Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * builtin-stat.c
4 *
5 * Builtin stat command: Give a precise performance counters summary
6 * overview about any workload, CPU or specific PID.
7 *
8 * Sample output:
9
10 $ perf stat ./hackbench 10
11
12 Time: 0.118
13
14 Performance counter stats for './hackbench 10':
15
16 1708.761321 task-clock # 11.037 CPUs utilized
17 41,190 context-switches # 0.024 M/sec
18 6,735 CPU-migrations # 0.004 M/sec
19 17,318 page-faults # 0.010 M/sec
20 5,205,202,243 cycles # 3.046 GHz
21 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
22 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
23 2,603,501,247 instructions # 0.50 insns per cycle
24 # 1.48 stalled cycles per insn
25 484,357,498 branches # 283.455 M/sec
26 6,388,934 branch-misses # 1.32% of all branches
27
28 0.154822978 seconds time elapsed
29
30 *
31 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 *
33 * Improvements and fixes by:
34 *
35 * Arjan van de Ven <arjan@linux.intel.com>
36 * Yanmin Zhang <yanmin.zhang@intel.com>
37 * Wu Fengguang <fengguang.wu@intel.com>
38 * Mike Galbraith <efault@gmx.de>
39 * Paul Mackerras <paulus@samba.org>
40 * Jaswinder Singh Rajput <jaswinder@kernel.org>
41 */
42
43#include "builtin.h"
44#include "perf.h"
45#include "util/cgroup.h"
46#include <subcmd/parse-options.h>
47#include "util/parse-events.h"
48#include "util/pmu.h"
49#include "util/event.h"
50#include "util/evlist.h"
51#include "util/evlist-hybrid.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_map.h"
59#include "util/counts.h"
60#include "util/topdown.h"
61#include "util/session.h"
62#include "util/tool.h"
63#include "util/string2.h"
64#include "util/metricgroup.h"
65#include "util/synthetic-events.h"
66#include "util/target.h"
67#include "util/time-utils.h"
68#include "util/top.h"
69#include "util/affinity.h"
70#include "util/pfm.h"
71#include "util/bpf_counter.h"
72#include "util/iostat.h"
73#include "util/pmu-hybrid.h"
74#include "asm/bug.h"
75
76#include <linux/time64.h>
77#include <linux/zalloc.h>
78#include <api/fs/fs.h>
79#include <errno.h>
80#include <signal.h>
81#include <stdlib.h>
82#include <sys/prctl.h>
83#include <inttypes.h>
84#include <locale.h>
85#include <math.h>
86#include <sys/types.h>
87#include <sys/stat.h>
88#include <sys/wait.h>
89#include <unistd.h>
90#include <sys/time.h>
91#include <sys/resource.h>
92#include <linux/err.h>
93
94#include <linux/ctype.h>
95#include <perf/evlist.h>
96#include <internal/threadmap.h>
97
98#define DEFAULT_SEPARATOR " "
99#define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
100
101static void print_counters(struct timespec *ts, int argc, const char **argv);
102
103/* Default events used for perf stat -T */
104static const char *transaction_attrs = {
105 "task-clock,"
106 "{"
107 "instructions,"
108 "cycles,"
109 "cpu/cycles-t/,"
110 "cpu/tx-start/,"
111 "cpu/el-start/,"
112 "cpu/cycles-ct/"
113 "}"
114};
115
116/* More limited version when the CPU does not have all events. */
117static const char * transaction_limited_attrs = {
118 "task-clock,"
119 "{"
120 "instructions,"
121 "cycles,"
122 "cpu/cycles-t/,"
123 "cpu/tx-start/"
124 "}"
125};
126
127static const char * topdown_attrs[] = {
128 "topdown-total-slots",
129 "topdown-slots-retired",
130 "topdown-recovery-bubbles",
131 "topdown-fetch-bubbles",
132 "topdown-slots-issued",
133 NULL,
134};
135
136static const char *topdown_metric_attrs[] = {
137 "slots",
138 "topdown-retiring",
139 "topdown-bad-spec",
140 "topdown-fe-bound",
141 "topdown-be-bound",
142 NULL,
143};
144
145static const char *topdown_metric_L2_attrs[] = {
146 "slots",
147 "topdown-retiring",
148 "topdown-bad-spec",
149 "topdown-fe-bound",
150 "topdown-be-bound",
151 "topdown-heavy-ops",
152 "topdown-br-mispredict",
153 "topdown-fetch-lat",
154 "topdown-mem-bound",
155 NULL,
156};
157
158#define TOPDOWN_MAX_LEVEL 2
159
160static const char *smi_cost_attrs = {
161 "{"
162 "msr/aperf/,"
163 "msr/smi/,"
164 "cycles"
165 "}"
166};
167
168static struct evlist *evsel_list;
169static bool all_counters_use_bpf = true;
170
171static struct target target = {
172 .uid = UINT_MAX,
173};
174
175#define METRIC_ONLY_LEN 20
176
177static volatile sig_atomic_t child_pid = -1;
178static int detailed_run = 0;
179static bool transaction_run;
180static bool topdown_run = false;
181static bool smi_cost = false;
182static bool smi_reset = false;
183static int big_num_opt = -1;
184static const char *pre_cmd = NULL;
185static const char *post_cmd = NULL;
186static bool sync_run = false;
187static bool forever = false;
188static bool force_metric_only = false;
189static struct timespec ref_time;
190static bool append_file;
191static bool interval_count;
192static const char *output_name;
193static int output_fd;
194static char *metrics;
195
196struct perf_stat {
197 bool record;
198 struct perf_data data;
199 struct perf_session *session;
200 u64 bytes_written;
201 struct perf_tool tool;
202 bool maps_allocated;
203 struct perf_cpu_map *cpus;
204 struct perf_thread_map *threads;
205 enum aggr_mode aggr_mode;
206};
207
208static struct perf_stat perf_stat;
209#define STAT_RECORD perf_stat.record
210
211static volatile sig_atomic_t done = 0;
212
213static struct perf_stat_config stat_config = {
214 .aggr_mode = AGGR_GLOBAL,
215 .scale = true,
216 .unit_width = 4, /* strlen("unit") */
217 .run_count = 1,
218 .metric_only_len = METRIC_ONLY_LEN,
219 .walltime_nsecs_stats = &walltime_nsecs_stats,
220 .ru_stats = &ru_stats,
221 .big_num = true,
222 .ctl_fd = -1,
223 .ctl_fd_ack = -1,
224 .iostat_run = false,
225};
226
227static bool cpus_map_matched(struct evsel *a, struct evsel *b)
228{
229 if (!a->core.cpus && !b->core.cpus)
230 return true;
231
232 if (!a->core.cpus || !b->core.cpus)
233 return false;
234
235 if (perf_cpu_map__nr(a->core.cpus) != perf_cpu_map__nr(b->core.cpus))
236 return false;
237
238 for (int i = 0; i < perf_cpu_map__nr(a->core.cpus); i++) {
239 if (perf_cpu_map__cpu(a->core.cpus, i).cpu !=
240 perf_cpu_map__cpu(b->core.cpus, i).cpu)
241 return false;
242 }
243
244 return true;
245}
246
247static void evlist__check_cpu_maps(struct evlist *evlist)
248{
249 struct evsel *evsel, *pos, *leader;
250 char buf[1024];
251
252 if (evlist__has_hybrid(evlist))
253 evlist__warn_hybrid_group(evlist);
254
255 evlist__for_each_entry(evlist, evsel) {
256 leader = evsel__leader(evsel);
257
258 /* Check that leader matches cpus with each member. */
259 if (leader == evsel)
260 continue;
261 if (cpus_map_matched(leader, evsel))
262 continue;
263
264 /* If there's mismatch disable the group and warn user. */
265 WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
266 evsel__group_desc(leader, buf, sizeof(buf));
267 pr_warning(" %s\n", buf);
268
269 if (verbose > 0) {
270 cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
271 pr_warning(" %s: %s\n", leader->name, buf);
272 cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
273 pr_warning(" %s: %s\n", evsel->name, buf);
274 }
275
276 for_each_group_evsel(pos, leader)
277 evsel__remove_from_group(pos, leader);
278 }
279}
280
281static inline void diff_timespec(struct timespec *r, struct timespec *a,
282 struct timespec *b)
283{
284 r->tv_sec = a->tv_sec - b->tv_sec;
285 if (a->tv_nsec < b->tv_nsec) {
286 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
287 r->tv_sec--;
288 } else {
289 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
290 }
291}
292
293static void perf_stat__reset_stats(void)
294{
295 evlist__reset_stats(evsel_list);
296 perf_stat__reset_shadow_stats();
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->core.sample_id, x, y)
324
325static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
326 struct perf_counts_values *count)
327{
328 struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
329 struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
330
331 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
332 process_synthesized_event, NULL);
333}
334
335static int read_single_counter(struct evsel *counter, int cpu_map_idx,
336 int thread, struct timespec *rs)
337{
338 switch(counter->tool_event) {
339 case PERF_TOOL_DURATION_TIME: {
340 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
341 struct perf_counts_values *count =
342 perf_counts(counter->counts, cpu_map_idx, thread);
343 count->ena = count->run = val;
344 count->val = val;
345 return 0;
346 }
347 case PERF_TOOL_USER_TIME:
348 case PERF_TOOL_SYSTEM_TIME: {
349 u64 val;
350 struct perf_counts_values *count =
351 perf_counts(counter->counts, cpu_map_idx, thread);
352 if (counter->tool_event == PERF_TOOL_USER_TIME)
353 val = ru_stats.ru_utime_usec_stat.mean;
354 else
355 val = ru_stats.ru_stime_usec_stat.mean;
356 count->ena = count->run = val;
357 count->val = val;
358 return 0;
359 }
360 default:
361 case PERF_TOOL_NONE:
362 return evsel__read_counter(counter, cpu_map_idx, thread);
363 case PERF_TOOL_MAX:
364 /* This should never be reached */
365 return 0;
366 }
367}
368
369/*
370 * Read out the results of a single counter:
371 * do not aggregate counts across CPUs in system-wide mode
372 */
373static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx)
374{
375 int nthreads = perf_thread_map__nr(evsel_list->core.threads);
376 int thread;
377
378 if (!counter->supported)
379 return -ENOENT;
380
381 for (thread = 0; thread < nthreads; thread++) {
382 struct perf_counts_values *count;
383
384 count = perf_counts(counter->counts, cpu_map_idx, thread);
385
386 /*
387 * The leader's group read loads data into its group members
388 * (via evsel__read_counter()) and sets their count->loaded.
389 */
390 if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
391 read_single_counter(counter, cpu_map_idx, thread, rs)) {
392 counter->counts->scaled = -1;
393 perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
394 perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
395 return -1;
396 }
397
398 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
399
400 if (STAT_RECORD) {
401 if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
402 pr_err("failed to write stat event\n");
403 return -1;
404 }
405 }
406
407 if (verbose > 1) {
408 fprintf(stat_config.output,
409 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
410 evsel__name(counter),
411 perf_cpu_map__cpu(evsel__cpus(counter),
412 cpu_map_idx).cpu,
413 count->val, count->ena, count->run);
414 }
415 }
416
417 return 0;
418}
419
420static int read_affinity_counters(struct timespec *rs)
421{
422 struct evlist_cpu_iterator evlist_cpu_itr;
423 struct affinity saved_affinity, *affinity;
424
425 if (all_counters_use_bpf)
426 return 0;
427
428 if (!target__has_cpu(&target) || target__has_per_thread(&target))
429 affinity = NULL;
430 else if (affinity__setup(&saved_affinity) < 0)
431 return -1;
432 else
433 affinity = &saved_affinity;
434
435 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
436 struct evsel *counter = evlist_cpu_itr.evsel;
437
438 if (evsel__is_bpf(counter))
439 continue;
440
441 if (!counter->err) {
442 counter->err = read_counter_cpu(counter, rs,
443 evlist_cpu_itr.cpu_map_idx);
444 }
445 }
446 if (affinity)
447 affinity__cleanup(&saved_affinity);
448
449 return 0;
450}
451
452static int read_bpf_map_counters(void)
453{
454 struct evsel *counter;
455 int err;
456
457 evlist__for_each_entry(evsel_list, counter) {
458 if (!evsel__is_bpf(counter))
459 continue;
460
461 err = bpf_counter__read(counter);
462 if (err)
463 return err;
464 }
465 return 0;
466}
467
468static int read_counters(struct timespec *rs)
469{
470 if (!stat_config.stop_read_counter) {
471 if (read_bpf_map_counters() ||
472 read_affinity_counters(rs))
473 return -1;
474 }
475 return 0;
476}
477
478static void process_counters(void)
479{
480 struct evsel *counter;
481
482 evlist__for_each_entry(evsel_list, counter) {
483 if (counter->err)
484 pr_debug("failed to read counter %s\n", counter->name);
485 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
486 pr_warning("failed to process counter %s\n", counter->name);
487 counter->err = 0;
488 }
489
490 perf_stat_merge_counters(&stat_config, evsel_list);
491 perf_stat_process_percore(&stat_config, evsel_list);
492 perf_stat_process_shadow_stats(&stat_config, evsel_list);
493}
494
495static void process_interval(void)
496{
497 struct timespec ts, rs;
498
499 clock_gettime(CLOCK_MONOTONIC, &ts);
500 diff_timespec(&rs, &ts, &ref_time);
501
502 perf_stat__reset_shadow_per_stat(&rt_stat);
503 evlist__reset_aggr_stats(evsel_list);
504
505 if (read_counters(&rs) == 0)
506 process_counters();
507
508 if (STAT_RECORD) {
509 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
510 pr_err("failed to write stat round event\n");
511 }
512
513 init_stats(&walltime_nsecs_stats);
514 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
515 print_counters(&rs, 0, NULL);
516}
517
518static bool handle_interval(unsigned int interval, int *times)
519{
520 if (interval) {
521 process_interval();
522 if (interval_count && !(--(*times)))
523 return true;
524 }
525 return false;
526}
527
528static int enable_counters(void)
529{
530 struct evsel *evsel;
531 int err;
532
533 evlist__for_each_entry(evsel_list, evsel) {
534 if (!evsel__is_bpf(evsel))
535 continue;
536
537 err = bpf_counter__enable(evsel);
538 if (err)
539 return err;
540 }
541
542 /*
543 * We need to enable counters only if:
544 * - we don't have tracee (attaching to task or cpu)
545 * - we have initial delay configured
546 */
547 if (!target__none(&target)) {
548 if (!all_counters_use_bpf)
549 evlist__enable(evsel_list);
550 }
551 return 0;
552}
553
554static void disable_counters(void)
555{
556 struct evsel *counter;
557
558 /*
559 * If we don't have tracee (attaching to task or cpu), counters may
560 * still be running. To get accurate group ratios, we must stop groups
561 * from counting before reading their constituent counters.
562 */
563 if (!target__none(&target)) {
564 evlist__for_each_entry(evsel_list, counter)
565 bpf_counter__disable(counter);
566 if (!all_counters_use_bpf)
567 evlist__disable(evsel_list);
568 }
569}
570
571static volatile sig_atomic_t workload_exec_errno;
572
573/*
574 * evlist__prepare_workload will send a SIGUSR1
575 * if the fork fails, since we asked by setting its
576 * want_signal to true.
577 */
578static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
579 void *ucontext __maybe_unused)
580{
581 workload_exec_errno = info->si_value.sival_int;
582}
583
584static bool evsel__should_store_id(struct evsel *counter)
585{
586 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
587}
588
589static bool is_target_alive(struct target *_target,
590 struct perf_thread_map *threads)
591{
592 struct stat st;
593 int i;
594
595 if (!target__has_task(_target))
596 return true;
597
598 for (i = 0; i < threads->nr; i++) {
599 char path[PATH_MAX];
600
601 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
602 threads->map[i].pid);
603
604 if (!stat(path, &st))
605 return true;
606 }
607
608 return false;
609}
610
611static void process_evlist(struct evlist *evlist, unsigned int interval)
612{
613 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
614
615 if (evlist__ctlfd_process(evlist, &cmd) > 0) {
616 switch (cmd) {
617 case EVLIST_CTL_CMD_ENABLE:
618 __fallthrough;
619 case EVLIST_CTL_CMD_DISABLE:
620 if (interval)
621 process_interval();
622 break;
623 case EVLIST_CTL_CMD_SNAPSHOT:
624 case EVLIST_CTL_CMD_ACK:
625 case EVLIST_CTL_CMD_UNSUPPORTED:
626 case EVLIST_CTL_CMD_EVLIST:
627 case EVLIST_CTL_CMD_STOP:
628 case EVLIST_CTL_CMD_PING:
629 default:
630 break;
631 }
632 }
633}
634
635static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
636 int *time_to_sleep)
637{
638 int tts = *time_to_sleep;
639 struct timespec time_diff;
640
641 diff_timespec(&time_diff, time_stop, time_start);
642
643 tts -= time_diff.tv_sec * MSEC_PER_SEC +
644 time_diff.tv_nsec / NSEC_PER_MSEC;
645
646 if (tts < 0)
647 tts = 0;
648
649 *time_to_sleep = tts;
650}
651
652static int dispatch_events(bool forks, int timeout, int interval, int *times)
653{
654 int child_exited = 0, status = 0;
655 int time_to_sleep, sleep_time;
656 struct timespec time_start, time_stop;
657
658 if (interval)
659 sleep_time = interval;
660 else if (timeout)
661 sleep_time = timeout;
662 else
663 sleep_time = 1000;
664
665 time_to_sleep = sleep_time;
666
667 while (!done) {
668 if (forks)
669 child_exited = waitpid(child_pid, &status, WNOHANG);
670 else
671 child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
672
673 if (child_exited)
674 break;
675
676 clock_gettime(CLOCK_MONOTONIC, &time_start);
677 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
678 if (timeout || handle_interval(interval, times))
679 break;
680 time_to_sleep = sleep_time;
681 } else { /* fd revent */
682 process_evlist(evsel_list, interval);
683 clock_gettime(CLOCK_MONOTONIC, &time_stop);
684 compute_tts(&time_start, &time_stop, &time_to_sleep);
685 }
686 }
687
688 return status;
689}
690
691enum counter_recovery {
692 COUNTER_SKIP,
693 COUNTER_RETRY,
694 COUNTER_FATAL,
695};
696
697static enum counter_recovery stat_handle_error(struct evsel *counter)
698{
699 char msg[BUFSIZ];
700 /*
701 * PPC returns ENXIO for HW counters until 2.6.37
702 * (behavior changed with commit b0a873e).
703 */
704 if (errno == EINVAL || errno == ENOSYS ||
705 errno == ENOENT || errno == EOPNOTSUPP ||
706 errno == ENXIO) {
707 if (verbose > 0)
708 ui__warning("%s event is not supported by the kernel.\n",
709 evsel__name(counter));
710 counter->supported = false;
711 /*
712 * errored is a sticky flag that means one of the counter's
713 * cpu event had a problem and needs to be reexamined.
714 */
715 counter->errored = true;
716
717 if ((evsel__leader(counter) != counter) ||
718 !(counter->core.leader->nr_members > 1))
719 return COUNTER_SKIP;
720 } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
721 if (verbose > 0)
722 ui__warning("%s\n", msg);
723 return COUNTER_RETRY;
724 } else if (target__has_per_thread(&target) &&
725 evsel_list->core.threads &&
726 evsel_list->core.threads->err_thread != -1) {
727 /*
728 * For global --per-thread case, skip current
729 * error thread.
730 */
731 if (!thread_map__remove(evsel_list->core.threads,
732 evsel_list->core.threads->err_thread)) {
733 evsel_list->core.threads->err_thread = -1;
734 return COUNTER_RETRY;
735 }
736 }
737
738 evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
739 ui__error("%s\n", msg);
740
741 if (child_pid != -1)
742 kill(child_pid, SIGTERM);
743 return COUNTER_FATAL;
744}
745
746static int __run_perf_stat(int argc, const char **argv, int run_idx)
747{
748 int interval = stat_config.interval;
749 int times = stat_config.times;
750 int timeout = stat_config.timeout;
751 char msg[BUFSIZ];
752 unsigned long long t0, t1;
753 struct evsel *counter;
754 size_t l;
755 int status = 0;
756 const bool forks = (argc > 0);
757 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
758 struct evlist_cpu_iterator evlist_cpu_itr;
759 struct affinity saved_affinity, *affinity = NULL;
760 int err;
761 bool second_pass = false;
762
763 if (forks) {
764 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
765 perror("failed to prepare workload");
766 return -1;
767 }
768 child_pid = evsel_list->workload.pid;
769 }
770
771 if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
772 if (affinity__setup(&saved_affinity) < 0)
773 return -1;
774 affinity = &saved_affinity;
775 }
776
777 evlist__for_each_entry(evsel_list, counter) {
778 counter->reset_group = false;
779 if (bpf_counter__load(counter, &target))
780 return -1;
781 if (!evsel__is_bpf(counter))
782 all_counters_use_bpf = false;
783 }
784
785 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
786 counter = evlist_cpu_itr.evsel;
787
788 /*
789 * bperf calls evsel__open_per_cpu() in bperf__load(), so
790 * no need to call it again here.
791 */
792 if (target.use_bpf)
793 break;
794
795 if (counter->reset_group || counter->errored)
796 continue;
797 if (evsel__is_bpf(counter))
798 continue;
799try_again:
800 if (create_perf_stat_counter(counter, &stat_config, &target,
801 evlist_cpu_itr.cpu_map_idx) < 0) {
802
803 /*
804 * Weak group failed. We cannot just undo this here
805 * because earlier CPUs might be in group mode, and the kernel
806 * doesn't support mixing group and non group reads. Defer
807 * it to later.
808 * Don't close here because we're in the wrong affinity.
809 */
810 if ((errno == EINVAL || errno == EBADF) &&
811 evsel__leader(counter) != counter &&
812 counter->weak_group) {
813 evlist__reset_weak_group(evsel_list, counter, false);
814 assert(counter->reset_group);
815 second_pass = true;
816 continue;
817 }
818
819 switch (stat_handle_error(counter)) {
820 case COUNTER_FATAL:
821 return -1;
822 case COUNTER_RETRY:
823 goto try_again;
824 case COUNTER_SKIP:
825 continue;
826 default:
827 break;
828 }
829
830 }
831 counter->supported = true;
832 }
833
834 if (second_pass) {
835 /*
836 * Now redo all the weak group after closing them,
837 * and also close errored counters.
838 */
839
840 /* First close errored or weak retry */
841 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
842 counter = evlist_cpu_itr.evsel;
843
844 if (!counter->reset_group && !counter->errored)
845 continue;
846
847 perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
848 }
849 /* Now reopen weak */
850 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
851 counter = evlist_cpu_itr.evsel;
852
853 if (!counter->reset_group)
854 continue;
855try_again_reset:
856 pr_debug2("reopening weak %s\n", evsel__name(counter));
857 if (create_perf_stat_counter(counter, &stat_config, &target,
858 evlist_cpu_itr.cpu_map_idx) < 0) {
859
860 switch (stat_handle_error(counter)) {
861 case COUNTER_FATAL:
862 return -1;
863 case COUNTER_RETRY:
864 goto try_again_reset;
865 case COUNTER_SKIP:
866 continue;
867 default:
868 break;
869 }
870 }
871 counter->supported = true;
872 }
873 }
874 affinity__cleanup(affinity);
875
876 evlist__for_each_entry(evsel_list, counter) {
877 if (!counter->supported) {
878 perf_evsel__free_fd(&counter->core);
879 continue;
880 }
881
882 l = strlen(counter->unit);
883 if (l > stat_config.unit_width)
884 stat_config.unit_width = l;
885
886 if (evsel__should_store_id(counter) &&
887 evsel__store_ids(counter, evsel_list))
888 return -1;
889 }
890
891 if (evlist__apply_filters(evsel_list, &counter)) {
892 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
893 counter->filter, evsel__name(counter), errno,
894 str_error_r(errno, msg, sizeof(msg)));
895 return -1;
896 }
897
898 if (STAT_RECORD) {
899 int fd = perf_data__fd(&perf_stat.data);
900
901 if (is_pipe) {
902 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
903 } else {
904 err = perf_session__write_header(perf_stat.session, evsel_list,
905 fd, false);
906 }
907
908 if (err < 0)
909 return err;
910
911 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
912 process_synthesized_event, is_pipe);
913 if (err < 0)
914 return err;
915 }
916
917 if (stat_config.initial_delay) {
918 pr_info(EVLIST_DISABLED_MSG);
919 } else {
920 err = enable_counters();
921 if (err)
922 return -1;
923 }
924
925 /* Exec the command, if any */
926 if (forks)
927 evlist__start_workload(evsel_list);
928
929 if (stat_config.initial_delay > 0) {
930 usleep(stat_config.initial_delay * USEC_PER_MSEC);
931 err = enable_counters();
932 if (err)
933 return -1;
934
935 pr_info(EVLIST_ENABLED_MSG);
936 }
937
938 t0 = rdclock();
939 clock_gettime(CLOCK_MONOTONIC, &ref_time);
940
941 if (forks) {
942 if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
943 status = dispatch_events(forks, timeout, interval, ×);
944 if (child_pid != -1) {
945 if (timeout)
946 kill(child_pid, SIGTERM);
947 wait4(child_pid, &status, 0, &stat_config.ru_data);
948 }
949
950 if (workload_exec_errno) {
951 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
952 pr_err("Workload failed: %s\n", emsg);
953 return -1;
954 }
955
956 if (WIFSIGNALED(status))
957 psignal(WTERMSIG(status), argv[0]);
958 } else {
959 status = dispatch_events(forks, timeout, interval, ×);
960 }
961
962 disable_counters();
963
964 t1 = rdclock();
965
966 if (stat_config.walltime_run_table)
967 stat_config.walltime_run[run_idx] = t1 - t0;
968
969 if (interval && stat_config.summary) {
970 stat_config.interval = 0;
971 stat_config.stop_read_counter = true;
972 init_stats(&walltime_nsecs_stats);
973 update_stats(&walltime_nsecs_stats, t1 - t0);
974
975 evlist__copy_prev_raw_counts(evsel_list);
976 evlist__reset_prev_raw_counts(evsel_list);
977 evlist__reset_aggr_stats(evsel_list);
978 perf_stat__reset_shadow_per_stat(&rt_stat);
979 } else {
980 update_stats(&walltime_nsecs_stats, t1 - t0);
981 update_rusage_stats(&ru_stats, &stat_config.ru_data);
982 }
983
984 /*
985 * Closing a group leader splits the group, and as we only disable
986 * group leaders, results in remaining events becoming enabled. To
987 * avoid arbitrary skew, we must read all counters before closing any
988 * group leaders.
989 */
990 if (read_counters(&(struct timespec) { .tv_nsec = t1-t0 }) == 0)
991 process_counters();
992
993 /*
994 * We need to keep evsel_list alive, because it's processed
995 * later the evsel_list will be closed after.
996 */
997 if (!STAT_RECORD)
998 evlist__close(evsel_list);
999
1000 return WEXITSTATUS(status);
1001}
1002
1003static int run_perf_stat(int argc, const char **argv, int run_idx)
1004{
1005 int ret;
1006
1007 if (pre_cmd) {
1008 ret = system(pre_cmd);
1009 if (ret)
1010 return ret;
1011 }
1012
1013 if (sync_run)
1014 sync();
1015
1016 ret = __run_perf_stat(argc, argv, run_idx);
1017 if (ret)
1018 return ret;
1019
1020 if (post_cmd) {
1021 ret = system(post_cmd);
1022 if (ret)
1023 return ret;
1024 }
1025
1026 return ret;
1027}
1028
1029static void print_counters(struct timespec *ts, int argc, const char **argv)
1030{
1031 /* Do not print anything if we record to the pipe. */
1032 if (STAT_RECORD && perf_stat.data.is_pipe)
1033 return;
1034 if (quiet)
1035 return;
1036
1037 evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1038}
1039
1040static volatile sig_atomic_t signr = -1;
1041
1042static void skip_signal(int signo)
1043{
1044 if ((child_pid == -1) || stat_config.interval)
1045 done = 1;
1046
1047 signr = signo;
1048 /*
1049 * render child_pid harmless
1050 * won't send SIGTERM to a random
1051 * process in case of race condition
1052 * and fast PID recycling
1053 */
1054 child_pid = -1;
1055}
1056
1057static void sig_atexit(void)
1058{
1059 sigset_t set, oset;
1060
1061 /*
1062 * avoid race condition with SIGCHLD handler
1063 * in skip_signal() which is modifying child_pid
1064 * goal is to avoid send SIGTERM to a random
1065 * process
1066 */
1067 sigemptyset(&set);
1068 sigaddset(&set, SIGCHLD);
1069 sigprocmask(SIG_BLOCK, &set, &oset);
1070
1071 if (child_pid != -1)
1072 kill(child_pid, SIGTERM);
1073
1074 sigprocmask(SIG_SETMASK, &oset, NULL);
1075
1076 if (signr == -1)
1077 return;
1078
1079 signal(signr, SIG_DFL);
1080 kill(getpid(), signr);
1081}
1082
1083void perf_stat__set_big_num(int set)
1084{
1085 stat_config.big_num = (set != 0);
1086}
1087
1088void perf_stat__set_no_csv_summary(int set)
1089{
1090 stat_config.no_csv_summary = (set != 0);
1091}
1092
1093static int stat__set_big_num(const struct option *opt __maybe_unused,
1094 const char *s __maybe_unused, int unset)
1095{
1096 big_num_opt = unset ? 0 : 1;
1097 perf_stat__set_big_num(!unset);
1098 return 0;
1099}
1100
1101static int enable_metric_only(const struct option *opt __maybe_unused,
1102 const char *s __maybe_unused, int unset)
1103{
1104 force_metric_only = true;
1105 stat_config.metric_only = !unset;
1106 return 0;
1107}
1108
1109static int append_metric_groups(const struct option *opt __maybe_unused,
1110 const char *str,
1111 int unset __maybe_unused)
1112{
1113 if (metrics) {
1114 char *tmp;
1115
1116 if (asprintf(&tmp, "%s,%s", metrics, str) < 0)
1117 return -ENOMEM;
1118 free(metrics);
1119 metrics = tmp;
1120 } else {
1121 metrics = strdup(str);
1122 if (!metrics)
1123 return -ENOMEM;
1124 }
1125 return 0;
1126}
1127
1128static int parse_control_option(const struct option *opt,
1129 const char *str,
1130 int unset __maybe_unused)
1131{
1132 struct perf_stat_config *config = opt->value;
1133
1134 return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1135}
1136
1137static int parse_stat_cgroups(const struct option *opt,
1138 const char *str, int unset)
1139{
1140 if (stat_config.cgroup_list) {
1141 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1142 return -1;
1143 }
1144
1145 return parse_cgroups(opt, str, unset);
1146}
1147
1148static int parse_hybrid_type(const struct option *opt,
1149 const char *str,
1150 int unset __maybe_unused)
1151{
1152 struct evlist *evlist = *(struct evlist **)opt->value;
1153
1154 if (!list_empty(&evlist->core.entries)) {
1155 fprintf(stderr, "Must define cputype before events/metrics\n");
1156 return -1;
1157 }
1158
1159 evlist->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu(str);
1160 if (!evlist->hybrid_pmu_name) {
1161 fprintf(stderr, "--cputype %s is not supported!\n", str);
1162 return -1;
1163 }
1164
1165 return 0;
1166}
1167
1168static struct option stat_options[] = {
1169 OPT_BOOLEAN('T', "transaction", &transaction_run,
1170 "hardware transaction statistics"),
1171 OPT_CALLBACK('e', "event", &evsel_list, "event",
1172 "event selector. use 'perf list' to list available events",
1173 parse_events_option),
1174 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1175 "event filter", parse_filter),
1176 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1177 "child tasks do not inherit counters"),
1178 OPT_STRING('p', "pid", &target.pid, "pid",
1179 "stat events on existing process id"),
1180 OPT_STRING('t', "tid", &target.tid, "tid",
1181 "stat events on existing thread id"),
1182#ifdef HAVE_BPF_SKEL
1183 OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1184 "stat events on existing bpf program id"),
1185 OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1186 "use bpf program to count events"),
1187 OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1188 "path to perf_event_attr map"),
1189#endif
1190 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1191 "system-wide collection from all CPUs"),
1192 OPT_BOOLEAN(0, "scale", &stat_config.scale,
1193 "Use --no-scale to disable counter scaling for multiplexing"),
1194 OPT_INCR('v', "verbose", &verbose,
1195 "be more verbose (show counter open errors, etc)"),
1196 OPT_INTEGER('r', "repeat", &stat_config.run_count,
1197 "repeat command and print average + stddev (max: 100, forever: 0)"),
1198 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1199 "display details about each run (only with -r option)"),
1200 OPT_BOOLEAN('n', "null", &stat_config.null_run,
1201 "null run - dont start any counters"),
1202 OPT_INCR('d', "detailed", &detailed_run,
1203 "detailed run - start a lot of events"),
1204 OPT_BOOLEAN('S', "sync", &sync_run,
1205 "call sync() before starting a run"),
1206 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1207 "print large numbers with thousands\' separators",
1208 stat__set_big_num),
1209 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1210 "list of cpus to monitor in system-wide"),
1211 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1212 "disable CPU count aggregation", AGGR_NONE),
1213 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1214 OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
1215 "Merge identical named hybrid events"),
1216 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1217 "print counts with custom separator"),
1218 OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
1219 "print counts in JSON format"),
1220 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1221 "monitor event in cgroup name only", parse_stat_cgroups),
1222 OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1223 "expand events for each cgroup"),
1224 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1225 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1226 OPT_INTEGER(0, "log-fd", &output_fd,
1227 "log output to fd, instead of stderr"),
1228 OPT_STRING(0, "pre", &pre_cmd, "command",
1229 "command to run prior to the measured command"),
1230 OPT_STRING(0, "post", &post_cmd, "command",
1231 "command to run after to the measured command"),
1232 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1233 "print counts at regular interval in ms "
1234 "(overhead is possible for values <= 100ms)"),
1235 OPT_INTEGER(0, "interval-count", &stat_config.times,
1236 "print counts for fixed number of times"),
1237 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1238 "clear screen in between new interval"),
1239 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1240 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1241 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1242 "aggregate counts per processor socket", AGGR_SOCKET),
1243 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1244 "aggregate counts per processor die", AGGR_DIE),
1245 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1246 "aggregate counts per physical processor core", AGGR_CORE),
1247 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1248 "aggregate counts per thread", AGGR_THREAD),
1249 OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1250 "aggregate counts per numa node", AGGR_NODE),
1251 OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1252 "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1253 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1254 "Only print computed metrics. No raw values", enable_metric_only),
1255 OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1256 "don't group metric events, impacts multiplexing"),
1257 OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1258 "don't try to share events between metrics in a group"),
1259 OPT_BOOLEAN(0, "topdown", &topdown_run,
1260 "measure top-down statistics"),
1261 OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1262 "Set the metrics level for the top-down statistics (0: max level)"),
1263 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1264 "measure SMI cost"),
1265 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1266 "monitor specified metrics or metric groups (separated by ,)",
1267 append_metric_groups),
1268 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1269 "Configure all used events to run in kernel space.",
1270 PARSE_OPT_EXCLUSIVE),
1271 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1272 "Configure all used events to run in user space.",
1273 PARSE_OPT_EXCLUSIVE),
1274 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1275 "Use with 'percore' event qualifier to show the event "
1276 "counts of one hardware thread by sum up total hardware "
1277 "threads of same physical core"),
1278 OPT_BOOLEAN(0, "summary", &stat_config.summary,
1279 "print summary for interval mode"),
1280 OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1281 "don't print 'summary' for CSV summary output"),
1282 OPT_BOOLEAN(0, "quiet", &quiet,
1283 "don't print any output, messages or warnings (useful with record)"),
1284 OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1285 "Only enable events on applying cpu with this type "
1286 "for hybrid platform (e.g. core or atom)",
1287 parse_hybrid_type),
1288#ifdef HAVE_LIBPFM
1289 OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1290 "libpfm4 event selector. use 'perf list' to list available events",
1291 parse_libpfm_events_option),
1292#endif
1293 OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1294 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1295 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1296 "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1297 parse_control_option),
1298 OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1299 "measure I/O performance metrics provided by arch/platform",
1300 iostat_parse),
1301 OPT_END()
1302};
1303
1304static const char *const aggr_mode__string[] = {
1305 [AGGR_CORE] = "core",
1306 [AGGR_DIE] = "die",
1307 [AGGR_GLOBAL] = "global",
1308 [AGGR_NODE] = "node",
1309 [AGGR_NONE] = "none",
1310 [AGGR_SOCKET] = "socket",
1311 [AGGR_THREAD] = "thread",
1312 [AGGR_UNSET] = "unset",
1313};
1314
1315static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1316 struct perf_cpu cpu)
1317{
1318 return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1319}
1320
1321static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1322 struct perf_cpu cpu)
1323{
1324 return aggr_cpu_id__die(cpu, /*data=*/NULL);
1325}
1326
1327static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1328 struct perf_cpu cpu)
1329{
1330 return aggr_cpu_id__core(cpu, /*data=*/NULL);
1331}
1332
1333static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1334 struct perf_cpu cpu)
1335{
1336 return aggr_cpu_id__node(cpu, /*data=*/NULL);
1337}
1338
1339static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused,
1340 struct perf_cpu cpu)
1341{
1342 return aggr_cpu_id__global(cpu, /*data=*/NULL);
1343}
1344
1345static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused,
1346 struct perf_cpu cpu)
1347{
1348 return aggr_cpu_id__cpu(cpu, /*data=*/NULL);
1349}
1350
1351static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1352 aggr_get_id_t get_id, struct perf_cpu cpu)
1353{
1354 struct aggr_cpu_id id;
1355
1356 /* per-process mode - should use global aggr mode */
1357 if (cpu.cpu == -1)
1358 return get_id(config, cpu);
1359
1360 if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1361 config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1362
1363 id = config->cpus_aggr_map->map[cpu.cpu];
1364 return id;
1365}
1366
1367static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1368 struct perf_cpu cpu)
1369{
1370 return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1371}
1372
1373static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1374 struct perf_cpu cpu)
1375{
1376 return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1377}
1378
1379static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1380 struct perf_cpu cpu)
1381{
1382 return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1383}
1384
1385static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1386 struct perf_cpu cpu)
1387{
1388 return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1389}
1390
1391static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config,
1392 struct perf_cpu cpu)
1393{
1394 return perf_stat__get_aggr(config, perf_stat__get_global, cpu);
1395}
1396
1397static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config,
1398 struct perf_cpu cpu)
1399{
1400 return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu);
1401}
1402
1403static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1404{
1405 switch (aggr_mode) {
1406 case AGGR_SOCKET:
1407 return aggr_cpu_id__socket;
1408 case AGGR_DIE:
1409 return aggr_cpu_id__die;
1410 case AGGR_CORE:
1411 return aggr_cpu_id__core;
1412 case AGGR_NODE:
1413 return aggr_cpu_id__node;
1414 case AGGR_NONE:
1415 return aggr_cpu_id__cpu;
1416 case AGGR_GLOBAL:
1417 return aggr_cpu_id__global;
1418 case AGGR_THREAD:
1419 case AGGR_UNSET:
1420 case AGGR_MAX:
1421 default:
1422 return NULL;
1423 }
1424}
1425
1426static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1427{
1428 switch (aggr_mode) {
1429 case AGGR_SOCKET:
1430 return perf_stat__get_socket_cached;
1431 case AGGR_DIE:
1432 return perf_stat__get_die_cached;
1433 case AGGR_CORE:
1434 return perf_stat__get_core_cached;
1435 case AGGR_NODE:
1436 return perf_stat__get_node_cached;
1437 case AGGR_NONE:
1438 return perf_stat__get_cpu_cached;
1439 case AGGR_GLOBAL:
1440 return perf_stat__get_global_cached;
1441 case AGGR_THREAD:
1442 case AGGR_UNSET:
1443 case AGGR_MAX:
1444 default:
1445 return NULL;
1446 }
1447}
1448
1449static int perf_stat_init_aggr_mode(void)
1450{
1451 int nr;
1452 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1453
1454 if (get_id) {
1455 bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1456 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1457 get_id, /*data=*/NULL, needs_sort);
1458 if (!stat_config.aggr_map) {
1459 pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1460 return -1;
1461 }
1462 stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1463 }
1464
1465 if (stat_config.aggr_mode == AGGR_THREAD) {
1466 nr = perf_thread_map__nr(evsel_list->core.threads);
1467 stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1468 if (stat_config.aggr_map == NULL)
1469 return -ENOMEM;
1470
1471 for (int s = 0; s < nr; s++) {
1472 struct aggr_cpu_id id = aggr_cpu_id__empty();
1473
1474 id.thread_idx = s;
1475 stat_config.aggr_map->map[s] = id;
1476 }
1477 return 0;
1478 }
1479
1480 /*
1481 * The evsel_list->cpus is the base we operate on,
1482 * taking the highest cpu number to be the size of
1483 * the aggregation translate cpumap.
1484 */
1485 if (evsel_list->core.user_requested_cpus)
1486 nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1487 else
1488 nr = 0;
1489 stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1490 return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1491}
1492
1493static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1494{
1495 if (map) {
1496 WARN_ONCE(refcount_read(&map->refcnt) != 0,
1497 "cpu_aggr_map refcnt unbalanced\n");
1498 free(map);
1499 }
1500}
1501
1502static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1503{
1504 if (map && refcount_dec_and_test(&map->refcnt))
1505 cpu_aggr_map__delete(map);
1506}
1507
1508static void perf_stat__exit_aggr_mode(void)
1509{
1510 cpu_aggr_map__put(stat_config.aggr_map);
1511 cpu_aggr_map__put(stat_config.cpus_aggr_map);
1512 stat_config.aggr_map = NULL;
1513 stat_config.cpus_aggr_map = NULL;
1514}
1515
1516static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1517{
1518 struct perf_env *env = data;
1519 struct aggr_cpu_id id = aggr_cpu_id__empty();
1520
1521 if (cpu.cpu != -1)
1522 id.socket = env->cpu[cpu.cpu].socket_id;
1523
1524 return id;
1525}
1526
1527static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1528{
1529 struct perf_env *env = data;
1530 struct aggr_cpu_id id = aggr_cpu_id__empty();
1531
1532 if (cpu.cpu != -1) {
1533 /*
1534 * die_id is relative to socket, so start
1535 * with the socket ID and then add die to
1536 * make a unique ID.
1537 */
1538 id.socket = env->cpu[cpu.cpu].socket_id;
1539 id.die = env->cpu[cpu.cpu].die_id;
1540 }
1541
1542 return id;
1543}
1544
1545static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1546{
1547 struct perf_env *env = data;
1548 struct aggr_cpu_id id = aggr_cpu_id__empty();
1549
1550 if (cpu.cpu != -1) {
1551 /*
1552 * core_id is relative to socket and die,
1553 * we need a global id. So we set
1554 * socket, die id and core id
1555 */
1556 id.socket = env->cpu[cpu.cpu].socket_id;
1557 id.die = env->cpu[cpu.cpu].die_id;
1558 id.core = env->cpu[cpu.cpu].core_id;
1559 }
1560
1561 return id;
1562}
1563
1564static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data)
1565{
1566 struct perf_env *env = data;
1567 struct aggr_cpu_id id = aggr_cpu_id__empty();
1568
1569 if (cpu.cpu != -1) {
1570 /*
1571 * core_id is relative to socket and die,
1572 * we need a global id. So we set
1573 * socket, die id and core id
1574 */
1575 id.socket = env->cpu[cpu.cpu].socket_id;
1576 id.die = env->cpu[cpu.cpu].die_id;
1577 id.core = env->cpu[cpu.cpu].core_id;
1578 id.cpu = cpu;
1579 }
1580
1581 return id;
1582}
1583
1584static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1585{
1586 struct aggr_cpu_id id = aggr_cpu_id__empty();
1587
1588 id.node = perf_env__numa_node(data, cpu);
1589 return id;
1590}
1591
1592static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused,
1593 void *data __maybe_unused)
1594{
1595 struct aggr_cpu_id id = aggr_cpu_id__empty();
1596
1597 /* it always aggregates to the cpu 0 */
1598 id.cpu = (struct perf_cpu){ .cpu = 0 };
1599 return id;
1600}
1601
1602static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1603 struct perf_cpu cpu)
1604{
1605 return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1606}
1607static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1608 struct perf_cpu cpu)
1609{
1610 return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1611}
1612
1613static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1614 struct perf_cpu cpu)
1615{
1616 return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1617}
1618
1619static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused,
1620 struct perf_cpu cpu)
1621{
1622 return perf_env__get_cpu_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1623}
1624
1625static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1626 struct perf_cpu cpu)
1627{
1628 return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1629}
1630
1631static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused,
1632 struct perf_cpu cpu)
1633{
1634 return perf_env__get_global_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1635}
1636
1637static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1638{
1639 switch (aggr_mode) {
1640 case AGGR_SOCKET:
1641 return perf_env__get_socket_aggr_by_cpu;
1642 case AGGR_DIE:
1643 return perf_env__get_die_aggr_by_cpu;
1644 case AGGR_CORE:
1645 return perf_env__get_core_aggr_by_cpu;
1646 case AGGR_NODE:
1647 return perf_env__get_node_aggr_by_cpu;
1648 case AGGR_GLOBAL:
1649 return perf_env__get_global_aggr_by_cpu;
1650 case AGGR_NONE:
1651 return perf_env__get_cpu_aggr_by_cpu;
1652 case AGGR_THREAD:
1653 case AGGR_UNSET:
1654 case AGGR_MAX:
1655 default:
1656 return NULL;
1657 }
1658}
1659
1660static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1661{
1662 switch (aggr_mode) {
1663 case AGGR_SOCKET:
1664 return perf_stat__get_socket_file;
1665 case AGGR_DIE:
1666 return perf_stat__get_die_file;
1667 case AGGR_CORE:
1668 return perf_stat__get_core_file;
1669 case AGGR_NODE:
1670 return perf_stat__get_node_file;
1671 case AGGR_GLOBAL:
1672 return perf_stat__get_global_file;
1673 case AGGR_NONE:
1674 return perf_stat__get_cpu_file;
1675 case AGGR_THREAD:
1676 case AGGR_UNSET:
1677 case AGGR_MAX:
1678 default:
1679 return NULL;
1680 }
1681}
1682
1683static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1684{
1685 struct perf_env *env = &st->session->header.env;
1686 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1687 bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1688
1689 if (stat_config.aggr_mode == AGGR_THREAD) {
1690 int nr = perf_thread_map__nr(evsel_list->core.threads);
1691
1692 stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1693 if (stat_config.aggr_map == NULL)
1694 return -ENOMEM;
1695
1696 for (int s = 0; s < nr; s++) {
1697 struct aggr_cpu_id id = aggr_cpu_id__empty();
1698
1699 id.thread_idx = s;
1700 stat_config.aggr_map->map[s] = id;
1701 }
1702 return 0;
1703 }
1704
1705 if (!get_id)
1706 return 0;
1707
1708 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1709 get_id, env, needs_sort);
1710 if (!stat_config.aggr_map) {
1711 pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1712 return -1;
1713 }
1714 stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1715 return 0;
1716}
1717
1718/*
1719 * Add default attributes, if there were no attributes specified or
1720 * if -d/--detailed, -d -d or -d -d -d is used:
1721 */
1722static int add_default_attributes(void)
1723{
1724 int err;
1725 struct perf_event_attr default_attrs0[] = {
1726
1727 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1728 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1729 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1730 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1731
1732 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1733};
1734 struct perf_event_attr frontend_attrs[] = {
1735 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1736};
1737 struct perf_event_attr backend_attrs[] = {
1738 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1739};
1740 struct perf_event_attr default_attrs1[] = {
1741 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1742 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1743 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1744
1745};
1746
1747/*
1748 * Detailed stats (-d), covering the L1 and last level data caches:
1749 */
1750 struct perf_event_attr detailed_attrs[] = {
1751
1752 { .type = PERF_TYPE_HW_CACHE,
1753 .config =
1754 PERF_COUNT_HW_CACHE_L1D << 0 |
1755 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1756 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1757
1758 { .type = PERF_TYPE_HW_CACHE,
1759 .config =
1760 PERF_COUNT_HW_CACHE_L1D << 0 |
1761 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1762 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1763
1764 { .type = PERF_TYPE_HW_CACHE,
1765 .config =
1766 PERF_COUNT_HW_CACHE_LL << 0 |
1767 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1768 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1769
1770 { .type = PERF_TYPE_HW_CACHE,
1771 .config =
1772 PERF_COUNT_HW_CACHE_LL << 0 |
1773 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1774 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1775};
1776
1777/*
1778 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1779 */
1780 struct perf_event_attr very_detailed_attrs[] = {
1781
1782 { .type = PERF_TYPE_HW_CACHE,
1783 .config =
1784 PERF_COUNT_HW_CACHE_L1I << 0 |
1785 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1786 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1787
1788 { .type = PERF_TYPE_HW_CACHE,
1789 .config =
1790 PERF_COUNT_HW_CACHE_L1I << 0 |
1791 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1792 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1793
1794 { .type = PERF_TYPE_HW_CACHE,
1795 .config =
1796 PERF_COUNT_HW_CACHE_DTLB << 0 |
1797 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1798 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1799
1800 { .type = PERF_TYPE_HW_CACHE,
1801 .config =
1802 PERF_COUNT_HW_CACHE_DTLB << 0 |
1803 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1804 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1805
1806 { .type = PERF_TYPE_HW_CACHE,
1807 .config =
1808 PERF_COUNT_HW_CACHE_ITLB << 0 |
1809 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1810 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1811
1812 { .type = PERF_TYPE_HW_CACHE,
1813 .config =
1814 PERF_COUNT_HW_CACHE_ITLB << 0 |
1815 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1816 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1817
1818};
1819
1820/*
1821 * Very, very detailed stats (-d -d -d), adding prefetch events:
1822 */
1823 struct perf_event_attr very_very_detailed_attrs[] = {
1824
1825 { .type = PERF_TYPE_HW_CACHE,
1826 .config =
1827 PERF_COUNT_HW_CACHE_L1D << 0 |
1828 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1829 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1830
1831 { .type = PERF_TYPE_HW_CACHE,
1832 .config =
1833 PERF_COUNT_HW_CACHE_L1D << 0 |
1834 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1835 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1836};
1837
1838 struct perf_event_attr default_null_attrs[] = {};
1839
1840 /* Set attrs if no event is selected and !null_run: */
1841 if (stat_config.null_run)
1842 return 0;
1843
1844 if (transaction_run) {
1845 struct parse_events_error errinfo;
1846 /* Handle -T as -M transaction. Once platform specific metrics
1847 * support has been added to the json files, all architectures
1848 * will use this approach. To determine transaction support
1849 * on an architecture test for such a metric name.
1850 */
1851 if (metricgroup__has_metric("transaction")) {
1852 return metricgroup__parse_groups(evsel_list, "transaction",
1853 stat_config.metric_no_group,
1854 stat_config.metric_no_merge,
1855 stat_config.user_requested_cpu_list,
1856 stat_config.system_wide,
1857 &stat_config.metric_events);
1858 }
1859
1860 parse_events_error__init(&errinfo);
1861 if (pmu_have_event("cpu", "cycles-ct") &&
1862 pmu_have_event("cpu", "el-start"))
1863 err = parse_events(evsel_list, transaction_attrs,
1864 &errinfo);
1865 else
1866 err = parse_events(evsel_list,
1867 transaction_limited_attrs,
1868 &errinfo);
1869 if (err) {
1870 fprintf(stderr, "Cannot set up transaction events\n");
1871 parse_events_error__print(&errinfo, transaction_attrs);
1872 }
1873 parse_events_error__exit(&errinfo);
1874 return err ? -1 : 0;
1875 }
1876
1877 if (smi_cost) {
1878 struct parse_events_error errinfo;
1879 int smi;
1880
1881 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1882 fprintf(stderr, "freeze_on_smi is not supported.\n");
1883 return -1;
1884 }
1885
1886 if (!smi) {
1887 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1888 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1889 return -1;
1890 }
1891 smi_reset = true;
1892 }
1893
1894 if (!pmu_have_event("msr", "aperf") ||
1895 !pmu_have_event("msr", "smi")) {
1896 fprintf(stderr, "To measure SMI cost, it needs "
1897 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1898 return -1;
1899 }
1900 if (!force_metric_only)
1901 stat_config.metric_only = true;
1902
1903 parse_events_error__init(&errinfo);
1904 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1905 if (err) {
1906 parse_events_error__print(&errinfo, smi_cost_attrs);
1907 fprintf(stderr, "Cannot set up SMI cost events\n");
1908 }
1909 parse_events_error__exit(&errinfo);
1910 return err ? -1 : 0;
1911 }
1912
1913 if (topdown_run) {
1914 const char **metric_attrs = topdown_metric_attrs;
1915 unsigned int max_level = 1;
1916 char *str = NULL;
1917 bool warn = false;
1918 const char *pmu_name = arch_get_topdown_pmu_name(evsel_list, true);
1919
1920 if (!force_metric_only)
1921 stat_config.metric_only = true;
1922
1923 if (pmu_have_event(pmu_name, topdown_metric_L2_attrs[5])) {
1924 metric_attrs = topdown_metric_L2_attrs;
1925 max_level = 2;
1926 }
1927
1928 if (stat_config.topdown_level > max_level) {
1929 pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1930 return -1;
1931 } else if (!stat_config.topdown_level)
1932 stat_config.topdown_level = max_level;
1933
1934 if (topdown_filter_events(metric_attrs, &str, 1, pmu_name) < 0) {
1935 pr_err("Out of memory\n");
1936 return -1;
1937 }
1938
1939 if (metric_attrs[0] && str) {
1940 if (!stat_config.interval && !stat_config.metric_only) {
1941 fprintf(stat_config.output,
1942 "Topdown accuracy may decrease when measuring long periods.\n"
1943 "Please print the result regularly, e.g. -I1000\n");
1944 }
1945 goto setup_metrics;
1946 }
1947
1948 zfree(&str);
1949
1950 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1951 stat_config.aggr_mode != AGGR_CORE) {
1952 pr_err("top down event configuration requires --per-core mode\n");
1953 return -1;
1954 }
1955 stat_config.aggr_mode = AGGR_CORE;
1956 if (nr_cgroups || !target__has_cpu(&target)) {
1957 pr_err("top down event configuration requires system-wide mode (-a)\n");
1958 return -1;
1959 }
1960
1961 if (topdown_filter_events(topdown_attrs, &str,
1962 arch_topdown_check_group(&warn),
1963 pmu_name) < 0) {
1964 pr_err("Out of memory\n");
1965 return -1;
1966 }
1967
1968 if (topdown_attrs[0] && str) {
1969 struct parse_events_error errinfo;
1970 if (warn)
1971 arch_topdown_group_warn();
1972setup_metrics:
1973 parse_events_error__init(&errinfo);
1974 err = parse_events(evsel_list, str, &errinfo);
1975 if (err) {
1976 fprintf(stderr,
1977 "Cannot set up top down events %s: %d\n",
1978 str, err);
1979 parse_events_error__print(&errinfo, str);
1980 parse_events_error__exit(&errinfo);
1981 free(str);
1982 return -1;
1983 }
1984 parse_events_error__exit(&errinfo);
1985 } else {
1986 fprintf(stderr, "System does not support topdown\n");
1987 return -1;
1988 }
1989 free(str);
1990 }
1991
1992 if (!stat_config.topdown_level)
1993 stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1994
1995 if (!evsel_list->core.nr_entries) {
1996 if (target__has_cpu(&target))
1997 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1998
1999 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2000 return -1;
2001 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2002 if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
2003 return -1;
2004 }
2005 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2006 if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
2007 return -1;
2008 }
2009 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2010 return -1;
2011 /* Platform specific attrs */
2012 if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0)
2013 return -1;
2014 }
2015
2016 /* Detailed events get appended to the event list: */
2017
2018 if (detailed_run < 1)
2019 return 0;
2020
2021 /* Append detailed run extra attributes: */
2022 if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2023 return -1;
2024
2025 if (detailed_run < 2)
2026 return 0;
2027
2028 /* Append very detailed run extra attributes: */
2029 if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2030 return -1;
2031
2032 if (detailed_run < 3)
2033 return 0;
2034
2035 /* Append very, very detailed run extra attributes: */
2036 return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2037}
2038
2039static const char * const stat_record_usage[] = {
2040 "perf stat record [<options>]",
2041 NULL,
2042};
2043
2044static void init_features(struct perf_session *session)
2045{
2046 int feat;
2047
2048 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2049 perf_header__set_feat(&session->header, feat);
2050
2051 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
2052 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2053 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2054 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2055 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2056}
2057
2058static int __cmd_record(int argc, const char **argv)
2059{
2060 struct perf_session *session;
2061 struct perf_data *data = &perf_stat.data;
2062
2063 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2064 PARSE_OPT_STOP_AT_NON_OPTION);
2065
2066 if (output_name)
2067 data->path = output_name;
2068
2069 if (stat_config.run_count != 1 || forever) {
2070 pr_err("Cannot use -r option with perf stat record.\n");
2071 return -1;
2072 }
2073
2074 session = perf_session__new(data, NULL);
2075 if (IS_ERR(session)) {
2076 pr_err("Perf session creation failed\n");
2077 return PTR_ERR(session);
2078 }
2079
2080 init_features(session);
2081
2082 session->evlist = evsel_list;
2083 perf_stat.session = session;
2084 perf_stat.record = true;
2085 return argc;
2086}
2087
2088static int process_stat_round_event(struct perf_session *session,
2089 union perf_event *event)
2090{
2091 struct perf_record_stat_round *stat_round = &event->stat_round;
2092 struct timespec tsh, *ts = NULL;
2093 const char **argv = session->header.env.cmdline_argv;
2094 int argc = session->header.env.nr_cmdline;
2095
2096 process_counters();
2097
2098 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2099 update_stats(&walltime_nsecs_stats, stat_round->time);
2100
2101 if (stat_config.interval && stat_round->time) {
2102 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2103 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2104 ts = &tsh;
2105 }
2106
2107 print_counters(ts, argc, argv);
2108 return 0;
2109}
2110
2111static
2112int process_stat_config_event(struct perf_session *session,
2113 union perf_event *event)
2114{
2115 struct perf_tool *tool = session->tool;
2116 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2117
2118 perf_event__read_stat_config(&stat_config, &event->stat_config);
2119
2120 if (perf_cpu_map__empty(st->cpus)) {
2121 if (st->aggr_mode != AGGR_UNSET)
2122 pr_warning("warning: processing task data, aggregation mode not set\n");
2123 } else if (st->aggr_mode != AGGR_UNSET) {
2124 stat_config.aggr_mode = st->aggr_mode;
2125 }
2126
2127 if (perf_stat.data.is_pipe)
2128 perf_stat_init_aggr_mode();
2129 else
2130 perf_stat_init_aggr_mode_file(st);
2131
2132 if (stat_config.aggr_map) {
2133 int nr_aggr = stat_config.aggr_map->nr;
2134
2135 if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) {
2136 pr_err("cannot allocate aggr counts\n");
2137 return -1;
2138 }
2139 }
2140 return 0;
2141}
2142
2143static int set_maps(struct perf_stat *st)
2144{
2145 if (!st->cpus || !st->threads)
2146 return 0;
2147
2148 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2149 return -EINVAL;
2150
2151 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2152
2153 if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true))
2154 return -ENOMEM;
2155
2156 st->maps_allocated = true;
2157 return 0;
2158}
2159
2160static
2161int process_thread_map_event(struct perf_session *session,
2162 union perf_event *event)
2163{
2164 struct perf_tool *tool = session->tool;
2165 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2166
2167 if (st->threads) {
2168 pr_warning("Extra thread map event, ignoring.\n");
2169 return 0;
2170 }
2171
2172 st->threads = thread_map__new_event(&event->thread_map);
2173 if (!st->threads)
2174 return -ENOMEM;
2175
2176 return set_maps(st);
2177}
2178
2179static
2180int process_cpu_map_event(struct perf_session *session,
2181 union perf_event *event)
2182{
2183 struct perf_tool *tool = session->tool;
2184 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2185 struct perf_cpu_map *cpus;
2186
2187 if (st->cpus) {
2188 pr_warning("Extra cpu map event, ignoring.\n");
2189 return 0;
2190 }
2191
2192 cpus = cpu_map__new_data(&event->cpu_map.data);
2193 if (!cpus)
2194 return -ENOMEM;
2195
2196 st->cpus = cpus;
2197 return set_maps(st);
2198}
2199
2200static const char * const stat_report_usage[] = {
2201 "perf stat report [<options>]",
2202 NULL,
2203};
2204
2205static struct perf_stat perf_stat = {
2206 .tool = {
2207 .attr = perf_event__process_attr,
2208 .event_update = perf_event__process_event_update,
2209 .thread_map = process_thread_map_event,
2210 .cpu_map = process_cpu_map_event,
2211 .stat_config = process_stat_config_event,
2212 .stat = perf_event__process_stat_event,
2213 .stat_round = process_stat_round_event,
2214 },
2215 .aggr_mode = AGGR_UNSET,
2216};
2217
2218static int __cmd_report(int argc, const char **argv)
2219{
2220 struct perf_session *session;
2221 const struct option options[] = {
2222 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2223 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2224 "aggregate counts per processor socket", AGGR_SOCKET),
2225 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2226 "aggregate counts per processor die", AGGR_DIE),
2227 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2228 "aggregate counts per physical processor core", AGGR_CORE),
2229 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2230 "aggregate counts per numa node", AGGR_NODE),
2231 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2232 "disable CPU count aggregation", AGGR_NONE),
2233 OPT_END()
2234 };
2235 struct stat st;
2236 int ret;
2237
2238 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2239
2240 if (!input_name || !strlen(input_name)) {
2241 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2242 input_name = "-";
2243 else
2244 input_name = "perf.data";
2245 }
2246
2247 perf_stat__init_shadow_stats();
2248
2249 perf_stat.data.path = input_name;
2250 perf_stat.data.mode = PERF_DATA_MODE_READ;
2251
2252 session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2253 if (IS_ERR(session))
2254 return PTR_ERR(session);
2255
2256 perf_stat.session = session;
2257 stat_config.output = stderr;
2258 evsel_list = session->evlist;
2259
2260 ret = perf_session__process_events(session);
2261 if (ret)
2262 return ret;
2263
2264 perf_session__delete(session);
2265 return 0;
2266}
2267
2268static void setup_system_wide(int forks)
2269{
2270 /*
2271 * Make system wide (-a) the default target if
2272 * no target was specified and one of following
2273 * conditions is met:
2274 *
2275 * - there's no workload specified
2276 * - there is workload specified but all requested
2277 * events are system wide events
2278 */
2279 if (!target__none(&target))
2280 return;
2281
2282 if (!forks)
2283 target.system_wide = true;
2284 else {
2285 struct evsel *counter;
2286
2287 evlist__for_each_entry(evsel_list, counter) {
2288 if (!counter->core.requires_cpu &&
2289 strcmp(counter->name, "duration_time")) {
2290 return;
2291 }
2292 }
2293
2294 if (evsel_list->core.nr_entries)
2295 target.system_wide = true;
2296 }
2297}
2298
2299int cmd_stat(int argc, const char **argv)
2300{
2301 const char * const stat_usage[] = {
2302 "perf stat [<options>] [<command>]",
2303 NULL
2304 };
2305 int status = -EINVAL, run_idx, err;
2306 const char *mode;
2307 FILE *output = stderr;
2308 unsigned int interval, timeout;
2309 const char * const stat_subcommands[] = { "record", "report" };
2310 char errbuf[BUFSIZ];
2311
2312 setlocale(LC_ALL, "");
2313
2314 evsel_list = evlist__new();
2315 if (evsel_list == NULL)
2316 return -ENOMEM;
2317
2318 parse_events__shrink_config_terms();
2319
2320 /* String-parsing callback-based options would segfault when negated */
2321 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2322 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2323 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2324
2325 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2326 (const char **) stat_usage,
2327 PARSE_OPT_STOP_AT_NON_OPTION);
2328
2329 if (stat_config.csv_sep) {
2330 stat_config.csv_output = true;
2331 if (!strcmp(stat_config.csv_sep, "\\t"))
2332 stat_config.csv_sep = "\t";
2333 } else
2334 stat_config.csv_sep = DEFAULT_SEPARATOR;
2335
2336 if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2337 argc = __cmd_record(argc, argv);
2338 if (argc < 0)
2339 return -1;
2340 } else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2341 return __cmd_report(argc, argv);
2342
2343 interval = stat_config.interval;
2344 timeout = stat_config.timeout;
2345
2346 /*
2347 * For record command the -o is already taken care of.
2348 */
2349 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2350 output = NULL;
2351
2352 if (output_name && output_fd) {
2353 fprintf(stderr, "cannot use both --output and --log-fd\n");
2354 parse_options_usage(stat_usage, stat_options, "o", 1);
2355 parse_options_usage(NULL, stat_options, "log-fd", 0);
2356 goto out;
2357 }
2358
2359 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2360 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2361 goto out;
2362 }
2363
2364 if (stat_config.metric_only && stat_config.run_count > 1) {
2365 fprintf(stderr, "--metric-only is not supported with -r\n");
2366 goto out;
2367 }
2368
2369 if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2370 fprintf(stderr, "--table is only supported with -r\n");
2371 parse_options_usage(stat_usage, stat_options, "r", 1);
2372 parse_options_usage(NULL, stat_options, "table", 0);
2373 goto out;
2374 }
2375
2376 if (output_fd < 0) {
2377 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2378 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2379 goto out;
2380 }
2381
2382 if (!output && !quiet) {
2383 struct timespec tm;
2384 mode = append_file ? "a" : "w";
2385
2386 output = fopen(output_name, mode);
2387 if (!output) {
2388 perror("failed to create output file");
2389 return -1;
2390 }
2391 clock_gettime(CLOCK_REALTIME, &tm);
2392 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2393 } else if (output_fd > 0) {
2394 mode = append_file ? "a" : "w";
2395 output = fdopen(output_fd, mode);
2396 if (!output) {
2397 perror("Failed opening logfd");
2398 return -errno;
2399 }
2400 }
2401
2402 if (stat_config.interval_clear && !isatty(fileno(output))) {
2403 fprintf(stderr, "--interval-clear does not work with output\n");
2404 parse_options_usage(stat_usage, stat_options, "o", 1);
2405 parse_options_usage(NULL, stat_options, "log-fd", 0);
2406 parse_options_usage(NULL, stat_options, "interval-clear", 0);
2407 return -1;
2408 }
2409
2410 stat_config.output = output;
2411
2412 /*
2413 * let the spreadsheet do the pretty-printing
2414 */
2415 if (stat_config.csv_output) {
2416 /* User explicitly passed -B? */
2417 if (big_num_opt == 1) {
2418 fprintf(stderr, "-B option not supported with -x\n");
2419 parse_options_usage(stat_usage, stat_options, "B", 1);
2420 parse_options_usage(NULL, stat_options, "x", 1);
2421 goto out;
2422 } else /* Nope, so disable big number formatting */
2423 stat_config.big_num = false;
2424 } else if (big_num_opt == 0) /* User passed --no-big-num */
2425 stat_config.big_num = false;
2426
2427 err = target__validate(&target);
2428 if (err) {
2429 target__strerror(&target, err, errbuf, BUFSIZ);
2430 pr_warning("%s\n", errbuf);
2431 }
2432
2433 setup_system_wide(argc);
2434
2435 /*
2436 * Display user/system times only for single
2437 * run and when there's specified tracee.
2438 */
2439 if ((stat_config.run_count == 1) && target__none(&target))
2440 stat_config.ru_display = true;
2441
2442 if (stat_config.run_count < 0) {
2443 pr_err("Run count must be a positive number\n");
2444 parse_options_usage(stat_usage, stat_options, "r", 1);
2445 goto out;
2446 } else if (stat_config.run_count == 0) {
2447 forever = true;
2448 stat_config.run_count = 1;
2449 }
2450
2451 if (stat_config.walltime_run_table) {
2452 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2453 if (!stat_config.walltime_run) {
2454 pr_err("failed to setup -r option");
2455 goto out;
2456 }
2457 }
2458
2459 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2460 !target__has_task(&target)) {
2461 if (!target.system_wide || target.cpu_list) {
2462 fprintf(stderr, "The --per-thread option is only "
2463 "available when monitoring via -p -t -a "
2464 "options or only --per-thread.\n");
2465 parse_options_usage(NULL, stat_options, "p", 1);
2466 parse_options_usage(NULL, stat_options, "t", 1);
2467 goto out;
2468 }
2469 }
2470
2471 /*
2472 * no_aggr, cgroup are for system-wide only
2473 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2474 */
2475 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2476 stat_config.aggr_mode != AGGR_THREAD) ||
2477 (nr_cgroups || stat_config.cgroup_list)) &&
2478 !target__has_cpu(&target)) {
2479 fprintf(stderr, "both cgroup and no-aggregation "
2480 "modes only available in system-wide mode\n");
2481
2482 parse_options_usage(stat_usage, stat_options, "G", 1);
2483 parse_options_usage(NULL, stat_options, "A", 1);
2484 parse_options_usage(NULL, stat_options, "a", 1);
2485 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2486 goto out;
2487 }
2488
2489 if (stat_config.iostat_run) {
2490 status = iostat_prepare(evsel_list, &stat_config);
2491 if (status)
2492 goto out;
2493 if (iostat_mode == IOSTAT_LIST) {
2494 iostat_list(evsel_list, &stat_config);
2495 goto out;
2496 } else if (verbose > 0)
2497 iostat_list(evsel_list, &stat_config);
2498 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2499 target.system_wide = true;
2500 }
2501
2502 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2503 target.per_thread = true;
2504
2505 stat_config.system_wide = target.system_wide;
2506 if (target.cpu_list) {
2507 stat_config.user_requested_cpu_list = strdup(target.cpu_list);
2508 if (!stat_config.user_requested_cpu_list) {
2509 status = -ENOMEM;
2510 goto out;
2511 }
2512 }
2513
2514 /*
2515 * Metric parsing needs to be delayed as metrics may optimize events
2516 * knowing the target is system-wide.
2517 */
2518 if (metrics) {
2519 metricgroup__parse_groups(evsel_list, metrics,
2520 stat_config.metric_no_group,
2521 stat_config.metric_no_merge,
2522 stat_config.user_requested_cpu_list,
2523 stat_config.system_wide,
2524 &stat_config.metric_events);
2525 zfree(&metrics);
2526 }
2527 perf_stat__collect_metric_expr(evsel_list);
2528 perf_stat__init_shadow_stats();
2529
2530 if (add_default_attributes())
2531 goto out;
2532
2533 if (stat_config.cgroup_list) {
2534 if (nr_cgroups > 0) {
2535 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2536 parse_options_usage(stat_usage, stat_options, "G", 1);
2537 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2538 goto out;
2539 }
2540
2541 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2542 &stat_config.metric_events, true) < 0) {
2543 parse_options_usage(stat_usage, stat_options,
2544 "for-each-cgroup", 0);
2545 goto out;
2546 }
2547 }
2548
2549 if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2550 pr_err("failed to use cpu list %s\n", target.cpu_list);
2551 goto out;
2552 }
2553
2554 target.hybrid = perf_pmu__has_hybrid();
2555 if (evlist__create_maps(evsel_list, &target) < 0) {
2556 if (target__has_task(&target)) {
2557 pr_err("Problems finding threads of monitor\n");
2558 parse_options_usage(stat_usage, stat_options, "p", 1);
2559 parse_options_usage(NULL, stat_options, "t", 1);
2560 } else if (target__has_cpu(&target)) {
2561 perror("failed to parse CPUs map");
2562 parse_options_usage(stat_usage, stat_options, "C", 1);
2563 parse_options_usage(NULL, stat_options, "a", 1);
2564 }
2565 goto out;
2566 }
2567
2568 evlist__check_cpu_maps(evsel_list);
2569
2570 /*
2571 * Initialize thread_map with comm names,
2572 * so we could print it out on output.
2573 */
2574 if (stat_config.aggr_mode == AGGR_THREAD) {
2575 thread_map__read_comms(evsel_list->core.threads);
2576 }
2577
2578 if (stat_config.aggr_mode == AGGR_NODE)
2579 cpu__setup_cpunode_map();
2580
2581 if (stat_config.times && interval)
2582 interval_count = true;
2583 else if (stat_config.times && !interval) {
2584 pr_err("interval-count option should be used together with "
2585 "interval-print.\n");
2586 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2587 parse_options_usage(stat_usage, stat_options, "I", 1);
2588 goto out;
2589 }
2590
2591 if (timeout && timeout < 100) {
2592 if (timeout < 10) {
2593 pr_err("timeout must be >= 10ms.\n");
2594 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2595 goto out;
2596 } else
2597 pr_warning("timeout < 100ms. "
2598 "The overhead percentage could be high in some cases. "
2599 "Please proceed with caution.\n");
2600 }
2601 if (timeout && interval) {
2602 pr_err("timeout option is not supported with interval-print.\n");
2603 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2604 parse_options_usage(stat_usage, stat_options, "I", 1);
2605 goto out;
2606 }
2607
2608 if (perf_stat_init_aggr_mode())
2609 goto out;
2610
2611 if (evlist__alloc_stats(&stat_config, evsel_list, interval))
2612 goto out;
2613
2614 /*
2615 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2616 * while avoiding that older tools show confusing messages.
2617 *
2618 * However for pipe sessions we need to keep it zero,
2619 * because script's perf_evsel__check_attr is triggered
2620 * by attr->sample_type != 0, and we can't run it on
2621 * stat sessions.
2622 */
2623 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2624
2625 /*
2626 * We dont want to block the signals - that would cause
2627 * child tasks to inherit that and Ctrl-C would not work.
2628 * What we want is for Ctrl-C to work in the exec()-ed
2629 * task, but being ignored by perf stat itself:
2630 */
2631 atexit(sig_atexit);
2632 if (!forever)
2633 signal(SIGINT, skip_signal);
2634 signal(SIGCHLD, skip_signal);
2635 signal(SIGALRM, skip_signal);
2636 signal(SIGABRT, skip_signal);
2637
2638 if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2639 goto out;
2640
2641 /* Enable ignoring missing threads when -p option is defined. */
2642 evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2643 status = 0;
2644 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2645 if (stat_config.run_count != 1 && verbose > 0)
2646 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2647 run_idx + 1);
2648
2649 if (run_idx != 0)
2650 evlist__reset_prev_raw_counts(evsel_list);
2651
2652 status = run_perf_stat(argc, argv, run_idx);
2653 if (forever && status != -1 && !interval) {
2654 print_counters(NULL, argc, argv);
2655 perf_stat__reset_stats();
2656 }
2657 }
2658
2659 if (!forever && status != -1 && (!interval || stat_config.summary))
2660 print_counters(NULL, argc, argv);
2661
2662 evlist__finalize_ctlfd(evsel_list);
2663
2664 if (STAT_RECORD) {
2665 /*
2666 * We synthesize the kernel mmap record just so that older tools
2667 * don't emit warnings about not being able to resolve symbols
2668 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2669 * a saner message about no samples being in the perf.data file.
2670 *
2671 * This also serves to suppress a warning about f_header.data.size == 0
2672 * in header.c at the moment 'perf stat record' gets introduced, which
2673 * is not really needed once we start adding the stat specific PERF_RECORD_
2674 * records, but the need to suppress the kptr_restrict messages in older
2675 * tools remain -acme
2676 */
2677 int fd = perf_data__fd(&perf_stat.data);
2678
2679 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2680 process_synthesized_event,
2681 &perf_stat.session->machines.host);
2682 if (err) {
2683 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2684 "older tools may produce warnings about this file\n.");
2685 }
2686
2687 if (!interval) {
2688 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2689 pr_err("failed to write stat round event\n");
2690 }
2691
2692 if (!perf_stat.data.is_pipe) {
2693 perf_stat.session->header.data_size += perf_stat.bytes_written;
2694 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2695 }
2696
2697 evlist__close(evsel_list);
2698 perf_session__delete(perf_stat.session);
2699 }
2700
2701 perf_stat__exit_aggr_mode();
2702 evlist__free_stats(evsel_list);
2703out:
2704 if (stat_config.iostat_run)
2705 iostat_release(evsel_list);
2706
2707 zfree(&stat_config.walltime_run);
2708 zfree(&stat_config.user_requested_cpu_list);
2709
2710 if (smi_cost && smi_reset)
2711 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2712
2713 evlist__delete(evsel_list);
2714
2715 metricgroup__rblist_exit(&stat_config.metric_events);
2716 evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2717
2718 return status;
2719}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * builtin-stat.c
4 *
5 * Builtin stat command: Give a precise performance counters summary
6 * overview about any workload, CPU or specific PID.
7 *
8 * Sample output:
9
10 $ perf stat ./hackbench 10
11
12 Time: 0.118
13
14 Performance counter stats for './hackbench 10':
15
16 1708.761321 task-clock # 11.037 CPUs utilized
17 41,190 context-switches # 0.024 M/sec
18 6,735 CPU-migrations # 0.004 M/sec
19 17,318 page-faults # 0.010 M/sec
20 5,205,202,243 cycles # 3.046 GHz
21 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
22 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
23 2,603,501,247 instructions # 0.50 insns per cycle
24 # 1.48 stalled cycles per insn
25 484,357,498 branches # 283.455 M/sec
26 6,388,934 branch-misses # 1.32% of all branches
27
28 0.154822978 seconds time elapsed
29
30 *
31 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 *
33 * Improvements and fixes by:
34 *
35 * Arjan van de Ven <arjan@linux.intel.com>
36 * Yanmin Zhang <yanmin.zhang@intel.com>
37 * Wu Fengguang <fengguang.wu@intel.com>
38 * Mike Galbraith <efault@gmx.de>
39 * Paul Mackerras <paulus@samba.org>
40 * Jaswinder Singh Rajput <jaswinder@kernel.org>
41 */
42
43#include "builtin.h"
44#include "perf.h"
45#include "util/cgroup.h"
46#include <subcmd/parse-options.h>
47#include "util/parse-events.h"
48#include "util/pmu.h"
49#include "util/event.h"
50#include "util/evlist.h"
51#include "util/evlist-hybrid.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_map.h"
59#include "util/counts.h"
60#include "util/topdown.h"
61#include "util/session.h"
62#include "util/tool.h"
63#include "util/string2.h"
64#include "util/metricgroup.h"
65#include "util/synthetic-events.h"
66#include "util/target.h"
67#include "util/time-utils.h"
68#include "util/top.h"
69#include "util/affinity.h"
70#include "util/pfm.h"
71#include "util/bpf_counter.h"
72#include "util/iostat.h"
73#include "util/pmu-hybrid.h"
74#include "asm/bug.h"
75
76#include <linux/time64.h>
77#include <linux/zalloc.h>
78#include <api/fs/fs.h>
79#include <errno.h>
80#include <signal.h>
81#include <stdlib.h>
82#include <sys/prctl.h>
83#include <inttypes.h>
84#include <locale.h>
85#include <math.h>
86#include <sys/types.h>
87#include <sys/stat.h>
88#include <sys/wait.h>
89#include <unistd.h>
90#include <sys/time.h>
91#include <sys/resource.h>
92#include <linux/err.h>
93
94#include <linux/ctype.h>
95#include <perf/evlist.h>
96
97#define DEFAULT_SEPARATOR " "
98#define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
99
100static void print_counters(struct timespec *ts, int argc, const char **argv);
101
102/* Default events used for perf stat -T */
103static const char *transaction_attrs = {
104 "task-clock,"
105 "{"
106 "instructions,"
107 "cycles,"
108 "cpu/cycles-t/,"
109 "cpu/tx-start/,"
110 "cpu/el-start/,"
111 "cpu/cycles-ct/"
112 "}"
113};
114
115/* More limited version when the CPU does not have all events. */
116static const char * transaction_limited_attrs = {
117 "task-clock,"
118 "{"
119 "instructions,"
120 "cycles,"
121 "cpu/cycles-t/,"
122 "cpu/tx-start/"
123 "}"
124};
125
126static const char * topdown_attrs[] = {
127 "topdown-total-slots",
128 "topdown-slots-retired",
129 "topdown-recovery-bubbles",
130 "topdown-fetch-bubbles",
131 "topdown-slots-issued",
132 NULL,
133};
134
135static const char *topdown_metric_attrs[] = {
136 "slots",
137 "topdown-retiring",
138 "topdown-bad-spec",
139 "topdown-fe-bound",
140 "topdown-be-bound",
141 NULL,
142};
143
144static const char *topdown_metric_L2_attrs[] = {
145 "slots",
146 "topdown-retiring",
147 "topdown-bad-spec",
148 "topdown-fe-bound",
149 "topdown-be-bound",
150 "topdown-heavy-ops",
151 "topdown-br-mispredict",
152 "topdown-fetch-lat",
153 "topdown-mem-bound",
154 NULL,
155};
156
157#define TOPDOWN_MAX_LEVEL 2
158
159static const char *smi_cost_attrs = {
160 "{"
161 "msr/aperf/,"
162 "msr/smi/,"
163 "cycles"
164 "}"
165};
166
167static struct evlist *evsel_list;
168static bool all_counters_use_bpf = true;
169
170static struct target target = {
171 .uid = UINT_MAX,
172};
173
174#define METRIC_ONLY_LEN 20
175
176static volatile pid_t child_pid = -1;
177static int detailed_run = 0;
178static bool transaction_run;
179static bool topdown_run = false;
180static bool smi_cost = false;
181static bool smi_reset = false;
182static int big_num_opt = -1;
183static bool group = false;
184static const char *pre_cmd = NULL;
185static const char *post_cmd = NULL;
186static bool sync_run = false;
187static bool forever = false;
188static bool force_metric_only = false;
189static struct timespec ref_time;
190static bool append_file;
191static bool interval_count;
192static const char *output_name;
193static int output_fd;
194
195struct perf_stat {
196 bool record;
197 struct perf_data data;
198 struct perf_session *session;
199 u64 bytes_written;
200 struct perf_tool tool;
201 bool maps_allocated;
202 struct perf_cpu_map *cpus;
203 struct perf_thread_map *threads;
204 enum aggr_mode aggr_mode;
205};
206
207static struct perf_stat perf_stat;
208#define STAT_RECORD perf_stat.record
209
210static volatile int done = 0;
211
212static struct perf_stat_config stat_config = {
213 .aggr_mode = AGGR_GLOBAL,
214 .scale = true,
215 .unit_width = 4, /* strlen("unit") */
216 .run_count = 1,
217 .metric_only_len = METRIC_ONLY_LEN,
218 .walltime_nsecs_stats = &walltime_nsecs_stats,
219 .big_num = true,
220 .ctl_fd = -1,
221 .ctl_fd_ack = -1,
222 .iostat_run = false,
223};
224
225static bool cpus_map_matched(struct evsel *a, struct evsel *b)
226{
227 if (!a->core.cpus && !b->core.cpus)
228 return true;
229
230 if (!a->core.cpus || !b->core.cpus)
231 return false;
232
233 if (a->core.cpus->nr != b->core.cpus->nr)
234 return false;
235
236 for (int i = 0; i < a->core.cpus->nr; i++) {
237 if (a->core.cpus->map[i] != b->core.cpus->map[i])
238 return false;
239 }
240
241 return true;
242}
243
244static void evlist__check_cpu_maps(struct evlist *evlist)
245{
246 struct evsel *evsel, *pos, *leader;
247 char buf[1024];
248
249 if (evlist__has_hybrid(evlist))
250 evlist__warn_hybrid_group(evlist);
251
252 evlist__for_each_entry(evlist, evsel) {
253 leader = evsel__leader(evsel);
254
255 /* Check that leader matches cpus with each member. */
256 if (leader == evsel)
257 continue;
258 if (cpus_map_matched(leader, evsel))
259 continue;
260
261 /* If there's mismatch disable the group and warn user. */
262 WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
263 evsel__group_desc(leader, buf, sizeof(buf));
264 pr_warning(" %s\n", buf);
265
266 if (verbose) {
267 cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
268 pr_warning(" %s: %s\n", leader->name, buf);
269 cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
270 pr_warning(" %s: %s\n", evsel->name, buf);
271 }
272
273 for_each_group_evsel(pos, leader) {
274 evsel__set_leader(pos, pos);
275 pos->core.nr_members = 0;
276 }
277 evsel->core.leader->nr_members = 0;
278 }
279}
280
281static inline void diff_timespec(struct timespec *r, struct timespec *a,
282 struct timespec *b)
283{
284 r->tv_sec = a->tv_sec - b->tv_sec;
285 if (a->tv_nsec < b->tv_nsec) {
286 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
287 r->tv_sec--;
288 } else {
289 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
290 }
291}
292
293static void perf_stat__reset_stats(void)
294{
295 int i;
296
297 evlist__reset_stats(evsel_list);
298 perf_stat__reset_shadow_stats();
299
300 for (i = 0; i < stat_config.stats_num; i++)
301 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
302}
303
304static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
305 union perf_event *event,
306 struct perf_sample *sample __maybe_unused,
307 struct machine *machine __maybe_unused)
308{
309 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
310 pr_err("failed to write perf data, error: %m\n");
311 return -1;
312 }
313
314 perf_stat.bytes_written += event->header.size;
315 return 0;
316}
317
318static int write_stat_round_event(u64 tm, u64 type)
319{
320 return perf_event__synthesize_stat_round(NULL, tm, type,
321 process_synthesized_event,
322 NULL);
323}
324
325#define WRITE_STAT_ROUND_EVENT(time, interval) \
326 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
327
328#define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
329
330static int evsel__write_stat_event(struct evsel *counter, u32 cpu, u32 thread,
331 struct perf_counts_values *count)
332{
333 struct perf_sample_id *sid = SID(counter, cpu, thread);
334
335 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
336 process_synthesized_event, NULL);
337}
338
339static int read_single_counter(struct evsel *counter, int cpu,
340 int thread, struct timespec *rs)
341{
342 if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
343 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
344 struct perf_counts_values *count =
345 perf_counts(counter->counts, cpu, thread);
346 count->ena = count->run = val;
347 count->val = val;
348 return 0;
349 }
350 return evsel__read_counter(counter, cpu, thread);
351}
352
353/*
354 * Read out the results of a single counter:
355 * do not aggregate counts across CPUs in system-wide mode
356 */
357static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu)
358{
359 int nthreads = perf_thread_map__nr(evsel_list->core.threads);
360 int thread;
361
362 if (!counter->supported)
363 return -ENOENT;
364
365 if (counter->core.system_wide)
366 nthreads = 1;
367
368 for (thread = 0; thread < nthreads; thread++) {
369 struct perf_counts_values *count;
370
371 count = perf_counts(counter->counts, cpu, thread);
372
373 /*
374 * The leader's group read loads data into its group members
375 * (via evsel__read_counter()) and sets their count->loaded.
376 */
377 if (!perf_counts__is_loaded(counter->counts, cpu, thread) &&
378 read_single_counter(counter, cpu, thread, rs)) {
379 counter->counts->scaled = -1;
380 perf_counts(counter->counts, cpu, thread)->ena = 0;
381 perf_counts(counter->counts, cpu, thread)->run = 0;
382 return -1;
383 }
384
385 perf_counts__set_loaded(counter->counts, cpu, thread, false);
386
387 if (STAT_RECORD) {
388 if (evsel__write_stat_event(counter, cpu, thread, count)) {
389 pr_err("failed to write stat event\n");
390 return -1;
391 }
392 }
393
394 if (verbose > 1) {
395 fprintf(stat_config.output,
396 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
397 evsel__name(counter),
398 cpu,
399 count->val, count->ena, count->run);
400 }
401 }
402
403 return 0;
404}
405
406static int read_affinity_counters(struct timespec *rs)
407{
408 struct evsel *counter;
409 struct affinity affinity;
410 int i, ncpus, cpu;
411
412 if (all_counters_use_bpf)
413 return 0;
414
415 if (affinity__setup(&affinity) < 0)
416 return -1;
417
418 ncpus = perf_cpu_map__nr(evsel_list->core.all_cpus);
419 if (!target__has_cpu(&target) || target__has_per_thread(&target))
420 ncpus = 1;
421 evlist__for_each_cpu(evsel_list, i, cpu) {
422 if (i >= ncpus)
423 break;
424 affinity__set(&affinity, cpu);
425
426 evlist__for_each_entry(evsel_list, counter) {
427 if (evsel__cpu_iter_skip(counter, cpu))
428 continue;
429 if (evsel__is_bpf(counter))
430 continue;
431 if (!counter->err) {
432 counter->err = read_counter_cpu(counter, rs,
433 counter->cpu_iter - 1);
434 }
435 }
436 }
437 affinity__cleanup(&affinity);
438 return 0;
439}
440
441static int read_bpf_map_counters(void)
442{
443 struct evsel *counter;
444 int err;
445
446 evlist__for_each_entry(evsel_list, counter) {
447 if (!evsel__is_bpf(counter))
448 continue;
449
450 err = bpf_counter__read(counter);
451 if (err)
452 return err;
453 }
454 return 0;
455}
456
457static void read_counters(struct timespec *rs)
458{
459 struct evsel *counter;
460
461 if (!stat_config.stop_read_counter) {
462 if (read_bpf_map_counters() ||
463 read_affinity_counters(rs))
464 return;
465 }
466
467 evlist__for_each_entry(evsel_list, counter) {
468 if (counter->err)
469 pr_debug("failed to read counter %s\n", counter->name);
470 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
471 pr_warning("failed to process counter %s\n", counter->name);
472 counter->err = 0;
473 }
474}
475
476static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
477{
478 int i;
479
480 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
481 if (!config->stats)
482 return -1;
483
484 config->stats_num = nthreads;
485
486 for (i = 0; i < nthreads; i++)
487 runtime_stat__init(&config->stats[i]);
488
489 return 0;
490}
491
492static void runtime_stat_delete(struct perf_stat_config *config)
493{
494 int i;
495
496 if (!config->stats)
497 return;
498
499 for (i = 0; i < config->stats_num; i++)
500 runtime_stat__exit(&config->stats[i]);
501
502 zfree(&config->stats);
503}
504
505static void runtime_stat_reset(struct perf_stat_config *config)
506{
507 int i;
508
509 if (!config->stats)
510 return;
511
512 for (i = 0; i < config->stats_num; i++)
513 perf_stat__reset_shadow_per_stat(&config->stats[i]);
514}
515
516static void process_interval(void)
517{
518 struct timespec ts, rs;
519
520 clock_gettime(CLOCK_MONOTONIC, &ts);
521 diff_timespec(&rs, &ts, &ref_time);
522
523 perf_stat__reset_shadow_per_stat(&rt_stat);
524 runtime_stat_reset(&stat_config);
525 read_counters(&rs);
526
527 if (STAT_RECORD) {
528 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
529 pr_err("failed to write stat round event\n");
530 }
531
532 init_stats(&walltime_nsecs_stats);
533 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
534 print_counters(&rs, 0, NULL);
535}
536
537static bool handle_interval(unsigned int interval, int *times)
538{
539 if (interval) {
540 process_interval();
541 if (interval_count && !(--(*times)))
542 return true;
543 }
544 return false;
545}
546
547static int enable_counters(void)
548{
549 struct evsel *evsel;
550 int err;
551
552 evlist__for_each_entry(evsel_list, evsel) {
553 if (!evsel__is_bpf(evsel))
554 continue;
555
556 err = bpf_counter__enable(evsel);
557 if (err)
558 return err;
559 }
560
561 if (stat_config.initial_delay < 0) {
562 pr_info(EVLIST_DISABLED_MSG);
563 return 0;
564 }
565
566 if (stat_config.initial_delay > 0) {
567 pr_info(EVLIST_DISABLED_MSG);
568 usleep(stat_config.initial_delay * USEC_PER_MSEC);
569 }
570
571 /*
572 * We need to enable counters only if:
573 * - we don't have tracee (attaching to task or cpu)
574 * - we have initial delay configured
575 */
576 if (!target__none(&target) || stat_config.initial_delay) {
577 if (!all_counters_use_bpf)
578 evlist__enable(evsel_list);
579 if (stat_config.initial_delay > 0)
580 pr_info(EVLIST_ENABLED_MSG);
581 }
582 return 0;
583}
584
585static void disable_counters(void)
586{
587 struct evsel *counter;
588
589 /*
590 * If we don't have tracee (attaching to task or cpu), counters may
591 * still be running. To get accurate group ratios, we must stop groups
592 * from counting before reading their constituent counters.
593 */
594 if (!target__none(&target)) {
595 evlist__for_each_entry(evsel_list, counter)
596 bpf_counter__disable(counter);
597 if (!all_counters_use_bpf)
598 evlist__disable(evsel_list);
599 }
600}
601
602static volatile int workload_exec_errno;
603
604/*
605 * evlist__prepare_workload will send a SIGUSR1
606 * if the fork fails, since we asked by setting its
607 * want_signal to true.
608 */
609static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
610 void *ucontext __maybe_unused)
611{
612 workload_exec_errno = info->si_value.sival_int;
613}
614
615static bool evsel__should_store_id(struct evsel *counter)
616{
617 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
618}
619
620static bool is_target_alive(struct target *_target,
621 struct perf_thread_map *threads)
622{
623 struct stat st;
624 int i;
625
626 if (!target__has_task(_target))
627 return true;
628
629 for (i = 0; i < threads->nr; i++) {
630 char path[PATH_MAX];
631
632 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
633 threads->map[i].pid);
634
635 if (!stat(path, &st))
636 return true;
637 }
638
639 return false;
640}
641
642static void process_evlist(struct evlist *evlist, unsigned int interval)
643{
644 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
645
646 if (evlist__ctlfd_process(evlist, &cmd) > 0) {
647 switch (cmd) {
648 case EVLIST_CTL_CMD_ENABLE:
649 if (interval)
650 process_interval();
651 break;
652 case EVLIST_CTL_CMD_DISABLE:
653 if (interval)
654 process_interval();
655 break;
656 case EVLIST_CTL_CMD_SNAPSHOT:
657 case EVLIST_CTL_CMD_ACK:
658 case EVLIST_CTL_CMD_UNSUPPORTED:
659 case EVLIST_CTL_CMD_EVLIST:
660 case EVLIST_CTL_CMD_STOP:
661 case EVLIST_CTL_CMD_PING:
662 default:
663 break;
664 }
665 }
666}
667
668static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
669 int *time_to_sleep)
670{
671 int tts = *time_to_sleep;
672 struct timespec time_diff;
673
674 diff_timespec(&time_diff, time_stop, time_start);
675
676 tts -= time_diff.tv_sec * MSEC_PER_SEC +
677 time_diff.tv_nsec / NSEC_PER_MSEC;
678
679 if (tts < 0)
680 tts = 0;
681
682 *time_to_sleep = tts;
683}
684
685static int dispatch_events(bool forks, int timeout, int interval, int *times)
686{
687 int child_exited = 0, status = 0;
688 int time_to_sleep, sleep_time;
689 struct timespec time_start, time_stop;
690
691 if (interval)
692 sleep_time = interval;
693 else if (timeout)
694 sleep_time = timeout;
695 else
696 sleep_time = 1000;
697
698 time_to_sleep = sleep_time;
699
700 while (!done) {
701 if (forks)
702 child_exited = waitpid(child_pid, &status, WNOHANG);
703 else
704 child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
705
706 if (child_exited)
707 break;
708
709 clock_gettime(CLOCK_MONOTONIC, &time_start);
710 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
711 if (timeout || handle_interval(interval, times))
712 break;
713 time_to_sleep = sleep_time;
714 } else { /* fd revent */
715 process_evlist(evsel_list, interval);
716 clock_gettime(CLOCK_MONOTONIC, &time_stop);
717 compute_tts(&time_start, &time_stop, &time_to_sleep);
718 }
719 }
720
721 return status;
722}
723
724enum counter_recovery {
725 COUNTER_SKIP,
726 COUNTER_RETRY,
727 COUNTER_FATAL,
728};
729
730static enum counter_recovery stat_handle_error(struct evsel *counter)
731{
732 char msg[BUFSIZ];
733 /*
734 * PPC returns ENXIO for HW counters until 2.6.37
735 * (behavior changed with commit b0a873e).
736 */
737 if (errno == EINVAL || errno == ENOSYS ||
738 errno == ENOENT || errno == EOPNOTSUPP ||
739 errno == ENXIO) {
740 if (verbose > 0)
741 ui__warning("%s event is not supported by the kernel.\n",
742 evsel__name(counter));
743 counter->supported = false;
744 /*
745 * errored is a sticky flag that means one of the counter's
746 * cpu event had a problem and needs to be reexamined.
747 */
748 counter->errored = true;
749
750 if ((evsel__leader(counter) != counter) ||
751 !(counter->core.leader->nr_members > 1))
752 return COUNTER_SKIP;
753 } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
754 if (verbose > 0)
755 ui__warning("%s\n", msg);
756 return COUNTER_RETRY;
757 } else if (target__has_per_thread(&target) &&
758 evsel_list->core.threads &&
759 evsel_list->core.threads->err_thread != -1) {
760 /*
761 * For global --per-thread case, skip current
762 * error thread.
763 */
764 if (!thread_map__remove(evsel_list->core.threads,
765 evsel_list->core.threads->err_thread)) {
766 evsel_list->core.threads->err_thread = -1;
767 return COUNTER_RETRY;
768 }
769 }
770
771 evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
772 ui__error("%s\n", msg);
773
774 if (child_pid != -1)
775 kill(child_pid, SIGTERM);
776 return COUNTER_FATAL;
777}
778
779static int __run_perf_stat(int argc, const char **argv, int run_idx)
780{
781 int interval = stat_config.interval;
782 int times = stat_config.times;
783 int timeout = stat_config.timeout;
784 char msg[BUFSIZ];
785 unsigned long long t0, t1;
786 struct evsel *counter;
787 size_t l;
788 int status = 0;
789 const bool forks = (argc > 0);
790 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
791 struct affinity affinity;
792 int i, cpu, err;
793 bool second_pass = false;
794
795 if (forks) {
796 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
797 perror("failed to prepare workload");
798 return -1;
799 }
800 child_pid = evsel_list->workload.pid;
801 }
802
803 if (group)
804 evlist__set_leader(evsel_list);
805
806 if (affinity__setup(&affinity) < 0)
807 return -1;
808
809 evlist__for_each_entry(evsel_list, counter) {
810 if (bpf_counter__load(counter, &target))
811 return -1;
812 if (!evsel__is_bpf(counter))
813 all_counters_use_bpf = false;
814 }
815
816 evlist__for_each_cpu (evsel_list, i, cpu) {
817 /*
818 * bperf calls evsel__open_per_cpu() in bperf__load(), so
819 * no need to call it again here.
820 */
821 if (target.use_bpf)
822 break;
823 affinity__set(&affinity, cpu);
824
825 evlist__for_each_entry(evsel_list, counter) {
826 if (evsel__cpu_iter_skip(counter, cpu))
827 continue;
828 if (counter->reset_group || counter->errored)
829 continue;
830 if (evsel__is_bpf(counter))
831 continue;
832try_again:
833 if (create_perf_stat_counter(counter, &stat_config, &target,
834 counter->cpu_iter - 1) < 0) {
835
836 /*
837 * Weak group failed. We cannot just undo this here
838 * because earlier CPUs might be in group mode, and the kernel
839 * doesn't support mixing group and non group reads. Defer
840 * it to later.
841 * Don't close here because we're in the wrong affinity.
842 */
843 if ((errno == EINVAL || errno == EBADF) &&
844 evsel__leader(counter) != counter &&
845 counter->weak_group) {
846 evlist__reset_weak_group(evsel_list, counter, false);
847 assert(counter->reset_group);
848 second_pass = true;
849 continue;
850 }
851
852 switch (stat_handle_error(counter)) {
853 case COUNTER_FATAL:
854 return -1;
855 case COUNTER_RETRY:
856 goto try_again;
857 case COUNTER_SKIP:
858 continue;
859 default:
860 break;
861 }
862
863 }
864 counter->supported = true;
865 }
866 }
867
868 if (second_pass) {
869 /*
870 * Now redo all the weak group after closing them,
871 * and also close errored counters.
872 */
873
874 evlist__for_each_cpu(evsel_list, i, cpu) {
875 affinity__set(&affinity, cpu);
876 /* First close errored or weak retry */
877 evlist__for_each_entry(evsel_list, counter) {
878 if (!counter->reset_group && !counter->errored)
879 continue;
880 if (evsel__cpu_iter_skip_no_inc(counter, cpu))
881 continue;
882 perf_evsel__close_cpu(&counter->core, counter->cpu_iter);
883 }
884 /* Now reopen weak */
885 evlist__for_each_entry(evsel_list, counter) {
886 if (!counter->reset_group && !counter->errored)
887 continue;
888 if (evsel__cpu_iter_skip(counter, cpu))
889 continue;
890 if (!counter->reset_group)
891 continue;
892try_again_reset:
893 pr_debug2("reopening weak %s\n", evsel__name(counter));
894 if (create_perf_stat_counter(counter, &stat_config, &target,
895 counter->cpu_iter - 1) < 0) {
896
897 switch (stat_handle_error(counter)) {
898 case COUNTER_FATAL:
899 return -1;
900 case COUNTER_RETRY:
901 goto try_again_reset;
902 case COUNTER_SKIP:
903 continue;
904 default:
905 break;
906 }
907 }
908 counter->supported = true;
909 }
910 }
911 }
912 affinity__cleanup(&affinity);
913
914 evlist__for_each_entry(evsel_list, counter) {
915 if (!counter->supported) {
916 perf_evsel__free_fd(&counter->core);
917 continue;
918 }
919
920 l = strlen(counter->unit);
921 if (l > stat_config.unit_width)
922 stat_config.unit_width = l;
923
924 if (evsel__should_store_id(counter) &&
925 evsel__store_ids(counter, evsel_list))
926 return -1;
927 }
928
929 if (evlist__apply_filters(evsel_list, &counter)) {
930 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
931 counter->filter, evsel__name(counter), errno,
932 str_error_r(errno, msg, sizeof(msg)));
933 return -1;
934 }
935
936 if (STAT_RECORD) {
937 int fd = perf_data__fd(&perf_stat.data);
938
939 if (is_pipe) {
940 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
941 } else {
942 err = perf_session__write_header(perf_stat.session, evsel_list,
943 fd, false);
944 }
945
946 if (err < 0)
947 return err;
948
949 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
950 process_synthesized_event, is_pipe);
951 if (err < 0)
952 return err;
953 }
954
955 /*
956 * Enable counters and exec the command:
957 */
958 if (forks) {
959 evlist__start_workload(evsel_list);
960 err = enable_counters();
961 if (err)
962 return -1;
963
964 t0 = rdclock();
965 clock_gettime(CLOCK_MONOTONIC, &ref_time);
966
967 if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
968 status = dispatch_events(forks, timeout, interval, ×);
969 if (child_pid != -1) {
970 if (timeout)
971 kill(child_pid, SIGTERM);
972 wait4(child_pid, &status, 0, &stat_config.ru_data);
973 }
974
975 if (workload_exec_errno) {
976 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
977 pr_err("Workload failed: %s\n", emsg);
978 return -1;
979 }
980
981 if (WIFSIGNALED(status))
982 psignal(WTERMSIG(status), argv[0]);
983 } else {
984 err = enable_counters();
985 if (err)
986 return -1;
987
988 t0 = rdclock();
989 clock_gettime(CLOCK_MONOTONIC, &ref_time);
990
991 status = dispatch_events(forks, timeout, interval, ×);
992 }
993
994 disable_counters();
995
996 t1 = rdclock();
997
998 if (stat_config.walltime_run_table)
999 stat_config.walltime_run[run_idx] = t1 - t0;
1000
1001 if (interval && stat_config.summary) {
1002 stat_config.interval = 0;
1003 stat_config.stop_read_counter = true;
1004 init_stats(&walltime_nsecs_stats);
1005 update_stats(&walltime_nsecs_stats, t1 - t0);
1006
1007 if (stat_config.aggr_mode == AGGR_GLOBAL)
1008 evlist__save_aggr_prev_raw_counts(evsel_list);
1009
1010 evlist__copy_prev_raw_counts(evsel_list);
1011 evlist__reset_prev_raw_counts(evsel_list);
1012 runtime_stat_reset(&stat_config);
1013 perf_stat__reset_shadow_per_stat(&rt_stat);
1014 } else
1015 update_stats(&walltime_nsecs_stats, t1 - t0);
1016
1017 /*
1018 * Closing a group leader splits the group, and as we only disable
1019 * group leaders, results in remaining events becoming enabled. To
1020 * avoid arbitrary skew, we must read all counters before closing any
1021 * group leaders.
1022 */
1023 read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
1024
1025 /*
1026 * We need to keep evsel_list alive, because it's processed
1027 * later the evsel_list will be closed after.
1028 */
1029 if (!STAT_RECORD)
1030 evlist__close(evsel_list);
1031
1032 return WEXITSTATUS(status);
1033}
1034
1035static int run_perf_stat(int argc, const char **argv, int run_idx)
1036{
1037 int ret;
1038
1039 if (pre_cmd) {
1040 ret = system(pre_cmd);
1041 if (ret)
1042 return ret;
1043 }
1044
1045 if (sync_run)
1046 sync();
1047
1048 ret = __run_perf_stat(argc, argv, run_idx);
1049 if (ret)
1050 return ret;
1051
1052 if (post_cmd) {
1053 ret = system(post_cmd);
1054 if (ret)
1055 return ret;
1056 }
1057
1058 return ret;
1059}
1060
1061static void print_counters(struct timespec *ts, int argc, const char **argv)
1062{
1063 /* Do not print anything if we record to the pipe. */
1064 if (STAT_RECORD && perf_stat.data.is_pipe)
1065 return;
1066 if (stat_config.quiet)
1067 return;
1068
1069 evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1070}
1071
1072static volatile int signr = -1;
1073
1074static void skip_signal(int signo)
1075{
1076 if ((child_pid == -1) || stat_config.interval)
1077 done = 1;
1078
1079 signr = signo;
1080 /*
1081 * render child_pid harmless
1082 * won't send SIGTERM to a random
1083 * process in case of race condition
1084 * and fast PID recycling
1085 */
1086 child_pid = -1;
1087}
1088
1089static void sig_atexit(void)
1090{
1091 sigset_t set, oset;
1092
1093 /*
1094 * avoid race condition with SIGCHLD handler
1095 * in skip_signal() which is modifying child_pid
1096 * goal is to avoid send SIGTERM to a random
1097 * process
1098 */
1099 sigemptyset(&set);
1100 sigaddset(&set, SIGCHLD);
1101 sigprocmask(SIG_BLOCK, &set, &oset);
1102
1103 if (child_pid != -1)
1104 kill(child_pid, SIGTERM);
1105
1106 sigprocmask(SIG_SETMASK, &oset, NULL);
1107
1108 if (signr == -1)
1109 return;
1110
1111 signal(signr, SIG_DFL);
1112 kill(getpid(), signr);
1113}
1114
1115void perf_stat__set_big_num(int set)
1116{
1117 stat_config.big_num = (set != 0);
1118}
1119
1120void perf_stat__set_no_csv_summary(int set)
1121{
1122 stat_config.no_csv_summary = (set != 0);
1123}
1124
1125static int stat__set_big_num(const struct option *opt __maybe_unused,
1126 const char *s __maybe_unused, int unset)
1127{
1128 big_num_opt = unset ? 0 : 1;
1129 perf_stat__set_big_num(!unset);
1130 return 0;
1131}
1132
1133static int enable_metric_only(const struct option *opt __maybe_unused,
1134 const char *s __maybe_unused, int unset)
1135{
1136 force_metric_only = true;
1137 stat_config.metric_only = !unset;
1138 return 0;
1139}
1140
1141static int parse_metric_groups(const struct option *opt,
1142 const char *str,
1143 int unset __maybe_unused)
1144{
1145 return metricgroup__parse_groups(opt, str,
1146 stat_config.metric_no_group,
1147 stat_config.metric_no_merge,
1148 &stat_config.metric_events);
1149}
1150
1151static int parse_control_option(const struct option *opt,
1152 const char *str,
1153 int unset __maybe_unused)
1154{
1155 struct perf_stat_config *config = opt->value;
1156
1157 return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1158}
1159
1160static int parse_stat_cgroups(const struct option *opt,
1161 const char *str, int unset)
1162{
1163 if (stat_config.cgroup_list) {
1164 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1165 return -1;
1166 }
1167
1168 return parse_cgroups(opt, str, unset);
1169}
1170
1171static struct option stat_options[] = {
1172 OPT_BOOLEAN('T', "transaction", &transaction_run,
1173 "hardware transaction statistics"),
1174 OPT_CALLBACK('e', "event", &evsel_list, "event",
1175 "event selector. use 'perf list' to list available events",
1176 parse_events_option),
1177 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1178 "event filter", parse_filter),
1179 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1180 "child tasks do not inherit counters"),
1181 OPT_STRING('p', "pid", &target.pid, "pid",
1182 "stat events on existing process id"),
1183 OPT_STRING('t', "tid", &target.tid, "tid",
1184 "stat events on existing thread id"),
1185#ifdef HAVE_BPF_SKEL
1186 OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1187 "stat events on existing bpf program id"),
1188 OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1189 "use bpf program to count events"),
1190 OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1191 "path to perf_event_attr map"),
1192#endif
1193 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1194 "system-wide collection from all CPUs"),
1195 OPT_BOOLEAN('g', "group", &group,
1196 "put the counters into a counter group"),
1197 OPT_BOOLEAN(0, "scale", &stat_config.scale,
1198 "Use --no-scale to disable counter scaling for multiplexing"),
1199 OPT_INCR('v', "verbose", &verbose,
1200 "be more verbose (show counter open errors, etc)"),
1201 OPT_INTEGER('r', "repeat", &stat_config.run_count,
1202 "repeat command and print average + stddev (max: 100, forever: 0)"),
1203 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1204 "display details about each run (only with -r option)"),
1205 OPT_BOOLEAN('n', "null", &stat_config.null_run,
1206 "null run - dont start any counters"),
1207 OPT_INCR('d', "detailed", &detailed_run,
1208 "detailed run - start a lot of events"),
1209 OPT_BOOLEAN('S', "sync", &sync_run,
1210 "call sync() before starting a run"),
1211 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1212 "print large numbers with thousands\' separators",
1213 stat__set_big_num),
1214 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1215 "list of cpus to monitor in system-wide"),
1216 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1217 "disable CPU count aggregation", AGGR_NONE),
1218 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1219 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1220 "print counts with custom separator"),
1221 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1222 "monitor event in cgroup name only", parse_stat_cgroups),
1223 OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1224 "expand events for each cgroup"),
1225 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1226 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1227 OPT_INTEGER(0, "log-fd", &output_fd,
1228 "log output to fd, instead of stderr"),
1229 OPT_STRING(0, "pre", &pre_cmd, "command",
1230 "command to run prior to the measured command"),
1231 OPT_STRING(0, "post", &post_cmd, "command",
1232 "command to run after to the measured command"),
1233 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1234 "print counts at regular interval in ms "
1235 "(overhead is possible for values <= 100ms)"),
1236 OPT_INTEGER(0, "interval-count", &stat_config.times,
1237 "print counts for fixed number of times"),
1238 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1239 "clear screen in between new interval"),
1240 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1241 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1242 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1243 "aggregate counts per processor socket", AGGR_SOCKET),
1244 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1245 "aggregate counts per processor die", AGGR_DIE),
1246 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1247 "aggregate counts per physical processor core", AGGR_CORE),
1248 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1249 "aggregate counts per thread", AGGR_THREAD),
1250 OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1251 "aggregate counts per numa node", AGGR_NODE),
1252 OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1253 "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1254 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1255 "Only print computed metrics. No raw values", enable_metric_only),
1256 OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1257 "don't group metric events, impacts multiplexing"),
1258 OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1259 "don't try to share events between metrics in a group"),
1260 OPT_BOOLEAN(0, "topdown", &topdown_run,
1261 "measure top-down statistics"),
1262 OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1263 "Set the metrics level for the top-down statistics (0: max level)"),
1264 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1265 "measure SMI cost"),
1266 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1267 "monitor specified metrics or metric groups (separated by ,)",
1268 parse_metric_groups),
1269 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1270 "Configure all used events to run in kernel space.",
1271 PARSE_OPT_EXCLUSIVE),
1272 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1273 "Configure all used events to run in user space.",
1274 PARSE_OPT_EXCLUSIVE),
1275 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1276 "Use with 'percore' event qualifier to show the event "
1277 "counts of one hardware thread by sum up total hardware "
1278 "threads of same physical core"),
1279 OPT_BOOLEAN(0, "summary", &stat_config.summary,
1280 "print summary for interval mode"),
1281 OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1282 "don't print 'summary' for CSV summary output"),
1283 OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1284 "don't print output (useful with record)"),
1285#ifdef HAVE_LIBPFM
1286 OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1287 "libpfm4 event selector. use 'perf list' to list available events",
1288 parse_libpfm_events_option),
1289#endif
1290 OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1291 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1292 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1293 "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1294 parse_control_option),
1295 OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1296 "measure I/O performance metrics provided by arch/platform",
1297 iostat_parse),
1298 OPT_END()
1299};
1300
1301static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1302 struct perf_cpu_map *map, int cpu)
1303{
1304 return cpu_map__get_socket(map, cpu, NULL);
1305}
1306
1307static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1308 struct perf_cpu_map *map, int cpu)
1309{
1310 return cpu_map__get_die(map, cpu, NULL);
1311}
1312
1313static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1314 struct perf_cpu_map *map, int cpu)
1315{
1316 return cpu_map__get_core(map, cpu, NULL);
1317}
1318
1319static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1320 struct perf_cpu_map *map, int cpu)
1321{
1322 return cpu_map__get_node(map, cpu, NULL);
1323}
1324
1325static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1326 aggr_get_id_t get_id, struct perf_cpu_map *map, int idx)
1327{
1328 int cpu;
1329 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1330
1331 if (idx >= map->nr)
1332 return id;
1333
1334 cpu = map->map[idx];
1335
1336 if (cpu_map__aggr_cpu_id_is_empty(config->cpus_aggr_map->map[cpu]))
1337 config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
1338
1339 id = config->cpus_aggr_map->map[cpu];
1340 return id;
1341}
1342
1343static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1344 struct perf_cpu_map *map, int idx)
1345{
1346 return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
1347}
1348
1349static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1350 struct perf_cpu_map *map, int idx)
1351{
1352 return perf_stat__get_aggr(config, perf_stat__get_die, map, idx);
1353}
1354
1355static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1356 struct perf_cpu_map *map, int idx)
1357{
1358 return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
1359}
1360
1361static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1362 struct perf_cpu_map *map, int idx)
1363{
1364 return perf_stat__get_aggr(config, perf_stat__get_node, map, idx);
1365}
1366
1367static bool term_percore_set(void)
1368{
1369 struct evsel *counter;
1370
1371 evlist__for_each_entry(evsel_list, counter) {
1372 if (counter->percore)
1373 return true;
1374 }
1375
1376 return false;
1377}
1378
1379static int perf_stat_init_aggr_mode(void)
1380{
1381 int nr;
1382
1383 switch (stat_config.aggr_mode) {
1384 case AGGR_SOCKET:
1385 if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1386 perror("cannot build socket map");
1387 return -1;
1388 }
1389 stat_config.aggr_get_id = perf_stat__get_socket_cached;
1390 break;
1391 case AGGR_DIE:
1392 if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1393 perror("cannot build die map");
1394 return -1;
1395 }
1396 stat_config.aggr_get_id = perf_stat__get_die_cached;
1397 break;
1398 case AGGR_CORE:
1399 if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1400 perror("cannot build core map");
1401 return -1;
1402 }
1403 stat_config.aggr_get_id = perf_stat__get_core_cached;
1404 break;
1405 case AGGR_NODE:
1406 if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1407 perror("cannot build core map");
1408 return -1;
1409 }
1410 stat_config.aggr_get_id = perf_stat__get_node_cached;
1411 break;
1412 case AGGR_NONE:
1413 if (term_percore_set()) {
1414 if (cpu_map__build_core_map(evsel_list->core.cpus,
1415 &stat_config.aggr_map)) {
1416 perror("cannot build core map");
1417 return -1;
1418 }
1419 stat_config.aggr_get_id = perf_stat__get_core_cached;
1420 }
1421 break;
1422 case AGGR_GLOBAL:
1423 case AGGR_THREAD:
1424 case AGGR_UNSET:
1425 default:
1426 break;
1427 }
1428
1429 /*
1430 * The evsel_list->cpus is the base we operate on,
1431 * taking the highest cpu number to be the size of
1432 * the aggregation translate cpumap.
1433 */
1434 nr = perf_cpu_map__max(evsel_list->core.cpus);
1435 stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1436 return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1437}
1438
1439static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1440{
1441 if (map) {
1442 WARN_ONCE(refcount_read(&map->refcnt) != 0,
1443 "cpu_aggr_map refcnt unbalanced\n");
1444 free(map);
1445 }
1446}
1447
1448static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1449{
1450 if (map && refcount_dec_and_test(&map->refcnt))
1451 cpu_aggr_map__delete(map);
1452}
1453
1454static void perf_stat__exit_aggr_mode(void)
1455{
1456 cpu_aggr_map__put(stat_config.aggr_map);
1457 cpu_aggr_map__put(stat_config.cpus_aggr_map);
1458 stat_config.aggr_map = NULL;
1459 stat_config.cpus_aggr_map = NULL;
1460}
1461
1462static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx)
1463{
1464 int cpu;
1465
1466 if (idx > map->nr)
1467 return -1;
1468
1469 cpu = map->map[idx];
1470
1471 if (cpu >= env->nr_cpus_avail)
1472 return -1;
1473
1474 return cpu;
1475}
1476
1477static struct aggr_cpu_id perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data)
1478{
1479 struct perf_env *env = data;
1480 int cpu = perf_env__get_cpu(env, map, idx);
1481 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1482
1483 if (cpu != -1)
1484 id.socket = env->cpu[cpu].socket_id;
1485
1486 return id;
1487}
1488
1489static struct aggr_cpu_id perf_env__get_die(struct perf_cpu_map *map, int idx, void *data)
1490{
1491 struct perf_env *env = data;
1492 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1493 int cpu = perf_env__get_cpu(env, map, idx);
1494
1495 if (cpu != -1) {
1496 /*
1497 * die_id is relative to socket, so start
1498 * with the socket ID and then add die to
1499 * make a unique ID.
1500 */
1501 id.socket = env->cpu[cpu].socket_id;
1502 id.die = env->cpu[cpu].die_id;
1503 }
1504
1505 return id;
1506}
1507
1508static struct aggr_cpu_id perf_env__get_core(struct perf_cpu_map *map, int idx, void *data)
1509{
1510 struct perf_env *env = data;
1511 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1512 int cpu = perf_env__get_cpu(env, map, idx);
1513
1514 if (cpu != -1) {
1515 /*
1516 * core_id is relative to socket and die,
1517 * we need a global id. So we set
1518 * socket, die id and core id
1519 */
1520 id.socket = env->cpu[cpu].socket_id;
1521 id.die = env->cpu[cpu].die_id;
1522 id.core = env->cpu[cpu].core_id;
1523 }
1524
1525 return id;
1526}
1527
1528static struct aggr_cpu_id perf_env__get_node(struct perf_cpu_map *map, int idx, void *data)
1529{
1530 int cpu = perf_env__get_cpu(data, map, idx);
1531 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1532
1533 id.node = perf_env__numa_node(data, cpu);
1534 return id;
1535}
1536
1537static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus,
1538 struct cpu_aggr_map **sockp)
1539{
1540 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1541}
1542
1543static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus,
1544 struct cpu_aggr_map **diep)
1545{
1546 return cpu_map__build_map(cpus, diep, perf_env__get_die, env);
1547}
1548
1549static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus,
1550 struct cpu_aggr_map **corep)
1551{
1552 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1553}
1554
1555static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus,
1556 struct cpu_aggr_map **nodep)
1557{
1558 return cpu_map__build_map(cpus, nodep, perf_env__get_node, env);
1559}
1560
1561static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1562 struct perf_cpu_map *map, int idx)
1563{
1564 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1565}
1566static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1567 struct perf_cpu_map *map, int idx)
1568{
1569 return perf_env__get_die(map, idx, &perf_stat.session->header.env);
1570}
1571
1572static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1573 struct perf_cpu_map *map, int idx)
1574{
1575 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1576}
1577
1578static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1579 struct perf_cpu_map *map, int idx)
1580{
1581 return perf_env__get_node(map, idx, &perf_stat.session->header.env);
1582}
1583
1584static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1585{
1586 struct perf_env *env = &st->session->header.env;
1587
1588 switch (stat_config.aggr_mode) {
1589 case AGGR_SOCKET:
1590 if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1591 perror("cannot build socket map");
1592 return -1;
1593 }
1594 stat_config.aggr_get_id = perf_stat__get_socket_file;
1595 break;
1596 case AGGR_DIE:
1597 if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1598 perror("cannot build die map");
1599 return -1;
1600 }
1601 stat_config.aggr_get_id = perf_stat__get_die_file;
1602 break;
1603 case AGGR_CORE:
1604 if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1605 perror("cannot build core map");
1606 return -1;
1607 }
1608 stat_config.aggr_get_id = perf_stat__get_core_file;
1609 break;
1610 case AGGR_NODE:
1611 if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1612 perror("cannot build core map");
1613 return -1;
1614 }
1615 stat_config.aggr_get_id = perf_stat__get_node_file;
1616 break;
1617 case AGGR_NONE:
1618 case AGGR_GLOBAL:
1619 case AGGR_THREAD:
1620 case AGGR_UNSET:
1621 default:
1622 break;
1623 }
1624
1625 return 0;
1626}
1627
1628/*
1629 * Add default attributes, if there were no attributes specified or
1630 * if -d/--detailed, -d -d or -d -d -d is used:
1631 */
1632static int add_default_attributes(void)
1633{
1634 int err;
1635 struct perf_event_attr default_attrs0[] = {
1636
1637 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1638 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1639 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1640 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1641
1642 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1643};
1644 struct perf_event_attr frontend_attrs[] = {
1645 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1646};
1647 struct perf_event_attr backend_attrs[] = {
1648 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1649};
1650 struct perf_event_attr default_attrs1[] = {
1651 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1652 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1653 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1654
1655};
1656 struct perf_event_attr default_sw_attrs[] = {
1657 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1658 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1659 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1660 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1661};
1662
1663/*
1664 * Detailed stats (-d), covering the L1 and last level data caches:
1665 */
1666 struct perf_event_attr detailed_attrs[] = {
1667
1668 { .type = PERF_TYPE_HW_CACHE,
1669 .config =
1670 PERF_COUNT_HW_CACHE_L1D << 0 |
1671 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1672 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1673
1674 { .type = PERF_TYPE_HW_CACHE,
1675 .config =
1676 PERF_COUNT_HW_CACHE_L1D << 0 |
1677 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1678 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1679
1680 { .type = PERF_TYPE_HW_CACHE,
1681 .config =
1682 PERF_COUNT_HW_CACHE_LL << 0 |
1683 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1684 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1685
1686 { .type = PERF_TYPE_HW_CACHE,
1687 .config =
1688 PERF_COUNT_HW_CACHE_LL << 0 |
1689 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1690 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1691};
1692
1693/*
1694 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1695 */
1696 struct perf_event_attr very_detailed_attrs[] = {
1697
1698 { .type = PERF_TYPE_HW_CACHE,
1699 .config =
1700 PERF_COUNT_HW_CACHE_L1I << 0 |
1701 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1702 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1703
1704 { .type = PERF_TYPE_HW_CACHE,
1705 .config =
1706 PERF_COUNT_HW_CACHE_L1I << 0 |
1707 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1708 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1709
1710 { .type = PERF_TYPE_HW_CACHE,
1711 .config =
1712 PERF_COUNT_HW_CACHE_DTLB << 0 |
1713 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1714 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1715
1716 { .type = PERF_TYPE_HW_CACHE,
1717 .config =
1718 PERF_COUNT_HW_CACHE_DTLB << 0 |
1719 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1720 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1721
1722 { .type = PERF_TYPE_HW_CACHE,
1723 .config =
1724 PERF_COUNT_HW_CACHE_ITLB << 0 |
1725 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1726 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1727
1728 { .type = PERF_TYPE_HW_CACHE,
1729 .config =
1730 PERF_COUNT_HW_CACHE_ITLB << 0 |
1731 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1732 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1733
1734};
1735
1736/*
1737 * Very, very detailed stats (-d -d -d), adding prefetch events:
1738 */
1739 struct perf_event_attr very_very_detailed_attrs[] = {
1740
1741 { .type = PERF_TYPE_HW_CACHE,
1742 .config =
1743 PERF_COUNT_HW_CACHE_L1D << 0 |
1744 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1745 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1746
1747 { .type = PERF_TYPE_HW_CACHE,
1748 .config =
1749 PERF_COUNT_HW_CACHE_L1D << 0 |
1750 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1751 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1752};
1753 struct parse_events_error errinfo;
1754
1755 /* Set attrs if no event is selected and !null_run: */
1756 if (stat_config.null_run)
1757 return 0;
1758
1759 bzero(&errinfo, sizeof(errinfo));
1760 if (transaction_run) {
1761 /* Handle -T as -M transaction. Once platform specific metrics
1762 * support has been added to the json files, all architectures
1763 * will use this approach. To determine transaction support
1764 * on an architecture test for such a metric name.
1765 */
1766 if (metricgroup__has_metric("transaction")) {
1767 struct option opt = { .value = &evsel_list };
1768
1769 return metricgroup__parse_groups(&opt, "transaction",
1770 stat_config.metric_no_group,
1771 stat_config.metric_no_merge,
1772 &stat_config.metric_events);
1773 }
1774
1775 if (pmu_have_event("cpu", "cycles-ct") &&
1776 pmu_have_event("cpu", "el-start"))
1777 err = parse_events(evsel_list, transaction_attrs,
1778 &errinfo);
1779 else
1780 err = parse_events(evsel_list,
1781 transaction_limited_attrs,
1782 &errinfo);
1783 if (err) {
1784 fprintf(stderr, "Cannot set up transaction events\n");
1785 parse_events_print_error(&errinfo, transaction_attrs);
1786 return -1;
1787 }
1788 return 0;
1789 }
1790
1791 if (smi_cost) {
1792 int smi;
1793
1794 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1795 fprintf(stderr, "freeze_on_smi is not supported.\n");
1796 return -1;
1797 }
1798
1799 if (!smi) {
1800 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1801 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1802 return -1;
1803 }
1804 smi_reset = true;
1805 }
1806
1807 if (pmu_have_event("msr", "aperf") &&
1808 pmu_have_event("msr", "smi")) {
1809 if (!force_metric_only)
1810 stat_config.metric_only = true;
1811 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1812 } else {
1813 fprintf(stderr, "To measure SMI cost, it needs "
1814 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1815 parse_events_print_error(&errinfo, smi_cost_attrs);
1816 return -1;
1817 }
1818 if (err) {
1819 parse_events_print_error(&errinfo, smi_cost_attrs);
1820 fprintf(stderr, "Cannot set up SMI cost events\n");
1821 return -1;
1822 }
1823 return 0;
1824 }
1825
1826 if (topdown_run) {
1827 const char **metric_attrs = topdown_metric_attrs;
1828 unsigned int max_level = 1;
1829 char *str = NULL;
1830 bool warn = false;
1831
1832 if (!force_metric_only)
1833 stat_config.metric_only = true;
1834
1835 if (pmu_have_event("cpu", topdown_metric_L2_attrs[5])) {
1836 metric_attrs = topdown_metric_L2_attrs;
1837 max_level = 2;
1838 }
1839
1840 if (stat_config.topdown_level > max_level) {
1841 pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1842 return -1;
1843 } else if (!stat_config.topdown_level)
1844 stat_config.topdown_level = max_level;
1845
1846 if (topdown_filter_events(metric_attrs, &str, 1) < 0) {
1847 pr_err("Out of memory\n");
1848 return -1;
1849 }
1850 if (metric_attrs[0] && str) {
1851 if (!stat_config.interval && !stat_config.metric_only) {
1852 fprintf(stat_config.output,
1853 "Topdown accuracy may decrease when measuring long periods.\n"
1854 "Please print the result regularly, e.g. -I1000\n");
1855 }
1856 goto setup_metrics;
1857 }
1858
1859 zfree(&str);
1860
1861 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1862 stat_config.aggr_mode != AGGR_CORE) {
1863 pr_err("top down event configuration requires --per-core mode\n");
1864 return -1;
1865 }
1866 stat_config.aggr_mode = AGGR_CORE;
1867 if (nr_cgroups || !target__has_cpu(&target)) {
1868 pr_err("top down event configuration requires system-wide mode (-a)\n");
1869 return -1;
1870 }
1871
1872 if (topdown_filter_events(topdown_attrs, &str,
1873 arch_topdown_check_group(&warn)) < 0) {
1874 pr_err("Out of memory\n");
1875 return -1;
1876 }
1877 if (topdown_attrs[0] && str) {
1878 if (warn)
1879 arch_topdown_group_warn();
1880setup_metrics:
1881 err = parse_events(evsel_list, str, &errinfo);
1882 if (err) {
1883 fprintf(stderr,
1884 "Cannot set up top down events %s: %d\n",
1885 str, err);
1886 parse_events_print_error(&errinfo, str);
1887 free(str);
1888 return -1;
1889 }
1890 } else {
1891 fprintf(stderr, "System does not support topdown\n");
1892 return -1;
1893 }
1894 free(str);
1895 }
1896
1897 if (!evsel_list->core.nr_entries) {
1898 if (perf_pmu__has_hybrid()) {
1899 const char *hybrid_str = "cycles,instructions,branches,branch-misses";
1900
1901 if (target__has_cpu(&target))
1902 default_sw_attrs[0].config = PERF_COUNT_SW_CPU_CLOCK;
1903
1904 if (evlist__add_default_attrs(evsel_list,
1905 default_sw_attrs) < 0) {
1906 return -1;
1907 }
1908
1909 err = parse_events(evsel_list, hybrid_str, &errinfo);
1910 if (err) {
1911 fprintf(stderr,
1912 "Cannot set up hybrid events %s: %d\n",
1913 hybrid_str, err);
1914 parse_events_print_error(&errinfo, hybrid_str);
1915 return -1;
1916 }
1917 return err;
1918 }
1919
1920 if (target__has_cpu(&target))
1921 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1922
1923 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1924 return -1;
1925 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1926 if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1927 return -1;
1928 }
1929 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1930 if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1931 return -1;
1932 }
1933 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1934 return -1;
1935
1936 stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1937 if (arch_evlist__add_default_attrs(evsel_list) < 0)
1938 return -1;
1939 }
1940
1941 /* Detailed events get appended to the event list: */
1942
1943 if (detailed_run < 1)
1944 return 0;
1945
1946 /* Append detailed run extra attributes: */
1947 if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1948 return -1;
1949
1950 if (detailed_run < 2)
1951 return 0;
1952
1953 /* Append very detailed run extra attributes: */
1954 if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1955 return -1;
1956
1957 if (detailed_run < 3)
1958 return 0;
1959
1960 /* Append very, very detailed run extra attributes: */
1961 return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1962}
1963
1964static const char * const stat_record_usage[] = {
1965 "perf stat record [<options>]",
1966 NULL,
1967};
1968
1969static void init_features(struct perf_session *session)
1970{
1971 int feat;
1972
1973 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1974 perf_header__set_feat(&session->header, feat);
1975
1976 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1977 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1978 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1979 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1980 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1981}
1982
1983static int __cmd_record(int argc, const char **argv)
1984{
1985 struct perf_session *session;
1986 struct perf_data *data = &perf_stat.data;
1987
1988 argc = parse_options(argc, argv, stat_options, stat_record_usage,
1989 PARSE_OPT_STOP_AT_NON_OPTION);
1990
1991 if (output_name)
1992 data->path = output_name;
1993
1994 if (stat_config.run_count != 1 || forever) {
1995 pr_err("Cannot use -r option with perf stat record.\n");
1996 return -1;
1997 }
1998
1999 session = perf_session__new(data, false, NULL);
2000 if (IS_ERR(session)) {
2001 pr_err("Perf session creation failed\n");
2002 return PTR_ERR(session);
2003 }
2004
2005 init_features(session);
2006
2007 session->evlist = evsel_list;
2008 perf_stat.session = session;
2009 perf_stat.record = true;
2010 return argc;
2011}
2012
2013static int process_stat_round_event(struct perf_session *session,
2014 union perf_event *event)
2015{
2016 struct perf_record_stat_round *stat_round = &event->stat_round;
2017 struct evsel *counter;
2018 struct timespec tsh, *ts = NULL;
2019 const char **argv = session->header.env.cmdline_argv;
2020 int argc = session->header.env.nr_cmdline;
2021
2022 evlist__for_each_entry(evsel_list, counter)
2023 perf_stat_process_counter(&stat_config, counter);
2024
2025 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2026 update_stats(&walltime_nsecs_stats, stat_round->time);
2027
2028 if (stat_config.interval && stat_round->time) {
2029 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2030 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2031 ts = &tsh;
2032 }
2033
2034 print_counters(ts, argc, argv);
2035 return 0;
2036}
2037
2038static
2039int process_stat_config_event(struct perf_session *session,
2040 union perf_event *event)
2041{
2042 struct perf_tool *tool = session->tool;
2043 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2044
2045 perf_event__read_stat_config(&stat_config, &event->stat_config);
2046
2047 if (perf_cpu_map__empty(st->cpus)) {
2048 if (st->aggr_mode != AGGR_UNSET)
2049 pr_warning("warning: processing task data, aggregation mode not set\n");
2050 return 0;
2051 }
2052
2053 if (st->aggr_mode != AGGR_UNSET)
2054 stat_config.aggr_mode = st->aggr_mode;
2055
2056 if (perf_stat.data.is_pipe)
2057 perf_stat_init_aggr_mode();
2058 else
2059 perf_stat_init_aggr_mode_file(st);
2060
2061 return 0;
2062}
2063
2064static int set_maps(struct perf_stat *st)
2065{
2066 if (!st->cpus || !st->threads)
2067 return 0;
2068
2069 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2070 return -EINVAL;
2071
2072 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2073
2074 if (evlist__alloc_stats(evsel_list, true))
2075 return -ENOMEM;
2076
2077 st->maps_allocated = true;
2078 return 0;
2079}
2080
2081static
2082int process_thread_map_event(struct perf_session *session,
2083 union perf_event *event)
2084{
2085 struct perf_tool *tool = session->tool;
2086 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2087
2088 if (st->threads) {
2089 pr_warning("Extra thread map event, ignoring.\n");
2090 return 0;
2091 }
2092
2093 st->threads = thread_map__new_event(&event->thread_map);
2094 if (!st->threads)
2095 return -ENOMEM;
2096
2097 return set_maps(st);
2098}
2099
2100static
2101int process_cpu_map_event(struct perf_session *session,
2102 union perf_event *event)
2103{
2104 struct perf_tool *tool = session->tool;
2105 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2106 struct perf_cpu_map *cpus;
2107
2108 if (st->cpus) {
2109 pr_warning("Extra cpu map event, ignoring.\n");
2110 return 0;
2111 }
2112
2113 cpus = cpu_map__new_data(&event->cpu_map.data);
2114 if (!cpus)
2115 return -ENOMEM;
2116
2117 st->cpus = cpus;
2118 return set_maps(st);
2119}
2120
2121static const char * const stat_report_usage[] = {
2122 "perf stat report [<options>]",
2123 NULL,
2124};
2125
2126static struct perf_stat perf_stat = {
2127 .tool = {
2128 .attr = perf_event__process_attr,
2129 .event_update = perf_event__process_event_update,
2130 .thread_map = process_thread_map_event,
2131 .cpu_map = process_cpu_map_event,
2132 .stat_config = process_stat_config_event,
2133 .stat = perf_event__process_stat_event,
2134 .stat_round = process_stat_round_event,
2135 },
2136 .aggr_mode = AGGR_UNSET,
2137};
2138
2139static int __cmd_report(int argc, const char **argv)
2140{
2141 struct perf_session *session;
2142 const struct option options[] = {
2143 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2144 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2145 "aggregate counts per processor socket", AGGR_SOCKET),
2146 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2147 "aggregate counts per processor die", AGGR_DIE),
2148 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2149 "aggregate counts per physical processor core", AGGR_CORE),
2150 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2151 "aggregate counts per numa node", AGGR_NODE),
2152 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2153 "disable CPU count aggregation", AGGR_NONE),
2154 OPT_END()
2155 };
2156 struct stat st;
2157 int ret;
2158
2159 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2160
2161 if (!input_name || !strlen(input_name)) {
2162 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2163 input_name = "-";
2164 else
2165 input_name = "perf.data";
2166 }
2167
2168 perf_stat.data.path = input_name;
2169 perf_stat.data.mode = PERF_DATA_MODE_READ;
2170
2171 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2172 if (IS_ERR(session))
2173 return PTR_ERR(session);
2174
2175 perf_stat.session = session;
2176 stat_config.output = stderr;
2177 evsel_list = session->evlist;
2178
2179 ret = perf_session__process_events(session);
2180 if (ret)
2181 return ret;
2182
2183 perf_session__delete(session);
2184 return 0;
2185}
2186
2187static void setup_system_wide(int forks)
2188{
2189 /*
2190 * Make system wide (-a) the default target if
2191 * no target was specified and one of following
2192 * conditions is met:
2193 *
2194 * - there's no workload specified
2195 * - there is workload specified but all requested
2196 * events are system wide events
2197 */
2198 if (!target__none(&target))
2199 return;
2200
2201 if (!forks)
2202 target.system_wide = true;
2203 else {
2204 struct evsel *counter;
2205
2206 evlist__for_each_entry(evsel_list, counter) {
2207 if (!counter->core.system_wide &&
2208 strcmp(counter->name, "duration_time")) {
2209 return;
2210 }
2211 }
2212
2213 if (evsel_list->core.nr_entries)
2214 target.system_wide = true;
2215 }
2216}
2217
2218int cmd_stat(int argc, const char **argv)
2219{
2220 const char * const stat_usage[] = {
2221 "perf stat [<options>] [<command>]",
2222 NULL
2223 };
2224 int status = -EINVAL, run_idx, err;
2225 const char *mode;
2226 FILE *output = stderr;
2227 unsigned int interval, timeout;
2228 const char * const stat_subcommands[] = { "record", "report" };
2229 char errbuf[BUFSIZ];
2230
2231 setlocale(LC_ALL, "");
2232
2233 evsel_list = evlist__new();
2234 if (evsel_list == NULL)
2235 return -ENOMEM;
2236
2237 parse_events__shrink_config_terms();
2238
2239 /* String-parsing callback-based options would segfault when negated */
2240 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2241 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2242 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2243
2244 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2245 (const char **) stat_usage,
2246 PARSE_OPT_STOP_AT_NON_OPTION);
2247 perf_stat__collect_metric_expr(evsel_list);
2248 perf_stat__init_shadow_stats();
2249
2250 if (stat_config.csv_sep) {
2251 stat_config.csv_output = true;
2252 if (!strcmp(stat_config.csv_sep, "\\t"))
2253 stat_config.csv_sep = "\t";
2254 } else
2255 stat_config.csv_sep = DEFAULT_SEPARATOR;
2256
2257 if (argc && !strncmp(argv[0], "rec", 3)) {
2258 argc = __cmd_record(argc, argv);
2259 if (argc < 0)
2260 return -1;
2261 } else if (argc && !strncmp(argv[0], "rep", 3))
2262 return __cmd_report(argc, argv);
2263
2264 interval = stat_config.interval;
2265 timeout = stat_config.timeout;
2266
2267 /*
2268 * For record command the -o is already taken care of.
2269 */
2270 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2271 output = NULL;
2272
2273 if (output_name && output_fd) {
2274 fprintf(stderr, "cannot use both --output and --log-fd\n");
2275 parse_options_usage(stat_usage, stat_options, "o", 1);
2276 parse_options_usage(NULL, stat_options, "log-fd", 0);
2277 goto out;
2278 }
2279
2280 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2281 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2282 goto out;
2283 }
2284
2285 if (stat_config.metric_only && stat_config.run_count > 1) {
2286 fprintf(stderr, "--metric-only is not supported with -r\n");
2287 goto out;
2288 }
2289
2290 if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2291 fprintf(stderr, "--table is only supported with -r\n");
2292 parse_options_usage(stat_usage, stat_options, "r", 1);
2293 parse_options_usage(NULL, stat_options, "table", 0);
2294 goto out;
2295 }
2296
2297 if (output_fd < 0) {
2298 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2299 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2300 goto out;
2301 }
2302
2303 if (!output && !stat_config.quiet) {
2304 struct timespec tm;
2305 mode = append_file ? "a" : "w";
2306
2307 output = fopen(output_name, mode);
2308 if (!output) {
2309 perror("failed to create output file");
2310 return -1;
2311 }
2312 clock_gettime(CLOCK_REALTIME, &tm);
2313 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2314 } else if (output_fd > 0) {
2315 mode = append_file ? "a" : "w";
2316 output = fdopen(output_fd, mode);
2317 if (!output) {
2318 perror("Failed opening logfd");
2319 return -errno;
2320 }
2321 }
2322
2323 stat_config.output = output;
2324
2325 /*
2326 * let the spreadsheet do the pretty-printing
2327 */
2328 if (stat_config.csv_output) {
2329 /* User explicitly passed -B? */
2330 if (big_num_opt == 1) {
2331 fprintf(stderr, "-B option not supported with -x\n");
2332 parse_options_usage(stat_usage, stat_options, "B", 1);
2333 parse_options_usage(NULL, stat_options, "x", 1);
2334 goto out;
2335 } else /* Nope, so disable big number formatting */
2336 stat_config.big_num = false;
2337 } else if (big_num_opt == 0) /* User passed --no-big-num */
2338 stat_config.big_num = false;
2339
2340 err = target__validate(&target);
2341 if (err) {
2342 target__strerror(&target, err, errbuf, BUFSIZ);
2343 pr_warning("%s\n", errbuf);
2344 }
2345
2346 setup_system_wide(argc);
2347
2348 /*
2349 * Display user/system times only for single
2350 * run and when there's specified tracee.
2351 */
2352 if ((stat_config.run_count == 1) && target__none(&target))
2353 stat_config.ru_display = true;
2354
2355 if (stat_config.run_count < 0) {
2356 pr_err("Run count must be a positive number\n");
2357 parse_options_usage(stat_usage, stat_options, "r", 1);
2358 goto out;
2359 } else if (stat_config.run_count == 0) {
2360 forever = true;
2361 stat_config.run_count = 1;
2362 }
2363
2364 if (stat_config.walltime_run_table) {
2365 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2366 if (!stat_config.walltime_run) {
2367 pr_err("failed to setup -r option");
2368 goto out;
2369 }
2370 }
2371
2372 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2373 !target__has_task(&target)) {
2374 if (!target.system_wide || target.cpu_list) {
2375 fprintf(stderr, "The --per-thread option is only "
2376 "available when monitoring via -p -t -a "
2377 "options or only --per-thread.\n");
2378 parse_options_usage(NULL, stat_options, "p", 1);
2379 parse_options_usage(NULL, stat_options, "t", 1);
2380 goto out;
2381 }
2382 }
2383
2384 /*
2385 * no_aggr, cgroup are for system-wide only
2386 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2387 */
2388 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2389 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2390 !target__has_cpu(&target)) {
2391 fprintf(stderr, "both cgroup and no-aggregation "
2392 "modes only available in system-wide mode\n");
2393
2394 parse_options_usage(stat_usage, stat_options, "G", 1);
2395 parse_options_usage(NULL, stat_options, "A", 1);
2396 parse_options_usage(NULL, stat_options, "a", 1);
2397 goto out;
2398 }
2399
2400 if (stat_config.iostat_run) {
2401 status = iostat_prepare(evsel_list, &stat_config);
2402 if (status)
2403 goto out;
2404 if (iostat_mode == IOSTAT_LIST) {
2405 iostat_list(evsel_list, &stat_config);
2406 goto out;
2407 } else if (verbose)
2408 iostat_list(evsel_list, &stat_config);
2409 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2410 target.system_wide = true;
2411 }
2412
2413 if (add_default_attributes())
2414 goto out;
2415
2416 if (stat_config.cgroup_list) {
2417 if (nr_cgroups > 0) {
2418 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2419 parse_options_usage(stat_usage, stat_options, "G", 1);
2420 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2421 goto out;
2422 }
2423
2424 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2425 &stat_config.metric_events, true) < 0) {
2426 parse_options_usage(stat_usage, stat_options,
2427 "for-each-cgroup", 0);
2428 goto out;
2429 }
2430 }
2431
2432 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2433 target.per_thread = true;
2434
2435 if (evlist__create_maps(evsel_list, &target) < 0) {
2436 if (target__has_task(&target)) {
2437 pr_err("Problems finding threads of monitor\n");
2438 parse_options_usage(stat_usage, stat_options, "p", 1);
2439 parse_options_usage(NULL, stat_options, "t", 1);
2440 } else if (target__has_cpu(&target)) {
2441 perror("failed to parse CPUs map");
2442 parse_options_usage(stat_usage, stat_options, "C", 1);
2443 parse_options_usage(NULL, stat_options, "a", 1);
2444 }
2445 goto out;
2446 }
2447
2448 evlist__check_cpu_maps(evsel_list);
2449
2450 /*
2451 * Initialize thread_map with comm names,
2452 * so we could print it out on output.
2453 */
2454 if (stat_config.aggr_mode == AGGR_THREAD) {
2455 thread_map__read_comms(evsel_list->core.threads);
2456 if (target.system_wide) {
2457 if (runtime_stat_new(&stat_config,
2458 perf_thread_map__nr(evsel_list->core.threads))) {
2459 goto out;
2460 }
2461 }
2462 }
2463
2464 if (stat_config.aggr_mode == AGGR_NODE)
2465 cpu__setup_cpunode_map();
2466
2467 if (stat_config.times && interval)
2468 interval_count = true;
2469 else if (stat_config.times && !interval) {
2470 pr_err("interval-count option should be used together with "
2471 "interval-print.\n");
2472 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2473 parse_options_usage(stat_usage, stat_options, "I", 1);
2474 goto out;
2475 }
2476
2477 if (timeout && timeout < 100) {
2478 if (timeout < 10) {
2479 pr_err("timeout must be >= 10ms.\n");
2480 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2481 goto out;
2482 } else
2483 pr_warning("timeout < 100ms. "
2484 "The overhead percentage could be high in some cases. "
2485 "Please proceed with caution.\n");
2486 }
2487 if (timeout && interval) {
2488 pr_err("timeout option is not supported with interval-print.\n");
2489 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2490 parse_options_usage(stat_usage, stat_options, "I", 1);
2491 goto out;
2492 }
2493
2494 if (evlist__alloc_stats(evsel_list, interval))
2495 goto out;
2496
2497 if (perf_stat_init_aggr_mode())
2498 goto out;
2499
2500 /*
2501 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2502 * while avoiding that older tools show confusing messages.
2503 *
2504 * However for pipe sessions we need to keep it zero,
2505 * because script's perf_evsel__check_attr is triggered
2506 * by attr->sample_type != 0, and we can't run it on
2507 * stat sessions.
2508 */
2509 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2510
2511 /*
2512 * We dont want to block the signals - that would cause
2513 * child tasks to inherit that and Ctrl-C would not work.
2514 * What we want is for Ctrl-C to work in the exec()-ed
2515 * task, but being ignored by perf stat itself:
2516 */
2517 atexit(sig_atexit);
2518 if (!forever)
2519 signal(SIGINT, skip_signal);
2520 signal(SIGCHLD, skip_signal);
2521 signal(SIGALRM, skip_signal);
2522 signal(SIGABRT, skip_signal);
2523
2524 if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2525 goto out;
2526
2527 status = 0;
2528 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2529 if (stat_config.run_count != 1 && verbose > 0)
2530 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2531 run_idx + 1);
2532
2533 if (run_idx != 0)
2534 evlist__reset_prev_raw_counts(evsel_list);
2535
2536 status = run_perf_stat(argc, argv, run_idx);
2537 if (forever && status != -1 && !interval) {
2538 print_counters(NULL, argc, argv);
2539 perf_stat__reset_stats();
2540 }
2541 }
2542
2543 if (!forever && status != -1 && (!interval || stat_config.summary))
2544 print_counters(NULL, argc, argv);
2545
2546 evlist__finalize_ctlfd(evsel_list);
2547
2548 if (STAT_RECORD) {
2549 /*
2550 * We synthesize the kernel mmap record just so that older tools
2551 * don't emit warnings about not being able to resolve symbols
2552 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2553 * a saner message about no samples being in the perf.data file.
2554 *
2555 * This also serves to suppress a warning about f_header.data.size == 0
2556 * in header.c at the moment 'perf stat record' gets introduced, which
2557 * is not really needed once we start adding the stat specific PERF_RECORD_
2558 * records, but the need to suppress the kptr_restrict messages in older
2559 * tools remain -acme
2560 */
2561 int fd = perf_data__fd(&perf_stat.data);
2562
2563 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2564 process_synthesized_event,
2565 &perf_stat.session->machines.host);
2566 if (err) {
2567 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2568 "older tools may produce warnings about this file\n.");
2569 }
2570
2571 if (!interval) {
2572 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2573 pr_err("failed to write stat round event\n");
2574 }
2575
2576 if (!perf_stat.data.is_pipe) {
2577 perf_stat.session->header.data_size += perf_stat.bytes_written;
2578 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2579 }
2580
2581 evlist__close(evsel_list);
2582 perf_session__delete(perf_stat.session);
2583 }
2584
2585 perf_stat__exit_aggr_mode();
2586 evlist__free_stats(evsel_list);
2587out:
2588 if (stat_config.iostat_run)
2589 iostat_release(evsel_list);
2590
2591 zfree(&stat_config.walltime_run);
2592
2593 if (smi_cost && smi_reset)
2594 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2595
2596 evlist__delete(evsel_list);
2597
2598 metricgroup__rblist_exit(&stat_config.metric_events);
2599 runtime_stat_delete(&stat_config);
2600 evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2601
2602 return status;
2603}