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>
70 #define DEFAULT_SEPARATOR " "
71 #define CNTR_NOT_SUPPORTED "<not supported>"
72 #define CNTR_NOT_COUNTED "<not counted>"
74 static void print_counters(struct timespec *ts, int argc, const char **argv);
76 /* Default events used for perf stat -T */
77 static const char *transaction_attrs = {
89 /* More limited version when the CPU does not have all events. */
90 static const char * transaction_limited_attrs = {
100 static struct perf_evlist *evsel_list;
102 static struct target target = {
106 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
108 static int run_count = 1;
109 static bool no_inherit = false;
110 static volatile pid_t child_pid = -1;
111 static bool null_run = false;
112 static int detailed_run = 0;
113 static bool transaction_run;
114 static bool big_num = true;
115 static int big_num_opt = -1;
116 static const char *csv_sep = NULL;
117 static bool csv_output = false;
118 static bool group = false;
119 static const char *pre_cmd = NULL;
120 static const char *post_cmd = NULL;
121 static bool sync_run = false;
122 static unsigned int initial_delay = 0;
123 static unsigned int unit_width = 4; /* strlen("unit") */
124 static bool forever = false;
125 static struct timespec ref_time;
126 static struct cpu_map *aggr_map;
127 static aggr_get_id_t aggr_get_id;
128 static bool append_file;
129 static const char *output_name;
130 static int output_fd;
134 struct perf_data_file file;
135 struct perf_session *session;
137 struct perf_tool tool;
139 struct cpu_map *cpus;
140 struct thread_map *threads;
141 enum aggr_mode aggr_mode;
144 static struct perf_stat perf_stat;
145 #define STAT_RECORD perf_stat.record
147 static volatile int done = 0;
149 static struct perf_stat_config stat_config = {
150 .aggr_mode = AGGR_GLOBAL,
154 static inline void diff_timespec(struct timespec *r, struct timespec *a,
157 r->tv_sec = a->tv_sec - b->tv_sec;
158 if (a->tv_nsec < b->tv_nsec) {
159 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
162 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
166 static void perf_stat__reset_stats(void)
168 perf_evlist__reset_stats(evsel_list);
169 perf_stat__reset_shadow_stats();
172 static int create_perf_stat_counter(struct perf_evsel *evsel)
174 struct perf_event_attr *attr = &evsel->attr;
176 if (stat_config.scale)
177 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
178 PERF_FORMAT_TOTAL_TIME_RUNNING;
180 attr->inherit = !no_inherit;
183 * Some events get initialized with sample_(period/type) set,
184 * like tracepoints. Clear it up for counting.
186 attr->sample_period = 0;
189 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
190 * while avoiding that older tools show confusing messages.
192 * However for pipe sessions we need to keep it zero,
193 * because script's perf_evsel__check_attr is triggered
194 * by attr->sample_type != 0, and we can't run it on
197 if (!(STAT_RECORD && perf_stat.file.is_pipe))
198 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
201 * Disabling all counters initially, they will be enabled
202 * either manually by us or by kernel via enable_on_exec
205 if (perf_evsel__is_group_leader(evsel)) {
209 * In case of initial_delay we enable tracee
212 if (target__none(&target) && !initial_delay)
213 attr->enable_on_exec = 1;
216 if (target__has_cpu(&target))
217 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
219 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
223 * Does the counter have nsecs as a unit?
225 static inline int nsec_counter(struct perf_evsel *evsel)
227 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
228 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
234 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
235 union perf_event *event,
236 struct perf_sample *sample __maybe_unused,
237 struct machine *machine __maybe_unused)
239 if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
240 pr_err("failed to write perf data, error: %m\n");
244 perf_stat.bytes_written += event->header.size;
248 static int write_stat_round_event(u64 tm, u64 type)
250 return perf_event__synthesize_stat_round(NULL, tm, type,
251 process_synthesized_event,
255 #define WRITE_STAT_ROUND_EVENT(time, interval) \
256 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
258 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
261 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
262 struct perf_counts_values *count)
264 struct perf_sample_id *sid = SID(counter, cpu, thread);
266 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
267 process_synthesized_event, NULL);
271 * Read out the results of a single counter:
272 * do not aggregate counts across CPUs in system-wide mode
274 static int read_counter(struct perf_evsel *counter)
276 int nthreads = thread_map__nr(evsel_list->threads);
277 int ncpus = perf_evsel__nr_cpus(counter);
280 if (!counter->supported)
283 if (counter->system_wide)
286 for (thread = 0; thread < nthreads; thread++) {
287 for (cpu = 0; cpu < ncpus; cpu++) {
288 struct perf_counts_values *count;
290 count = perf_counts(counter->counts, cpu, thread);
291 if (perf_evsel__read(counter, cpu, thread, count))
295 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
296 pr_err("failed to write stat event\n");
306 static void read_counters(bool close_counters)
308 struct perf_evsel *counter;
310 evlist__for_each(evsel_list, counter) {
311 if (read_counter(counter))
312 pr_debug("failed to read counter %s\n", counter->name);
314 if (perf_stat_process_counter(&stat_config, counter))
315 pr_warning("failed to process counter %s\n", counter->name);
317 if (close_counters) {
318 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
319 thread_map__nr(evsel_list->threads));
324 static void process_interval(void)
326 struct timespec ts, rs;
328 read_counters(false);
330 clock_gettime(CLOCK_MONOTONIC, &ts);
331 diff_timespec(&rs, &ts, &ref_time);
334 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSECS_PER_SEC + rs.tv_nsec, INTERVAL))
335 pr_err("failed to write stat round event\n");
338 print_counters(&rs, 0, NULL);
341 static void enable_counters(void)
344 usleep(initial_delay * 1000);
347 * We need to enable counters only if:
348 * - we don't have tracee (attaching to task or cpu)
349 * - we have initial delay configured
351 if (!target__none(&target) || initial_delay)
352 perf_evlist__enable(evsel_list);
355 static volatile int workload_exec_errno;
358 * perf_evlist__prepare_workload will send a SIGUSR1
359 * if the fork fails, since we asked by setting its
360 * want_signal to true.
362 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
363 void *ucontext __maybe_unused)
365 workload_exec_errno = info->si_value.sival_int;
368 static bool has_unit(struct perf_evsel *counter)
370 return counter->unit && *counter->unit;
373 static bool has_scale(struct perf_evsel *counter)
375 return counter->scale != 1;
378 static int perf_stat_synthesize_config(bool is_pipe)
380 struct perf_evsel *counter;
384 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
385 process_synthesized_event);
387 pr_err("Couldn't synthesize attrs.\n");
393 * Synthesize other events stuff not carried within
394 * attr event - unit, scale, name
396 evlist__for_each(evsel_list, counter) {
397 if (!counter->supported)
401 * Synthesize unit and scale only if it's defined.
403 if (has_unit(counter)) {
404 err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
406 pr_err("Couldn't synthesize evsel unit.\n");
411 if (has_scale(counter)) {
412 err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
414 pr_err("Couldn't synthesize evsel scale.\n");
419 if (counter->own_cpus) {
420 err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
422 pr_err("Couldn't synthesize evsel scale.\n");
428 * Name is needed only for pipe output,
429 * perf.data carries event names.
432 err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
434 pr_err("Couldn't synthesize evsel name.\n");
440 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
441 process_synthesized_event,
444 pr_err("Couldn't synthesize thread map.\n");
448 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
449 process_synthesized_event, NULL);
451 pr_err("Couldn't synthesize thread map.\n");
455 err = perf_event__synthesize_stat_config(NULL, &stat_config,
456 process_synthesized_event, NULL);
458 pr_err("Couldn't synthesize config.\n");
465 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
467 static int __store_counter_ids(struct perf_evsel *counter,
468 struct cpu_map *cpus,
469 struct thread_map *threads)
473 for (cpu = 0; cpu < cpus->nr; cpu++) {
474 for (thread = 0; thread < threads->nr; thread++) {
475 int fd = FD(counter, cpu, thread);
477 if (perf_evlist__id_add_fd(evsel_list, counter,
478 cpu, thread, fd) < 0)
486 static int store_counter_ids(struct perf_evsel *counter)
488 struct cpu_map *cpus = counter->cpus;
489 struct thread_map *threads = counter->threads;
491 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
494 return __store_counter_ids(counter, cpus, threads);
497 static int __run_perf_stat(int argc, const char **argv)
499 int interval = stat_config.interval;
501 unsigned long long t0, t1;
502 struct perf_evsel *counter;
506 const bool forks = (argc > 0);
507 bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
510 ts.tv_sec = interval / 1000;
511 ts.tv_nsec = (interval % 1000) * 1000000;
518 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
519 workload_exec_failed_signal) < 0) {
520 perror("failed to prepare workload");
523 child_pid = evsel_list->workload.pid;
527 perf_evlist__set_leader(evsel_list);
529 evlist__for_each(evsel_list, counter) {
530 if (create_perf_stat_counter(counter) < 0) {
532 * PPC returns ENXIO for HW counters until 2.6.37
533 * (behavior changed with commit b0a873e).
535 if (errno == EINVAL || errno == ENOSYS ||
536 errno == ENOENT || errno == EOPNOTSUPP ||
539 ui__warning("%s event is not supported by the kernel.\n",
540 perf_evsel__name(counter));
541 counter->supported = false;
543 if ((counter->leader != counter) ||
544 !(counter->leader->nr_members > 1))
548 perf_evsel__open_strerror(counter, &target,
549 errno, msg, sizeof(msg));
550 ui__error("%s\n", msg);
553 kill(child_pid, SIGTERM);
557 counter->supported = true;
559 l = strlen(counter->unit);
563 if (STAT_RECORD && store_counter_ids(counter))
567 if (perf_evlist__apply_filters(evsel_list, &counter)) {
568 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
569 counter->filter, perf_evsel__name(counter), errno,
570 strerror_r(errno, msg, sizeof(msg)));
575 int err, fd = perf_data_file__fd(&perf_stat.file);
578 err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
580 err = perf_session__write_header(perf_stat.session, evsel_list,
587 err = perf_stat_synthesize_config(is_pipe);
593 * Enable counters and exec the command:
596 clock_gettime(CLOCK_MONOTONIC, &ref_time);
599 perf_evlist__start_workload(evsel_list);
603 while (!waitpid(child_pid, &status, WNOHANG)) {
604 nanosleep(&ts, NULL);
610 if (workload_exec_errno) {
611 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
612 pr_err("Workload failed: %s\n", emsg);
616 if (WIFSIGNALED(status))
617 psignal(WTERMSIG(status), argv[0]);
621 nanosleep(&ts, NULL);
629 update_stats(&walltime_nsecs_stats, t1 - t0);
633 return WEXITSTATUS(status);
636 static int run_perf_stat(int argc, const char **argv)
641 ret = system(pre_cmd);
649 ret = __run_perf_stat(argc, argv);
654 ret = system(post_cmd);
662 static void print_running(u64 run, u64 ena)
665 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
669 ena ? 100.0 * run / ena : 100.0);
670 } else if (run != ena) {
671 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
675 static void print_noise_pct(double total, double avg)
677 double pct = rel_stddev_stats(total, avg);
680 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
682 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
685 static void print_noise(struct perf_evsel *evsel, double avg)
687 struct perf_stat_evsel *ps;
693 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
696 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
698 switch (stat_config.aggr_mode) {
700 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
701 cpu_map__id_to_socket(id),
703 cpu_map__id_to_cpu(id),
710 fprintf(stat_config.output, "S%*d%s%*d%s",
719 fprintf(stat_config.output, "CPU%*d%s",
721 perf_evsel__cpus(evsel)->map[id], csv_sep);
724 fprintf(stat_config.output, "%*s-%*d%s",
726 thread_map__comm(evsel->threads, id),
728 thread_map__pid(evsel->threads, id),
744 struct perf_evsel *evsel;
747 #define METRIC_LEN 35
749 static void new_line_std(void *ctx)
751 struct outstate *os = ctx;
756 static void do_new_line_std(struct outstate *os)
759 fputs(os->prefix, os->fh);
760 aggr_printout(os->evsel, os->id, os->nr);
761 if (stat_config.aggr_mode == AGGR_NONE)
762 fprintf(os->fh, " ");
763 fprintf(os->fh, " ");
766 static void print_metric_std(void *ctx, const char *color, const char *fmt,
767 const char *unit, double val)
769 struct outstate *os = ctx;
772 bool newline = os->newline;
776 if (unit == NULL || fmt == NULL) {
777 fprintf(out, "%-*s", METRIC_LEN, "");
784 n = fprintf(out, " # ");
786 n += color_fprintf(out, color, fmt, val);
788 n += fprintf(out, fmt, val);
789 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
792 static void new_line_csv(void *ctx)
794 struct outstate *os = ctx;
799 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
800 aggr_printout(os->evsel, os->id, os->nr);
801 for (i = 0; i < os->nfields; i++)
802 fputs(csv_sep, os->fh);
805 static void print_metric_csv(void *ctx,
806 const char *color __maybe_unused,
807 const char *fmt, const char *unit, double val)
809 struct outstate *os = ctx;
811 char buf[64], *vals, *ends;
813 if (unit == NULL || fmt == NULL) {
814 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
817 snprintf(buf, sizeof(buf), fmt, val);
819 while (isspace(*vals))
822 while (isdigit(*ends) || *ends == '.')
825 while (isspace(*unit))
827 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
830 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
832 FILE *output = stat_config.output;
833 double msecs = avg / 1e6;
834 const char *fmt_v, *fmt_n;
837 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
838 fmt_n = csv_output ? "%s" : "%-25s";
840 aggr_printout(evsel, id, nr);
842 scnprintf(name, sizeof(name), "%s%s",
843 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
845 fprintf(output, fmt_v, msecs, csv_sep);
848 fprintf(output, "%s%s", evsel->unit, csv_sep);
850 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
852 fprintf(output, fmt_n, name);
855 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
858 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
865 if (stat_config.aggr_mode == AGGR_NONE)
868 if (stat_config.aggr_mode == AGGR_GLOBAL)
871 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
872 int cpu2 = perf_evsel__cpus(evsel)->map[i];
874 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
880 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
882 FILE *output = stat_config.output;
883 double sc = evsel->scale;
887 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
890 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
892 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
895 aggr_printout(evsel, id, nr);
897 fprintf(output, fmt, avg, csv_sep);
900 fprintf(output, "%-*s%s",
901 csv_output ? 0 : unit_width,
902 evsel->unit, csv_sep);
904 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
907 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
910 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
911 char *prefix, u64 run, u64 ena, double noise)
913 struct perf_stat_output_ctx out;
914 struct outstate os = {
915 .fh = stat_config.output,
916 .prefix = prefix ? prefix : "",
921 print_metric_t pm = print_metric_std;
927 static int aggr_fields[] = {
935 pm = print_metric_csv;
938 os.nfields += aggr_fields[stat_config.aggr_mode];
942 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
943 aggr_printout(counter, id, nr);
945 fprintf(stat_config.output, "%*s%s",
947 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
950 fprintf(stat_config.output, "%-*s%s",
951 csv_output ? 0 : unit_width,
952 counter->unit, csv_sep);
954 fprintf(stat_config.output, "%*s",
955 csv_output ? 0 : -25,
956 perf_evsel__name(counter));
959 fprintf(stat_config.output, "%s%s",
960 csv_sep, counter->cgrp->name);
963 pm(&os, NULL, NULL, "", 0);
964 print_noise(counter, noise);
965 print_running(run, ena);
967 pm(&os, NULL, NULL, "", 0);
971 if (nsec_counter(counter))
972 nsec_printout(id, nr, counter, uval);
974 abs_printout(id, nr, counter, uval);
976 out.print_metric = pm;
981 print_noise(counter, noise);
982 print_running(run, ena);
985 perf_stat__print_shadow_stats(counter, uval,
986 first_shadow_cpu(counter, id),
989 print_noise(counter, noise);
990 print_running(run, ena);
994 static void aggr_update_shadow(void)
998 struct perf_evsel *counter;
1000 for (s = 0; s < aggr_map->nr; s++) {
1001 id = aggr_map->map[s];
1002 evlist__for_each(evsel_list, counter) {
1004 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1005 s2 = aggr_get_id(evsel_list->cpus, cpu);
1008 val += perf_counts(counter->counts, cpu, 0)->val;
1010 val = val * counter->scale;
1011 perf_stat__update_shadow_stats(counter, &val,
1012 first_shadow_cpu(counter, id));
1017 static void print_aggr(char *prefix)
1019 FILE *output = stat_config.output;
1020 struct perf_evsel *counter;
1021 int cpu, s, s2, id, nr;
1025 if (!(aggr_map || aggr_get_id))
1028 aggr_update_shadow();
1030 for (s = 0; s < aggr_map->nr; s++) {
1031 id = aggr_map->map[s];
1032 evlist__for_each(evsel_list, counter) {
1033 val = ena = run = 0;
1035 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1036 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1039 val += perf_counts(counter->counts, cpu, 0)->val;
1040 ena += perf_counts(counter->counts, cpu, 0)->ena;
1041 run += perf_counts(counter->counts, cpu, 0)->run;
1045 fprintf(output, "%s", prefix);
1047 uval = val * counter->scale;
1048 printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1049 fputc('\n', output);
1054 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1056 FILE *output = stat_config.output;
1057 int nthreads = thread_map__nr(counter->threads);
1058 int ncpus = cpu_map__nr(counter->cpus);
1062 for (thread = 0; thread < nthreads; thread++) {
1063 u64 ena = 0, run = 0, val = 0;
1065 for (cpu = 0; cpu < ncpus; cpu++) {
1066 val += perf_counts(counter->counts, cpu, thread)->val;
1067 ena += perf_counts(counter->counts, cpu, thread)->ena;
1068 run += perf_counts(counter->counts, cpu, thread)->run;
1072 fprintf(output, "%s", prefix);
1074 uval = val * counter->scale;
1075 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1076 fputc('\n', output);
1081 * Print out the results of a single counter:
1082 * aggregated counts in system-wide mode
1084 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1086 FILE *output = stat_config.output;
1087 struct perf_stat_evsel *ps = counter->priv;
1088 double avg = avg_stats(&ps->res_stats[0]);
1090 double avg_enabled, avg_running;
1092 avg_enabled = avg_stats(&ps->res_stats[1]);
1093 avg_running = avg_stats(&ps->res_stats[2]);
1096 fprintf(output, "%s", prefix);
1098 uval = avg * counter->scale;
1099 printout(-1, 0, counter, uval, prefix, avg_running, avg_enabled, avg);
1100 fprintf(output, "\n");
1104 * Print out the results of a single counter:
1105 * does not use aggregated count in system-wide
1107 static void print_counter(struct perf_evsel *counter, char *prefix)
1109 FILE *output = stat_config.output;
1114 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1115 val = perf_counts(counter->counts, cpu, 0)->val;
1116 ena = perf_counts(counter->counts, cpu, 0)->ena;
1117 run = perf_counts(counter->counts, cpu, 0)->run;
1120 fprintf(output, "%s", prefix);
1122 uval = val * counter->scale;
1123 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1125 fputc('\n', output);
1129 static void print_interval(char *prefix, struct timespec *ts)
1131 FILE *output = stat_config.output;
1132 static int num_print_interval;
1134 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1136 if (num_print_interval == 0 && !csv_output) {
1137 switch (stat_config.aggr_mode) {
1139 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
1142 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
1145 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
1148 fprintf(output, "# time comm-pid counts %*s events\n", unit_width, "unit");
1152 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
1158 if (++num_print_interval == 25)
1159 num_print_interval = 0;
1162 static void print_header(int argc, const char **argv)
1164 FILE *output = stat_config.output;
1170 fprintf(output, "\n");
1171 fprintf(output, " Performance counter stats for ");
1172 if (target.system_wide)
1173 fprintf(output, "\'system wide");
1174 else if (target.cpu_list)
1175 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1176 else if (!target__has_task(&target)) {
1177 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1178 for (i = 1; argv && (i < argc); i++)
1179 fprintf(output, " %s", argv[i]);
1180 } else if (target.pid)
1181 fprintf(output, "process id \'%s", target.pid);
1183 fprintf(output, "thread id \'%s", target.tid);
1185 fprintf(output, "\'");
1187 fprintf(output, " (%d runs)", run_count);
1188 fprintf(output, ":\n\n");
1192 static void print_footer(void)
1194 FILE *output = stat_config.output;
1197 fprintf(output, "\n");
1198 fprintf(output, " %17.9f seconds time elapsed",
1199 avg_stats(&walltime_nsecs_stats)/1e9);
1200 if (run_count > 1) {
1201 fprintf(output, " ");
1202 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1203 avg_stats(&walltime_nsecs_stats));
1205 fprintf(output, "\n\n");
1208 static void print_counters(struct timespec *ts, int argc, const char **argv)
1210 int interval = stat_config.interval;
1211 struct perf_evsel *counter;
1212 char buf[64], *prefix = NULL;
1214 /* Do not print anything if we record to the pipe. */
1215 if (STAT_RECORD && perf_stat.file.is_pipe)
1219 print_interval(prefix = buf, ts);
1221 print_header(argc, argv);
1223 switch (stat_config.aggr_mode) {
1229 evlist__for_each(evsel_list, counter)
1230 print_aggr_thread(counter, prefix);
1233 evlist__for_each(evsel_list, counter)
1234 print_counter_aggr(counter, prefix);
1237 evlist__for_each(evsel_list, counter)
1238 print_counter(counter, prefix);
1245 if (!interval && !csv_output)
1248 fflush(stat_config.output);
1251 static volatile int signr = -1;
1253 static void skip_signal(int signo)
1255 if ((child_pid == -1) || stat_config.interval)
1260 * render child_pid harmless
1261 * won't send SIGTERM to a random
1262 * process in case of race condition
1263 * and fast PID recycling
1268 static void sig_atexit(void)
1273 * avoid race condition with SIGCHLD handler
1274 * in skip_signal() which is modifying child_pid
1275 * goal is to avoid send SIGTERM to a random
1279 sigaddset(&set, SIGCHLD);
1280 sigprocmask(SIG_BLOCK, &set, &oset);
1282 if (child_pid != -1)
1283 kill(child_pid, SIGTERM);
1285 sigprocmask(SIG_SETMASK, &oset, NULL);
1290 signal(signr, SIG_DFL);
1291 kill(getpid(), signr);
1294 static int stat__set_big_num(const struct option *opt __maybe_unused,
1295 const char *s __maybe_unused, int unset)
1297 big_num_opt = unset ? 0 : 1;
1301 static const struct option stat_options[] = {
1302 OPT_BOOLEAN('T', "transaction", &transaction_run,
1303 "hardware transaction statistics"),
1304 OPT_CALLBACK('e', "event", &evsel_list, "event",
1305 "event selector. use 'perf list' to list available events",
1306 parse_events_option),
1307 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1308 "event filter", parse_filter),
1309 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1310 "child tasks do not inherit counters"),
1311 OPT_STRING('p', "pid", &target.pid, "pid",
1312 "stat events on existing process id"),
1313 OPT_STRING('t', "tid", &target.tid, "tid",
1314 "stat events on existing thread id"),
1315 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1316 "system-wide collection from all CPUs"),
1317 OPT_BOOLEAN('g', "group", &group,
1318 "put the counters into a counter group"),
1319 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1320 OPT_INCR('v', "verbose", &verbose,
1321 "be more verbose (show counter open errors, etc)"),
1322 OPT_INTEGER('r', "repeat", &run_count,
1323 "repeat command and print average + stddev (max: 100, forever: 0)"),
1324 OPT_BOOLEAN('n', "null", &null_run,
1325 "null run - dont start any counters"),
1326 OPT_INCR('d', "detailed", &detailed_run,
1327 "detailed run - start a lot of events"),
1328 OPT_BOOLEAN('S', "sync", &sync_run,
1329 "call sync() before starting a run"),
1330 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1331 "print large numbers with thousands\' separators",
1333 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1334 "list of cpus to monitor in system-wide"),
1335 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1336 "disable CPU count aggregation", AGGR_NONE),
1337 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1338 "print counts with custom separator"),
1339 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1340 "monitor event in cgroup name only", parse_cgroups),
1341 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1342 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1343 OPT_INTEGER(0, "log-fd", &output_fd,
1344 "log output to fd, instead of stderr"),
1345 OPT_STRING(0, "pre", &pre_cmd, "command",
1346 "command to run prior to the measured command"),
1347 OPT_STRING(0, "post", &post_cmd, "command",
1348 "command to run after to the measured command"),
1349 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1350 "print counts at regular interval in ms (>= 10)"),
1351 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1352 "aggregate counts per processor socket", AGGR_SOCKET),
1353 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1354 "aggregate counts per physical processor core", AGGR_CORE),
1355 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1356 "aggregate counts per thread", AGGR_THREAD),
1357 OPT_UINTEGER('D', "delay", &initial_delay,
1358 "ms to wait before starting measurement after program start"),
1362 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1364 return cpu_map__get_socket(map, cpu, NULL);
1367 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1369 return cpu_map__get_core(map, cpu, NULL);
1372 static int cpu_map__get_max(struct cpu_map *map)
1376 for (i = 0; i < map->nr; i++) {
1377 if (map->map[i] > max)
1384 static struct cpu_map *cpus_aggr_map;
1386 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1393 cpu = map->map[idx];
1395 if (cpus_aggr_map->map[cpu] == -1)
1396 cpus_aggr_map->map[cpu] = get_id(map, idx);
1398 return cpus_aggr_map->map[cpu];
1401 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1403 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1406 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1408 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1411 static int perf_stat_init_aggr_mode(void)
1415 switch (stat_config.aggr_mode) {
1417 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1418 perror("cannot build socket map");
1421 aggr_get_id = perf_stat__get_socket_cached;
1424 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1425 perror("cannot build core map");
1428 aggr_get_id = perf_stat__get_core_cached;
1439 * The evsel_list->cpus is the base we operate on,
1440 * taking the highest cpu number to be the size of
1441 * the aggregation translate cpumap.
1443 nr = cpu_map__get_max(evsel_list->cpus);
1444 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1445 return cpus_aggr_map ? 0 : -ENOMEM;
1448 static void perf_stat__exit_aggr_mode(void)
1450 cpu_map__put(aggr_map);
1451 cpu_map__put(cpus_aggr_map);
1453 cpus_aggr_map = NULL;
1456 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1463 cpu = map->map[idx];
1465 if (cpu >= env->nr_cpus_online)
1471 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1473 struct perf_env *env = data;
1474 int cpu = perf_env__get_cpu(env, map, idx);
1476 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1479 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1481 struct perf_env *env = data;
1482 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1485 int socket_id = env->cpu[cpu].socket_id;
1488 * Encode socket in upper 16 bits
1489 * core_id is relative to socket, and
1490 * we need a global id. So we combine
1493 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1499 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1500 struct cpu_map **sockp)
1502 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1505 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1506 struct cpu_map **corep)
1508 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1511 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1513 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1516 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1518 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1521 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1523 struct perf_env *env = &st->session->header.env;
1525 switch (stat_config.aggr_mode) {
1527 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1528 perror("cannot build socket map");
1531 aggr_get_id = perf_stat__get_socket_file;
1534 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1535 perror("cannot build core map");
1538 aggr_get_id = perf_stat__get_core_file;
1552 * Add default attributes, if there were no attributes specified or
1553 * if -d/--detailed, -d -d or -d -d -d is used:
1555 static int add_default_attributes(void)
1557 struct perf_event_attr default_attrs0[] = {
1559 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1560 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1561 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1562 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1564 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1566 struct perf_event_attr frontend_attrs[] = {
1567 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1569 struct perf_event_attr backend_attrs[] = {
1570 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1572 struct perf_event_attr default_attrs1[] = {
1573 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1574 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1575 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1580 * Detailed stats (-d), covering the L1 and last level data caches:
1582 struct perf_event_attr detailed_attrs[] = {
1584 { .type = PERF_TYPE_HW_CACHE,
1586 PERF_COUNT_HW_CACHE_L1D << 0 |
1587 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1588 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1590 { .type = PERF_TYPE_HW_CACHE,
1592 PERF_COUNT_HW_CACHE_L1D << 0 |
1593 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1594 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1596 { .type = PERF_TYPE_HW_CACHE,
1598 PERF_COUNT_HW_CACHE_LL << 0 |
1599 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1600 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1602 { .type = PERF_TYPE_HW_CACHE,
1604 PERF_COUNT_HW_CACHE_LL << 0 |
1605 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1606 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1610 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1612 struct perf_event_attr very_detailed_attrs[] = {
1614 { .type = PERF_TYPE_HW_CACHE,
1616 PERF_COUNT_HW_CACHE_L1I << 0 |
1617 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1618 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1620 { .type = PERF_TYPE_HW_CACHE,
1622 PERF_COUNT_HW_CACHE_L1I << 0 |
1623 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1624 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1626 { .type = PERF_TYPE_HW_CACHE,
1628 PERF_COUNT_HW_CACHE_DTLB << 0 |
1629 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1630 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1632 { .type = PERF_TYPE_HW_CACHE,
1634 PERF_COUNT_HW_CACHE_DTLB << 0 |
1635 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1636 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1638 { .type = PERF_TYPE_HW_CACHE,
1640 PERF_COUNT_HW_CACHE_ITLB << 0 |
1641 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1642 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1644 { .type = PERF_TYPE_HW_CACHE,
1646 PERF_COUNT_HW_CACHE_ITLB << 0 |
1647 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1648 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1653 * Very, very detailed stats (-d -d -d), adding prefetch events:
1655 struct perf_event_attr very_very_detailed_attrs[] = {
1657 { .type = PERF_TYPE_HW_CACHE,
1659 PERF_COUNT_HW_CACHE_L1D << 0 |
1660 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1661 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1663 { .type = PERF_TYPE_HW_CACHE,
1665 PERF_COUNT_HW_CACHE_L1D << 0 |
1666 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1667 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1670 /* Set attrs if no event is selected and !null_run: */
1674 if (transaction_run) {
1676 if (pmu_have_event("cpu", "cycles-ct") &&
1677 pmu_have_event("cpu", "el-start"))
1678 err = parse_events(evsel_list, transaction_attrs, NULL);
1680 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1682 fprintf(stderr, "Cannot set up transaction events\n");
1688 if (!evsel_list->nr_entries) {
1689 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1691 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1692 if (perf_evlist__add_default_attrs(evsel_list,
1693 frontend_attrs) < 0)
1696 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1697 if (perf_evlist__add_default_attrs(evsel_list,
1701 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1705 /* Detailed events get appended to the event list: */
1707 if (detailed_run < 1)
1710 /* Append detailed run extra attributes: */
1711 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1714 if (detailed_run < 2)
1717 /* Append very detailed run extra attributes: */
1718 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1721 if (detailed_run < 3)
1724 /* Append very, very detailed run extra attributes: */
1725 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1728 static const char * const stat_record_usage[] = {
1729 "perf stat record [<options>]",
1733 static void init_features(struct perf_session *session)
1737 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1738 perf_header__set_feat(&session->header, feat);
1740 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1741 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1742 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1743 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1746 static int __cmd_record(int argc, const char **argv)
1748 struct perf_session *session;
1749 struct perf_data_file *file = &perf_stat.file;
1751 argc = parse_options(argc, argv, stat_options, stat_record_usage,
1752 PARSE_OPT_STOP_AT_NON_OPTION);
1755 file->path = output_name;
1757 if (run_count != 1 || forever) {
1758 pr_err("Cannot use -r option with perf stat record.\n");
1762 session = perf_session__new(file, false, NULL);
1763 if (session == NULL) {
1764 pr_err("Perf session creation failed.\n");
1768 init_features(session);
1770 session->evlist = evsel_list;
1771 perf_stat.session = session;
1772 perf_stat.record = true;
1776 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
1777 union perf_event *event,
1778 struct perf_session *session)
1780 struct stat_round_event *round = &event->stat_round;
1781 struct perf_evsel *counter;
1782 struct timespec tsh, *ts = NULL;
1783 const char **argv = session->header.env.cmdline_argv;
1784 int argc = session->header.env.nr_cmdline;
1786 evlist__for_each(evsel_list, counter)
1787 perf_stat_process_counter(&stat_config, counter);
1789 if (round->type == PERF_STAT_ROUND_TYPE__FINAL)
1790 update_stats(&walltime_nsecs_stats, round->time);
1792 if (stat_config.interval && round->time) {
1793 tsh.tv_sec = round->time / NSECS_PER_SEC;
1794 tsh.tv_nsec = round->time % NSECS_PER_SEC;
1798 print_counters(ts, argc, argv);
1803 int process_stat_config_event(struct perf_tool *tool __maybe_unused,
1804 union perf_event *event,
1805 struct perf_session *session __maybe_unused)
1807 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1809 perf_event__read_stat_config(&stat_config, &event->stat_config);
1811 if (cpu_map__empty(st->cpus)) {
1812 if (st->aggr_mode != AGGR_UNSET)
1813 pr_warning("warning: processing task data, aggregation mode not set\n");
1817 if (st->aggr_mode != AGGR_UNSET)
1818 stat_config.aggr_mode = st->aggr_mode;
1820 if (perf_stat.file.is_pipe)
1821 perf_stat_init_aggr_mode();
1823 perf_stat_init_aggr_mode_file(st);
1828 static int set_maps(struct perf_stat *st)
1830 if (!st->cpus || !st->threads)
1833 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1836 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
1838 if (perf_evlist__alloc_stats(evsel_list, true))
1841 st->maps_allocated = true;
1846 int process_thread_map_event(struct perf_tool *tool __maybe_unused,
1847 union perf_event *event,
1848 struct perf_session *session __maybe_unused)
1850 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1853 pr_warning("Extra thread map event, ignoring.\n");
1857 st->threads = thread_map__new_event(&event->thread_map);
1861 return set_maps(st);
1865 int process_cpu_map_event(struct perf_tool *tool __maybe_unused,
1866 union perf_event *event,
1867 struct perf_session *session __maybe_unused)
1869 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1870 struct cpu_map *cpus;
1873 pr_warning("Extra cpu map event, ignoring.\n");
1877 cpus = cpu_map__new_data(&event->cpu_map.data);
1882 return set_maps(st);
1885 static const char * const stat_report_usage[] = {
1886 "perf stat report [<options>]",
1890 static struct perf_stat perf_stat = {
1892 .attr = perf_event__process_attr,
1893 .event_update = perf_event__process_event_update,
1894 .thread_map = process_thread_map_event,
1895 .cpu_map = process_cpu_map_event,
1896 .stat_config = process_stat_config_event,
1897 .stat = perf_event__process_stat_event,
1898 .stat_round = process_stat_round_event,
1900 .aggr_mode = AGGR_UNSET,
1903 static int __cmd_report(int argc, const char **argv)
1905 struct perf_session *session;
1906 const struct option options[] = {
1907 OPT_STRING('i', "input", &input_name, "file", "input file name"),
1908 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1909 "aggregate counts per processor socket", AGGR_SOCKET),
1910 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1911 "aggregate counts per physical processor core", AGGR_CORE),
1912 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1913 "disable CPU count aggregation", AGGR_NONE),
1919 argc = parse_options(argc, argv, options, stat_report_usage, 0);
1921 if (!input_name || !strlen(input_name)) {
1922 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1925 input_name = "perf.data";
1928 perf_stat.file.path = input_name;
1929 perf_stat.file.mode = PERF_DATA_MODE_READ;
1931 session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
1932 if (session == NULL)
1935 perf_stat.session = session;
1936 stat_config.output = stderr;
1937 evsel_list = session->evlist;
1939 ret = perf_session__process_events(session);
1943 perf_session__delete(session);
1947 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1949 const char * const stat_usage[] = {
1950 "perf stat [<options>] [<command>]",
1953 int status = -EINVAL, run_idx;
1955 FILE *output = stderr;
1956 unsigned int interval;
1957 const char * const stat_subcommands[] = { "record", "report" };
1959 setlocale(LC_ALL, "");
1961 evsel_list = perf_evlist__new();
1962 if (evsel_list == NULL)
1965 parse_events__shrink_config_terms();
1966 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1967 (const char **) stat_usage,
1968 PARSE_OPT_STOP_AT_NON_OPTION);
1969 perf_stat__init_shadow_stats();
1973 if (!strcmp(csv_sep, "\\t"))
1976 csv_sep = DEFAULT_SEPARATOR;
1978 if (argc && !strncmp(argv[0], "rec", 3)) {
1979 argc = __cmd_record(argc, argv);
1982 } else if (argc && !strncmp(argv[0], "rep", 3))
1983 return __cmd_report(argc, argv);
1985 interval = stat_config.interval;
1988 * For record command the -o is already taken care of.
1990 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1993 if (output_name && output_fd) {
1994 fprintf(stderr, "cannot use both --output and --log-fd\n");
1995 parse_options_usage(stat_usage, stat_options, "o", 1);
1996 parse_options_usage(NULL, stat_options, "log-fd", 0);
2000 if (output_fd < 0) {
2001 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2002 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2008 mode = append_file ? "a" : "w";
2010 output = fopen(output_name, mode);
2012 perror("failed to create output file");
2015 clock_gettime(CLOCK_REALTIME, &tm);
2016 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2017 } else if (output_fd > 0) {
2018 mode = append_file ? "a" : "w";
2019 output = fdopen(output_fd, mode);
2021 perror("Failed opening logfd");
2026 stat_config.output = output;
2029 * let the spreadsheet do the pretty-printing
2032 /* User explicitly passed -B? */
2033 if (big_num_opt == 1) {
2034 fprintf(stderr, "-B option not supported with -x\n");
2035 parse_options_usage(stat_usage, stat_options, "B", 1);
2036 parse_options_usage(NULL, stat_options, "x", 1);
2038 } else /* Nope, so disable big number formatting */
2040 } else if (big_num_opt == 0) /* User passed --no-big-num */
2043 if (!argc && target__none(&target))
2044 usage_with_options(stat_usage, stat_options);
2046 if (run_count < 0) {
2047 pr_err("Run count must be a positive number\n");
2048 parse_options_usage(stat_usage, stat_options, "r", 1);
2050 } else if (run_count == 0) {
2055 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2056 fprintf(stderr, "The --per-thread option is only available "
2057 "when monitoring via -p -t options.\n");
2058 parse_options_usage(NULL, stat_options, "p", 1);
2059 parse_options_usage(NULL, stat_options, "t", 1);
2064 * no_aggr, cgroup are for system-wide only
2065 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2067 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2068 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2069 !target__has_cpu(&target)) {
2070 fprintf(stderr, "both cgroup and no-aggregation "
2071 "modes only available in system-wide mode\n");
2073 parse_options_usage(stat_usage, stat_options, "G", 1);
2074 parse_options_usage(NULL, stat_options, "A", 1);
2075 parse_options_usage(NULL, stat_options, "a", 1);
2079 if (add_default_attributes())
2082 target__validate(&target);
2084 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2085 if (target__has_task(&target)) {
2086 pr_err("Problems finding threads of monitor\n");
2087 parse_options_usage(stat_usage, stat_options, "p", 1);
2088 parse_options_usage(NULL, stat_options, "t", 1);
2089 } else if (target__has_cpu(&target)) {
2090 perror("failed to parse CPUs map");
2091 parse_options_usage(stat_usage, stat_options, "C", 1);
2092 parse_options_usage(NULL, stat_options, "a", 1);
2098 * Initialize thread_map with comm names,
2099 * so we could print it out on output.
2101 if (stat_config.aggr_mode == AGGR_THREAD)
2102 thread_map__read_comms(evsel_list->threads);
2104 if (interval && interval < 100) {
2105 if (interval < 10) {
2106 pr_err("print interval must be >= 10ms\n");
2107 parse_options_usage(stat_usage, stat_options, "I", 1);
2110 pr_warning("print interval < 100ms. "
2111 "The overhead percentage could be high in some cases. "
2112 "Please proceed with caution.\n");
2115 if (perf_evlist__alloc_stats(evsel_list, interval))
2118 if (perf_stat_init_aggr_mode())
2122 * We dont want to block the signals - that would cause
2123 * child tasks to inherit that and Ctrl-C would not work.
2124 * What we want is for Ctrl-C to work in the exec()-ed
2125 * task, but being ignored by perf stat itself:
2129 signal(SIGINT, skip_signal);
2130 signal(SIGCHLD, skip_signal);
2131 signal(SIGALRM, skip_signal);
2132 signal(SIGABRT, skip_signal);
2135 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2136 if (run_count != 1 && verbose)
2137 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2140 status = run_perf_stat(argc, argv);
2141 if (forever && status != -1) {
2142 print_counters(NULL, argc, argv);
2143 perf_stat__reset_stats();
2147 if (!forever && status != -1 && !interval)
2148 print_counters(NULL, argc, argv);
2152 * We synthesize the kernel mmap record just so that older tools
2153 * don't emit warnings about not being able to resolve symbols
2154 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2155 * a saner message about no samples being in the perf.data file.
2157 * This also serves to suppress a warning about f_header.data.size == 0
2158 * in header.c at the moment 'perf stat record' gets introduced, which
2159 * is not really needed once we start adding the stat specific PERF_RECORD_
2160 * records, but the need to suppress the kptr_restrict messages in older
2161 * tools remain -acme
2163 int fd = perf_data_file__fd(&perf_stat.file);
2164 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2165 process_synthesized_event,
2166 &perf_stat.session->machines.host);
2168 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2169 "older tools may produce warnings about this file\n.");
2173 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2174 pr_err("failed to write stat round event\n");
2177 if (!perf_stat.file.is_pipe) {
2178 perf_stat.session->header.data_size += perf_stat.bytes_written;
2179 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2182 perf_session__delete(perf_stat.session);
2185 perf_stat__exit_aggr_mode();
2186 perf_evlist__free_stats(evsel_list);
2188 perf_evlist__delete(evsel_list);