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