4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ./hackbench 10
13 Performance counter stats for './hackbench 10':
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
27 0.154822978 seconds time elapsed
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 * Improvements and fixes by:
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
41 * Released under the GPL v2. (and only v2, not any later version)
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include <subcmd/parse-options.h>
49 #include "util/parse-events.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/drv_configs.h"
56 #include "util/color.h"
57 #include "util/stat.h"
58 #include "util/header.h"
59 #include "util/cpumap.h"
60 #include "util/thread.h"
61 #include "util/thread_map.h"
62 #include "util/counts.h"
63 #include "util/group.h"
64 #include "util/session.h"
65 #include "util/tool.h"
66 #include "util/string2.h"
67 #include "util/metricgroup.h"
70 #include <linux/time64.h>
71 #include <api/fs/fs.h>
75 #include <sys/prctl.h>
79 #include <sys/types.h>
84 #include "sane_ctype.h"
86 #define DEFAULT_SEPARATOR " "
87 #define CNTR_NOT_SUPPORTED "<not supported>"
88 #define CNTR_NOT_COUNTED "<not counted>"
89 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
91 static void print_counters(struct timespec *ts, int argc, const char **argv);
93 /* Default events used for perf stat -T */
94 static const char *transaction_attrs = {
106 /* More limited version when the CPU does not have all events. */
107 static const char * transaction_limited_attrs = {
117 static const char * topdown_attrs[] = {
118 "topdown-total-slots",
119 "topdown-slots-retired",
120 "topdown-recovery-bubbles",
121 "topdown-fetch-bubbles",
122 "topdown-slots-issued",
126 static const char *smi_cost_attrs = {
134 static struct perf_evlist *evsel_list;
136 static struct rblist metric_events;
138 static struct target target = {
142 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
144 static int run_count = 1;
145 static bool no_inherit = false;
146 static volatile pid_t child_pid = -1;
147 static bool null_run = false;
148 static int detailed_run = 0;
149 static bool transaction_run;
150 static bool topdown_run = false;
151 static bool smi_cost = false;
152 static bool smi_reset = false;
153 static bool big_num = true;
154 static int big_num_opt = -1;
155 static const char *csv_sep = NULL;
156 static bool csv_output = false;
157 static bool group = false;
158 static const char *pre_cmd = NULL;
159 static const char *post_cmd = NULL;
160 static bool sync_run = false;
161 static unsigned int initial_delay = 0;
162 static unsigned int unit_width = 4; /* strlen("unit") */
163 static bool forever = false;
164 static bool metric_only = false;
165 static bool force_metric_only = false;
166 static bool no_merge = false;
167 static struct timespec ref_time;
168 static struct cpu_map *aggr_map;
169 static aggr_get_id_t aggr_get_id;
170 static bool append_file;
171 static bool interval_count;
172 static const char *output_name;
173 static int output_fd;
174 static int print_free_counters_hint;
175 static int print_mixed_hw_group_error;
179 struct perf_data data;
180 struct perf_session *session;
182 struct perf_tool tool;
184 struct cpu_map *cpus;
185 struct thread_map *threads;
186 enum aggr_mode aggr_mode;
189 static struct perf_stat perf_stat;
190 #define STAT_RECORD perf_stat.record
192 static volatile int done = 0;
194 static struct perf_stat_config stat_config = {
195 .aggr_mode = AGGR_GLOBAL,
199 static bool is_duration_time(struct perf_evsel *evsel)
201 return !strcmp(evsel->name, "duration_time");
204 static inline void diff_timespec(struct timespec *r, struct timespec *a,
207 r->tv_sec = a->tv_sec - b->tv_sec;
208 if (a->tv_nsec < b->tv_nsec) {
209 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
212 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
216 static void perf_stat__reset_stats(void)
220 perf_evlist__reset_stats(evsel_list);
221 perf_stat__reset_shadow_stats();
223 for (i = 0; i < stat_config.stats_num; i++)
224 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
227 static int create_perf_stat_counter(struct perf_evsel *evsel)
229 struct perf_event_attr *attr = &evsel->attr;
230 struct perf_evsel *leader = evsel->leader;
232 if (stat_config.scale) {
233 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
234 PERF_FORMAT_TOTAL_TIME_RUNNING;
238 * The event is part of non trivial group, let's enable
239 * the group read (for leader) and ID retrieval for all
242 if (leader->nr_members > 1)
243 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
245 attr->inherit = !no_inherit;
248 * Some events get initialized with sample_(period/type) set,
249 * like tracepoints. Clear it up for counting.
251 attr->sample_period = 0;
254 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
255 * while avoiding that older tools show confusing messages.
257 * However for pipe sessions we need to keep it zero,
258 * because script's perf_evsel__check_attr is triggered
259 * by attr->sample_type != 0, and we can't run it on
262 if (!(STAT_RECORD && perf_stat.data.is_pipe))
263 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
266 * Disabling all counters initially, they will be enabled
267 * either manually by us or by kernel via enable_on_exec
270 if (perf_evsel__is_group_leader(evsel)) {
274 * In case of initial_delay we enable tracee
277 if (target__none(&target) && !initial_delay)
278 attr->enable_on_exec = 1;
281 if (target__has_cpu(&target) && !target__has_per_thread(&target))
282 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
284 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
288 * Does the counter have nsecs as a unit?
290 static inline int nsec_counter(struct perf_evsel *evsel)
292 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
293 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
299 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
300 union perf_event *event,
301 struct perf_sample *sample __maybe_unused,
302 struct machine *machine __maybe_unused)
304 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
305 pr_err("failed to write perf data, error: %m\n");
309 perf_stat.bytes_written += event->header.size;
313 static int write_stat_round_event(u64 tm, u64 type)
315 return perf_event__synthesize_stat_round(NULL, tm, type,
316 process_synthesized_event,
320 #define WRITE_STAT_ROUND_EVENT(time, interval) \
321 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
323 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
326 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
327 struct perf_counts_values *count)
329 struct perf_sample_id *sid = SID(counter, cpu, thread);
331 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
332 process_synthesized_event, NULL);
336 * Read out the results of a single counter:
337 * do not aggregate counts across CPUs in system-wide mode
339 static int read_counter(struct perf_evsel *counter)
341 int nthreads = thread_map__nr(evsel_list->threads);
342 int ncpus, cpu, thread;
344 if (target__has_cpu(&target) && !target__has_per_thread(&target))
345 ncpus = perf_evsel__nr_cpus(counter);
349 if (!counter->supported)
352 if (counter->system_wide)
355 for (thread = 0; thread < nthreads; thread++) {
356 for (cpu = 0; cpu < ncpus; cpu++) {
357 struct perf_counts_values *count;
359 count = perf_counts(counter->counts, cpu, thread);
362 * The leader's group read loads data into its group members
363 * (via perf_evsel__read_counter) and sets threir count->loaded.
365 if (!count->loaded &&
366 perf_evsel__read_counter(counter, cpu, thread)) {
367 counter->counts->scaled = -1;
368 perf_counts(counter->counts, cpu, thread)->ena = 0;
369 perf_counts(counter->counts, cpu, thread)->run = 0;
373 count->loaded = false;
376 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
377 pr_err("failed to write stat event\n");
383 fprintf(stat_config.output,
384 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
385 perf_evsel__name(counter),
387 count->val, count->ena, count->run);
395 static void read_counters(void)
397 struct perf_evsel *counter;
400 evlist__for_each_entry(evsel_list, counter) {
401 ret = read_counter(counter);
403 pr_debug("failed to read counter %s\n", counter->name);
405 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
406 pr_warning("failed to process counter %s\n", counter->name);
410 static void process_interval(void)
412 struct timespec ts, rs;
416 clock_gettime(CLOCK_MONOTONIC, &ts);
417 diff_timespec(&rs, &ts, &ref_time);
420 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
421 pr_err("failed to write stat round event\n");
424 init_stats(&walltime_nsecs_stats);
425 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
426 print_counters(&rs, 0, NULL);
429 static void enable_counters(void)
432 usleep(initial_delay * USEC_PER_MSEC);
435 * We need to enable counters only if:
436 * - we don't have tracee (attaching to task or cpu)
437 * - we have initial delay configured
439 if (!target__none(&target) || initial_delay)
440 perf_evlist__enable(evsel_list);
443 static void disable_counters(void)
446 * If we don't have tracee (attaching to task or cpu), counters may
447 * still be running. To get accurate group ratios, we must stop groups
448 * from counting before reading their constituent counters.
450 if (!target__none(&target))
451 perf_evlist__disable(evsel_list);
454 static volatile int workload_exec_errno;
457 * perf_evlist__prepare_workload will send a SIGUSR1
458 * if the fork fails, since we asked by setting its
459 * want_signal to true.
461 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
462 void *ucontext __maybe_unused)
464 workload_exec_errno = info->si_value.sival_int;
467 static int perf_stat_synthesize_config(bool is_pipe)
472 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
473 process_synthesized_event);
475 pr_err("Couldn't synthesize attrs.\n");
480 err = perf_event__synthesize_extra_attr(NULL,
482 process_synthesized_event,
485 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
486 process_synthesized_event,
489 pr_err("Couldn't synthesize thread map.\n");
493 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
494 process_synthesized_event, NULL);
496 pr_err("Couldn't synthesize thread map.\n");
500 err = perf_event__synthesize_stat_config(NULL, &stat_config,
501 process_synthesized_event, NULL);
503 pr_err("Couldn't synthesize config.\n");
510 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
512 static int __store_counter_ids(struct perf_evsel *counter)
516 for (cpu = 0; cpu < xyarray__max_x(counter->fd); cpu++) {
517 for (thread = 0; thread < xyarray__max_y(counter->fd);
519 int fd = FD(counter, cpu, thread);
521 if (perf_evlist__id_add_fd(evsel_list, counter,
522 cpu, thread, fd) < 0)
530 static int store_counter_ids(struct perf_evsel *counter)
532 struct cpu_map *cpus = counter->cpus;
533 struct thread_map *threads = counter->threads;
535 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
538 return __store_counter_ids(counter);
541 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
543 return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
546 static struct perf_evsel *perf_evsel__reset_weak_group(struct perf_evsel *evsel)
548 struct perf_evsel *c2, *leader;
551 leader = evsel->leader;
552 pr_debug("Weak group for %s/%d failed\n",
553 leader->name, leader->nr_members);
556 * for_each_group_member doesn't work here because it doesn't
557 * include the first entry.
559 evlist__for_each_entry(evsel_list, c2) {
562 if (c2->leader == leader) {
564 perf_evsel__close(c2);
572 static int __run_perf_stat(int argc, const char **argv)
574 int interval = stat_config.interval;
575 int times = stat_config.times;
576 int timeout = stat_config.timeout;
578 unsigned long long t0, t1;
579 struct perf_evsel *counter;
583 const bool forks = (argc > 0);
584 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
585 struct perf_evsel_config_term *err_term;
588 ts.tv_sec = interval / USEC_PER_MSEC;
589 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
590 } else if (timeout) {
591 ts.tv_sec = timeout / USEC_PER_MSEC;
592 ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
599 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
600 workload_exec_failed_signal) < 0) {
601 perror("failed to prepare workload");
604 child_pid = evsel_list->workload.pid;
608 perf_evlist__set_leader(evsel_list);
610 evlist__for_each_entry(evsel_list, counter) {
612 if (create_perf_stat_counter(counter) < 0) {
614 /* Weak group failed. Reset the group. */
615 if ((errno == EINVAL || errno == EBADF) &&
616 counter->leader != counter &&
617 counter->weak_group) {
618 counter = perf_evsel__reset_weak_group(counter);
623 * PPC returns ENXIO for HW counters until 2.6.37
624 * (behavior changed with commit b0a873e).
626 if (errno == EINVAL || errno == ENOSYS ||
627 errno == ENOENT || errno == EOPNOTSUPP ||
630 ui__warning("%s event is not supported by the kernel.\n",
631 perf_evsel__name(counter));
632 counter->supported = false;
634 if ((counter->leader != counter) ||
635 !(counter->leader->nr_members > 1))
637 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
639 ui__warning("%s\n", msg);
641 } else if (target__has_per_thread(&target) &&
642 evsel_list->threads &&
643 evsel_list->threads->err_thread != -1) {
645 * For global --per-thread case, skip current
648 if (!thread_map__remove(evsel_list->threads,
649 evsel_list->threads->err_thread)) {
650 evsel_list->threads->err_thread = -1;
655 perf_evsel__open_strerror(counter, &target,
656 errno, msg, sizeof(msg));
657 ui__error("%s\n", msg);
660 kill(child_pid, SIGTERM);
664 counter->supported = true;
666 l = strlen(counter->unit);
670 if (perf_evsel__should_store_id(counter) &&
671 store_counter_ids(counter))
675 if (perf_evlist__apply_filters(evsel_list, &counter)) {
676 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
677 counter->filter, perf_evsel__name(counter), errno,
678 str_error_r(errno, msg, sizeof(msg)));
682 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
683 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
684 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
685 str_error_r(errno, msg, sizeof(msg)));
690 int err, fd = perf_data__fd(&perf_stat.data);
693 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
695 err = perf_session__write_header(perf_stat.session, evsel_list,
702 err = perf_stat_synthesize_config(is_pipe);
708 * Enable counters and exec the command:
711 clock_gettime(CLOCK_MONOTONIC, &ref_time);
714 perf_evlist__start_workload(evsel_list);
717 if (interval || timeout) {
718 while (!waitpid(child_pid, &status, WNOHANG)) {
719 nanosleep(&ts, NULL);
723 if (interval_count && !(--times))
727 waitpid(child_pid, &status, 0);
729 if (workload_exec_errno) {
730 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
731 pr_err("Workload failed: %s\n", emsg);
735 if (WIFSIGNALED(status))
736 psignal(WTERMSIG(status), argv[0]);
740 nanosleep(&ts, NULL);
745 if (interval_count && !(--times))
755 update_stats(&walltime_nsecs_stats, t1 - t0);
758 * Closing a group leader splits the group, and as we only disable
759 * group leaders, results in remaining events becoming enabled. To
760 * avoid arbitrary skew, we must read all counters before closing any
764 perf_evlist__close(evsel_list);
766 return WEXITSTATUS(status);
769 static int run_perf_stat(int argc, const char **argv)
774 ret = system(pre_cmd);
782 ret = __run_perf_stat(argc, argv);
787 ret = system(post_cmd);
795 static void print_running(u64 run, u64 ena)
798 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
802 ena ? 100.0 * run / ena : 100.0);
803 } else if (run != ena) {
804 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
808 static void print_noise_pct(double total, double avg)
810 double pct = rel_stddev_stats(total, avg);
813 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
815 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
818 static void print_noise(struct perf_evsel *evsel, double avg)
820 struct perf_stat_evsel *ps;
826 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
829 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
831 switch (stat_config.aggr_mode) {
833 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
834 cpu_map__id_to_socket(id),
836 cpu_map__id_to_cpu(id),
843 fprintf(stat_config.output, "S%*d%s%*d%s",
852 fprintf(stat_config.output, "CPU%*d%s",
854 perf_evsel__cpus(evsel)->map[id], csv_sep);
857 fprintf(stat_config.output, "%*s-%*d%s",
859 thread_map__comm(evsel->threads, id),
861 thread_map__pid(evsel->threads, id),
877 struct perf_evsel *evsel;
880 #define METRIC_LEN 35
882 static void new_line_std(void *ctx)
884 struct outstate *os = ctx;
889 static void do_new_line_std(struct outstate *os)
892 fputs(os->prefix, os->fh);
893 aggr_printout(os->evsel, os->id, os->nr);
894 if (stat_config.aggr_mode == AGGR_NONE)
895 fprintf(os->fh, " ");
896 fprintf(os->fh, " ");
899 static void print_metric_std(void *ctx, const char *color, const char *fmt,
900 const char *unit, double val)
902 struct outstate *os = ctx;
905 bool newline = os->newline;
909 if (unit == NULL || fmt == NULL) {
910 fprintf(out, "%-*s", METRIC_LEN, "");
917 n = fprintf(out, " # ");
919 n += color_fprintf(out, color, fmt, val);
921 n += fprintf(out, fmt, val);
922 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
925 static void new_line_csv(void *ctx)
927 struct outstate *os = ctx;
932 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
933 aggr_printout(os->evsel, os->id, os->nr);
934 for (i = 0; i < os->nfields; i++)
935 fputs(csv_sep, os->fh);
938 static void print_metric_csv(void *ctx,
939 const char *color __maybe_unused,
940 const char *fmt, const char *unit, double val)
942 struct outstate *os = ctx;
944 char buf[64], *vals, *ends;
946 if (unit == NULL || fmt == NULL) {
947 fprintf(out, "%s%s", csv_sep, csv_sep);
950 snprintf(buf, sizeof(buf), fmt, val);
951 ends = vals = ltrim(buf);
952 while (isdigit(*ends) || *ends == '.')
955 while (isspace(*unit))
957 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
960 #define METRIC_ONLY_LEN 20
962 /* Filter out some columns that don't work well in metrics only mode */
964 static bool valid_only_metric(const char *unit)
968 if (strstr(unit, "/sec") ||
969 strstr(unit, "hz") ||
970 strstr(unit, "Hz") ||
971 strstr(unit, "CPUs utilized"))
976 static const char *fixunit(char *buf, struct perf_evsel *evsel,
979 if (!strncmp(unit, "of all", 6)) {
980 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
987 static void print_metric_only(void *ctx, const char *color, const char *fmt,
988 const char *unit, double val)
990 struct outstate *os = ctx;
994 unsigned mlen = METRIC_ONLY_LEN;
996 if (!valid_only_metric(unit))
998 unit = fixunit(buf, os->evsel, unit);
1000 n = color_fprintf(out, color, fmt, val);
1002 n = fprintf(out, fmt, val);
1003 if (n > METRIC_ONLY_LEN)
1004 n = METRIC_ONLY_LEN;
1005 if (mlen < strlen(unit))
1006 mlen = strlen(unit) + 1;
1007 fprintf(out, "%*s", mlen - n, "");
1010 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
1012 const char *unit, double val)
1014 struct outstate *os = ctx;
1016 char buf[64], *vals, *ends;
1019 if (!valid_only_metric(unit))
1021 unit = fixunit(tbuf, os->evsel, unit);
1022 snprintf(buf, sizeof buf, fmt, val);
1023 ends = vals = ltrim(buf);
1024 while (isdigit(*ends) || *ends == '.')
1027 fprintf(out, "%s%s", vals, csv_sep);
1030 static void new_line_metric(void *ctx __maybe_unused)
1034 static void print_metric_header(void *ctx, const char *color __maybe_unused,
1035 const char *fmt __maybe_unused,
1036 const char *unit, double val __maybe_unused)
1038 struct outstate *os = ctx;
1041 if (!valid_only_metric(unit))
1043 unit = fixunit(tbuf, os->evsel, unit);
1045 fprintf(os->fh, "%s%s", unit, csv_sep);
1047 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
1050 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1052 FILE *output = stat_config.output;
1053 double msecs = avg / NSEC_PER_MSEC;
1054 const char *fmt_v, *fmt_n;
1057 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
1058 fmt_n = csv_output ? "%s" : "%-25s";
1060 aggr_printout(evsel, id, nr);
1062 scnprintf(name, sizeof(name), "%s%s",
1063 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
1065 fprintf(output, fmt_v, msecs, csv_sep);
1068 fprintf(output, "%s%s", evsel->unit, csv_sep);
1070 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
1072 fprintf(output, fmt_n, name);
1075 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1078 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1085 if (stat_config.aggr_mode == AGGR_NONE)
1088 if (stat_config.aggr_mode == AGGR_GLOBAL)
1091 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1092 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1094 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1100 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1102 FILE *output = stat_config.output;
1103 double sc = evsel->scale;
1107 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
1110 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1112 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1115 aggr_printout(evsel, id, nr);
1117 fprintf(output, fmt, avg, csv_sep);
1120 fprintf(output, "%-*s%s",
1121 csv_output ? 0 : unit_width,
1122 evsel->unit, csv_sep);
1124 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1127 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1130 static bool is_mixed_hw_group(struct perf_evsel *counter)
1132 struct perf_evlist *evlist = counter->evlist;
1133 u32 pmu_type = counter->attr.type;
1134 struct perf_evsel *pos;
1136 if (counter->nr_members < 2)
1139 evlist__for_each_entry(evlist, pos) {
1140 /* software events can be part of any hardware group */
1141 if (pos->attr.type == PERF_TYPE_SOFTWARE)
1143 if (pmu_type == PERF_TYPE_SOFTWARE) {
1144 pmu_type = pos->attr.type;
1147 if (pmu_type != pos->attr.type)
1154 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1155 char *prefix, u64 run, u64 ena, double noise,
1156 struct runtime_stat *st)
1158 struct perf_stat_output_ctx out;
1159 struct outstate os = {
1160 .fh = stat_config.output,
1161 .prefix = prefix ? prefix : "",
1166 print_metric_t pm = print_metric_std;
1170 nl = new_line_metric;
1172 pm = print_metric_only_csv;
1174 pm = print_metric_only;
1178 if (csv_output && !metric_only) {
1179 static int aggr_fields[] = {
1187 pm = print_metric_csv;
1190 os.nfields += aggr_fields[stat_config.aggr_mode];
1194 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1196 pm(&os, NULL, "", "", 0);
1199 aggr_printout(counter, id, nr);
1201 fprintf(stat_config.output, "%*s%s",
1202 csv_output ? 0 : 18,
1203 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1206 if (counter->supported) {
1207 print_free_counters_hint = 1;
1208 if (is_mixed_hw_group(counter))
1209 print_mixed_hw_group_error = 1;
1212 fprintf(stat_config.output, "%-*s%s",
1213 csv_output ? 0 : unit_width,
1214 counter->unit, csv_sep);
1216 fprintf(stat_config.output, "%*s",
1217 csv_output ? 0 : -25,
1218 perf_evsel__name(counter));
1221 fprintf(stat_config.output, "%s%s",
1222 csv_sep, counter->cgrp->name);
1225 pm(&os, NULL, NULL, "", 0);
1226 print_noise(counter, noise);
1227 print_running(run, ena);
1229 pm(&os, NULL, NULL, "", 0);
1235 else if (nsec_counter(counter))
1236 nsec_printout(id, nr, counter, uval);
1238 abs_printout(id, nr, counter, uval);
1240 out.print_metric = pm;
1243 out.force_header = false;
1245 if (csv_output && !metric_only) {
1246 print_noise(counter, noise);
1247 print_running(run, ena);
1250 perf_stat__print_shadow_stats(counter, uval,
1251 first_shadow_cpu(counter, id),
1252 &out, &metric_events, st);
1253 if (!csv_output && !metric_only) {
1254 print_noise(counter, noise);
1255 print_running(run, ena);
1259 static void aggr_update_shadow(void)
1263 struct perf_evsel *counter;
1265 for (s = 0; s < aggr_map->nr; s++) {
1266 id = aggr_map->map[s];
1267 evlist__for_each_entry(evsel_list, counter) {
1269 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1270 s2 = aggr_get_id(evsel_list->cpus, cpu);
1273 val += perf_counts(counter->counts, cpu, 0)->val;
1275 perf_stat__update_shadow_stats(counter, val,
1276 first_shadow_cpu(counter, id),
1282 static void uniquify_event_name(struct perf_evsel *counter)
1287 if (counter->uniquified_name ||
1288 !counter->pmu_name || !strncmp(counter->name, counter->pmu_name,
1289 strlen(counter->pmu_name)))
1292 config = strchr(counter->name, '/');
1294 if (asprintf(&new_name,
1295 "%s%s", counter->pmu_name, config) > 0) {
1296 free(counter->name);
1297 counter->name = new_name;
1300 if (asprintf(&new_name,
1301 "%s [%s]", counter->name, counter->pmu_name) > 0) {
1302 free(counter->name);
1303 counter->name = new_name;
1307 counter->uniquified_name = true;
1310 static void collect_all_aliases(struct perf_evsel *counter,
1311 void (*cb)(struct perf_evsel *counter, void *data,
1315 struct perf_evsel *alias;
1317 alias = list_prepare_entry(counter, &(evsel_list->entries), node);
1318 list_for_each_entry_continue (alias, &evsel_list->entries, node) {
1319 if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) ||
1320 alias->scale != counter->scale ||
1321 alias->cgrp != counter->cgrp ||
1322 strcmp(alias->unit, counter->unit) ||
1323 nsec_counter(alias) != nsec_counter(counter))
1325 alias->merged_stat = true;
1326 cb(alias, data, false);
1330 static bool collect_data(struct perf_evsel *counter,
1331 void (*cb)(struct perf_evsel *counter, void *data,
1335 if (counter->merged_stat)
1337 cb(counter, data, true);
1339 uniquify_event_name(counter);
1340 else if (counter->auto_merge_stats)
1341 collect_all_aliases(counter, cb, data);
1352 static void aggr_cb(struct perf_evsel *counter, void *data, bool first)
1354 struct aggr_data *ad = data;
1357 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1358 struct perf_counts_values *counts;
1360 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1365 counts = perf_counts(counter->counts, cpu, 0);
1367 * When any result is bad, make them all to give
1368 * consistent output in interval mode.
1370 if (counts->ena == 0 || counts->run == 0 ||
1371 counter->counts->scaled == -1) {
1376 ad->val += counts->val;
1377 ad->ena += counts->ena;
1378 ad->run += counts->run;
1382 static void print_aggr(char *prefix)
1384 FILE *output = stat_config.output;
1385 struct perf_evsel *counter;
1391 if (!(aggr_map || aggr_get_id))
1394 aggr_update_shadow();
1397 * With metric_only everything is on a single line.
1398 * Without each counter has its own line.
1400 for (s = 0; s < aggr_map->nr; s++) {
1401 struct aggr_data ad;
1402 if (prefix && metric_only)
1403 fprintf(output, "%s", prefix);
1405 ad.id = id = aggr_map->map[s];
1407 evlist__for_each_entry(evsel_list, counter) {
1408 if (is_duration_time(counter))
1411 ad.val = ad.ena = ad.run = 0;
1413 if (!collect_data(counter, aggr_cb, &ad))
1419 if (first && metric_only) {
1421 aggr_printout(counter, id, nr);
1423 if (prefix && !metric_only)
1424 fprintf(output, "%s", prefix);
1426 uval = val * counter->scale;
1427 printout(id, nr, counter, uval, prefix, run, ena, 1.0,
1430 fputc('\n', output);
1433 fputc('\n', output);
1437 static int cmp_val(const void *a, const void *b)
1439 return ((struct perf_aggr_thread_value *)b)->val -
1440 ((struct perf_aggr_thread_value *)a)->val;
1443 static struct perf_aggr_thread_value *sort_aggr_thread(
1444 struct perf_evsel *counter,
1445 int nthreads, int ncpus,
1448 int cpu, thread, i = 0;
1450 struct perf_aggr_thread_value *buf;
1452 buf = calloc(nthreads, sizeof(struct perf_aggr_thread_value));
1456 for (thread = 0; thread < nthreads; thread++) {
1457 u64 ena = 0, run = 0, val = 0;
1459 for (cpu = 0; cpu < ncpus; cpu++) {
1460 val += perf_counts(counter->counts, cpu, thread)->val;
1461 ena += perf_counts(counter->counts, cpu, thread)->ena;
1462 run += perf_counts(counter->counts, cpu, thread)->run;
1465 uval = val * counter->scale;
1468 * Skip value 0 when enabling --per-thread globally,
1469 * otherwise too many 0 output.
1471 if (uval == 0.0 && target__has_per_thread(&target))
1474 buf[i].counter = counter;
1483 qsort(buf, i, sizeof(struct perf_aggr_thread_value), cmp_val);
1491 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1493 FILE *output = stat_config.output;
1494 int nthreads = thread_map__nr(counter->threads);
1495 int ncpus = cpu_map__nr(counter->cpus);
1496 int thread, sorted_threads, id;
1497 struct perf_aggr_thread_value *buf;
1499 buf = sort_aggr_thread(counter, nthreads, ncpus, &sorted_threads);
1501 perror("cannot sort aggr thread");
1505 for (thread = 0; thread < sorted_threads; thread++) {
1507 fprintf(output, "%s", prefix);
1509 id = buf[thread].id;
1510 if (stat_config.stats)
1511 printout(id, 0, buf[thread].counter, buf[thread].uval,
1512 prefix, buf[thread].run, buf[thread].ena, 1.0,
1513 &stat_config.stats[id]);
1515 printout(id, 0, buf[thread].counter, buf[thread].uval,
1516 prefix, buf[thread].run, buf[thread].ena, 1.0,
1518 fputc('\n', output);
1525 double avg, avg_enabled, avg_running;
1528 static void counter_aggr_cb(struct perf_evsel *counter, void *data,
1529 bool first __maybe_unused)
1531 struct caggr_data *cd = data;
1532 struct perf_stat_evsel *ps = counter->stats;
1534 cd->avg += avg_stats(&ps->res_stats[0]);
1535 cd->avg_enabled += avg_stats(&ps->res_stats[1]);
1536 cd->avg_running += avg_stats(&ps->res_stats[2]);
1540 * Print out the results of a single counter:
1541 * aggregated counts in system-wide mode
1543 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1545 FILE *output = stat_config.output;
1547 struct caggr_data cd = { .avg = 0.0 };
1549 if (!collect_data(counter, counter_aggr_cb, &cd))
1552 if (prefix && !metric_only)
1553 fprintf(output, "%s", prefix);
1555 uval = cd.avg * counter->scale;
1556 printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled,
1559 fprintf(output, "\n");
1562 static void counter_cb(struct perf_evsel *counter, void *data,
1563 bool first __maybe_unused)
1565 struct aggr_data *ad = data;
1567 ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
1568 ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
1569 ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
1573 * Print out the results of a single counter:
1574 * does not use aggregated count in system-wide
1576 static void print_counter(struct perf_evsel *counter, char *prefix)
1578 FILE *output = stat_config.output;
1583 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1584 struct aggr_data ad = { .cpu = cpu };
1586 if (!collect_data(counter, counter_cb, &ad))
1593 fprintf(output, "%s", prefix);
1595 uval = val * counter->scale;
1596 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0,
1599 fputc('\n', output);
1603 static void print_no_aggr_metric(char *prefix)
1607 struct perf_evsel *counter;
1611 nrcpus = evsel_list->cpus->nr;
1612 for (cpu = 0; cpu < nrcpus; cpu++) {
1616 fputs(prefix, stat_config.output);
1617 evlist__for_each_entry(evsel_list, counter) {
1618 if (is_duration_time(counter))
1621 aggr_printout(counter, cpu, 0);
1624 val = perf_counts(counter->counts, cpu, 0)->val;
1625 ena = perf_counts(counter->counts, cpu, 0)->ena;
1626 run = perf_counts(counter->counts, cpu, 0)->run;
1628 uval = val * counter->scale;
1629 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0,
1632 fputc('\n', stat_config.output);
1636 static int aggr_header_lens[] = {
1644 static const char *aggr_header_csv[] = {
1645 [AGGR_CORE] = "core,cpus,",
1646 [AGGR_SOCKET] = "socket,cpus",
1647 [AGGR_NONE] = "cpu,",
1648 [AGGR_THREAD] = "comm-pid,",
1652 static void print_metric_headers(const char *prefix, bool no_indent)
1654 struct perf_stat_output_ctx out;
1655 struct perf_evsel *counter;
1656 struct outstate os = {
1657 .fh = stat_config.output
1661 fprintf(stat_config.output, "%s", prefix);
1663 if (!csv_output && !no_indent)
1664 fprintf(stat_config.output, "%*s",
1665 aggr_header_lens[stat_config.aggr_mode], "");
1667 if (stat_config.interval)
1668 fputs("time,", stat_config.output);
1669 fputs(aggr_header_csv[stat_config.aggr_mode],
1670 stat_config.output);
1673 /* Print metrics headers only */
1674 evlist__for_each_entry(evsel_list, counter) {
1675 if (is_duration_time(counter))
1679 out.print_metric = print_metric_header;
1680 out.new_line = new_line_metric;
1681 out.force_header = true;
1683 perf_stat__print_shadow_stats(counter, 0,
1689 fputc('\n', stat_config.output);
1692 static void print_interval(char *prefix, struct timespec *ts)
1694 FILE *output = stat_config.output;
1695 static int num_print_interval;
1697 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1699 if (num_print_interval == 0 && !csv_output) {
1700 switch (stat_config.aggr_mode) {
1702 fprintf(output, "# time socket cpus");
1704 fprintf(output, " counts %*s events\n", unit_width, "unit");
1707 fprintf(output, "# time core cpus");
1709 fprintf(output, " counts %*s events\n", unit_width, "unit");
1712 fprintf(output, "# time CPU");
1714 fprintf(output, " counts %*s events\n", unit_width, "unit");
1717 fprintf(output, "# time comm-pid");
1719 fprintf(output, " counts %*s events\n", unit_width, "unit");
1723 fprintf(output, "# time");
1725 fprintf(output, " counts %*s events\n", unit_width, "unit");
1731 if (num_print_interval == 0 && metric_only)
1732 print_metric_headers(" ", true);
1733 if (++num_print_interval == 25)
1734 num_print_interval = 0;
1737 static void print_header(int argc, const char **argv)
1739 FILE *output = stat_config.output;
1745 fprintf(output, "\n");
1746 fprintf(output, " Performance counter stats for ");
1747 if (target.system_wide)
1748 fprintf(output, "\'system wide");
1749 else if (target.cpu_list)
1750 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1751 else if (!target__has_task(&target)) {
1752 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1753 for (i = 1; argv && (i < argc); i++)
1754 fprintf(output, " %s", argv[i]);
1755 } else if (target.pid)
1756 fprintf(output, "process id \'%s", target.pid);
1758 fprintf(output, "thread id \'%s", target.tid);
1760 fprintf(output, "\'");
1762 fprintf(output, " (%d runs)", run_count);
1763 fprintf(output, ":\n\n");
1767 static void print_footer(void)
1769 FILE *output = stat_config.output;
1773 fprintf(output, "\n");
1774 fprintf(output, " %17.9f seconds time elapsed",
1775 avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC);
1776 if (run_count > 1) {
1777 fprintf(output, " ");
1778 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1779 avg_stats(&walltime_nsecs_stats));
1781 fprintf(output, "\n\n");
1783 if (print_free_counters_hint &&
1784 sysctl__read_int("kernel/nmi_watchdog", &n) >= 0 &&
1787 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1788 " echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1790 " echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1792 if (print_mixed_hw_group_error)
1794 "The events in group usually have to be from "
1795 "the same PMU. Try reorganizing the group.\n");
1798 static void print_counters(struct timespec *ts, int argc, const char **argv)
1800 int interval = stat_config.interval;
1801 struct perf_evsel *counter;
1802 char buf[64], *prefix = NULL;
1804 /* Do not print anything if we record to the pipe. */
1805 if (STAT_RECORD && perf_stat.data.is_pipe)
1809 print_interval(prefix = buf, ts);
1811 print_header(argc, argv);
1814 static int num_print_iv;
1816 if (num_print_iv == 0 && !interval)
1817 print_metric_headers(prefix, false);
1818 if (num_print_iv++ == 25)
1820 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1821 fprintf(stat_config.output, "%s", prefix);
1824 switch (stat_config.aggr_mode) {
1830 evlist__for_each_entry(evsel_list, counter) {
1831 if (is_duration_time(counter))
1833 print_aggr_thread(counter, prefix);
1837 evlist__for_each_entry(evsel_list, counter) {
1838 if (is_duration_time(counter))
1840 print_counter_aggr(counter, prefix);
1843 fputc('\n', stat_config.output);
1847 print_no_aggr_metric(prefix);
1849 evlist__for_each_entry(evsel_list, counter) {
1850 if (is_duration_time(counter))
1852 print_counter(counter, prefix);
1861 if (!interval && !csv_output)
1864 fflush(stat_config.output);
1867 static volatile int signr = -1;
1869 static void skip_signal(int signo)
1871 if ((child_pid == -1) || stat_config.interval)
1876 * render child_pid harmless
1877 * won't send SIGTERM to a random
1878 * process in case of race condition
1879 * and fast PID recycling
1884 static void sig_atexit(void)
1889 * avoid race condition with SIGCHLD handler
1890 * in skip_signal() which is modifying child_pid
1891 * goal is to avoid send SIGTERM to a random
1895 sigaddset(&set, SIGCHLD);
1896 sigprocmask(SIG_BLOCK, &set, &oset);
1898 if (child_pid != -1)
1899 kill(child_pid, SIGTERM);
1901 sigprocmask(SIG_SETMASK, &oset, NULL);
1906 signal(signr, SIG_DFL);
1907 kill(getpid(), signr);
1910 static int stat__set_big_num(const struct option *opt __maybe_unused,
1911 const char *s __maybe_unused, int unset)
1913 big_num_opt = unset ? 0 : 1;
1917 static int enable_metric_only(const struct option *opt __maybe_unused,
1918 const char *s __maybe_unused, int unset)
1920 force_metric_only = true;
1921 metric_only = !unset;
1925 static int parse_metric_groups(const struct option *opt,
1927 int unset __maybe_unused)
1929 return metricgroup__parse_groups(opt, str, &metric_events);
1932 static const struct option stat_options[] = {
1933 OPT_BOOLEAN('T', "transaction", &transaction_run,
1934 "hardware transaction statistics"),
1935 OPT_CALLBACK('e', "event", &evsel_list, "event",
1936 "event selector. use 'perf list' to list available events",
1937 parse_events_option),
1938 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1939 "event filter", parse_filter),
1940 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1941 "child tasks do not inherit counters"),
1942 OPT_STRING('p', "pid", &target.pid, "pid",
1943 "stat events on existing process id"),
1944 OPT_STRING('t', "tid", &target.tid, "tid",
1945 "stat events on existing thread id"),
1946 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1947 "system-wide collection from all CPUs"),
1948 OPT_BOOLEAN('g', "group", &group,
1949 "put the counters into a counter group"),
1950 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1951 OPT_INCR('v', "verbose", &verbose,
1952 "be more verbose (show counter open errors, etc)"),
1953 OPT_INTEGER('r', "repeat", &run_count,
1954 "repeat command and print average + stddev (max: 100, forever: 0)"),
1955 OPT_BOOLEAN('n', "null", &null_run,
1956 "null run - dont start any counters"),
1957 OPT_INCR('d', "detailed", &detailed_run,
1958 "detailed run - start a lot of events"),
1959 OPT_BOOLEAN('S', "sync", &sync_run,
1960 "call sync() before starting a run"),
1961 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1962 "print large numbers with thousands\' separators",
1964 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1965 "list of cpus to monitor in system-wide"),
1966 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1967 "disable CPU count aggregation", AGGR_NONE),
1968 OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"),
1969 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1970 "print counts with custom separator"),
1971 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1972 "monitor event in cgroup name only", parse_cgroups),
1973 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1974 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1975 OPT_INTEGER(0, "log-fd", &output_fd,
1976 "log output to fd, instead of stderr"),
1977 OPT_STRING(0, "pre", &pre_cmd, "command",
1978 "command to run prior to the measured command"),
1979 OPT_STRING(0, "post", &post_cmd, "command",
1980 "command to run after to the measured command"),
1981 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1982 "print counts at regular interval in ms "
1983 "(overhead is possible for values <= 100ms)"),
1984 OPT_INTEGER(0, "interval-count", &stat_config.times,
1985 "print counts for fixed number of times"),
1986 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1987 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1988 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1989 "aggregate counts per processor socket", AGGR_SOCKET),
1990 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1991 "aggregate counts per physical processor core", AGGR_CORE),
1992 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1993 "aggregate counts per thread", AGGR_THREAD),
1994 OPT_UINTEGER('D', "delay", &initial_delay,
1995 "ms to wait before starting measurement after program start"),
1996 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1997 "Only print computed metrics. No raw values", enable_metric_only),
1998 OPT_BOOLEAN(0, "topdown", &topdown_run,
1999 "measure topdown level 1 statistics"),
2000 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
2001 "measure SMI cost"),
2002 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
2003 "monitor specified metrics or metric groups (separated by ,)",
2004 parse_metric_groups),
2008 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
2010 return cpu_map__get_socket(map, cpu, NULL);
2013 static int perf_stat__get_core(struct cpu_map *map, int cpu)
2015 return cpu_map__get_core(map, cpu, NULL);
2018 static int cpu_map__get_max(struct cpu_map *map)
2022 for (i = 0; i < map->nr; i++) {
2023 if (map->map[i] > max)
2030 static struct cpu_map *cpus_aggr_map;
2032 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
2039 cpu = map->map[idx];
2041 if (cpus_aggr_map->map[cpu] == -1)
2042 cpus_aggr_map->map[cpu] = get_id(map, idx);
2044 return cpus_aggr_map->map[cpu];
2047 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
2049 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
2052 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
2054 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
2057 static int perf_stat_init_aggr_mode(void)
2061 switch (stat_config.aggr_mode) {
2063 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
2064 perror("cannot build socket map");
2067 aggr_get_id = perf_stat__get_socket_cached;
2070 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
2071 perror("cannot build core map");
2074 aggr_get_id = perf_stat__get_core_cached;
2085 * The evsel_list->cpus is the base we operate on,
2086 * taking the highest cpu number to be the size of
2087 * the aggregation translate cpumap.
2089 nr = cpu_map__get_max(evsel_list->cpus);
2090 cpus_aggr_map = cpu_map__empty_new(nr + 1);
2091 return cpus_aggr_map ? 0 : -ENOMEM;
2094 static void perf_stat__exit_aggr_mode(void)
2096 cpu_map__put(aggr_map);
2097 cpu_map__put(cpus_aggr_map);
2099 cpus_aggr_map = NULL;
2102 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
2109 cpu = map->map[idx];
2111 if (cpu >= env->nr_cpus_avail)
2117 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
2119 struct perf_env *env = data;
2120 int cpu = perf_env__get_cpu(env, map, idx);
2122 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
2125 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
2127 struct perf_env *env = data;
2128 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
2131 int socket_id = env->cpu[cpu].socket_id;
2134 * Encode socket in upper 16 bits
2135 * core_id is relative to socket, and
2136 * we need a global id. So we combine
2139 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
2145 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
2146 struct cpu_map **sockp)
2148 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
2151 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
2152 struct cpu_map **corep)
2154 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
2157 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
2159 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
2162 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
2164 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
2167 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
2169 struct perf_env *env = &st->session->header.env;
2171 switch (stat_config.aggr_mode) {
2173 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
2174 perror("cannot build socket map");
2177 aggr_get_id = perf_stat__get_socket_file;
2180 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
2181 perror("cannot build core map");
2184 aggr_get_id = perf_stat__get_core_file;
2197 static int topdown_filter_events(const char **attr, char **str, bool use_group)
2204 for (i = 0; attr[i]; i++) {
2205 if (pmu_have_event("cpu", attr[i])) {
2206 len += strlen(attr[i]) + 1;
2207 attr[i - off] = attr[i];
2211 attr[i - off] = NULL;
2213 *str = malloc(len + 1 + 2);
2223 for (i = 0; attr[i]; i++) {
2236 __weak bool arch_topdown_check_group(bool *warn)
2242 __weak void arch_topdown_group_warn(void)
2247 * Add default attributes, if there were no attributes specified or
2248 * if -d/--detailed, -d -d or -d -d -d is used:
2250 static int add_default_attributes(void)
2253 struct perf_event_attr default_attrs0[] = {
2255 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
2256 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
2257 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
2258 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
2260 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
2262 struct perf_event_attr frontend_attrs[] = {
2263 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
2265 struct perf_event_attr backend_attrs[] = {
2266 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
2268 struct perf_event_attr default_attrs1[] = {
2269 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
2270 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
2271 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
2276 * Detailed stats (-d), covering the L1 and last level data caches:
2278 struct perf_event_attr detailed_attrs[] = {
2280 { .type = PERF_TYPE_HW_CACHE,
2282 PERF_COUNT_HW_CACHE_L1D << 0 |
2283 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2284 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2286 { .type = PERF_TYPE_HW_CACHE,
2288 PERF_COUNT_HW_CACHE_L1D << 0 |
2289 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2290 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2292 { .type = PERF_TYPE_HW_CACHE,
2294 PERF_COUNT_HW_CACHE_LL << 0 |
2295 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2296 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2298 { .type = PERF_TYPE_HW_CACHE,
2300 PERF_COUNT_HW_CACHE_LL << 0 |
2301 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2302 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2306 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2308 struct perf_event_attr very_detailed_attrs[] = {
2310 { .type = PERF_TYPE_HW_CACHE,
2312 PERF_COUNT_HW_CACHE_L1I << 0 |
2313 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2314 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2316 { .type = PERF_TYPE_HW_CACHE,
2318 PERF_COUNT_HW_CACHE_L1I << 0 |
2319 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2320 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2322 { .type = PERF_TYPE_HW_CACHE,
2324 PERF_COUNT_HW_CACHE_DTLB << 0 |
2325 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2326 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2328 { .type = PERF_TYPE_HW_CACHE,
2330 PERF_COUNT_HW_CACHE_DTLB << 0 |
2331 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2332 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2334 { .type = PERF_TYPE_HW_CACHE,
2336 PERF_COUNT_HW_CACHE_ITLB << 0 |
2337 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2338 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2340 { .type = PERF_TYPE_HW_CACHE,
2342 PERF_COUNT_HW_CACHE_ITLB << 0 |
2343 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2344 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2349 * Very, very detailed stats (-d -d -d), adding prefetch events:
2351 struct perf_event_attr very_very_detailed_attrs[] = {
2353 { .type = PERF_TYPE_HW_CACHE,
2355 PERF_COUNT_HW_CACHE_L1D << 0 |
2356 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2357 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2359 { .type = PERF_TYPE_HW_CACHE,
2361 PERF_COUNT_HW_CACHE_L1D << 0 |
2362 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2363 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2366 /* Set attrs if no event is selected and !null_run: */
2370 if (transaction_run) {
2371 struct parse_events_error errinfo;
2373 if (pmu_have_event("cpu", "cycles-ct") &&
2374 pmu_have_event("cpu", "el-start"))
2375 err = parse_events(evsel_list, transaction_attrs,
2378 err = parse_events(evsel_list,
2379 transaction_limited_attrs,
2382 fprintf(stderr, "Cannot set up transaction events\n");
2391 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2392 fprintf(stderr, "freeze_on_smi is not supported.\n");
2397 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2398 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2404 if (pmu_have_event("msr", "aperf") &&
2405 pmu_have_event("msr", "smi")) {
2406 if (!force_metric_only)
2408 err = parse_events(evsel_list, smi_cost_attrs, NULL);
2410 fprintf(stderr, "To measure SMI cost, it needs "
2411 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
2415 fprintf(stderr, "Cannot set up SMI cost events\n");
2425 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2426 stat_config.aggr_mode != AGGR_CORE) {
2427 pr_err("top down event configuration requires --per-core mode\n");
2430 stat_config.aggr_mode = AGGR_CORE;
2431 if (nr_cgroups || !target__has_cpu(&target)) {
2432 pr_err("top down event configuration requires system-wide mode (-a)\n");
2436 if (!force_metric_only)
2438 if (topdown_filter_events(topdown_attrs, &str,
2439 arch_topdown_check_group(&warn)) < 0) {
2440 pr_err("Out of memory\n");
2443 if (topdown_attrs[0] && str) {
2445 arch_topdown_group_warn();
2446 err = parse_events(evsel_list, str, NULL);
2449 "Cannot set up top down events %s: %d\n",
2455 fprintf(stderr, "System does not support topdown\n");
2461 if (!evsel_list->nr_entries) {
2462 if (target__has_cpu(&target))
2463 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2465 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2467 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2468 if (perf_evlist__add_default_attrs(evsel_list,
2469 frontend_attrs) < 0)
2472 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2473 if (perf_evlist__add_default_attrs(evsel_list,
2477 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2481 /* Detailed events get appended to the event list: */
2483 if (detailed_run < 1)
2486 /* Append detailed run extra attributes: */
2487 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2490 if (detailed_run < 2)
2493 /* Append very detailed run extra attributes: */
2494 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2497 if (detailed_run < 3)
2500 /* Append very, very detailed run extra attributes: */
2501 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2504 static const char * const stat_record_usage[] = {
2505 "perf stat record [<options>]",
2509 static void init_features(struct perf_session *session)
2513 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2514 perf_header__set_feat(&session->header, feat);
2516 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2517 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2518 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2519 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2522 static int __cmd_record(int argc, const char **argv)
2524 struct perf_session *session;
2525 struct perf_data *data = &perf_stat.data;
2527 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2528 PARSE_OPT_STOP_AT_NON_OPTION);
2531 data->file.path = output_name;
2533 if (run_count != 1 || forever) {
2534 pr_err("Cannot use -r option with perf stat record.\n");
2538 session = perf_session__new(data, false, NULL);
2539 if (session == NULL) {
2540 pr_err("Perf session creation failed.\n");
2544 init_features(session);
2546 session->evlist = evsel_list;
2547 perf_stat.session = session;
2548 perf_stat.record = true;
2552 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2553 union perf_event *event,
2554 struct perf_session *session)
2556 struct stat_round_event *stat_round = &event->stat_round;
2557 struct perf_evsel *counter;
2558 struct timespec tsh, *ts = NULL;
2559 const char **argv = session->header.env.cmdline_argv;
2560 int argc = session->header.env.nr_cmdline;
2562 evlist__for_each_entry(evsel_list, counter)
2563 perf_stat_process_counter(&stat_config, counter);
2565 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2566 update_stats(&walltime_nsecs_stats, stat_round->time);
2568 if (stat_config.interval && stat_round->time) {
2569 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2570 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2574 print_counters(ts, argc, argv);
2579 int process_stat_config_event(struct perf_tool *tool,
2580 union perf_event *event,
2581 struct perf_session *session __maybe_unused)
2583 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2585 perf_event__read_stat_config(&stat_config, &event->stat_config);
2587 if (cpu_map__empty(st->cpus)) {
2588 if (st->aggr_mode != AGGR_UNSET)
2589 pr_warning("warning: processing task data, aggregation mode not set\n");
2593 if (st->aggr_mode != AGGR_UNSET)
2594 stat_config.aggr_mode = st->aggr_mode;
2596 if (perf_stat.data.is_pipe)
2597 perf_stat_init_aggr_mode();
2599 perf_stat_init_aggr_mode_file(st);
2604 static int set_maps(struct perf_stat *st)
2606 if (!st->cpus || !st->threads)
2609 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2612 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2614 if (perf_evlist__alloc_stats(evsel_list, true))
2617 st->maps_allocated = true;
2622 int process_thread_map_event(struct perf_tool *tool,
2623 union perf_event *event,
2624 struct perf_session *session __maybe_unused)
2626 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2629 pr_warning("Extra thread map event, ignoring.\n");
2633 st->threads = thread_map__new_event(&event->thread_map);
2637 return set_maps(st);
2641 int process_cpu_map_event(struct perf_tool *tool,
2642 union perf_event *event,
2643 struct perf_session *session __maybe_unused)
2645 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2646 struct cpu_map *cpus;
2649 pr_warning("Extra cpu map event, ignoring.\n");
2653 cpus = cpu_map__new_data(&event->cpu_map.data);
2658 return set_maps(st);
2661 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
2665 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
2669 config->stats_num = nthreads;
2671 for (i = 0; i < nthreads; i++)
2672 runtime_stat__init(&config->stats[i]);
2677 static void runtime_stat_delete(struct perf_stat_config *config)
2684 for (i = 0; i < config->stats_num; i++)
2685 runtime_stat__exit(&config->stats[i]);
2687 free(config->stats);
2690 static const char * const stat_report_usage[] = {
2691 "perf stat report [<options>]",
2695 static struct perf_stat perf_stat = {
2697 .attr = perf_event__process_attr,
2698 .event_update = perf_event__process_event_update,
2699 .thread_map = process_thread_map_event,
2700 .cpu_map = process_cpu_map_event,
2701 .stat_config = process_stat_config_event,
2702 .stat = perf_event__process_stat_event,
2703 .stat_round = process_stat_round_event,
2705 .aggr_mode = AGGR_UNSET,
2708 static int __cmd_report(int argc, const char **argv)
2710 struct perf_session *session;
2711 const struct option options[] = {
2712 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2713 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2714 "aggregate counts per processor socket", AGGR_SOCKET),
2715 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2716 "aggregate counts per physical processor core", AGGR_CORE),
2717 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2718 "disable CPU count aggregation", AGGR_NONE),
2724 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2726 if (!input_name || !strlen(input_name)) {
2727 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2730 input_name = "perf.data";
2733 perf_stat.data.file.path = input_name;
2734 perf_stat.data.mode = PERF_DATA_MODE_READ;
2736 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2737 if (session == NULL)
2740 perf_stat.session = session;
2741 stat_config.output = stderr;
2742 evsel_list = session->evlist;
2744 ret = perf_session__process_events(session);
2748 perf_session__delete(session);
2752 static void setup_system_wide(int forks)
2755 * Make system wide (-a) the default target if
2756 * no target was specified and one of following
2757 * conditions is met:
2759 * - there's no workload specified
2760 * - there is workload specified but all requested
2761 * events are system wide events
2763 if (!target__none(&target))
2767 target.system_wide = true;
2769 struct perf_evsel *counter;
2771 evlist__for_each_entry(evsel_list, counter) {
2772 if (!counter->system_wide)
2776 if (evsel_list->nr_entries)
2777 target.system_wide = true;
2781 int cmd_stat(int argc, const char **argv)
2783 const char * const stat_usage[] = {
2784 "perf stat [<options>] [<command>]",
2787 int status = -EINVAL, run_idx;
2789 FILE *output = stderr;
2790 unsigned int interval, timeout;
2791 const char * const stat_subcommands[] = { "record", "report" };
2793 setlocale(LC_ALL, "");
2795 evsel_list = perf_evlist__new();
2796 if (evsel_list == NULL)
2799 parse_events__shrink_config_terms();
2800 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2801 (const char **) stat_usage,
2802 PARSE_OPT_STOP_AT_NON_OPTION);
2803 perf_stat__collect_metric_expr(evsel_list);
2804 perf_stat__init_shadow_stats();
2808 if (!strcmp(csv_sep, "\\t"))
2811 csv_sep = DEFAULT_SEPARATOR;
2813 if (argc && !strncmp(argv[0], "rec", 3)) {
2814 argc = __cmd_record(argc, argv);
2817 } else if (argc && !strncmp(argv[0], "rep", 3))
2818 return __cmd_report(argc, argv);
2820 interval = stat_config.interval;
2821 timeout = stat_config.timeout;
2824 * For record command the -o is already taken care of.
2826 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2829 if (output_name && output_fd) {
2830 fprintf(stderr, "cannot use both --output and --log-fd\n");
2831 parse_options_usage(stat_usage, stat_options, "o", 1);
2832 parse_options_usage(NULL, stat_options, "log-fd", 0);
2836 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2837 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2841 if (metric_only && run_count > 1) {
2842 fprintf(stderr, "--metric-only is not supported with -r\n");
2846 if (output_fd < 0) {
2847 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2848 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2854 mode = append_file ? "a" : "w";
2856 output = fopen(output_name, mode);
2858 perror("failed to create output file");
2861 clock_gettime(CLOCK_REALTIME, &tm);
2862 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2863 } else if (output_fd > 0) {
2864 mode = append_file ? "a" : "w";
2865 output = fdopen(output_fd, mode);
2867 perror("Failed opening logfd");
2872 stat_config.output = output;
2875 * let the spreadsheet do the pretty-printing
2878 /* User explicitly passed -B? */
2879 if (big_num_opt == 1) {
2880 fprintf(stderr, "-B option not supported with -x\n");
2881 parse_options_usage(stat_usage, stat_options, "B", 1);
2882 parse_options_usage(NULL, stat_options, "x", 1);
2884 } else /* Nope, so disable big number formatting */
2886 } else if (big_num_opt == 0) /* User passed --no-big-num */
2889 setup_system_wide(argc);
2891 if (run_count < 0) {
2892 pr_err("Run count must be a positive number\n");
2893 parse_options_usage(stat_usage, stat_options, "r", 1);
2895 } else if (run_count == 0) {
2900 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2901 !target__has_task(&target)) {
2902 if (!target.system_wide || target.cpu_list) {
2903 fprintf(stderr, "The --per-thread option is only "
2904 "available when monitoring via -p -t -a "
2905 "options or only --per-thread.\n");
2906 parse_options_usage(NULL, stat_options, "p", 1);
2907 parse_options_usage(NULL, stat_options, "t", 1);
2913 * no_aggr, cgroup are for system-wide only
2914 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2916 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2917 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2918 !target__has_cpu(&target)) {
2919 fprintf(stderr, "both cgroup and no-aggregation "
2920 "modes only available in system-wide mode\n");
2922 parse_options_usage(stat_usage, stat_options, "G", 1);
2923 parse_options_usage(NULL, stat_options, "A", 1);
2924 parse_options_usage(NULL, stat_options, "a", 1);
2928 if (add_default_attributes())
2931 target__validate(&target);
2933 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2934 target.per_thread = true;
2936 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2937 if (target__has_task(&target)) {
2938 pr_err("Problems finding threads of monitor\n");
2939 parse_options_usage(stat_usage, stat_options, "p", 1);
2940 parse_options_usage(NULL, stat_options, "t", 1);
2941 } else if (target__has_cpu(&target)) {
2942 perror("failed to parse CPUs map");
2943 parse_options_usage(stat_usage, stat_options, "C", 1);
2944 parse_options_usage(NULL, stat_options, "a", 1);
2950 * Initialize thread_map with comm names,
2951 * so we could print it out on output.
2953 if (stat_config.aggr_mode == AGGR_THREAD) {
2954 thread_map__read_comms(evsel_list->threads);
2955 if (target.system_wide) {
2956 if (runtime_stat_new(&stat_config,
2957 thread_map__nr(evsel_list->threads))) {
2963 if (stat_config.times && interval)
2964 interval_count = true;
2965 else if (stat_config.times && !interval) {
2966 pr_err("interval-count option should be used together with "
2967 "interval-print.\n");
2968 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2969 parse_options_usage(stat_usage, stat_options, "I", 1);
2973 if (timeout && timeout < 100) {
2975 pr_err("timeout must be >= 10ms.\n");
2976 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2979 pr_warning("timeout < 100ms. "
2980 "The overhead percentage could be high in some cases. "
2981 "Please proceed with caution.\n");
2983 if (timeout && interval) {
2984 pr_err("timeout option is not supported with interval-print.\n");
2985 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2986 parse_options_usage(stat_usage, stat_options, "I", 1);
2990 if (perf_evlist__alloc_stats(evsel_list, interval))
2993 if (perf_stat_init_aggr_mode())
2997 * We dont want to block the signals - that would cause
2998 * child tasks to inherit that and Ctrl-C would not work.
2999 * What we want is for Ctrl-C to work in the exec()-ed
3000 * task, but being ignored by perf stat itself:
3004 signal(SIGINT, skip_signal);
3005 signal(SIGCHLD, skip_signal);
3006 signal(SIGALRM, skip_signal);
3007 signal(SIGABRT, skip_signal);
3010 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
3011 if (run_count != 1 && verbose > 0)
3012 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
3015 status = run_perf_stat(argc, argv);
3016 if (forever && status != -1) {
3017 print_counters(NULL, argc, argv);
3018 perf_stat__reset_stats();
3022 if (!forever && status != -1 && !interval)
3023 print_counters(NULL, argc, argv);
3027 * We synthesize the kernel mmap record just so that older tools
3028 * don't emit warnings about not being able to resolve symbols
3029 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
3030 * a saner message about no samples being in the perf.data file.
3032 * This also serves to suppress a warning about f_header.data.size == 0
3033 * in header.c at the moment 'perf stat record' gets introduced, which
3034 * is not really needed once we start adding the stat specific PERF_RECORD_
3035 * records, but the need to suppress the kptr_restrict messages in older
3036 * tools remain -acme
3038 int fd = perf_data__fd(&perf_stat.data);
3039 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
3040 process_synthesized_event,
3041 &perf_stat.session->machines.host);
3043 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
3044 "older tools may produce warnings about this file\n.");
3048 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
3049 pr_err("failed to write stat round event\n");
3052 if (!perf_stat.data.is_pipe) {
3053 perf_stat.session->header.data_size += perf_stat.bytes_written;
3054 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
3057 perf_session__delete(perf_stat.session);
3060 perf_stat__exit_aggr_mode();
3061 perf_evlist__free_stats(evsel_list);
3063 if (smi_cost && smi_reset)
3064 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
3066 perf_evlist__delete(evsel_list);
3068 runtime_stat_delete(&stat_config);