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