles_stats[cpu]);
1070 total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1071 if (total2 < avg)
1072 total2 = avg;
1073 if (total)
1074 fprintf(output,
1075 " # %5.2f%% aborted cycles ",
1076 100.0 * ((total2-avg) / total));
1077 } else if (transaction_run &&
1078 perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1079 avg > 0 &&
1080 runtime_cycles_in_tx_stats[cpu].n != 0) {
1081 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1082
1083 if (total)
1084 ratio = total / avg;
1085
1086 fprintf(output, " # %8.0f cycles / transaction ", ratio);
1087 } else if (transaction_run &&
1088 perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1089 avg > 0 &&
1090 runtime_cycles_in_tx_stats[cpu].n != 0) {
1091 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1092
1093 if (total)
1094 ratio = total / avg;
1095
1096 fprintf(output, " # %8.0f cycles / elision ", ratio);
1097 } else if (runtime_nsecs_stats[cpu].n != 0) {
1098 char unit = 'M';
1099
1100 total = avg_stats(&runtime_nsecs_stats[cpu]);
1101
1102 if (total)
1103 ratio = 1000.0 * avg / total;
1104 if (ratio < 0.001) {
1105 ratio *= 1000;
1106 unit = 'K';
1107 }
1108
1109 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
1110 } else {
1111 fprintf(output, " ");
1112 }
1113}
1114
1115static void print_aggr(char *prefix)
1116{
1117 struct perf_evsel *counter;
1118 int cpu, cpu2, s, s2, id, nr;
1119 double uval;
1120 u64 ena, run, val;
1121
1122 if (!(aggr_map || aggr_get_id))
1123 return;
1124
1125 for (s = 0; s < aggr_map->nr; s++) {
1126 id = aggr_map->map[s];
1127 evlist__for_each(evsel_list, counter) {
1128 val = ena = run = 0;
1129 nr = 0;
1130 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1131 cpu2 = perf_evsel__cpus(counter)->map[cpu];
1132 s2 = aggr_get_id(evsel_list->cpus, cpu2);
1133 if (s2 != id)
1134 continue;
1135 val += counter->counts->cpu[cpu].val;
1136 ena += counter->counts->cpu[cpu].ena;
1137 run += counter->counts->cpu[cpu].run;
1138 nr++;
1139 }
1140 if (prefix)
1141 fprintf(output, "%s", prefix);
1142
1143 if (run == 0 || ena == 0) {
1144 aggr_printout(counter, id, nr);
1145
1146 fprintf(output, "%*s%s",
1147 csv_output ? 0 : 18,
1148 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1149 csv_sep);
1150
1151 fprintf(output, "%-*s%s",
1152 csv_output ? 0 : unit_width,
1153 counter->unit, csv_sep);
1154
1155 fprintf(output, "%*s",
1156 csv_output ? 0 : -25,
1157 perf_evsel__name(counter));
1158
1159 if (counter->cgrp)
1160 fprintf(output, "%s%s",
1161 csv_sep, counter->cgrp->name);
1162
1163 fputc('\n', output);
1164 continue;
1165 }
1166 uval = val * counter->scale;
1167
1168 if (nsec_counter(counter))
1169 nsec_printout(id, nr, counter, uval);
1170 else
1171 abs_printout(id, nr, counter, uval);
1172
1173 if (!csv_output) {
1174 print_noise(counter, 1.0);
1175
1176 if (run != ena)
1177 fprintf(output, " (%.2f%%)",
1178 100.0 * run / ena);
1179 }
1180 fputc('\n', output);
1181 }
1182 }
1183}
1184
1185/*
1186 * Print out the results of a single counter:
1187 * aggregated counts in system-wide mode
1188 */
1189static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1190{
1191 struct perf_stat *ps = counter->priv;
1192 double avg = avg_stats(&ps->res_stats[0]);
1193 int scaled = counter->counts->scaled;
1194 double uval;
1195
1196 if (prefix)
1197 fprintf(output, "%s", prefix);
1198
1199 if (scaled == -1) {
1200 fprintf(output, "%*s%s",
1201 csv_output ? 0 : 18,
1202 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1203 csv_sep);
1204 fprintf(output, "%-*s%s",
1205 csv_output ? 0 : unit_width,
1206 counter->unit, csv_sep);
1207 fprintf(output, "%*s",
1208 csv_output ? 0 : -25,
1209 perf_evsel__name(counter));
1210
1211 if (counter->cgrp)
1212 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1213
1214 fputc('\n', output);
1215 return;
1216 }
1217
1218 uval = avg * counter->scale;
1219
1220 if (nsec_counter(counter))
1221 nsec_printout(-1, 0, counter, uval);
1222 else
1223 abs_printout(-1, 0, counter, uval);
1224
1225 print_noise(counter, avg);
1226
1227 if (csv_output) {
1228 fputc('\n', output);
1229 return;
1230 }
1231
1232 if (scaled) {
1233 double avg_enabled, avg_running;
1234
1235 avg_enabled = avg_stats(&ps->res_stats[1]);
1236 avg_running = avg_stats(&ps->res_stats[2]);
1237
1238 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1239 }
1240 fprintf(output, "\n");
1241}
1242
1243/*
1244 * Print out the results of a single counter:
1245 * does not use aggregated count in system-wide
1246 */
1247static void print_counter(struct perf_evsel *counter, char *prefix)
1248{
1249 u64 ena, run, val;
1250 double uval;
1251 int cpu;
1252
1253 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1254 val = counter->counts->cpu[cpu].val;
1255 ena = counter->counts->cpu[cpu].ena;
1256 run = counter->counts->cpu[cpu].run;
1257
1258 if (prefix)
1259 fprintf(output, "%s", prefix);
1260
1261 if (run == 0 || ena == 0) {
1262 fprintf(output, "CPU%*d%s%*s%s",
1263 csv_output ? 0 : -4,
1264 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1265 csv_output ? 0 : 18,
1266 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1267 csv_sep);
1268
1269 fprintf(output, "%-*s%s",
1270 csv_output ? 0 : unit_width,
1271 counter->unit, csv_sep);
1272
1273 fprintf(output, "%*s",
1274 csv_output ? 0 : -25,
1275 perf_evsel__name(counter));
1276
1277 if (counter->cgrp)
1278 fprintf(output, "%s%s",
1279 csv_sep, counter->cgrp->name);
1280
1281 fputc('\n', output);
1282 continue;
1283 }
1284
1285 uval = val * counter->scale;
1286
1287 if (nsec_counter(counter))
1288 nsec_printout(cpu, 0, counter, uval);
1289 else
1290 abs_printout(cpu, 0, counter, uval);
1291
1292 if (!csv_output) {
1293 print_noise(counter, 1.0);
1294
1295 if (run != ena)
1296 fprintf(output, " (%.2f%%)",
1297 100.0 * run / ena);
1298 }
1299 fputc('\n', output);
1300 }
1301}
1302
1303static void print_stat(int argc, const char **argv)
1304{
1305 struct perf_evsel *counter;
1306 int i;
1307
1308 fflush(stdout);
1309
1310 if (!csv_output) {
1311 fprintf(output, "\n");
1312 fprintf(output, " Performance counter stats for ");
1313 if (target.system_wide)
1314 fprintf(output, "\'system wide");
1315 else if (target.cpu_list)
1316 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1317 else if (!target__has_task(&target)) {
1318 fprintf(output, "\'%s", argv[0]);
1319 for (i = 1; i < argc; i++)
1320 fprintf(output, " %s", argv[i]);
1321 } else if (target.pid)
1322 fprintf(output, "process id \'%s", target.pid);
1323 else
1324 fprintf(output, "thread id \'%s", target.tid);
1325
1326 fprintf(output, "\'");
1327 if (run_count > 1)
1328 fprintf(output, " (%d runs)", run_count);
1329 fprintf(output, ":\n\n");
1330 }
1331
1332 switch (aggr_mode) {
1333 case AGGR_CORE:
1334 case AGGR_SOCKET:
1335 print_aggr(NULL);
1336 break;
1337 case AGGR_GLOBAL:
1338 evlist__for_each(evsel_list, counter)
1339 print_counter_aggr(counter, NULL);
1340 break;
1341 case AGGR_NONE:
1342 evlist__for_each(evsel_list, counter)
1343 print_counter(counter, NULL);
1344 break;
1345 default:
1346 break;
1347 }
1348
1349 if (!csv_output) {
1350 if (!null_run)
1351 fprintf(output, "\n");
1352 fprintf(output, " %17.9f seconds time elapsed",
1353 avg_stats(&walltime_nsecs_stats)/1e9);
1354 if (run_count > 1) {
1355 fprintf(output, " ");
1356 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1357 avg_stats(&walltime_nsecs_stats));
1358 }
1359 fprintf(output, "\n\n");
1360 }
1361}
1362
1363static volatile int signr = -1;
1364
1365static void skip_signal(int signo)
1366{
1367 if ((child_pid == -1) || interval)
1368 done = 1;
1369
1370 signr = signo;
1371 /*
1372 * render child_pid harmless
1373 * won't send SIGTERM to a random
1374 * process in case of race condition
1375 * and fast PID recycling
1376 */
1377 child_pid = -1;
1378}
1379
1380static void sig_atexit(void)
1381{
1382 sigset_t set, oset;
1383
1384 /*
1385 * avoid race condition with SIGCHLD handler
1386 * in skip_signal() which is modifying child_pid
1387 * goal is to avoid send SIGTERM to a random
1388 * process
1389 */
1390 sigemptyset(&set);
1391 sigaddset(&set, SIGCHLD);
1392 sigprocmask(SIG_BLOCK, &set, &oset);
1393
1394 if (child_pid != -1)
1395 kill(child_pid, SIGTERM);
1396
1397 sigprocmask(SIG_SETMASK, &oset, NULL);
1398
1399 if (signr == -1)
1400 return;
1401
1402 signal(signr, SIG_DFL);
1403 kill(getpid(), signr);
1404}
1405
1406static int stat__set_big_num(const struct option *opt __maybe_unused,
1407 const char *s __maybe_unused, int unset)
1408{
1409 big_num_opt = unset ? 0 : 1;
1410 return 0;
1411}
1412
1413static int perf_stat_init_aggr_mode(void)
1414{
1415 switch (aggr_mode) {
1416 case AGGR_SOCKET:
1417 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1418 perror("cannot build socket map");
1419 return -1;
1420 }
1421 aggr_get_id = cpu_map__get_socket;
1422 break;
1423 case AGGR_CORE:
1424 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1425 perror("cannot build core map");
1426 return -1;
1427 }
1428 aggr_get_id = cpu_map__get_core;
1429 break;
1430 case AGGR_NONE:
1431 case AGGR_GLOBAL:
1432 default:
1433 break;
1434 }
1435 return 0;
1436}
1437
1438static int setup_events(const char * const *attrs, unsigned len)
1439{
1440 unsigned i;
1441
1442 for (i = 0; i < len; i++) {
1443 if (parse_events(evsel_list, attrs[i]))
1444 return -1;
1445 }
1446 return 0;
1447}
1448
1449/*
1450 * Add default attributes, if there were no attributes specified or
1451 * if -d/--detailed, -d -d or -d -d -d is used:
1452 */
1453static int add_default_attributes(void)
1454{
1455 struct perf_event_attr default_attrs[] = {
1456
1457 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1458 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1459 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1460 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1461
1462 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1463 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1464 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1465 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1466 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1467 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1468
1469};
1470
1471/*
1472 * Detailed stats (-d), covering the L1 and last level data caches:
1473 */
1474 struct perf_event_attr detailed_attrs[] = {
1475
1476 { .type = PERF_TYPE_HW_CACHE,
1477 .config =
1478 PERF_COUNT_HW_CACHE_L1D << 0 |
1479 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1480 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1481
1482 { .type = PERF_TYPE_HW_CACHE,
1483 .config =
1484 PERF_COUNT_HW_CACHE_L1D << 0 |
1485 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1486 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1487
1488 { .type = PERF_TYPE_HW_CACHE,
1489 .config =
1490 PERF_COUNT_HW_CACHE_LL << 0 |
1491 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1492 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1493
1494 { .type = PERF_TYPE_HW_CACHE,
1495 .config =
1496 PERF_COUNT_HW_CACHE_LL << 0 |
1497 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1498 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1499};
1500
1501/*
1502 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1503 */
1504 struct perf_event_attr very_detailed_attrs[] = {
1505
1506 { .type = PERF_TYPE_HW_CACHE,
1507 .config =
1508 PERF_COUNT_HW_CACHE_L1I << 0 |
1509 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1510 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1511
1512 { .type = PERF_TYPE_HW_CACHE,
1513 .config =
1514 PERF_COUNT_HW_CACHE_L1I << 0 |
1515 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1516 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1517
1518 { .type = PERF_TYPE_HW_CACHE,
1519 .config =
1520 PERF_COUNT_HW_CACHE_DTLB << 0 |
1521 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1522 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1523
1524 { .type = PERF_TYPE_HW_CACHE,
1525 .config =
1526 PERF_COUNT_HW_CACHE_DTLB << 0 |
1527 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1528 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1529
1530 { .type = PERF_TYPE_HW_CACHE,
1531 .config =
1532 PERF_COUNT_HW_CACHE_ITLB << 0 |
1533 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1534 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1535
1536 { .type = PERF_TYPE_HW_CACHE,
1537 .config =
1538 PERF_COUNT_HW_CACHE_ITLB << 0 |
1539 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1540 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1541
1542};
1543
1544/*
1545 * Very, very detailed stats (-d -d -d), adding prefetch events:
1546 */
1547 struct perf_event_attr very_very_detailed_attrs[] = {
1548
1549 { .type = PERF_TYPE_HW_CACHE,
1550 .config =
1551 PERF_COUNT_HW_CACHE_L1D << 0 |
1552 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1553 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1554
1555 { .type = PERF_TYPE_HW_CACHE,
1556 .config =
1557 PERF_COUNT_HW_CACHE_L1D << 0 |
1558 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1559 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1560};
1561
1562 /* Set attrs if no event is selected and !null_run: */
1563 if (null_run)
1564 return 0;
1565
1566 if (transaction_run) {
1567 int err;
1568 if (pmu_have_event("cpu", "cycles-ct") &&
1569 pmu_have_event("cpu", "el-start"))
1570 err = setup_events(transaction_attrs,
1571 ARRAY_SIZE(transaction_attrs));
1572 else
1573 err = setup_events(transaction_limited_attrs,
1574 ARRAY_SIZE(transaction_limited_attrs));
1575 if (err < 0) {
1576 fprintf(stderr, "Cannot set up transaction events\n");
1577 return -1;
1578 }
1579 return 0;
1580 }
1581
1582 if (!evsel_list->nr_entries) {
1583 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1584 return -1;
1585 }
1586
1587 /* Detailed events get appended to the event list: */
1588
1589 if (detailed_run < 1)
1590 return 0;
1591
1592 /* Append detailed run extra attributes: */
1593 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1594 return -1;
1595
1596 if (detailed_run < 2)
1597 return 0;
1598
1599 /* Append very detailed run extra attributes: */
1600 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1601 return -1;
1602
1603 if (detailed_run < 3)
1604 return 0;
1605
1606 /* Append very, very detailed run extra attributes: */
1607 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1608}
1609
1610int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1611{
1612 bool append_file = false;
1613 int output_fd = 0;
1614 const char *output_name = NULL;
1615 const struct option options[] = {
1616 OPT_BOOLEAN('T', "transaction", &transaction_run,
1617 "hardware transaction statistics"),
1618 OPT_CALLBACK('e', "event", &evsel_list, "event",
1619 "event selector. use 'perf list' to list available events",
1620 parse_events_option),
1621 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1622 "event filter", parse_filter),
1623 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1624 "child tasks do not inherit counters"),
1625 OPT_STRING('p', "pid", &target.pid, "pid",
1626 "stat events on existing process id"),
1627 OPT_STRING('t', "tid", &target.tid, "tid",
1628 "stat events on existing thread id"),
1629 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1630 "system-wide collection from all CPUs"),
1631 OPT_BOOLEAN('g', "group", &group,
1632 "put the counters into a counter group"),
1633 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1634 OPT_INCR('v', "verbose", &verbose,
1635 "be more verbose (show counter open errors, etc)"),
1636 OPT_INTEGER('r', "repeat", &run_count,
1637 "repeat command and print average + stddev (max: 100, forever: 0)"),
1638 OPT_BOOLEAN('n', "null", &null_run,
1639 "null run - dont start any counters"),
1640 OPT_INCR('d', "detailed", &detailed_run,
1641 "detailed run - start a lot of events"),
1642 OPT_BOOLEAN('S', "sync", &sync_run,
1643 "call sync() before starting a run"),
1644 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1645 "print large numbers with thousands\' separators",
1646 stat__set_big_num),
1647 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1648 "list of cpus to monitor in system-wide"),
1649 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1650 "disable CPU count aggregation", AGGR_NONE),
1651 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1652 "print counts with custom separator"),
1653 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1654 "monitor event in cgroup name only", parse_cgroups),
1655 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1656 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1657 OPT_INTEGER(0, "log-fd", &output_fd,
1658 "log output to fd, instead of stderr"),
1659 OPT_STRING(0, "pre", &pre_cmd, "command",
1660 "command to run prior to the measured command"),
1661 OPT_STRING(0, "post", &post_cmd, "command",
1662 "command to run after to the measured command"),
1663 OPT_UINTEGER('I', "interval-print", &interval,
1664 "print counts at regular interval in ms (>= 100)"),
1665 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1666 "aggregate counts per processor socket", AGGR_SOCKET),
1667 OPT_SET_UINT(0, "per-core", &aggr_mode,
1668 "aggregate counts per physical processor core", AGGR_CORE),
1669 OPT_UINTEGER('D', "delay", &initial_delay,
1670 "ms to wait before starting measurement after program start"),
1671 OPT_END()
1672 };
1673 const char * const stat_usage[] = {
1674 "perf stat [<options>] [<command>]",
1675 NULL
1676 };
1677 int status = -EINVAL, run_idx;
1678 const char *mode;
1679
1680 setlocale(LC_ALL, "");
1681
1682 evsel_list = perf_evlist__new();
1683 if (evsel_list == NULL)
1684 return -ENOMEM;
1685
1686 argc = parse_options(argc, argv, options, stat_usage,
1687 PARSE_OPT_STOP_AT_NON_OPTION);
1688
1689 output = stderr;
1690 if (output_name && strcmp(output_name, "-"))
1691 output = NULL;
1692
1693 if (output_name && output_fd) {
1694 fprintf(stderr, "cannot use both --output and --log-fd\n");
1695 parse_options_usage(stat_usage, options, "o", 1);
1696 parse_options_usage(NULL, options, "log-fd", 0);
1697 goto out;
1698 }
1699
1700 if (output_fd < 0) {
1701 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1702 parse_options_usage(stat_usage, options, "log-fd", 0);
1703 goto out;
1704 }
1705
1706 if (!output) {
1707 struct timespec tm;
1708 mode = append_file ? "a" : "w";
1709
1710 output = fopen(output_name, mode);
1711 if (!output) {
1712 perror("failed to create output file");
1713 return -1;
1714 }
1715 clock_gettime(CLOCK_REALTIME, &tm);
1716 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1717 } else if (output_fd > 0) {
1718 mode = append_file ? "a" : "w";
1719 output = fdopen(output_fd, mode);
1720 if (!output) {
1721 perror("Failed opening logfd");
1722 return -errno;
1723 }
1724 }
1725
1726 if (csv_sep) {
1727 csv_output = true;
1728 if (!strcmp(csv_sep, "\\t"))
1729 csv_sep = "\t";
1730 } else
1731 csv_sep = DEFAULT_SEPARATOR;
1732
1733 /*
1734 * let the spreadsheet do the pretty-printing
1735 */
1736 if (csv_output) {
1737 /* User explicitly passed -B? */
1738 if (big_num_opt == 1) {
1739 fprintf(stderr, "-B option not supported with -x\n");
1740 parse_options_usage(stat_usage, options, "B", 1);
1741 parse_options_usage(NULL, options, "x", 1);
1742 goto out;
1743 } else /* Nope, so disable big number formatting */
1744 big_num = false;
1745 } else if (big_num_opt == 0) /* User passed --no-big-num */
1746 big_num = false;
1747
1748 if (!argc && target__none(&target))
1749 usage_with_options(stat_usage, options);
1750
1751 if (run_count < 0) {
1752 pr_err("Run count must be a positive number\n");
1753 parse_options_usage(stat_usage, options, "r", 1);
1754 goto out;
1755 } else if (run_count == 0) {
1756 forever = true;
1757 run_count = 1;
1758 }
1759
1760 /* no_aggr, cgroup are for system-wide only */
1761 if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1762 !target__has_cpu(&target)) {
1763 fprintf(stderr, "both cgroup and no-aggregation "
1764 "modes only available in system-wide mode\n");
1765
1766 parse_options_usage(stat_usage, options, "G", 1);
1767 parse_options_usage(NULL, options, "A", 1);
1768 parse_options_usage(NULL, options, "a", 1);
1769 goto out;
1770 }
1771
1772 if (add_default_attributes())
1773 goto out;
1774
1775 target__validate(&target);
1776
1777 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1778 if (target__has_task(&target)) {
1779 pr_err("Problems finding threads of monitor\n");
1780 parse_options_usage(stat_usage, options, "p", 1);
1781 parse_options_usage(NULL, options, "t", 1);
1782 } else if (target__has_cpu(&target)) {
1783 perror("failed to parse CPUs map");
1784 parse_options_usage(stat_usage, options, "C", 1);
1785 parse_options_usage(NULL, options, "a", 1);
1786 }
1787 goto out;
1788 }
1789 if (interval && interval < 100) {
1790 pr_err("print interval must be >= 100ms\n");
1791 parse_options_usage(stat_usage, options, "I", 1);
1792 goto out;
1793 }
1794
1795 if (perf_evlist__alloc_stats(evsel_list, interval))
1796 goto out;
1797
1798 if (perf_stat_init_aggr_mode())
1799 goto out;
1800
1801 /*
1802 * We dont want to block the signals - that would cause
1803 * child tasks to inherit that and Ctrl-C would not work.
1804 * What we want is for Ctrl-C to work in the exec()-ed
1805 * task, but being ignored by perf stat itself:
1806 */
1807 atexit(sig_atexit);
1808 if (!forever)
1809 signal(SIGINT, skip_signal);
1810 signal(SIGCHLD, skip_signal);
1811 signal(SIGALRM, skip_signal);
1812 signal(SIGABRT, skip_signal);
1813
1814 status = 0;
1815 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1816 if (run_count != 1 && verbose)
1817 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1818 run_idx + 1);
1819
1820 status = run_perf_stat(argc, argv);
1821 if (forever && status != -1) {
1822 print_stat(argc, argv);
1823 perf_stat__reset_stats(evsel_list);
1824 }
1825 }
1826
1827 if (!forever && status != -1 && !interval)
1828 print_stat(argc, argv);
1829
1830 perf_evlist__free_stats(evsel_list);
1831out:
1832 perf_evlist__delete(evsel_list);
1833 return status;
1834}