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"
71 #include <linux/time64.h>
72 #include <api/fs/fs.h>
76 #include <sys/prctl.h>
80 #include <sys/types.h>
85 #include <sys/resource.h>
88 #include "sane_ctype.h"
90 #define DEFAULT_SEPARATOR " "
91 #define CNTR_NOT_SUPPORTED "<not supported>"
92 #define CNTR_NOT_COUNTED "<not counted>"
93 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
95 static void print_counters(struct timespec *ts, int argc, const char **argv);
97 /* Default events used for perf stat -T */
98 static const char *transaction_attrs = {
110 /* More limited version when the CPU does not have all events. */
111 static const char * transaction_limited_attrs = {
121 static const char * topdown_attrs[] = {
122 "topdown-total-slots",
123 "topdown-slots-retired",
124 "topdown-recovery-bubbles",
125 "topdown-fetch-bubbles",
126 "topdown-slots-issued",
130 static const char *smi_cost_attrs = {
138 static struct perf_evlist *evsel_list;
140 static struct rblist metric_events;
142 static struct target target = {
146 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
148 #define METRIC_ONLY_LEN 20
150 static int run_count = 1;
151 static volatile pid_t child_pid = -1;
152 static bool null_run = false;
153 static int detailed_run = 0;
154 static bool transaction_run;
155 static bool topdown_run = false;
156 static bool smi_cost = false;
157 static bool smi_reset = false;
158 static bool big_num = true;
159 static int big_num_opt = -1;
160 static const char *csv_sep = NULL;
161 static bool csv_output = false;
162 static bool group = false;
163 static const char *pre_cmd = NULL;
164 static const char *post_cmd = NULL;
165 static bool sync_run = false;
166 static unsigned int unit_width = 4; /* strlen("unit") */
167 static bool forever = false;
168 static bool metric_only = false;
169 static bool force_metric_only = false;
170 static bool no_merge = false;
171 static bool walltime_run_table = false;
172 static struct timespec ref_time;
173 static struct cpu_map *aggr_map;
174 static aggr_get_id_t aggr_get_id;
175 static bool append_file;
176 static bool interval_count;
177 static bool interval_clear;
178 static const char *output_name;
179 static int output_fd;
180 static int print_free_counters_hint;
181 static int print_mixed_hw_group_error;
182 static u64 *walltime_run;
183 static bool ru_display = false;
184 static struct rusage ru_data;
185 static unsigned int metric_only_len = METRIC_ONLY_LEN;
189 struct perf_data data;
190 struct perf_session *session;
192 struct perf_tool tool;
194 struct cpu_map *cpus;
195 struct thread_map *threads;
196 enum aggr_mode aggr_mode;
199 static struct perf_stat perf_stat;
200 #define STAT_RECORD perf_stat.record
202 static volatile int done = 0;
204 static struct perf_stat_config stat_config = {
205 .aggr_mode = AGGR_GLOBAL,
209 static bool is_duration_time(struct perf_evsel *evsel)
211 return !strcmp(evsel->name, "duration_time");
214 static inline void diff_timespec(struct timespec *r, struct timespec *a,
217 r->tv_sec = a->tv_sec - b->tv_sec;
218 if (a->tv_nsec < b->tv_nsec) {
219 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
222 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
226 static void perf_stat__reset_stats(void)
230 perf_evlist__reset_stats(evsel_list);
231 perf_stat__reset_shadow_stats();
233 for (i = 0; i < stat_config.stats_num; i++)
234 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
237 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
238 union perf_event *event,
239 struct perf_sample *sample __maybe_unused,
240 struct machine *machine __maybe_unused)
242 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
243 pr_err("failed to write perf data, error: %m\n");
247 perf_stat.bytes_written += event->header.size;
251 static int write_stat_round_event(u64 tm, u64 type)
253 return perf_event__synthesize_stat_round(NULL, tm, type,
254 process_synthesized_event,
258 #define WRITE_STAT_ROUND_EVENT(time, interval) \
259 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
261 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
264 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
265 struct perf_counts_values *count)
267 struct perf_sample_id *sid = SID(counter, cpu, thread);
269 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
270 process_synthesized_event, NULL);
274 * Read out the results of a single counter:
275 * do not aggregate counts across CPUs in system-wide mode
277 static int read_counter(struct perf_evsel *counter)
279 int nthreads = thread_map__nr(evsel_list->threads);
280 int ncpus, cpu, thread;
282 if (target__has_cpu(&target) && !target__has_per_thread(&target))
283 ncpus = perf_evsel__nr_cpus(counter);
287 if (!counter->supported)
290 if (counter->system_wide)
293 for (thread = 0; thread < nthreads; thread++) {
294 for (cpu = 0; cpu < ncpus; cpu++) {
295 struct perf_counts_values *count;
297 count = perf_counts(counter->counts, cpu, thread);
300 * The leader's group read loads data into its group members
301 * (via perf_evsel__read_counter) and sets threir count->loaded.
303 if (!count->loaded &&
304 perf_evsel__read_counter(counter, cpu, thread)) {
305 counter->counts->scaled = -1;
306 perf_counts(counter->counts, cpu, thread)->ena = 0;
307 perf_counts(counter->counts, cpu, thread)->run = 0;
311 count->loaded = false;
314 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
315 pr_err("failed to write stat event\n");
321 fprintf(stat_config.output,
322 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
323 perf_evsel__name(counter),
325 count->val, count->ena, count->run);
333 static void read_counters(void)
335 struct perf_evsel *counter;
338 evlist__for_each_entry(evsel_list, counter) {
339 ret = read_counter(counter);
341 pr_debug("failed to read counter %s\n", counter->name);
343 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
344 pr_warning("failed to process counter %s\n", counter->name);
348 static void process_interval(void)
350 struct timespec ts, rs;
354 clock_gettime(CLOCK_MONOTONIC, &ts);
355 diff_timespec(&rs, &ts, &ref_time);
358 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
359 pr_err("failed to write stat round event\n");
362 init_stats(&walltime_nsecs_stats);
363 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
364 print_counters(&rs, 0, NULL);
367 static void enable_counters(void)
369 if (stat_config.initial_delay)
370 usleep(stat_config.initial_delay * USEC_PER_MSEC);
373 * We need to enable counters only if:
374 * - we don't have tracee (attaching to task or cpu)
375 * - we have initial delay configured
377 if (!target__none(&target) || stat_config.initial_delay)
378 perf_evlist__enable(evsel_list);
381 static void disable_counters(void)
384 * If we don't have tracee (attaching to task or cpu), counters may
385 * still be running. To get accurate group ratios, we must stop groups
386 * from counting before reading their constituent counters.
388 if (!target__none(&target))
389 perf_evlist__disable(evsel_list);
392 static volatile int workload_exec_errno;
395 * perf_evlist__prepare_workload will send a SIGUSR1
396 * if the fork fails, since we asked by setting its
397 * want_signal to true.
399 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
400 void *ucontext __maybe_unused)
402 workload_exec_errno = info->si_value.sival_int;
405 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
407 return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
410 static struct perf_evsel *perf_evsel__reset_weak_group(struct perf_evsel *evsel)
412 struct perf_evsel *c2, *leader;
415 leader = evsel->leader;
416 pr_debug("Weak group for %s/%d failed\n",
417 leader->name, leader->nr_members);
420 * for_each_group_member doesn't work here because it doesn't
421 * include the first entry.
423 evlist__for_each_entry(evsel_list, c2) {
426 if (c2->leader == leader) {
428 perf_evsel__close(c2);
436 static int __run_perf_stat(int argc, const char **argv, int run_idx)
438 int interval = stat_config.interval;
439 int times = stat_config.times;
440 int timeout = stat_config.timeout;
442 unsigned long long t0, t1;
443 struct perf_evsel *counter;
447 const bool forks = (argc > 0);
448 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
449 struct perf_evsel_config_term *err_term;
452 ts.tv_sec = interval / USEC_PER_MSEC;
453 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
454 } else if (timeout) {
455 ts.tv_sec = timeout / USEC_PER_MSEC;
456 ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
463 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
464 workload_exec_failed_signal) < 0) {
465 perror("failed to prepare workload");
468 child_pid = evsel_list->workload.pid;
472 perf_evlist__set_leader(evsel_list);
474 evlist__for_each_entry(evsel_list, counter) {
476 if (create_perf_stat_counter(counter, &stat_config, &target) < 0) {
478 /* Weak group failed. Reset the group. */
479 if ((errno == EINVAL || errno == EBADF) &&
480 counter->leader != counter &&
481 counter->weak_group) {
482 counter = perf_evsel__reset_weak_group(counter);
487 * PPC returns ENXIO for HW counters until 2.6.37
488 * (behavior changed with commit b0a873e).
490 if (errno == EINVAL || errno == ENOSYS ||
491 errno == ENOENT || errno == EOPNOTSUPP ||
494 ui__warning("%s event is not supported by the kernel.\n",
495 perf_evsel__name(counter));
496 counter->supported = false;
498 if ((counter->leader != counter) ||
499 !(counter->leader->nr_members > 1))
501 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
503 ui__warning("%s\n", msg);
505 } else if (target__has_per_thread(&target) &&
506 evsel_list->threads &&
507 evsel_list->threads->err_thread != -1) {
509 * For global --per-thread case, skip current
512 if (!thread_map__remove(evsel_list->threads,
513 evsel_list->threads->err_thread)) {
514 evsel_list->threads->err_thread = -1;
519 perf_evsel__open_strerror(counter, &target,
520 errno, msg, sizeof(msg));
521 ui__error("%s\n", msg);
524 kill(child_pid, SIGTERM);
528 counter->supported = true;
530 l = strlen(counter->unit);
534 if (perf_evsel__should_store_id(counter) &&
535 perf_evsel__store_ids(counter, evsel_list))
539 if (perf_evlist__apply_filters(evsel_list, &counter)) {
540 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
541 counter->filter, perf_evsel__name(counter), errno,
542 str_error_r(errno, msg, sizeof(msg)));
546 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
547 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
548 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
549 str_error_r(errno, msg, sizeof(msg)));
554 int err, fd = perf_data__fd(&perf_stat.data);
557 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
559 err = perf_session__write_header(perf_stat.session, evsel_list,
566 err = perf_stat_synthesize_config(&stat_config, NULL, evsel_list,
567 process_synthesized_event, is_pipe);
573 * Enable counters and exec the command:
576 clock_gettime(CLOCK_MONOTONIC, &ref_time);
579 perf_evlist__start_workload(evsel_list);
582 if (interval || timeout) {
583 while (!waitpid(child_pid, &status, WNOHANG)) {
584 nanosleep(&ts, NULL);
588 if (interval_count && !(--times))
592 wait4(child_pid, &status, 0, &ru_data);
594 if (workload_exec_errno) {
595 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
596 pr_err("Workload failed: %s\n", emsg);
600 if (WIFSIGNALED(status))
601 psignal(WTERMSIG(status), argv[0]);
605 nanosleep(&ts, NULL);
610 if (interval_count && !(--times))
620 if (walltime_run_table)
621 walltime_run[run_idx] = t1 - t0;
623 update_stats(&walltime_nsecs_stats, t1 - t0);
626 * Closing a group leader splits the group, and as we only disable
627 * group leaders, results in remaining events becoming enabled. To
628 * avoid arbitrary skew, we must read all counters before closing any
632 perf_evlist__close(evsel_list);
634 return WEXITSTATUS(status);
637 static int run_perf_stat(int argc, const char **argv, int run_idx)
642 ret = system(pre_cmd);
650 ret = __run_perf_stat(argc, argv, run_idx);
655 ret = system(post_cmd);
663 static void print_running(u64 run, u64 ena)
666 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
670 ena ? 100.0 * run / ena : 100.0);
671 } else if (run != ena) {
672 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
676 static void print_noise_pct(double total, double avg)
678 double pct = rel_stddev_stats(total, avg);
681 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
683 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
686 static void print_noise(struct perf_evsel *evsel, double avg)
688 struct perf_stat_evsel *ps;
694 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
697 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
699 switch (stat_config.aggr_mode) {
701 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
702 cpu_map__id_to_socket(id),
704 cpu_map__id_to_cpu(id),
711 fprintf(stat_config.output, "S%*d%s%*d%s",
720 fprintf(stat_config.output, "CPU%*d%s",
722 perf_evsel__cpus(evsel)->map[id], csv_sep);
725 fprintf(stat_config.output, "%*s-%*d%s",
727 thread_map__comm(evsel->threads, id),
729 thread_map__pid(evsel->threads, id),
745 struct perf_evsel *evsel;
748 #define METRIC_LEN 35
750 static void new_line_std(void *ctx)
752 struct outstate *os = ctx;
757 static void do_new_line_std(struct outstate *os)
760 fputs(os->prefix, os->fh);
761 aggr_printout(os->evsel, os->id, os->nr);
762 if (stat_config.aggr_mode == AGGR_NONE)
763 fprintf(os->fh, " ");
764 fprintf(os->fh, " ");
767 static void print_metric_std(void *ctx, const char *color, const char *fmt,
768 const char *unit, double val)
770 struct outstate *os = ctx;
773 bool newline = os->newline;
777 if (unit == NULL || fmt == NULL) {
778 fprintf(out, "%-*s", METRIC_LEN, "");
785 n = fprintf(out, " # ");
787 n += color_fprintf(out, color, fmt, val);
789 n += fprintf(out, fmt, val);
790 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
793 static void new_line_csv(void *ctx)
795 struct outstate *os = ctx;
800 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
801 aggr_printout(os->evsel, os->id, os->nr);
802 for (i = 0; i < os->nfields; i++)
803 fputs(csv_sep, os->fh);
806 static void print_metric_csv(void *ctx,
807 const char *color __maybe_unused,
808 const char *fmt, const char *unit, double val)
810 struct outstate *os = ctx;
812 char buf[64], *vals, *ends;
814 if (unit == NULL || fmt == NULL) {
815 fprintf(out, "%s%s", csv_sep, csv_sep);
818 snprintf(buf, sizeof(buf), fmt, val);
819 ends = vals = ltrim(buf);
820 while (isdigit(*ends) || *ends == '.')
823 while (isspace(*unit))
825 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
828 /* Filter out some columns that don't work well in metrics only mode */
830 static bool valid_only_metric(const char *unit)
834 if (strstr(unit, "/sec") ||
835 strstr(unit, "hz") ||
836 strstr(unit, "Hz") ||
837 strstr(unit, "CPUs utilized"))
842 static const char *fixunit(char *buf, struct perf_evsel *evsel,
845 if (!strncmp(unit, "of all", 6)) {
846 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
853 static void print_metric_only(void *ctx, const char *color, const char *fmt,
854 const char *unit, double val)
856 struct outstate *os = ctx;
858 char buf[1024], str[1024];
859 unsigned mlen = metric_only_len;
861 if (!valid_only_metric(unit))
863 unit = fixunit(buf, os->evsel, unit);
864 if (mlen < strlen(unit))
865 mlen = strlen(unit) + 1;
868 mlen += strlen(color) + sizeof(PERF_COLOR_RESET) - 1;
870 color_snprintf(str, sizeof(str), color ?: "", fmt, val);
871 fprintf(out, "%*s ", mlen, str);
874 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
876 const char *unit, double val)
878 struct outstate *os = ctx;
880 char buf[64], *vals, *ends;
883 if (!valid_only_metric(unit))
885 unit = fixunit(tbuf, os->evsel, unit);
886 snprintf(buf, sizeof buf, fmt, val);
887 ends = vals = ltrim(buf);
888 while (isdigit(*ends) || *ends == '.')
891 fprintf(out, "%s%s", vals, csv_sep);
894 static void new_line_metric(void *ctx __maybe_unused)
898 static void print_metric_header(void *ctx, const char *color __maybe_unused,
899 const char *fmt __maybe_unused,
900 const char *unit, double val __maybe_unused)
902 struct outstate *os = ctx;
905 if (!valid_only_metric(unit))
907 unit = fixunit(tbuf, os->evsel, unit);
909 fprintf(os->fh, "%s%s", unit, csv_sep);
911 fprintf(os->fh, "%*s ", metric_only_len, unit);
914 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
921 if (stat_config.aggr_mode == AGGR_NONE)
924 if (stat_config.aggr_mode == AGGR_GLOBAL)
927 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
928 int cpu2 = perf_evsel__cpus(evsel)->map[i];
930 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
936 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
938 FILE *output = stat_config.output;
939 double sc = evsel->scale;
943 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
946 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
948 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
951 aggr_printout(evsel, id, nr);
953 fprintf(output, fmt, avg, csv_sep);
956 fprintf(output, "%-*s%s",
957 csv_output ? 0 : unit_width,
958 evsel->unit, csv_sep);
960 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
963 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
966 static bool is_mixed_hw_group(struct perf_evsel *counter)
968 struct perf_evlist *evlist = counter->evlist;
969 u32 pmu_type = counter->attr.type;
970 struct perf_evsel *pos;
972 if (counter->nr_members < 2)
975 evlist__for_each_entry(evlist, pos) {
976 /* software events can be part of any hardware group */
977 if (pos->attr.type == PERF_TYPE_SOFTWARE)
979 if (pmu_type == PERF_TYPE_SOFTWARE) {
980 pmu_type = pos->attr.type;
983 if (pmu_type != pos->attr.type)
990 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
991 char *prefix, u64 run, u64 ena, double noise,
992 struct runtime_stat *st)
994 struct perf_stat_output_ctx out;
995 struct outstate os = {
996 .fh = stat_config.output,
997 .prefix = prefix ? prefix : "",
1002 print_metric_t pm = print_metric_std;
1006 nl = new_line_metric;
1008 pm = print_metric_only_csv;
1010 pm = print_metric_only;
1014 if (csv_output && !metric_only) {
1015 static int aggr_fields[] = {
1023 pm = print_metric_csv;
1026 os.nfields += aggr_fields[stat_config.aggr_mode];
1030 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1032 pm(&os, NULL, "", "", 0);
1035 aggr_printout(counter, id, nr);
1037 fprintf(stat_config.output, "%*s%s",
1038 csv_output ? 0 : 18,
1039 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1042 if (counter->supported) {
1043 print_free_counters_hint = 1;
1044 if (is_mixed_hw_group(counter))
1045 print_mixed_hw_group_error = 1;
1048 fprintf(stat_config.output, "%-*s%s",
1049 csv_output ? 0 : unit_width,
1050 counter->unit, csv_sep);
1052 fprintf(stat_config.output, "%*s",
1053 csv_output ? 0 : -25,
1054 perf_evsel__name(counter));
1057 fprintf(stat_config.output, "%s%s",
1058 csv_sep, counter->cgrp->name);
1061 pm(&os, NULL, NULL, "", 0);
1062 print_noise(counter, noise);
1063 print_running(run, ena);
1065 pm(&os, NULL, NULL, "", 0);
1070 abs_printout(id, nr, counter, uval);
1072 out.print_metric = pm;
1075 out.force_header = false;
1077 if (csv_output && !metric_only) {
1078 print_noise(counter, noise);
1079 print_running(run, ena);
1082 perf_stat__print_shadow_stats(counter, uval,
1083 first_shadow_cpu(counter, id),
1084 &out, &metric_events, st);
1085 if (!csv_output && !metric_only) {
1086 print_noise(counter, noise);
1087 print_running(run, ena);
1091 static void aggr_update_shadow(void)
1095 struct perf_evsel *counter;
1097 for (s = 0; s < aggr_map->nr; s++) {
1098 id = aggr_map->map[s];
1099 evlist__for_each_entry(evsel_list, counter) {
1101 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1102 s2 = aggr_get_id(evsel_list->cpus, cpu);
1105 val += perf_counts(counter->counts, cpu, 0)->val;
1107 perf_stat__update_shadow_stats(counter, val,
1108 first_shadow_cpu(counter, id),
1114 static void uniquify_event_name(struct perf_evsel *counter)
1119 if (counter->uniquified_name ||
1120 !counter->pmu_name || !strncmp(counter->name, counter->pmu_name,
1121 strlen(counter->pmu_name)))
1124 config = strchr(counter->name, '/');
1126 if (asprintf(&new_name,
1127 "%s%s", counter->pmu_name, config) > 0) {
1128 free(counter->name);
1129 counter->name = new_name;
1132 if (asprintf(&new_name,
1133 "%s [%s]", counter->name, counter->pmu_name) > 0) {
1134 free(counter->name);
1135 counter->name = new_name;
1139 counter->uniquified_name = true;
1142 static void collect_all_aliases(struct perf_evsel *counter,
1143 void (*cb)(struct perf_evsel *counter, void *data,
1147 struct perf_evsel *alias;
1149 alias = list_prepare_entry(counter, &(evsel_list->entries), node);
1150 list_for_each_entry_continue (alias, &evsel_list->entries, node) {
1151 if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) ||
1152 alias->scale != counter->scale ||
1153 alias->cgrp != counter->cgrp ||
1154 strcmp(alias->unit, counter->unit) ||
1155 perf_evsel__is_clock(alias) != perf_evsel__is_clock(counter))
1157 alias->merged_stat = true;
1158 cb(alias, data, false);
1162 static bool collect_data(struct perf_evsel *counter,
1163 void (*cb)(struct perf_evsel *counter, void *data,
1167 if (counter->merged_stat)
1169 cb(counter, data, true);
1171 uniquify_event_name(counter);
1172 else if (counter->auto_merge_stats)
1173 collect_all_aliases(counter, cb, data);
1184 static void aggr_cb(struct perf_evsel *counter, void *data, bool first)
1186 struct aggr_data *ad = data;
1189 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1190 struct perf_counts_values *counts;
1192 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1197 counts = perf_counts(counter->counts, cpu, 0);
1199 * When any result is bad, make them all to give
1200 * consistent output in interval mode.
1202 if (counts->ena == 0 || counts->run == 0 ||
1203 counter->counts->scaled == -1) {
1208 ad->val += counts->val;
1209 ad->ena += counts->ena;
1210 ad->run += counts->run;
1214 static void print_aggr(char *prefix)
1216 FILE *output = stat_config.output;
1217 struct perf_evsel *counter;
1223 if (!(aggr_map || aggr_get_id))
1226 aggr_update_shadow();
1229 * With metric_only everything is on a single line.
1230 * Without each counter has its own line.
1232 for (s = 0; s < aggr_map->nr; s++) {
1233 struct aggr_data ad;
1234 if (prefix && metric_only)
1235 fprintf(output, "%s", prefix);
1237 ad.id = id = aggr_map->map[s];
1239 evlist__for_each_entry(evsel_list, counter) {
1240 if (is_duration_time(counter))
1243 ad.val = ad.ena = ad.run = 0;
1245 if (!collect_data(counter, aggr_cb, &ad))
1251 if (first && metric_only) {
1253 aggr_printout(counter, id, nr);
1255 if (prefix && !metric_only)
1256 fprintf(output, "%s", prefix);
1258 uval = val * counter->scale;
1259 printout(id, nr, counter, uval, prefix, run, ena, 1.0,
1262 fputc('\n', output);
1265 fputc('\n', output);
1269 static int cmp_val(const void *a, const void *b)
1271 return ((struct perf_aggr_thread_value *)b)->val -
1272 ((struct perf_aggr_thread_value *)a)->val;
1275 static struct perf_aggr_thread_value *sort_aggr_thread(
1276 struct perf_evsel *counter,
1277 int nthreads, int ncpus,
1280 int cpu, thread, i = 0;
1282 struct perf_aggr_thread_value *buf;
1284 buf = calloc(nthreads, sizeof(struct perf_aggr_thread_value));
1288 for (thread = 0; thread < nthreads; thread++) {
1289 u64 ena = 0, run = 0, val = 0;
1291 for (cpu = 0; cpu < ncpus; cpu++) {
1292 val += perf_counts(counter->counts, cpu, thread)->val;
1293 ena += perf_counts(counter->counts, cpu, thread)->ena;
1294 run += perf_counts(counter->counts, cpu, thread)->run;
1297 uval = val * counter->scale;
1300 * Skip value 0 when enabling --per-thread globally,
1301 * otherwise too many 0 output.
1303 if (uval == 0.0 && target__has_per_thread(&target))
1306 buf[i].counter = counter;
1315 qsort(buf, i, sizeof(struct perf_aggr_thread_value), cmp_val);
1323 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1325 FILE *output = stat_config.output;
1326 int nthreads = thread_map__nr(counter->threads);
1327 int ncpus = cpu_map__nr(counter->cpus);
1328 int thread, sorted_threads, id;
1329 struct perf_aggr_thread_value *buf;
1331 buf = sort_aggr_thread(counter, nthreads, ncpus, &sorted_threads);
1333 perror("cannot sort aggr thread");
1337 for (thread = 0; thread < sorted_threads; thread++) {
1339 fprintf(output, "%s", prefix);
1341 id = buf[thread].id;
1342 if (stat_config.stats)
1343 printout(id, 0, buf[thread].counter, buf[thread].uval,
1344 prefix, buf[thread].run, buf[thread].ena, 1.0,
1345 &stat_config.stats[id]);
1347 printout(id, 0, buf[thread].counter, buf[thread].uval,
1348 prefix, buf[thread].run, buf[thread].ena, 1.0,
1350 fputc('\n', output);
1357 double avg, avg_enabled, avg_running;
1360 static void counter_aggr_cb(struct perf_evsel *counter, void *data,
1361 bool first __maybe_unused)
1363 struct caggr_data *cd = data;
1364 struct perf_stat_evsel *ps = counter->stats;
1366 cd->avg += avg_stats(&ps->res_stats[0]);
1367 cd->avg_enabled += avg_stats(&ps->res_stats[1]);
1368 cd->avg_running += avg_stats(&ps->res_stats[2]);
1372 * Print out the results of a single counter:
1373 * aggregated counts in system-wide mode
1375 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1377 FILE *output = stat_config.output;
1379 struct caggr_data cd = { .avg = 0.0 };
1381 if (!collect_data(counter, counter_aggr_cb, &cd))
1384 if (prefix && !metric_only)
1385 fprintf(output, "%s", prefix);
1387 uval = cd.avg * counter->scale;
1388 printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled,
1391 fprintf(output, "\n");
1394 static void counter_cb(struct perf_evsel *counter, void *data,
1395 bool first __maybe_unused)
1397 struct aggr_data *ad = data;
1399 ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
1400 ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
1401 ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
1405 * Print out the results of a single counter:
1406 * does not use aggregated count in system-wide
1408 static void print_counter(struct perf_evsel *counter, char *prefix)
1410 FILE *output = stat_config.output;
1415 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1416 struct aggr_data ad = { .cpu = cpu };
1418 if (!collect_data(counter, counter_cb, &ad))
1425 fprintf(output, "%s", prefix);
1427 uval = val * counter->scale;
1428 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0,
1431 fputc('\n', output);
1435 static void print_no_aggr_metric(char *prefix)
1439 struct perf_evsel *counter;
1443 nrcpus = evsel_list->cpus->nr;
1444 for (cpu = 0; cpu < nrcpus; cpu++) {
1448 fputs(prefix, stat_config.output);
1449 evlist__for_each_entry(evsel_list, counter) {
1450 if (is_duration_time(counter))
1453 aggr_printout(counter, cpu, 0);
1456 val = perf_counts(counter->counts, cpu, 0)->val;
1457 ena = perf_counts(counter->counts, cpu, 0)->ena;
1458 run = perf_counts(counter->counts, cpu, 0)->run;
1460 uval = val * counter->scale;
1461 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0,
1464 fputc('\n', stat_config.output);
1468 static int aggr_header_lens[] = {
1476 static const char *aggr_header_csv[] = {
1477 [AGGR_CORE] = "core,cpus,",
1478 [AGGR_SOCKET] = "socket,cpus",
1479 [AGGR_NONE] = "cpu,",
1480 [AGGR_THREAD] = "comm-pid,",
1484 static void print_metric_headers(const char *prefix, bool no_indent)
1486 struct perf_stat_output_ctx out;
1487 struct perf_evsel *counter;
1488 struct outstate os = {
1489 .fh = stat_config.output
1493 fprintf(stat_config.output, "%s", prefix);
1495 if (!csv_output && !no_indent)
1496 fprintf(stat_config.output, "%*s",
1497 aggr_header_lens[stat_config.aggr_mode], "");
1499 if (stat_config.interval)
1500 fputs("time,", stat_config.output);
1501 fputs(aggr_header_csv[stat_config.aggr_mode],
1502 stat_config.output);
1505 /* Print metrics headers only */
1506 evlist__for_each_entry(evsel_list, counter) {
1507 if (is_duration_time(counter))
1511 out.print_metric = print_metric_header;
1512 out.new_line = new_line_metric;
1513 out.force_header = true;
1515 perf_stat__print_shadow_stats(counter, 0,
1521 fputc('\n', stat_config.output);
1524 static void print_interval(char *prefix, struct timespec *ts)
1526 FILE *output = stat_config.output;
1527 static int num_print_interval;
1530 puts(CONSOLE_CLEAR);
1532 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1534 if ((num_print_interval == 0 && !csv_output) || interval_clear) {
1535 switch (stat_config.aggr_mode) {
1537 fprintf(output, "# time socket cpus");
1539 fprintf(output, " counts %*s events\n", unit_width, "unit");
1542 fprintf(output, "# time core cpus");
1544 fprintf(output, " counts %*s events\n", unit_width, "unit");
1547 fprintf(output, "# time CPU ");
1549 fprintf(output, " counts %*s events\n", unit_width, "unit");
1552 fprintf(output, "# time comm-pid");
1554 fprintf(output, " counts %*s events\n", unit_width, "unit");
1558 fprintf(output, "# time");
1560 fprintf(output, " counts %*s events\n", unit_width, "unit");
1566 if ((num_print_interval == 0 || interval_clear) && metric_only)
1567 print_metric_headers(" ", true);
1568 if (++num_print_interval == 25)
1569 num_print_interval = 0;
1572 static void print_header(int argc, const char **argv)
1574 FILE *output = stat_config.output;
1580 fprintf(output, "\n");
1581 fprintf(output, " Performance counter stats for ");
1582 if (target.system_wide)
1583 fprintf(output, "\'system wide");
1584 else if (target.cpu_list)
1585 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1586 else if (!target__has_task(&target)) {
1587 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1588 for (i = 1; argv && (i < argc); i++)
1589 fprintf(output, " %s", argv[i]);
1590 } else if (target.pid)
1591 fprintf(output, "process id \'%s", target.pid);
1593 fprintf(output, "thread id \'%s", target.tid);
1595 fprintf(output, "\'");
1597 fprintf(output, " (%d runs)", run_count);
1598 fprintf(output, ":\n\n");
1602 static int get_precision(double num)
1607 return lround(ceil(-log10(num)));
1610 static void print_table(FILE *output, int precision, double avg)
1613 int idx, indent = 0;
1615 scnprintf(tmp, 64, " %17.*f", precision, avg);
1616 while (tmp[indent] == ' ')
1619 fprintf(output, "%*s# Table of individual measurements:\n", indent, "");
1621 for (idx = 0; idx < run_count; idx++) {
1622 double run = (double) walltime_run[idx] / NSEC_PER_SEC;
1623 int h, n = 1 + abs((int) (100.0 * (run - avg)/run) / 5);
1625 fprintf(output, " %17.*f (%+.*f) ",
1626 precision, run, precision, run - avg);
1628 for (h = 0; h < n; h++)
1629 fprintf(output, "#");
1631 fprintf(output, "\n");
1634 fprintf(output, "\n%*s# Final result:\n", indent, "");
1637 static double timeval2double(struct timeval *t)
1639 return t->tv_sec + (double) t->tv_usec/USEC_PER_SEC;
1642 static void print_footer(void)
1644 double avg = avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC;
1645 FILE *output = stat_config.output;
1649 fprintf(output, "\n");
1651 if (run_count == 1) {
1652 fprintf(output, " %17.9f seconds time elapsed", avg);
1655 double ru_utime = timeval2double(&ru_data.ru_utime);
1656 double ru_stime = timeval2double(&ru_data.ru_stime);
1658 fprintf(output, "\n\n");
1659 fprintf(output, " %17.9f seconds user\n", ru_utime);
1660 fprintf(output, " %17.9f seconds sys\n", ru_stime);
1663 double sd = stddev_stats(&walltime_nsecs_stats) / NSEC_PER_SEC;
1665 * Display at most 2 more significant
1666 * digits than the stddev inaccuracy.
1668 int precision = get_precision(sd) + 2;
1670 if (walltime_run_table)
1671 print_table(output, precision, avg);
1673 fprintf(output, " %17.*f +- %.*f seconds time elapsed",
1674 precision, avg, precision, sd);
1676 print_noise_pct(sd, avg);
1678 fprintf(output, "\n\n");
1680 if (print_free_counters_hint &&
1681 sysctl__read_int("kernel/nmi_watchdog", &n) >= 0 &&
1684 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1685 " echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1687 " echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1689 if (print_mixed_hw_group_error)
1691 "The events in group usually have to be from "
1692 "the same PMU. Try reorganizing the group.\n");
1696 perf_evlist__print_counters(struct perf_evlist *evlist,
1697 struct timespec *ts,
1698 int argc, const char **argv)
1700 int interval = stat_config.interval;
1701 struct perf_evsel *counter;
1702 char buf[64], *prefix = NULL;
1705 print_interval(prefix = buf, ts);
1707 print_header(argc, argv);
1710 static int num_print_iv;
1712 if (num_print_iv == 0 && !interval)
1713 print_metric_headers(prefix, false);
1714 if (num_print_iv++ == 25)
1716 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1717 fprintf(stat_config.output, "%s", prefix);
1720 switch (stat_config.aggr_mode) {
1726 evlist__for_each_entry(evlist, counter) {
1727 if (is_duration_time(counter))
1729 print_aggr_thread(counter, prefix);
1733 evlist__for_each_entry(evlist, counter) {
1734 if (is_duration_time(counter))
1736 print_counter_aggr(counter, prefix);
1739 fputc('\n', stat_config.output);
1743 print_no_aggr_metric(prefix);
1745 evlist__for_each_entry(evlist, counter) {
1746 if (is_duration_time(counter))
1748 print_counter(counter, prefix);
1757 if (!interval && !csv_output)
1760 fflush(stat_config.output);
1763 static void print_counters(struct timespec *ts, int argc, const char **argv)
1765 /* Do not print anything if we record to the pipe. */
1766 if (STAT_RECORD && perf_stat.data.is_pipe)
1769 perf_evlist__print_counters(evsel_list, ts, argc, argv);
1772 static volatile int signr = -1;
1774 static void skip_signal(int signo)
1776 if ((child_pid == -1) || stat_config.interval)
1781 * render child_pid harmless
1782 * won't send SIGTERM to a random
1783 * process in case of race condition
1784 * and fast PID recycling
1789 static void sig_atexit(void)
1794 * avoid race condition with SIGCHLD handler
1795 * in skip_signal() which is modifying child_pid
1796 * goal is to avoid send SIGTERM to a random
1800 sigaddset(&set, SIGCHLD);
1801 sigprocmask(SIG_BLOCK, &set, &oset);
1803 if (child_pid != -1)
1804 kill(child_pid, SIGTERM);
1806 sigprocmask(SIG_SETMASK, &oset, NULL);
1811 signal(signr, SIG_DFL);
1812 kill(getpid(), signr);
1815 static int stat__set_big_num(const struct option *opt __maybe_unused,
1816 const char *s __maybe_unused, int unset)
1818 big_num_opt = unset ? 0 : 1;
1822 static int enable_metric_only(const struct option *opt __maybe_unused,
1823 const char *s __maybe_unused, int unset)
1825 force_metric_only = true;
1826 metric_only = !unset;
1830 static int parse_metric_groups(const struct option *opt,
1832 int unset __maybe_unused)
1834 return metricgroup__parse_groups(opt, str, &metric_events);
1837 static const struct option stat_options[] = {
1838 OPT_BOOLEAN('T', "transaction", &transaction_run,
1839 "hardware transaction statistics"),
1840 OPT_CALLBACK('e', "event", &evsel_list, "event",
1841 "event selector. use 'perf list' to list available events",
1842 parse_events_option),
1843 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1844 "event filter", parse_filter),
1845 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1846 "child tasks do not inherit counters"),
1847 OPT_STRING('p', "pid", &target.pid, "pid",
1848 "stat events on existing process id"),
1849 OPT_STRING('t', "tid", &target.tid, "tid",
1850 "stat events on existing thread id"),
1851 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1852 "system-wide collection from all CPUs"),
1853 OPT_BOOLEAN('g', "group", &group,
1854 "put the counters into a counter group"),
1855 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1856 OPT_INCR('v', "verbose", &verbose,
1857 "be more verbose (show counter open errors, etc)"),
1858 OPT_INTEGER('r', "repeat", &run_count,
1859 "repeat command and print average + stddev (max: 100, forever: 0)"),
1860 OPT_BOOLEAN(0, "table", &walltime_run_table,
1861 "display details about each run (only with -r option)"),
1862 OPT_BOOLEAN('n', "null", &null_run,
1863 "null run - dont start any counters"),
1864 OPT_INCR('d', "detailed", &detailed_run,
1865 "detailed run - start a lot of events"),
1866 OPT_BOOLEAN('S', "sync", &sync_run,
1867 "call sync() before starting a run"),
1868 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1869 "print large numbers with thousands\' separators",
1871 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1872 "list of cpus to monitor in system-wide"),
1873 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1874 "disable CPU count aggregation", AGGR_NONE),
1875 OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"),
1876 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1877 "print counts with custom separator"),
1878 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1879 "monitor event in cgroup name only", parse_cgroups),
1880 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1881 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1882 OPT_INTEGER(0, "log-fd", &output_fd,
1883 "log output to fd, instead of stderr"),
1884 OPT_STRING(0, "pre", &pre_cmd, "command",
1885 "command to run prior to the measured command"),
1886 OPT_STRING(0, "post", &post_cmd, "command",
1887 "command to run after to the measured command"),
1888 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1889 "print counts at regular interval in ms "
1890 "(overhead is possible for values <= 100ms)"),
1891 OPT_INTEGER(0, "interval-count", &stat_config.times,
1892 "print counts for fixed number of times"),
1893 OPT_BOOLEAN(0, "interval-clear", &interval_clear,
1894 "clear screen in between new interval"),
1895 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1896 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1897 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1898 "aggregate counts per processor socket", AGGR_SOCKET),
1899 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1900 "aggregate counts per physical processor core", AGGR_CORE),
1901 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1902 "aggregate counts per thread", AGGR_THREAD),
1903 OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
1904 "ms to wait before starting measurement after program start"),
1905 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1906 "Only print computed metrics. No raw values", enable_metric_only),
1907 OPT_BOOLEAN(0, "topdown", &topdown_run,
1908 "measure topdown level 1 statistics"),
1909 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1910 "measure SMI cost"),
1911 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1912 "monitor specified metrics or metric groups (separated by ,)",
1913 parse_metric_groups),
1917 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1919 return cpu_map__get_socket(map, cpu, NULL);
1922 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1924 return cpu_map__get_core(map, cpu, NULL);
1927 static int cpu_map__get_max(struct cpu_map *map)
1931 for (i = 0; i < map->nr; i++) {
1932 if (map->map[i] > max)
1939 static struct cpu_map *cpus_aggr_map;
1941 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1948 cpu = map->map[idx];
1950 if (cpus_aggr_map->map[cpu] == -1)
1951 cpus_aggr_map->map[cpu] = get_id(map, idx);
1953 return cpus_aggr_map->map[cpu];
1956 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1958 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1961 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1963 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1966 static int perf_stat_init_aggr_mode(void)
1970 switch (stat_config.aggr_mode) {
1972 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1973 perror("cannot build socket map");
1976 aggr_get_id = perf_stat__get_socket_cached;
1979 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1980 perror("cannot build core map");
1983 aggr_get_id = perf_stat__get_core_cached;
1994 * The evsel_list->cpus is the base we operate on,
1995 * taking the highest cpu number to be the size of
1996 * the aggregation translate cpumap.
1998 nr = cpu_map__get_max(evsel_list->cpus);
1999 cpus_aggr_map = cpu_map__empty_new(nr + 1);
2000 return cpus_aggr_map ? 0 : -ENOMEM;
2003 static void perf_stat__exit_aggr_mode(void)
2005 cpu_map__put(aggr_map);
2006 cpu_map__put(cpus_aggr_map);
2008 cpus_aggr_map = NULL;
2011 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
2018 cpu = map->map[idx];
2020 if (cpu >= env->nr_cpus_avail)
2026 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
2028 struct perf_env *env = data;
2029 int cpu = perf_env__get_cpu(env, map, idx);
2031 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
2034 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
2036 struct perf_env *env = data;
2037 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
2040 int socket_id = env->cpu[cpu].socket_id;
2043 * Encode socket in upper 16 bits
2044 * core_id is relative to socket, and
2045 * we need a global id. So we combine
2048 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
2054 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
2055 struct cpu_map **sockp)
2057 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
2060 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
2061 struct cpu_map **corep)
2063 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
2066 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
2068 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
2071 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
2073 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
2076 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
2078 struct perf_env *env = &st->session->header.env;
2080 switch (stat_config.aggr_mode) {
2082 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
2083 perror("cannot build socket map");
2086 aggr_get_id = perf_stat__get_socket_file;
2089 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
2090 perror("cannot build core map");
2093 aggr_get_id = perf_stat__get_core_file;
2106 static int topdown_filter_events(const char **attr, char **str, bool use_group)
2113 for (i = 0; attr[i]; i++) {
2114 if (pmu_have_event("cpu", attr[i])) {
2115 len += strlen(attr[i]) + 1;
2116 attr[i - off] = attr[i];
2120 attr[i - off] = NULL;
2122 *str = malloc(len + 1 + 2);
2132 for (i = 0; attr[i]; i++) {
2145 __weak bool arch_topdown_check_group(bool *warn)
2151 __weak void arch_topdown_group_warn(void)
2156 * Add default attributes, if there were no attributes specified or
2157 * if -d/--detailed, -d -d or -d -d -d is used:
2159 static int add_default_attributes(void)
2162 struct perf_event_attr default_attrs0[] = {
2164 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
2165 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
2166 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
2167 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
2169 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
2171 struct perf_event_attr frontend_attrs[] = {
2172 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
2174 struct perf_event_attr backend_attrs[] = {
2175 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
2177 struct perf_event_attr default_attrs1[] = {
2178 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
2179 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
2180 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
2185 * Detailed stats (-d), covering the L1 and last level data caches:
2187 struct perf_event_attr detailed_attrs[] = {
2189 { .type = PERF_TYPE_HW_CACHE,
2191 PERF_COUNT_HW_CACHE_L1D << 0 |
2192 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2193 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2195 { .type = PERF_TYPE_HW_CACHE,
2197 PERF_COUNT_HW_CACHE_L1D << 0 |
2198 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2199 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2201 { .type = PERF_TYPE_HW_CACHE,
2203 PERF_COUNT_HW_CACHE_LL << 0 |
2204 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2205 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2207 { .type = PERF_TYPE_HW_CACHE,
2209 PERF_COUNT_HW_CACHE_LL << 0 |
2210 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2211 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2215 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2217 struct perf_event_attr very_detailed_attrs[] = {
2219 { .type = PERF_TYPE_HW_CACHE,
2221 PERF_COUNT_HW_CACHE_L1I << 0 |
2222 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2223 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2225 { .type = PERF_TYPE_HW_CACHE,
2227 PERF_COUNT_HW_CACHE_L1I << 0 |
2228 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2229 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2231 { .type = PERF_TYPE_HW_CACHE,
2233 PERF_COUNT_HW_CACHE_DTLB << 0 |
2234 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2235 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2237 { .type = PERF_TYPE_HW_CACHE,
2239 PERF_COUNT_HW_CACHE_DTLB << 0 |
2240 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2241 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2243 { .type = PERF_TYPE_HW_CACHE,
2245 PERF_COUNT_HW_CACHE_ITLB << 0 |
2246 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2247 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2249 { .type = PERF_TYPE_HW_CACHE,
2251 PERF_COUNT_HW_CACHE_ITLB << 0 |
2252 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2253 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2258 * Very, very detailed stats (-d -d -d), adding prefetch events:
2260 struct perf_event_attr very_very_detailed_attrs[] = {
2262 { .type = PERF_TYPE_HW_CACHE,
2264 PERF_COUNT_HW_CACHE_L1D << 0 |
2265 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2266 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2268 { .type = PERF_TYPE_HW_CACHE,
2270 PERF_COUNT_HW_CACHE_L1D << 0 |
2271 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2272 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2274 struct parse_events_error errinfo;
2276 /* Set attrs if no event is selected and !null_run: */
2280 if (transaction_run) {
2281 /* Handle -T as -M transaction. Once platform specific metrics
2282 * support has been added to the json files, all archictures
2283 * will use this approach. To determine transaction support
2284 * on an architecture test for such a metric name.
2286 if (metricgroup__has_metric("transaction")) {
2287 struct option opt = { .value = &evsel_list };
2289 return metricgroup__parse_groups(&opt, "transaction",
2293 if (pmu_have_event("cpu", "cycles-ct") &&
2294 pmu_have_event("cpu", "el-start"))
2295 err = parse_events(evsel_list, transaction_attrs,
2298 err = parse_events(evsel_list,
2299 transaction_limited_attrs,
2302 fprintf(stderr, "Cannot set up transaction events\n");
2303 parse_events_print_error(&errinfo, transaction_attrs);
2312 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2313 fprintf(stderr, "freeze_on_smi is not supported.\n");
2318 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2319 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2325 if (pmu_have_event("msr", "aperf") &&
2326 pmu_have_event("msr", "smi")) {
2327 if (!force_metric_only)
2329 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
2331 fprintf(stderr, "To measure SMI cost, it needs "
2332 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
2333 parse_events_print_error(&errinfo, smi_cost_attrs);
2337 fprintf(stderr, "Cannot set up SMI cost events\n");
2347 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2348 stat_config.aggr_mode != AGGR_CORE) {
2349 pr_err("top down event configuration requires --per-core mode\n");
2352 stat_config.aggr_mode = AGGR_CORE;
2353 if (nr_cgroups || !target__has_cpu(&target)) {
2354 pr_err("top down event configuration requires system-wide mode (-a)\n");
2358 if (!force_metric_only)
2360 if (topdown_filter_events(topdown_attrs, &str,
2361 arch_topdown_check_group(&warn)) < 0) {
2362 pr_err("Out of memory\n");
2365 if (topdown_attrs[0] && str) {
2367 arch_topdown_group_warn();
2368 err = parse_events(evsel_list, str, &errinfo);
2371 "Cannot set up top down events %s: %d\n",
2374 parse_events_print_error(&errinfo, str);
2378 fprintf(stderr, "System does not support topdown\n");
2384 if (!evsel_list->nr_entries) {
2385 if (target__has_cpu(&target))
2386 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2388 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2390 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2391 if (perf_evlist__add_default_attrs(evsel_list,
2392 frontend_attrs) < 0)
2395 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2396 if (perf_evlist__add_default_attrs(evsel_list,
2400 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2404 /* Detailed events get appended to the event list: */
2406 if (detailed_run < 1)
2409 /* Append detailed run extra attributes: */
2410 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2413 if (detailed_run < 2)
2416 /* Append very detailed run extra attributes: */
2417 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2420 if (detailed_run < 3)
2423 /* Append very, very detailed run extra attributes: */
2424 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2427 static const char * const stat_record_usage[] = {
2428 "perf stat record [<options>]",
2432 static void init_features(struct perf_session *session)
2436 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2437 perf_header__set_feat(&session->header, feat);
2439 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2440 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2441 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2442 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2445 static int __cmd_record(int argc, const char **argv)
2447 struct perf_session *session;
2448 struct perf_data *data = &perf_stat.data;
2450 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2451 PARSE_OPT_STOP_AT_NON_OPTION);
2454 data->file.path = output_name;
2456 if (run_count != 1 || forever) {
2457 pr_err("Cannot use -r option with perf stat record.\n");
2461 session = perf_session__new(data, false, NULL);
2462 if (session == NULL) {
2463 pr_err("Perf session creation failed.\n");
2467 init_features(session);
2469 session->evlist = evsel_list;
2470 perf_stat.session = session;
2471 perf_stat.record = true;
2475 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2476 union perf_event *event,
2477 struct perf_session *session)
2479 struct stat_round_event *stat_round = &event->stat_round;
2480 struct perf_evsel *counter;
2481 struct timespec tsh, *ts = NULL;
2482 const char **argv = session->header.env.cmdline_argv;
2483 int argc = session->header.env.nr_cmdline;
2485 evlist__for_each_entry(evsel_list, counter)
2486 perf_stat_process_counter(&stat_config, counter);
2488 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2489 update_stats(&walltime_nsecs_stats, stat_round->time);
2491 if (stat_config.interval && stat_round->time) {
2492 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2493 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2497 print_counters(ts, argc, argv);
2502 int process_stat_config_event(struct perf_tool *tool,
2503 union perf_event *event,
2504 struct perf_session *session __maybe_unused)
2506 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2508 perf_event__read_stat_config(&stat_config, &event->stat_config);
2510 if (cpu_map__empty(st->cpus)) {
2511 if (st->aggr_mode != AGGR_UNSET)
2512 pr_warning("warning: processing task data, aggregation mode not set\n");
2516 if (st->aggr_mode != AGGR_UNSET)
2517 stat_config.aggr_mode = st->aggr_mode;
2519 if (perf_stat.data.is_pipe)
2520 perf_stat_init_aggr_mode();
2522 perf_stat_init_aggr_mode_file(st);
2527 static int set_maps(struct perf_stat *st)
2529 if (!st->cpus || !st->threads)
2532 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2535 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2537 if (perf_evlist__alloc_stats(evsel_list, true))
2540 st->maps_allocated = true;
2545 int process_thread_map_event(struct perf_tool *tool,
2546 union perf_event *event,
2547 struct perf_session *session __maybe_unused)
2549 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2552 pr_warning("Extra thread map event, ignoring.\n");
2556 st->threads = thread_map__new_event(&event->thread_map);
2560 return set_maps(st);
2564 int process_cpu_map_event(struct perf_tool *tool,
2565 union perf_event *event,
2566 struct perf_session *session __maybe_unused)
2568 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2569 struct cpu_map *cpus;
2572 pr_warning("Extra cpu map event, ignoring.\n");
2576 cpus = cpu_map__new_data(&event->cpu_map.data);
2581 return set_maps(st);
2584 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
2588 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
2592 config->stats_num = nthreads;
2594 for (i = 0; i < nthreads; i++)
2595 runtime_stat__init(&config->stats[i]);
2600 static void runtime_stat_delete(struct perf_stat_config *config)
2607 for (i = 0; i < config->stats_num; i++)
2608 runtime_stat__exit(&config->stats[i]);
2610 free(config->stats);
2613 static const char * const stat_report_usage[] = {
2614 "perf stat report [<options>]",
2618 static struct perf_stat perf_stat = {
2620 .attr = perf_event__process_attr,
2621 .event_update = perf_event__process_event_update,
2622 .thread_map = process_thread_map_event,
2623 .cpu_map = process_cpu_map_event,
2624 .stat_config = process_stat_config_event,
2625 .stat = perf_event__process_stat_event,
2626 .stat_round = process_stat_round_event,
2628 .aggr_mode = AGGR_UNSET,
2631 static int __cmd_report(int argc, const char **argv)
2633 struct perf_session *session;
2634 const struct option options[] = {
2635 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2636 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2637 "aggregate counts per processor socket", AGGR_SOCKET),
2638 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2639 "aggregate counts per physical processor core", AGGR_CORE),
2640 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2641 "disable CPU count aggregation", AGGR_NONE),
2647 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2649 if (!input_name || !strlen(input_name)) {
2650 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2653 input_name = "perf.data";
2656 perf_stat.data.file.path = input_name;
2657 perf_stat.data.mode = PERF_DATA_MODE_READ;
2659 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2660 if (session == NULL)
2663 perf_stat.session = session;
2664 stat_config.output = stderr;
2665 evsel_list = session->evlist;
2667 ret = perf_session__process_events(session);
2671 perf_session__delete(session);
2675 static void setup_system_wide(int forks)
2678 * Make system wide (-a) the default target if
2679 * no target was specified and one of following
2680 * conditions is met:
2682 * - there's no workload specified
2683 * - there is workload specified but all requested
2684 * events are system wide events
2686 if (!target__none(&target))
2690 target.system_wide = true;
2692 struct perf_evsel *counter;
2694 evlist__for_each_entry(evsel_list, counter) {
2695 if (!counter->system_wide)
2699 if (evsel_list->nr_entries)
2700 target.system_wide = true;
2704 int cmd_stat(int argc, const char **argv)
2706 const char * const stat_usage[] = {
2707 "perf stat [<options>] [<command>]",
2710 int status = -EINVAL, run_idx;
2712 FILE *output = stderr;
2713 unsigned int interval, timeout;
2714 const char * const stat_subcommands[] = { "record", "report" };
2716 setlocale(LC_ALL, "");
2718 evsel_list = perf_evlist__new();
2719 if (evsel_list == NULL)
2722 parse_events__shrink_config_terms();
2723 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2724 (const char **) stat_usage,
2725 PARSE_OPT_STOP_AT_NON_OPTION);
2726 perf_stat__collect_metric_expr(evsel_list);
2727 perf_stat__init_shadow_stats();
2731 if (!strcmp(csv_sep, "\\t"))
2734 csv_sep = DEFAULT_SEPARATOR;
2736 if (argc && !strncmp(argv[0], "rec", 3)) {
2737 argc = __cmd_record(argc, argv);
2740 } else if (argc && !strncmp(argv[0], "rep", 3))
2741 return __cmd_report(argc, argv);
2743 interval = stat_config.interval;
2744 timeout = stat_config.timeout;
2747 * For record command the -o is already taken care of.
2749 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2752 if (output_name && output_fd) {
2753 fprintf(stderr, "cannot use both --output and --log-fd\n");
2754 parse_options_usage(stat_usage, stat_options, "o", 1);
2755 parse_options_usage(NULL, stat_options, "log-fd", 0);
2759 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2760 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2764 if (metric_only && run_count > 1) {
2765 fprintf(stderr, "--metric-only is not supported with -r\n");
2769 if (walltime_run_table && run_count <= 1) {
2770 fprintf(stderr, "--table is only supported with -r\n");
2771 parse_options_usage(stat_usage, stat_options, "r", 1);
2772 parse_options_usage(NULL, stat_options, "table", 0);
2776 if (output_fd < 0) {
2777 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2778 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2784 mode = append_file ? "a" : "w";
2786 output = fopen(output_name, mode);
2788 perror("failed to create output file");
2791 clock_gettime(CLOCK_REALTIME, &tm);
2792 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2793 } else if (output_fd > 0) {
2794 mode = append_file ? "a" : "w";
2795 output = fdopen(output_fd, mode);
2797 perror("Failed opening logfd");
2802 stat_config.output = output;
2805 * let the spreadsheet do the pretty-printing
2808 /* User explicitly passed -B? */
2809 if (big_num_opt == 1) {
2810 fprintf(stderr, "-B option not supported with -x\n");
2811 parse_options_usage(stat_usage, stat_options, "B", 1);
2812 parse_options_usage(NULL, stat_options, "x", 1);
2814 } else /* Nope, so disable big number formatting */
2816 } else if (big_num_opt == 0) /* User passed --no-big-num */
2819 setup_system_wide(argc);
2822 * Display user/system times only for single
2823 * run and when there's specified tracee.
2825 if ((run_count == 1) && target__none(&target))
2828 if (run_count < 0) {
2829 pr_err("Run count must be a positive number\n");
2830 parse_options_usage(stat_usage, stat_options, "r", 1);
2832 } else if (run_count == 0) {
2837 if (walltime_run_table) {
2838 walltime_run = zalloc(run_count * sizeof(walltime_run[0]));
2839 if (!walltime_run) {
2840 pr_err("failed to setup -r option");
2845 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2846 !target__has_task(&target)) {
2847 if (!target.system_wide || target.cpu_list) {
2848 fprintf(stderr, "The --per-thread option is only "
2849 "available when monitoring via -p -t -a "
2850 "options or only --per-thread.\n");
2851 parse_options_usage(NULL, stat_options, "p", 1);
2852 parse_options_usage(NULL, stat_options, "t", 1);
2858 * no_aggr, cgroup are for system-wide only
2859 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2861 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2862 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2863 !target__has_cpu(&target)) {
2864 fprintf(stderr, "both cgroup and no-aggregation "
2865 "modes only available in system-wide mode\n");
2867 parse_options_usage(stat_usage, stat_options, "G", 1);
2868 parse_options_usage(NULL, stat_options, "A", 1);
2869 parse_options_usage(NULL, stat_options, "a", 1);
2873 if (add_default_attributes())
2876 target__validate(&target);
2878 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2879 target.per_thread = true;
2881 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2882 if (target__has_task(&target)) {
2883 pr_err("Problems finding threads of monitor\n");
2884 parse_options_usage(stat_usage, stat_options, "p", 1);
2885 parse_options_usage(NULL, stat_options, "t", 1);
2886 } else if (target__has_cpu(&target)) {
2887 perror("failed to parse CPUs map");
2888 parse_options_usage(stat_usage, stat_options, "C", 1);
2889 parse_options_usage(NULL, stat_options, "a", 1);
2895 * Initialize thread_map with comm names,
2896 * so we could print it out on output.
2898 if (stat_config.aggr_mode == AGGR_THREAD) {
2899 thread_map__read_comms(evsel_list->threads);
2900 if (target.system_wide) {
2901 if (runtime_stat_new(&stat_config,
2902 thread_map__nr(evsel_list->threads))) {
2908 if (stat_config.times && interval)
2909 interval_count = true;
2910 else if (stat_config.times && !interval) {
2911 pr_err("interval-count option should be used together with "
2912 "interval-print.\n");
2913 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2914 parse_options_usage(stat_usage, stat_options, "I", 1);
2918 if (timeout && timeout < 100) {
2920 pr_err("timeout must be >= 10ms.\n");
2921 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2924 pr_warning("timeout < 100ms. "
2925 "The overhead percentage could be high in some cases. "
2926 "Please proceed with caution.\n");
2928 if (timeout && interval) {
2929 pr_err("timeout option is not supported with interval-print.\n");
2930 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2931 parse_options_usage(stat_usage, stat_options, "I", 1);
2935 if (perf_evlist__alloc_stats(evsel_list, interval))
2938 if (perf_stat_init_aggr_mode())
2942 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2943 * while avoiding that older tools show confusing messages.
2945 * However for pipe sessions we need to keep it zero,
2946 * because script's perf_evsel__check_attr is triggered
2947 * by attr->sample_type != 0, and we can't run it on
2950 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2953 * We dont want to block the signals - that would cause
2954 * child tasks to inherit that and Ctrl-C would not work.
2955 * What we want is for Ctrl-C to work in the exec()-ed
2956 * task, but being ignored by perf stat itself:
2960 signal(SIGINT, skip_signal);
2961 signal(SIGCHLD, skip_signal);
2962 signal(SIGALRM, skip_signal);
2963 signal(SIGABRT, skip_signal);
2966 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2967 if (run_count != 1 && verbose > 0)
2968 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2971 status = run_perf_stat(argc, argv, run_idx);
2972 if (forever && status != -1) {
2973 print_counters(NULL, argc, argv);
2974 perf_stat__reset_stats();
2978 if (!forever && status != -1 && !interval)
2979 print_counters(NULL, argc, argv);
2983 * We synthesize the kernel mmap record just so that older tools
2984 * don't emit warnings about not being able to resolve symbols
2985 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2986 * a saner message about no samples being in the perf.data file.
2988 * This also serves to suppress a warning about f_header.data.size == 0
2989 * in header.c at the moment 'perf stat record' gets introduced, which
2990 * is not really needed once we start adding the stat specific PERF_RECORD_
2991 * records, but the need to suppress the kptr_restrict messages in older
2992 * tools remain -acme
2994 int fd = perf_data__fd(&perf_stat.data);
2995 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2996 process_synthesized_event,
2997 &perf_stat.session->machines.host);
2999 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
3000 "older tools may produce warnings about this file\n.");
3004 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
3005 pr_err("failed to write stat round event\n");
3008 if (!perf_stat.data.is_pipe) {
3009 perf_stat.session->header.data_size += perf_stat.bytes_written;
3010 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
3013 perf_session__delete(perf_stat.session);
3016 perf_stat__exit_aggr_mode();
3017 perf_evlist__free_stats(evsel_list);
3021 if (smi_cost && smi_reset)
3022 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
3024 perf_evlist__delete(evsel_list);
3026 runtime_stat_delete(&stat_config);