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/color.h"
56 #include "util/stat.h"
57 #include "util/header.h"
58 #include "util/cpumap.h"
59 #include "util/thread.h"
60 #include "util/thread_map.h"
61 #include "util/counts.h"
62 #include "util/session.h"
63 #include "util/tool.h"
67 #include <sys/prctl.h>
71 #define DEFAULT_SEPARATOR " "
72 #define CNTR_NOT_SUPPORTED "<not supported>"
73 #define CNTR_NOT_COUNTED "<not counted>"
75 static void print_counters(struct timespec *ts, int argc, const char **argv);
77 /* Default events used for perf stat -T */
78 static const char *transaction_attrs = {
90 /* More limited version when the CPU does not have all events. */
91 static const char * transaction_limited_attrs = {
101 static struct perf_evlist *evsel_list;
103 static struct target target = {
107 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
109 static int run_count = 1;
110 static bool no_inherit = false;
111 static volatile pid_t child_pid = -1;
112 static bool null_run = false;
113 static int detailed_run = 0;
114 static bool transaction_run;
115 static bool big_num = true;
116 static int big_num_opt = -1;
117 static const char *csv_sep = NULL;
118 static bool csv_output = false;
119 static bool group = false;
120 static const char *pre_cmd = NULL;
121 static const char *post_cmd = NULL;
122 static bool sync_run = false;
123 static unsigned int initial_delay = 0;
124 static unsigned int unit_width = 4; /* strlen("unit") */
125 static bool forever = false;
126 static bool metric_only = false;
127 static struct timespec ref_time;
128 static struct cpu_map *aggr_map;
129 static aggr_get_id_t aggr_get_id;
130 static bool append_file;
131 static const char *output_name;
132 static int output_fd;
136 struct perf_data_file file;
137 struct perf_session *session;
139 struct perf_tool tool;
141 struct cpu_map *cpus;
142 struct thread_map *threads;
143 enum aggr_mode aggr_mode;
146 static struct perf_stat perf_stat;
147 #define STAT_RECORD perf_stat.record
149 static volatile int done = 0;
151 static struct perf_stat_config stat_config = {
152 .aggr_mode = AGGR_GLOBAL,
156 static inline void diff_timespec(struct timespec *r, struct timespec *a,
159 r->tv_sec = a->tv_sec - b->tv_sec;
160 if (a->tv_nsec < b->tv_nsec) {
161 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
164 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
168 static void perf_stat__reset_stats(void)
170 perf_evlist__reset_stats(evsel_list);
171 perf_stat__reset_shadow_stats();
174 static int create_perf_stat_counter(struct perf_evsel *evsel)
176 struct perf_event_attr *attr = &evsel->attr;
178 if (stat_config.scale)
179 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
180 PERF_FORMAT_TOTAL_TIME_RUNNING;
182 attr->inherit = !no_inherit;
185 * Some events get initialized with sample_(period/type) set,
186 * like tracepoints. Clear it up for counting.
188 attr->sample_period = 0;
191 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
192 * while avoiding that older tools show confusing messages.
194 * However for pipe sessions we need to keep it zero,
195 * because script's perf_evsel__check_attr is triggered
196 * by attr->sample_type != 0, and we can't run it on
199 if (!(STAT_RECORD && perf_stat.file.is_pipe))
200 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
203 * Disabling all counters initially, they will be enabled
204 * either manually by us or by kernel via enable_on_exec
207 if (perf_evsel__is_group_leader(evsel)) {
211 * In case of initial_delay we enable tracee
214 if (target__none(&target) && !initial_delay)
215 attr->enable_on_exec = 1;
218 if (target__has_cpu(&target))
219 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
221 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
225 * Does the counter have nsecs as a unit?
227 static inline int nsec_counter(struct perf_evsel *evsel)
229 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
230 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
236 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
237 union perf_event *event,
238 struct perf_sample *sample __maybe_unused,
239 struct machine *machine __maybe_unused)
241 if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
242 pr_err("failed to write perf data, error: %m\n");
246 perf_stat.bytes_written += event->header.size;
250 static int write_stat_round_event(u64 tm, u64 type)
252 return perf_event__synthesize_stat_round(NULL, tm, type,
253 process_synthesized_event,
257 #define WRITE_STAT_ROUND_EVENT(time, interval) \
258 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
260 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
263 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
264 struct perf_counts_values *count)
266 struct perf_sample_id *sid = SID(counter, cpu, thread);
268 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
269 process_synthesized_event, NULL);
273 * Read out the results of a single counter:
274 * do not aggregate counts across CPUs in system-wide mode
276 static int read_counter(struct perf_evsel *counter)
278 int nthreads = thread_map__nr(evsel_list->threads);
279 int ncpus = perf_evsel__nr_cpus(counter);
282 if (!counter->supported)
285 if (counter->system_wide)
288 for (thread = 0; thread < nthreads; thread++) {
289 for (cpu = 0; cpu < ncpus; cpu++) {
290 struct perf_counts_values *count;
292 count = perf_counts(counter->counts, cpu, thread);
293 if (perf_evsel__read(counter, cpu, thread, count))
297 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
298 pr_err("failed to write stat event\n");
304 fprintf(stat_config.output,
305 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
306 perf_evsel__name(counter),
308 count->val, count->ena, count->run);
316 static void read_counters(bool close_counters)
318 struct perf_evsel *counter;
320 evlist__for_each(evsel_list, counter) {
321 if (read_counter(counter))
322 pr_debug("failed to read counter %s\n", counter->name);
324 if (perf_stat_process_counter(&stat_config, counter))
325 pr_warning("failed to process counter %s\n", counter->name);
327 if (close_counters) {
328 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
329 thread_map__nr(evsel_list->threads));
334 static void process_interval(void)
336 struct timespec ts, rs;
338 read_counters(false);
340 clock_gettime(CLOCK_MONOTONIC, &ts);
341 diff_timespec(&rs, &ts, &ref_time);
344 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSECS_PER_SEC + rs.tv_nsec, INTERVAL))
345 pr_err("failed to write stat round event\n");
348 print_counters(&rs, 0, NULL);
351 static void enable_counters(void)
354 usleep(initial_delay * 1000);
357 * We need to enable counters only if:
358 * - we don't have tracee (attaching to task or cpu)
359 * - we have initial delay configured
361 if (!target__none(&target) || initial_delay)
362 perf_evlist__enable(evsel_list);
365 static volatile int workload_exec_errno;
368 * perf_evlist__prepare_workload will send a SIGUSR1
369 * if the fork fails, since we asked by setting its
370 * want_signal to true.
372 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
373 void *ucontext __maybe_unused)
375 workload_exec_errno = info->si_value.sival_int;
378 static bool has_unit(struct perf_evsel *counter)
380 return counter->unit && *counter->unit;
383 static bool has_scale(struct perf_evsel *counter)
385 return counter->scale != 1;
388 static int perf_stat_synthesize_config(bool is_pipe)
390 struct perf_evsel *counter;
394 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
395 process_synthesized_event);
397 pr_err("Couldn't synthesize attrs.\n");
403 * Synthesize other events stuff not carried within
404 * attr event - unit, scale, name
406 evlist__for_each(evsel_list, counter) {
407 if (!counter->supported)
411 * Synthesize unit and scale only if it's defined.
413 if (has_unit(counter)) {
414 err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
416 pr_err("Couldn't synthesize evsel unit.\n");
421 if (has_scale(counter)) {
422 err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
424 pr_err("Couldn't synthesize evsel scale.\n");
429 if (counter->own_cpus) {
430 err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
432 pr_err("Couldn't synthesize evsel scale.\n");
438 * Name is needed only for pipe output,
439 * perf.data carries event names.
442 err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
444 pr_err("Couldn't synthesize evsel name.\n");
450 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
451 process_synthesized_event,
454 pr_err("Couldn't synthesize thread map.\n");
458 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
459 process_synthesized_event, NULL);
461 pr_err("Couldn't synthesize thread map.\n");
465 err = perf_event__synthesize_stat_config(NULL, &stat_config,
466 process_synthesized_event, NULL);
468 pr_err("Couldn't synthesize config.\n");
475 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
477 static int __store_counter_ids(struct perf_evsel *counter,
478 struct cpu_map *cpus,
479 struct thread_map *threads)
483 for (cpu = 0; cpu < cpus->nr; cpu++) {
484 for (thread = 0; thread < threads->nr; thread++) {
485 int fd = FD(counter, cpu, thread);
487 if (perf_evlist__id_add_fd(evsel_list, counter,
488 cpu, thread, fd) < 0)
496 static int store_counter_ids(struct perf_evsel *counter)
498 struct cpu_map *cpus = counter->cpus;
499 struct thread_map *threads = counter->threads;
501 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
504 return __store_counter_ids(counter, cpus, threads);
507 static int __run_perf_stat(int argc, const char **argv)
509 int interval = stat_config.interval;
511 unsigned long long t0, t1;
512 struct perf_evsel *counter;
516 const bool forks = (argc > 0);
517 bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
520 ts.tv_sec = interval / 1000;
521 ts.tv_nsec = (interval % 1000) * 1000000;
528 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
529 workload_exec_failed_signal) < 0) {
530 perror("failed to prepare workload");
533 child_pid = evsel_list->workload.pid;
537 perf_evlist__set_leader(evsel_list);
539 evlist__for_each(evsel_list, counter) {
541 if (create_perf_stat_counter(counter) < 0) {
543 * PPC returns ENXIO for HW counters until 2.6.37
544 * (behavior changed with commit b0a873e).
546 if (errno == EINVAL || errno == ENOSYS ||
547 errno == ENOENT || errno == EOPNOTSUPP ||
550 ui__warning("%s event is not supported by the kernel.\n",
551 perf_evsel__name(counter));
552 counter->supported = false;
554 if ((counter->leader != counter) ||
555 !(counter->leader->nr_members > 1))
557 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
559 ui__warning("%s\n", msg);
563 perf_evsel__open_strerror(counter, &target,
564 errno, msg, sizeof(msg));
565 ui__error("%s\n", msg);
568 kill(child_pid, SIGTERM);
572 counter->supported = true;
574 l = strlen(counter->unit);
578 if (STAT_RECORD && store_counter_ids(counter))
582 if (perf_evlist__apply_filters(evsel_list, &counter)) {
583 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
584 counter->filter, perf_evsel__name(counter), errno,
585 strerror_r(errno, msg, sizeof(msg)));
590 int err, fd = perf_data_file__fd(&perf_stat.file);
593 err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
595 err = perf_session__write_header(perf_stat.session, evsel_list,
602 err = perf_stat_synthesize_config(is_pipe);
608 * Enable counters and exec the command:
611 clock_gettime(CLOCK_MONOTONIC, &ref_time);
614 perf_evlist__start_workload(evsel_list);
618 while (!waitpid(child_pid, &status, WNOHANG)) {
619 nanosleep(&ts, NULL);
625 if (workload_exec_errno) {
626 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
627 pr_err("Workload failed: %s\n", emsg);
631 if (WIFSIGNALED(status))
632 psignal(WTERMSIG(status), argv[0]);
636 nanosleep(&ts, NULL);
644 update_stats(&walltime_nsecs_stats, t1 - t0);
648 return WEXITSTATUS(status);
651 static int run_perf_stat(int argc, const char **argv)
656 ret = system(pre_cmd);
664 ret = __run_perf_stat(argc, argv);
669 ret = system(post_cmd);
677 static void print_running(u64 run, u64 ena)
680 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
684 ena ? 100.0 * run / ena : 100.0);
685 } else if (run != ena) {
686 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
690 static void print_noise_pct(double total, double avg)
692 double pct = rel_stddev_stats(total, avg);
695 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
697 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
700 static void print_noise(struct perf_evsel *evsel, double avg)
702 struct perf_stat_evsel *ps;
708 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
711 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
713 switch (stat_config.aggr_mode) {
715 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
716 cpu_map__id_to_socket(id),
718 cpu_map__id_to_cpu(id),
725 fprintf(stat_config.output, "S%*d%s%*d%s",
734 fprintf(stat_config.output, "CPU%*d%s",
736 perf_evsel__cpus(evsel)->map[id], csv_sep);
739 fprintf(stat_config.output, "%*s-%*d%s",
741 thread_map__comm(evsel->threads, id),
743 thread_map__pid(evsel->threads, id),
759 struct perf_evsel *evsel;
762 #define METRIC_LEN 35
764 static void new_line_std(void *ctx)
766 struct outstate *os = ctx;
771 static void do_new_line_std(struct outstate *os)
774 fputs(os->prefix, os->fh);
775 aggr_printout(os->evsel, os->id, os->nr);
776 if (stat_config.aggr_mode == AGGR_NONE)
777 fprintf(os->fh, " ");
778 fprintf(os->fh, " ");
781 static void print_metric_std(void *ctx, const char *color, const char *fmt,
782 const char *unit, double val)
784 struct outstate *os = ctx;
787 bool newline = os->newline;
791 if (unit == NULL || fmt == NULL) {
792 fprintf(out, "%-*s", METRIC_LEN, "");
799 n = fprintf(out, " # ");
801 n += color_fprintf(out, color, fmt, val);
803 n += fprintf(out, fmt, val);
804 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
807 static void new_line_csv(void *ctx)
809 struct outstate *os = ctx;
814 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
815 aggr_printout(os->evsel, os->id, os->nr);
816 for (i = 0; i < os->nfields; i++)
817 fputs(csv_sep, os->fh);
820 static void print_metric_csv(void *ctx,
821 const char *color __maybe_unused,
822 const char *fmt, const char *unit, double val)
824 struct outstate *os = ctx;
826 char buf[64], *vals, *ends;
828 if (unit == NULL || fmt == NULL) {
829 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
832 snprintf(buf, sizeof(buf), fmt, val);
834 while (isspace(*vals))
837 while (isdigit(*ends) || *ends == '.')
840 while (isspace(*unit))
842 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
845 #define METRIC_ONLY_LEN 20
847 /* Filter out some columns that don't work well in metrics only mode */
849 static bool valid_only_metric(const char *unit)
853 if (strstr(unit, "/sec") ||
854 strstr(unit, "hz") ||
855 strstr(unit, "Hz") ||
856 strstr(unit, "CPUs utilized"))
861 static const char *fixunit(char *buf, struct perf_evsel *evsel,
864 if (!strncmp(unit, "of all", 6)) {
865 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
872 static void print_metric_only(void *ctx, const char *color, const char *fmt,
873 const char *unit, double val)
875 struct outstate *os = ctx;
879 unsigned mlen = METRIC_ONLY_LEN;
881 if (!valid_only_metric(unit))
883 unit = fixunit(buf, os->evsel, unit);
885 n = color_fprintf(out, color, fmt, val);
887 n = fprintf(out, fmt, val);
888 if (n > METRIC_ONLY_LEN)
890 if (mlen < strlen(unit))
891 mlen = strlen(unit) + 1;
892 fprintf(out, "%*s", mlen - n, "");
895 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
897 const char *unit, double val)
899 struct outstate *os = ctx;
901 char buf[64], *vals, *ends;
904 if (!valid_only_metric(unit))
906 unit = fixunit(tbuf, os->evsel, unit);
907 snprintf(buf, sizeof buf, fmt, val);
909 while (isspace(*vals))
912 while (isdigit(*ends) || *ends == '.')
915 fprintf(out, "%s%s", vals, csv_sep);
918 static void new_line_metric(void *ctx __maybe_unused)
922 static void print_metric_header(void *ctx, const char *color __maybe_unused,
923 const char *fmt __maybe_unused,
924 const char *unit, double val __maybe_unused)
926 struct outstate *os = ctx;
929 if (!valid_only_metric(unit))
931 unit = fixunit(tbuf, os->evsel, unit);
933 fprintf(os->fh, "%s%s", unit, csv_sep);
935 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
938 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
940 FILE *output = stat_config.output;
941 double msecs = avg / 1e6;
942 const char *fmt_v, *fmt_n;
945 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
946 fmt_n = csv_output ? "%s" : "%-25s";
948 aggr_printout(evsel, id, nr);
950 scnprintf(name, sizeof(name), "%s%s",
951 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
953 fprintf(output, fmt_v, msecs, csv_sep);
956 fprintf(output, "%s%s", evsel->unit, csv_sep);
958 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
960 fprintf(output, fmt_n, name);
963 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
966 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
973 if (stat_config.aggr_mode == AGGR_NONE)
976 if (stat_config.aggr_mode == AGGR_GLOBAL)
979 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
980 int cpu2 = perf_evsel__cpus(evsel)->map[i];
982 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
988 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
990 FILE *output = stat_config.output;
991 double sc = evsel->scale;
995 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
998 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1000 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1003 aggr_printout(evsel, id, nr);
1005 fprintf(output, fmt, avg, csv_sep);
1008 fprintf(output, "%-*s%s",
1009 csv_output ? 0 : unit_width,
1010 evsel->unit, csv_sep);
1012 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1015 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1018 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1019 char *prefix, u64 run, u64 ena, double noise)
1021 struct perf_stat_output_ctx out;
1022 struct outstate os = {
1023 .fh = stat_config.output,
1024 .prefix = prefix ? prefix : "",
1029 print_metric_t pm = print_metric_std;
1033 nl = new_line_metric;
1035 pm = print_metric_only_csv;
1037 pm = print_metric_only;
1041 if (csv_output && !metric_only) {
1042 static int aggr_fields[] = {
1050 pm = print_metric_csv;
1053 os.nfields += aggr_fields[stat_config.aggr_mode];
1057 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1059 pm(&os, NULL, "", "", 0);
1062 aggr_printout(counter, id, nr);
1064 fprintf(stat_config.output, "%*s%s",
1065 csv_output ? 0 : 18,
1066 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1069 fprintf(stat_config.output, "%-*s%s",
1070 csv_output ? 0 : unit_width,
1071 counter->unit, csv_sep);
1073 fprintf(stat_config.output, "%*s",
1074 csv_output ? 0 : -25,
1075 perf_evsel__name(counter));
1078 fprintf(stat_config.output, "%s%s",
1079 csv_sep, counter->cgrp->name);
1082 pm(&os, NULL, NULL, "", 0);
1083 print_noise(counter, noise);
1084 print_running(run, ena);
1086 pm(&os, NULL, NULL, "", 0);
1092 else if (nsec_counter(counter))
1093 nsec_printout(id, nr, counter, uval);
1095 abs_printout(id, nr, counter, uval);
1097 out.print_metric = pm;
1101 if (csv_output && !metric_only) {
1102 print_noise(counter, noise);
1103 print_running(run, ena);
1106 perf_stat__print_shadow_stats(counter, uval,
1107 first_shadow_cpu(counter, id),
1109 if (!csv_output && !metric_only) {
1110 print_noise(counter, noise);
1111 print_running(run, ena);
1115 static void aggr_update_shadow(void)
1119 struct perf_evsel *counter;
1121 for (s = 0; s < aggr_map->nr; s++) {
1122 id = aggr_map->map[s];
1123 evlist__for_each(evsel_list, counter) {
1125 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1126 s2 = aggr_get_id(evsel_list->cpus, cpu);
1129 val += perf_counts(counter->counts, cpu, 0)->val;
1131 val = val * counter->scale;
1132 perf_stat__update_shadow_stats(counter, &val,
1133 first_shadow_cpu(counter, id));
1138 static void print_aggr(char *prefix)
1140 FILE *output = stat_config.output;
1141 struct perf_evsel *counter;
1142 int cpu, s, s2, id, nr;
1147 if (!(aggr_map || aggr_get_id))
1150 aggr_update_shadow();
1153 * With metric_only everything is on a single line.
1154 * Without each counter has its own line.
1156 for (s = 0; s < aggr_map->nr; s++) {
1157 if (prefix && metric_only)
1158 fprintf(output, "%s", prefix);
1160 id = aggr_map->map[s];
1162 evlist__for_each(evsel_list, counter) {
1163 val = ena = run = 0;
1165 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1166 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1169 val += perf_counts(counter->counts, cpu, 0)->val;
1170 ena += perf_counts(counter->counts, cpu, 0)->ena;
1171 run += perf_counts(counter->counts, cpu, 0)->run;
1174 if (first && metric_only) {
1176 aggr_printout(counter, id, nr);
1178 if (prefix && !metric_only)
1179 fprintf(output, "%s", prefix);
1181 uval = val * counter->scale;
1182 printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1184 fputc('\n', output);
1187 fputc('\n', output);
1191 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1193 FILE *output = stat_config.output;
1194 int nthreads = thread_map__nr(counter->threads);
1195 int ncpus = cpu_map__nr(counter->cpus);
1199 for (thread = 0; thread < nthreads; thread++) {
1200 u64 ena = 0, run = 0, val = 0;
1202 for (cpu = 0; cpu < ncpus; cpu++) {
1203 val += perf_counts(counter->counts, cpu, thread)->val;
1204 ena += perf_counts(counter->counts, cpu, thread)->ena;
1205 run += perf_counts(counter->counts, cpu, thread)->run;
1209 fprintf(output, "%s", prefix);
1211 uval = val * counter->scale;
1212 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1213 fputc('\n', output);
1218 * Print out the results of a single counter:
1219 * aggregated counts in system-wide mode
1221 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1223 FILE *output = stat_config.output;
1224 struct perf_stat_evsel *ps = counter->priv;
1225 double avg = avg_stats(&ps->res_stats[0]);
1227 double avg_enabled, avg_running;
1229 avg_enabled = avg_stats(&ps->res_stats[1]);
1230 avg_running = avg_stats(&ps->res_stats[2]);
1232 if (prefix && !metric_only)
1233 fprintf(output, "%s", prefix);
1235 uval = avg * counter->scale;
1236 printout(-1, 0, counter, uval, prefix, avg_running, avg_enabled, avg);
1238 fprintf(output, "\n");
1242 * Print out the results of a single counter:
1243 * does not use aggregated count in system-wide
1245 static void print_counter(struct perf_evsel *counter, char *prefix)
1247 FILE *output = stat_config.output;
1252 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1253 val = perf_counts(counter->counts, cpu, 0)->val;
1254 ena = perf_counts(counter->counts, cpu, 0)->ena;
1255 run = perf_counts(counter->counts, cpu, 0)->run;
1258 fprintf(output, "%s", prefix);
1260 uval = val * counter->scale;
1261 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1263 fputc('\n', output);
1267 static void print_no_aggr_metric(char *prefix)
1271 struct perf_evsel *counter;
1275 nrcpus = evsel_list->cpus->nr;
1276 for (cpu = 0; cpu < nrcpus; cpu++) {
1280 fputs(prefix, stat_config.output);
1281 evlist__for_each(evsel_list, counter) {
1283 aggr_printout(counter, cpu, 0);
1286 val = perf_counts(counter->counts, cpu, 0)->val;
1287 ena = perf_counts(counter->counts, cpu, 0)->ena;
1288 run = perf_counts(counter->counts, cpu, 0)->run;
1290 uval = val * counter->scale;
1291 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1293 fputc('\n', stat_config.output);
1297 static int aggr_header_lens[] = {
1305 static void print_metric_headers(char *prefix)
1307 struct perf_stat_output_ctx out;
1308 struct perf_evsel *counter;
1309 struct outstate os = {
1310 .fh = stat_config.output
1314 fprintf(stat_config.output, "%s", prefix);
1317 fprintf(stat_config.output, "%*s",
1318 aggr_header_lens[stat_config.aggr_mode], "");
1320 /* Print metrics headers only */
1321 evlist__for_each(evsel_list, counter) {
1324 out.print_metric = print_metric_header;
1325 out.new_line = new_line_metric;
1327 perf_stat__print_shadow_stats(counter, 0,
1331 fputc('\n', stat_config.output);
1334 static void print_interval(char *prefix, struct timespec *ts)
1336 FILE *output = stat_config.output;
1337 static int num_print_interval;
1339 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1341 if (num_print_interval == 0 && !csv_output && !metric_only) {
1342 switch (stat_config.aggr_mode) {
1344 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
1347 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
1350 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
1353 fprintf(output, "# time comm-pid counts %*s events\n", unit_width, "unit");
1357 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
1363 if (++num_print_interval == 25)
1364 num_print_interval = 0;
1367 static void print_header(int argc, const char **argv)
1369 FILE *output = stat_config.output;
1375 fprintf(output, "\n");
1376 fprintf(output, " Performance counter stats for ");
1377 if (target.system_wide)
1378 fprintf(output, "\'system wide");
1379 else if (target.cpu_list)
1380 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1381 else if (!target__has_task(&target)) {
1382 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1383 for (i = 1; argv && (i < argc); i++)
1384 fprintf(output, " %s", argv[i]);
1385 } else if (target.pid)
1386 fprintf(output, "process id \'%s", target.pid);
1388 fprintf(output, "thread id \'%s", target.tid);
1390 fprintf(output, "\'");
1392 fprintf(output, " (%d runs)", run_count);
1393 fprintf(output, ":\n\n");
1397 static void print_footer(void)
1399 FILE *output = stat_config.output;
1402 fprintf(output, "\n");
1403 fprintf(output, " %17.9f seconds time elapsed",
1404 avg_stats(&walltime_nsecs_stats)/1e9);
1405 if (run_count > 1) {
1406 fprintf(output, " ");
1407 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1408 avg_stats(&walltime_nsecs_stats));
1410 fprintf(output, "\n\n");
1413 static void print_counters(struct timespec *ts, int argc, const char **argv)
1415 int interval = stat_config.interval;
1416 struct perf_evsel *counter;
1417 char buf[64], *prefix = NULL;
1419 /* Do not print anything if we record to the pipe. */
1420 if (STAT_RECORD && perf_stat.file.is_pipe)
1424 print_interval(prefix = buf, ts);
1426 print_header(argc, argv);
1429 static int num_print_iv;
1431 if (num_print_iv == 0)
1432 print_metric_headers(prefix);
1433 if (num_print_iv++ == 25)
1435 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1436 fprintf(stat_config.output, "%s", prefix);
1439 switch (stat_config.aggr_mode) {
1445 evlist__for_each(evsel_list, counter)
1446 print_aggr_thread(counter, prefix);
1449 evlist__for_each(evsel_list, counter)
1450 print_counter_aggr(counter, prefix);
1452 fputc('\n', stat_config.output);
1456 print_no_aggr_metric(prefix);
1458 evlist__for_each(evsel_list, counter)
1459 print_counter(counter, prefix);
1467 if (!interval && !csv_output)
1470 fflush(stat_config.output);
1473 static volatile int signr = -1;
1475 static void skip_signal(int signo)
1477 if ((child_pid == -1) || stat_config.interval)
1482 * render child_pid harmless
1483 * won't send SIGTERM to a random
1484 * process in case of race condition
1485 * and fast PID recycling
1490 static void sig_atexit(void)
1495 * avoid race condition with SIGCHLD handler
1496 * in skip_signal() which is modifying child_pid
1497 * goal is to avoid send SIGTERM to a random
1501 sigaddset(&set, SIGCHLD);
1502 sigprocmask(SIG_BLOCK, &set, &oset);
1504 if (child_pid != -1)
1505 kill(child_pid, SIGTERM);
1507 sigprocmask(SIG_SETMASK, &oset, NULL);
1512 signal(signr, SIG_DFL);
1513 kill(getpid(), signr);
1516 static int stat__set_big_num(const struct option *opt __maybe_unused,
1517 const char *s __maybe_unused, int unset)
1519 big_num_opt = unset ? 0 : 1;
1523 static const struct option stat_options[] = {
1524 OPT_BOOLEAN('T', "transaction", &transaction_run,
1525 "hardware transaction statistics"),
1526 OPT_CALLBACK('e', "event", &evsel_list, "event",
1527 "event selector. use 'perf list' to list available events",
1528 parse_events_option),
1529 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1530 "event filter", parse_filter),
1531 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1532 "child tasks do not inherit counters"),
1533 OPT_STRING('p', "pid", &target.pid, "pid",
1534 "stat events on existing process id"),
1535 OPT_STRING('t', "tid", &target.tid, "tid",
1536 "stat events on existing thread id"),
1537 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1538 "system-wide collection from all CPUs"),
1539 OPT_BOOLEAN('g', "group", &group,
1540 "put the counters into a counter group"),
1541 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1542 OPT_INCR('v', "verbose", &verbose,
1543 "be more verbose (show counter open errors, etc)"),
1544 OPT_INTEGER('r', "repeat", &run_count,
1545 "repeat command and print average + stddev (max: 100, forever: 0)"),
1546 OPT_BOOLEAN('n', "null", &null_run,
1547 "null run - dont start any counters"),
1548 OPT_INCR('d', "detailed", &detailed_run,
1549 "detailed run - start a lot of events"),
1550 OPT_BOOLEAN('S', "sync", &sync_run,
1551 "call sync() before starting a run"),
1552 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1553 "print large numbers with thousands\' separators",
1555 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1556 "list of cpus to monitor in system-wide"),
1557 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1558 "disable CPU count aggregation", AGGR_NONE),
1559 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1560 "print counts with custom separator"),
1561 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1562 "monitor event in cgroup name only", parse_cgroups),
1563 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1564 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1565 OPT_INTEGER(0, "log-fd", &output_fd,
1566 "log output to fd, instead of stderr"),
1567 OPT_STRING(0, "pre", &pre_cmd, "command",
1568 "command to run prior to the measured command"),
1569 OPT_STRING(0, "post", &post_cmd, "command",
1570 "command to run after to the measured command"),
1571 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1572 "print counts at regular interval in ms (>= 10)"),
1573 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1574 "aggregate counts per processor socket", AGGR_SOCKET),
1575 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1576 "aggregate counts per physical processor core", AGGR_CORE),
1577 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1578 "aggregate counts per thread", AGGR_THREAD),
1579 OPT_UINTEGER('D', "delay", &initial_delay,
1580 "ms to wait before starting measurement after program start"),
1581 OPT_BOOLEAN(0, "metric-only", &metric_only,
1582 "Only print computed metrics. No raw values"),
1586 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1588 return cpu_map__get_socket(map, cpu, NULL);
1591 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1593 return cpu_map__get_core(map, cpu, NULL);
1596 static int cpu_map__get_max(struct cpu_map *map)
1600 for (i = 0; i < map->nr; i++) {
1601 if (map->map[i] > max)
1608 static struct cpu_map *cpus_aggr_map;
1610 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1617 cpu = map->map[idx];
1619 if (cpus_aggr_map->map[cpu] == -1)
1620 cpus_aggr_map->map[cpu] = get_id(map, idx);
1622 return cpus_aggr_map->map[cpu];
1625 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1627 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1630 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1632 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1635 static int perf_stat_init_aggr_mode(void)
1639 switch (stat_config.aggr_mode) {
1641 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1642 perror("cannot build socket map");
1645 aggr_get_id = perf_stat__get_socket_cached;
1648 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1649 perror("cannot build core map");
1652 aggr_get_id = perf_stat__get_core_cached;
1663 * The evsel_list->cpus is the base we operate on,
1664 * taking the highest cpu number to be the size of
1665 * the aggregation translate cpumap.
1667 nr = cpu_map__get_max(evsel_list->cpus);
1668 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1669 return cpus_aggr_map ? 0 : -ENOMEM;
1672 static void perf_stat__exit_aggr_mode(void)
1674 cpu_map__put(aggr_map);
1675 cpu_map__put(cpus_aggr_map);
1677 cpus_aggr_map = NULL;
1680 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1687 cpu = map->map[idx];
1689 if (cpu >= env->nr_cpus_online)
1695 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1697 struct perf_env *env = data;
1698 int cpu = perf_env__get_cpu(env, map, idx);
1700 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1703 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1705 struct perf_env *env = data;
1706 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1709 int socket_id = env->cpu[cpu].socket_id;
1712 * Encode socket in upper 16 bits
1713 * core_id is relative to socket, and
1714 * we need a global id. So we combine
1717 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1723 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1724 struct cpu_map **sockp)
1726 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1729 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1730 struct cpu_map **corep)
1732 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1735 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1737 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1740 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1742 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1745 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1747 struct perf_env *env = &st->session->header.env;
1749 switch (stat_config.aggr_mode) {
1751 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1752 perror("cannot build socket map");
1755 aggr_get_id = perf_stat__get_socket_file;
1758 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1759 perror("cannot build core map");
1762 aggr_get_id = perf_stat__get_core_file;
1776 * Add default attributes, if there were no attributes specified or
1777 * if -d/--detailed, -d -d or -d -d -d is used:
1779 static int add_default_attributes(void)
1781 struct perf_event_attr default_attrs0[] = {
1783 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1784 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1785 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1786 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1788 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1790 struct perf_event_attr frontend_attrs[] = {
1791 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1793 struct perf_event_attr backend_attrs[] = {
1794 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1796 struct perf_event_attr default_attrs1[] = {
1797 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1798 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1799 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1804 * Detailed stats (-d), covering the L1 and last level data caches:
1806 struct perf_event_attr detailed_attrs[] = {
1808 { .type = PERF_TYPE_HW_CACHE,
1810 PERF_COUNT_HW_CACHE_L1D << 0 |
1811 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1812 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1814 { .type = PERF_TYPE_HW_CACHE,
1816 PERF_COUNT_HW_CACHE_L1D << 0 |
1817 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1818 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1820 { .type = PERF_TYPE_HW_CACHE,
1822 PERF_COUNT_HW_CACHE_LL << 0 |
1823 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1824 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1826 { .type = PERF_TYPE_HW_CACHE,
1828 PERF_COUNT_HW_CACHE_LL << 0 |
1829 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1830 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1834 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1836 struct perf_event_attr very_detailed_attrs[] = {
1838 { .type = PERF_TYPE_HW_CACHE,
1840 PERF_COUNT_HW_CACHE_L1I << 0 |
1841 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1842 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1844 { .type = PERF_TYPE_HW_CACHE,
1846 PERF_COUNT_HW_CACHE_L1I << 0 |
1847 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1848 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1850 { .type = PERF_TYPE_HW_CACHE,
1852 PERF_COUNT_HW_CACHE_DTLB << 0 |
1853 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1854 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1856 { .type = PERF_TYPE_HW_CACHE,
1858 PERF_COUNT_HW_CACHE_DTLB << 0 |
1859 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1860 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1862 { .type = PERF_TYPE_HW_CACHE,
1864 PERF_COUNT_HW_CACHE_ITLB << 0 |
1865 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1866 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1868 { .type = PERF_TYPE_HW_CACHE,
1870 PERF_COUNT_HW_CACHE_ITLB << 0 |
1871 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1872 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1877 * Very, very detailed stats (-d -d -d), adding prefetch events:
1879 struct perf_event_attr very_very_detailed_attrs[] = {
1881 { .type = PERF_TYPE_HW_CACHE,
1883 PERF_COUNT_HW_CACHE_L1D << 0 |
1884 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1885 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1887 { .type = PERF_TYPE_HW_CACHE,
1889 PERF_COUNT_HW_CACHE_L1D << 0 |
1890 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1891 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1894 /* Set attrs if no event is selected and !null_run: */
1898 if (transaction_run) {
1900 if (pmu_have_event("cpu", "cycles-ct") &&
1901 pmu_have_event("cpu", "el-start"))
1902 err = parse_events(evsel_list, transaction_attrs, NULL);
1904 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1906 fprintf(stderr, "Cannot set up transaction events\n");
1912 if (!evsel_list->nr_entries) {
1913 if (target__has_cpu(&target))
1914 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1916 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1918 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1919 if (perf_evlist__add_default_attrs(evsel_list,
1920 frontend_attrs) < 0)
1923 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1924 if (perf_evlist__add_default_attrs(evsel_list,
1928 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1932 /* Detailed events get appended to the event list: */
1934 if (detailed_run < 1)
1937 /* Append detailed run extra attributes: */
1938 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1941 if (detailed_run < 2)
1944 /* Append very detailed run extra attributes: */
1945 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1948 if (detailed_run < 3)
1951 /* Append very, very detailed run extra attributes: */
1952 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1955 static const char * const stat_record_usage[] = {
1956 "perf stat record [<options>]",
1960 static void init_features(struct perf_session *session)
1964 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1965 perf_header__set_feat(&session->header, feat);
1967 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1968 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1969 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1970 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1973 static int __cmd_record(int argc, const char **argv)
1975 struct perf_session *session;
1976 struct perf_data_file *file = &perf_stat.file;
1978 argc = parse_options(argc, argv, stat_options, stat_record_usage,
1979 PARSE_OPT_STOP_AT_NON_OPTION);
1982 file->path = output_name;
1984 if (run_count != 1 || forever) {
1985 pr_err("Cannot use -r option with perf stat record.\n");
1989 session = perf_session__new(file, false, NULL);
1990 if (session == NULL) {
1991 pr_err("Perf session creation failed.\n");
1995 init_features(session);
1997 session->evlist = evsel_list;
1998 perf_stat.session = session;
1999 perf_stat.record = true;
2003 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2004 union perf_event *event,
2005 struct perf_session *session)
2007 struct stat_round_event *stat_round = &event->stat_round;
2008 struct perf_evsel *counter;
2009 struct timespec tsh, *ts = NULL;
2010 const char **argv = session->header.env.cmdline_argv;
2011 int argc = session->header.env.nr_cmdline;
2013 evlist__for_each(evsel_list, counter)
2014 perf_stat_process_counter(&stat_config, counter);
2016 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2017 update_stats(&walltime_nsecs_stats, stat_round->time);
2019 if (stat_config.interval && stat_round->time) {
2020 tsh.tv_sec = stat_round->time / NSECS_PER_SEC;
2021 tsh.tv_nsec = stat_round->time % NSECS_PER_SEC;
2025 print_counters(ts, argc, argv);
2030 int process_stat_config_event(struct perf_tool *tool __maybe_unused,
2031 union perf_event *event,
2032 struct perf_session *session __maybe_unused)
2034 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2036 perf_event__read_stat_config(&stat_config, &event->stat_config);
2038 if (cpu_map__empty(st->cpus)) {
2039 if (st->aggr_mode != AGGR_UNSET)
2040 pr_warning("warning: processing task data, aggregation mode not set\n");
2044 if (st->aggr_mode != AGGR_UNSET)
2045 stat_config.aggr_mode = st->aggr_mode;
2047 if (perf_stat.file.is_pipe)
2048 perf_stat_init_aggr_mode();
2050 perf_stat_init_aggr_mode_file(st);
2055 static int set_maps(struct perf_stat *st)
2057 if (!st->cpus || !st->threads)
2060 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2063 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2065 if (perf_evlist__alloc_stats(evsel_list, true))
2068 st->maps_allocated = true;
2073 int process_thread_map_event(struct perf_tool *tool __maybe_unused,
2074 union perf_event *event,
2075 struct perf_session *session __maybe_unused)
2077 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2080 pr_warning("Extra thread map event, ignoring.\n");
2084 st->threads = thread_map__new_event(&event->thread_map);
2088 return set_maps(st);
2092 int process_cpu_map_event(struct perf_tool *tool __maybe_unused,
2093 union perf_event *event,
2094 struct perf_session *session __maybe_unused)
2096 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2097 struct cpu_map *cpus;
2100 pr_warning("Extra cpu map event, ignoring.\n");
2104 cpus = cpu_map__new_data(&event->cpu_map.data);
2109 return set_maps(st);
2112 static const char * const stat_report_usage[] = {
2113 "perf stat report [<options>]",
2117 static struct perf_stat perf_stat = {
2119 .attr = perf_event__process_attr,
2120 .event_update = perf_event__process_event_update,
2121 .thread_map = process_thread_map_event,
2122 .cpu_map = process_cpu_map_event,
2123 .stat_config = process_stat_config_event,
2124 .stat = perf_event__process_stat_event,
2125 .stat_round = process_stat_round_event,
2127 .aggr_mode = AGGR_UNSET,
2130 static int __cmd_report(int argc, const char **argv)
2132 struct perf_session *session;
2133 const struct option options[] = {
2134 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2135 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2136 "aggregate counts per processor socket", AGGR_SOCKET),
2137 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2138 "aggregate counts per physical processor core", AGGR_CORE),
2139 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2140 "disable CPU count aggregation", AGGR_NONE),
2146 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2148 if (!input_name || !strlen(input_name)) {
2149 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2152 input_name = "perf.data";
2155 perf_stat.file.path = input_name;
2156 perf_stat.file.mode = PERF_DATA_MODE_READ;
2158 session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
2159 if (session == NULL)
2162 perf_stat.session = session;
2163 stat_config.output = stderr;
2164 evsel_list = session->evlist;
2166 ret = perf_session__process_events(session);
2170 perf_session__delete(session);
2174 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
2176 const char * const stat_usage[] = {
2177 "perf stat [<options>] [<command>]",
2180 int status = -EINVAL, run_idx;
2182 FILE *output = stderr;
2183 unsigned int interval;
2184 const char * const stat_subcommands[] = { "record", "report" };
2186 setlocale(LC_ALL, "");
2188 evsel_list = perf_evlist__new();
2189 if (evsel_list == NULL)
2192 parse_events__shrink_config_terms();
2193 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2194 (const char **) stat_usage,
2195 PARSE_OPT_STOP_AT_NON_OPTION);
2196 perf_stat__init_shadow_stats();
2200 if (!strcmp(csv_sep, "\\t"))
2203 csv_sep = DEFAULT_SEPARATOR;
2205 if (argc && !strncmp(argv[0], "rec", 3)) {
2206 argc = __cmd_record(argc, argv);
2209 } else if (argc && !strncmp(argv[0], "rep", 3))
2210 return __cmd_report(argc, argv);
2212 interval = stat_config.interval;
2215 * For record command the -o is already taken care of.
2217 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2220 if (output_name && output_fd) {
2221 fprintf(stderr, "cannot use both --output and --log-fd\n");
2222 parse_options_usage(stat_usage, stat_options, "o", 1);
2223 parse_options_usage(NULL, stat_options, "log-fd", 0);
2227 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2228 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2232 if (metric_only && run_count > 1) {
2233 fprintf(stderr, "--metric-only is not supported with -r\n");
2237 if (output_fd < 0) {
2238 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2239 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2245 mode = append_file ? "a" : "w";
2247 output = fopen(output_name, mode);
2249 perror("failed to create output file");
2252 clock_gettime(CLOCK_REALTIME, &tm);
2253 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2254 } else if (output_fd > 0) {
2255 mode = append_file ? "a" : "w";
2256 output = fdopen(output_fd, mode);
2258 perror("Failed opening logfd");
2263 stat_config.output = output;
2266 * let the spreadsheet do the pretty-printing
2269 /* User explicitly passed -B? */
2270 if (big_num_opt == 1) {
2271 fprintf(stderr, "-B option not supported with -x\n");
2272 parse_options_usage(stat_usage, stat_options, "B", 1);
2273 parse_options_usage(NULL, stat_options, "x", 1);
2275 } else /* Nope, so disable big number formatting */
2277 } else if (big_num_opt == 0) /* User passed --no-big-num */
2280 if (!argc && target__none(&target))
2281 usage_with_options(stat_usage, stat_options);
2283 if (run_count < 0) {
2284 pr_err("Run count must be a positive number\n");
2285 parse_options_usage(stat_usage, stat_options, "r", 1);
2287 } else if (run_count == 0) {
2292 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2293 fprintf(stderr, "The --per-thread option is only available "
2294 "when monitoring via -p -t options.\n");
2295 parse_options_usage(NULL, stat_options, "p", 1);
2296 parse_options_usage(NULL, stat_options, "t", 1);
2301 * no_aggr, cgroup are for system-wide only
2302 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2304 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2305 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2306 !target__has_cpu(&target)) {
2307 fprintf(stderr, "both cgroup and no-aggregation "
2308 "modes only available in system-wide mode\n");
2310 parse_options_usage(stat_usage, stat_options, "G", 1);
2311 parse_options_usage(NULL, stat_options, "A", 1);
2312 parse_options_usage(NULL, stat_options, "a", 1);
2316 if (add_default_attributes())
2319 target__validate(&target);
2321 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2322 if (target__has_task(&target)) {
2323 pr_err("Problems finding threads of monitor\n");
2324 parse_options_usage(stat_usage, stat_options, "p", 1);
2325 parse_options_usage(NULL, stat_options, "t", 1);
2326 } else if (target__has_cpu(&target)) {
2327 perror("failed to parse CPUs map");
2328 parse_options_usage(stat_usage, stat_options, "C", 1);
2329 parse_options_usage(NULL, stat_options, "a", 1);
2335 * Initialize thread_map with comm names,
2336 * so we could print it out on output.
2338 if (stat_config.aggr_mode == AGGR_THREAD)
2339 thread_map__read_comms(evsel_list->threads);
2341 if (interval && interval < 100) {
2342 if (interval < 10) {
2343 pr_err("print interval must be >= 10ms\n");
2344 parse_options_usage(stat_usage, stat_options, "I", 1);
2347 pr_warning("print interval < 100ms. "
2348 "The overhead percentage could be high in some cases. "
2349 "Please proceed with caution.\n");
2352 if (perf_evlist__alloc_stats(evsel_list, interval))
2355 if (perf_stat_init_aggr_mode())
2359 * We dont want to block the signals - that would cause
2360 * child tasks to inherit that and Ctrl-C would not work.
2361 * What we want is for Ctrl-C to work in the exec()-ed
2362 * task, but being ignored by perf stat itself:
2366 signal(SIGINT, skip_signal);
2367 signal(SIGCHLD, skip_signal);
2368 signal(SIGALRM, skip_signal);
2369 signal(SIGABRT, skip_signal);
2372 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2373 if (run_count != 1 && verbose)
2374 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2377 status = run_perf_stat(argc, argv);
2378 if (forever && status != -1) {
2379 print_counters(NULL, argc, argv);
2380 perf_stat__reset_stats();
2384 if (!forever && status != -1 && !interval)
2385 print_counters(NULL, argc, argv);
2389 * We synthesize the kernel mmap record just so that older tools
2390 * don't emit warnings about not being able to resolve symbols
2391 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2392 * a saner message about no samples being in the perf.data file.
2394 * This also serves to suppress a warning about f_header.data.size == 0
2395 * in header.c at the moment 'perf stat record' gets introduced, which
2396 * is not really needed once we start adding the stat specific PERF_RECORD_
2397 * records, but the need to suppress the kptr_restrict messages in older
2398 * tools remain -acme
2400 int fd = perf_data_file__fd(&perf_stat.file);
2401 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2402 process_synthesized_event,
2403 &perf_stat.session->machines.host);
2405 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2406 "older tools may produce warnings about this file\n.");
2410 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2411 pr_err("failed to write stat round event\n");
2414 if (!perf_stat.file.is_pipe) {
2415 perf_stat.session->header.data_size += perf_stat.bytes_written;
2416 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2419 perf_session__delete(perf_stat.session);
2422 perf_stat__exit_aggr_mode();
2423 perf_evlist__free_stats(evsel_list);
2425 perf_evlist__delete(evsel_list);