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