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 "util/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"
64 #include <sys/prctl.h>
67 #define DEFAULT_SEPARATOR " "
68 #define CNTR_NOT_SUPPORTED "<not supported>"
69 #define CNTR_NOT_COUNTED "<not counted>"
71 static void print_counters(struct timespec *ts, int argc, const char **argv);
73 /* Default events used for perf stat -T */
74 static const char *transaction_attrs = {
86 /* More limited version when the CPU does not have all events. */
87 static const char * transaction_limited_attrs = {
97 static struct perf_evlist *evsel_list;
99 static struct target target = {
103 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
105 static int run_count = 1;
106 static bool no_inherit = false;
107 static volatile pid_t child_pid = -1;
108 static bool null_run = false;
109 static int detailed_run = 0;
110 static bool transaction_run;
111 static bool big_num = true;
112 static int big_num_opt = -1;
113 static const char *csv_sep = NULL;
114 static bool csv_output = false;
115 static bool group = false;
116 static const char *pre_cmd = NULL;
117 static const char *post_cmd = NULL;
118 static bool sync_run = false;
119 static unsigned int initial_delay = 0;
120 static unsigned int unit_width = 4; /* strlen("unit") */
121 static bool forever = false;
122 static struct timespec ref_time;
123 static struct cpu_map *aggr_map;
124 static aggr_get_id_t aggr_get_id;
125 static bool append_file;
126 static const char *output_name;
127 static int output_fd;
129 static volatile int done = 0;
131 static struct perf_stat_config stat_config = {
132 .aggr_mode = AGGR_GLOBAL,
136 static inline void diff_timespec(struct timespec *r, struct timespec *a,
139 r->tv_sec = a->tv_sec - b->tv_sec;
140 if (a->tv_nsec < b->tv_nsec) {
141 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
144 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
148 static void perf_stat__reset_stats(void)
150 perf_evlist__reset_stats(evsel_list);
151 perf_stat__reset_shadow_stats();
154 static int create_perf_stat_counter(struct perf_evsel *evsel)
156 struct perf_event_attr *attr = &evsel->attr;
158 if (stat_config.scale)
159 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
160 PERF_FORMAT_TOTAL_TIME_RUNNING;
162 attr->inherit = !no_inherit;
165 * Some events get initialized with sample_(period/type) set,
166 * like tracepoints. Clear it up for counting.
168 attr->sample_period = 0;
169 attr->sample_type = 0;
171 if (target__has_cpu(&target))
172 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
174 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
177 attr->enable_on_exec = 1;
180 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
184 * Does the counter have nsecs as a unit?
186 static inline int nsec_counter(struct perf_evsel *evsel)
188 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
189 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
196 * Read out the results of a single counter:
197 * do not aggregate counts across CPUs in system-wide mode
199 static int read_counter(struct perf_evsel *counter)
201 int nthreads = thread_map__nr(evsel_list->threads);
202 int ncpus = perf_evsel__nr_cpus(counter);
205 if (!counter->supported)
208 if (counter->system_wide)
211 for (thread = 0; thread < nthreads; thread++) {
212 for (cpu = 0; cpu < ncpus; cpu++) {
213 struct perf_counts_values *count;
215 count = perf_counts(counter->counts, cpu, thread);
216 if (perf_evsel__read(counter, cpu, thread, count))
224 static void read_counters(bool close_counters)
226 struct perf_evsel *counter;
228 evlist__for_each(evsel_list, counter) {
229 if (read_counter(counter))
230 pr_debug("failed to read counter %s\n", counter->name);
232 if (perf_stat_process_counter(&stat_config, counter))
233 pr_warning("failed to process counter %s\n", counter->name);
235 if (close_counters) {
236 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
237 thread_map__nr(evsel_list->threads));
242 static void process_interval(void)
244 struct timespec ts, rs;
246 read_counters(false);
248 clock_gettime(CLOCK_MONOTONIC, &ts);
249 diff_timespec(&rs, &ts, &ref_time);
251 print_counters(&rs, 0, NULL);
254 static void handle_initial_delay(void)
257 usleep(initial_delay * 1000);
258 perf_evlist__enable(evsel_list);
262 static volatile int workload_exec_errno;
265 * perf_evlist__prepare_workload will send a SIGUSR1
266 * if the fork fails, since we asked by setting its
267 * want_signal to true.
269 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
270 void *ucontext __maybe_unused)
272 workload_exec_errno = info->si_value.sival_int;
275 static int __run_perf_stat(int argc, const char **argv)
277 int interval = stat_config.interval;
279 unsigned long long t0, t1;
280 struct perf_evsel *counter;
284 const bool forks = (argc > 0);
287 ts.tv_sec = interval / 1000;
288 ts.tv_nsec = (interval % 1000) * 1000000;
295 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
296 workload_exec_failed_signal) < 0) {
297 perror("failed to prepare workload");
300 child_pid = evsel_list->workload.pid;
304 perf_evlist__set_leader(evsel_list);
306 evlist__for_each(evsel_list, counter) {
307 if (create_perf_stat_counter(counter) < 0) {
309 * PPC returns ENXIO for HW counters until 2.6.37
310 * (behavior changed with commit b0a873e).
312 if (errno == EINVAL || errno == ENOSYS ||
313 errno == ENOENT || errno == EOPNOTSUPP ||
316 ui__warning("%s event is not supported by the kernel.\n",
317 perf_evsel__name(counter));
318 counter->supported = false;
320 if ((counter->leader != counter) ||
321 !(counter->leader->nr_members > 1))
325 perf_evsel__open_strerror(counter, &target,
326 errno, msg, sizeof(msg));
327 ui__error("%s\n", msg);
330 kill(child_pid, SIGTERM);
334 counter->supported = true;
336 l = strlen(counter->unit);
341 if (perf_evlist__apply_filters(evsel_list, &counter)) {
342 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
343 counter->filter, perf_evsel__name(counter), errno,
344 strerror_r(errno, msg, sizeof(msg)));
349 * Enable counters and exec the command:
352 clock_gettime(CLOCK_MONOTONIC, &ref_time);
355 perf_evlist__start_workload(evsel_list);
356 handle_initial_delay();
359 while (!waitpid(child_pid, &status, WNOHANG)) {
360 nanosleep(&ts, NULL);
366 if (workload_exec_errno) {
367 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
368 pr_err("Workload failed: %s\n", emsg);
372 if (WIFSIGNALED(status))
373 psignal(WTERMSIG(status), argv[0]);
375 handle_initial_delay();
377 nanosleep(&ts, NULL);
385 update_stats(&walltime_nsecs_stats, t1 - t0);
389 return WEXITSTATUS(status);
392 static int run_perf_stat(int argc, const char **argv)
397 ret = system(pre_cmd);
405 ret = __run_perf_stat(argc, argv);
410 ret = system(post_cmd);
418 static void print_running(u64 run, u64 ena)
421 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
425 ena ? 100.0 * run / ena : 100.0);
426 } else if (run != ena) {
427 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
431 static void print_noise_pct(double total, double avg)
433 double pct = rel_stddev_stats(total, avg);
436 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
438 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
441 static void print_noise(struct perf_evsel *evsel, double avg)
443 struct perf_stat_evsel *ps;
449 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
452 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
454 switch (stat_config.aggr_mode) {
456 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
457 cpu_map__id_to_socket(id),
459 cpu_map__id_to_cpu(id),
466 fprintf(stat_config.output, "S%*d%s%*d%s",
475 fprintf(stat_config.output, "CPU%*d%s",
477 perf_evsel__cpus(evsel)->map[id], csv_sep);
480 fprintf(stat_config.output, "%*s-%*d%s",
482 thread_map__comm(evsel->threads, id),
484 thread_map__pid(evsel->threads, id),
494 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
496 FILE *output = stat_config.output;
497 double msecs = avg / 1e6;
498 const char *fmt_v, *fmt_n;
501 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
502 fmt_n = csv_output ? "%s" : "%-25s";
504 aggr_printout(evsel, id, nr);
506 scnprintf(name, sizeof(name), "%s%s",
507 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
509 fprintf(output, fmt_v, msecs, csv_sep);
512 fprintf(output, "%s%s", evsel->unit, csv_sep);
514 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
516 fprintf(output, fmt_n, name);
519 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
522 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
524 FILE *output = stat_config.output;
525 double sc = evsel->scale;
529 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
532 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
534 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
537 aggr_printout(evsel, id, nr);
539 fprintf(output, fmt, avg, csv_sep);
542 fprintf(output, "%-*s%s",
543 csv_output ? 0 : unit_width,
544 evsel->unit, csv_sep);
546 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
549 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
552 static void printout(int id, int nr, struct perf_evsel *counter, double uval)
554 int cpu = cpu_map__id_to_cpu(id);
556 if (stat_config.aggr_mode == AGGR_GLOBAL)
559 if (nsec_counter(counter))
560 nsec_printout(id, nr, counter, uval);
562 abs_printout(id, nr, counter, uval);
564 if (!csv_output && !stat_config.interval)
565 perf_stat__print_shadow_stats(stat_config.output, counter,
567 stat_config.aggr_mode);
570 static void print_aggr(char *prefix)
572 FILE *output = stat_config.output;
573 struct perf_evsel *counter;
574 int cpu, s, s2, id, nr;
578 if (!(aggr_map || aggr_get_id))
581 for (s = 0; s < aggr_map->nr; s++) {
582 id = aggr_map->map[s];
583 evlist__for_each(evsel_list, counter) {
586 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
587 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
590 val += perf_counts(counter->counts, cpu, 0)->val;
591 ena += perf_counts(counter->counts, cpu, 0)->ena;
592 run += perf_counts(counter->counts, cpu, 0)->run;
596 fprintf(output, "%s", prefix);
598 if (run == 0 || ena == 0) {
599 aggr_printout(counter, id, nr);
601 fprintf(output, "%*s%s",
603 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
606 fprintf(output, "%-*s%s",
607 csv_output ? 0 : unit_width,
608 counter->unit, csv_sep);
610 fprintf(output, "%*s",
611 csv_output ? 0 : -25,
612 perf_evsel__name(counter));
615 fprintf(output, "%s%s",
616 csv_sep, counter->cgrp->name);
618 print_running(run, ena);
622 uval = val * counter->scale;
623 printout(id, nr, counter, uval);
625 print_noise(counter, 1.0);
627 print_running(run, ena);
633 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
635 FILE *output = stat_config.output;
636 int nthreads = thread_map__nr(counter->threads);
637 int ncpus = cpu_map__nr(counter->cpus);
641 for (thread = 0; thread < nthreads; thread++) {
642 u64 ena = 0, run = 0, val = 0;
644 for (cpu = 0; cpu < ncpus; cpu++) {
645 val += perf_counts(counter->counts, cpu, thread)->val;
646 ena += perf_counts(counter->counts, cpu, thread)->ena;
647 run += perf_counts(counter->counts, cpu, thread)->run;
651 fprintf(output, "%s", prefix);
653 uval = val * counter->scale;
654 printout(thread, 0, counter, uval);
657 print_noise(counter, 1.0);
659 print_running(run, ena);
665 * Print out the results of a single counter:
666 * aggregated counts in system-wide mode
668 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
670 FILE *output = stat_config.output;
671 struct perf_stat_evsel *ps = counter->priv;
672 double avg = avg_stats(&ps->res_stats[0]);
673 int scaled = counter->counts->scaled;
675 double avg_enabled, avg_running;
677 avg_enabled = avg_stats(&ps->res_stats[1]);
678 avg_running = avg_stats(&ps->res_stats[2]);
681 fprintf(output, "%s", prefix);
683 if (scaled == -1 || !counter->supported) {
684 fprintf(output, "%*s%s",
686 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
688 fprintf(output, "%-*s%s",
689 csv_output ? 0 : unit_width,
690 counter->unit, csv_sep);
691 fprintf(output, "%*s",
692 csv_output ? 0 : -25,
693 perf_evsel__name(counter));
696 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
698 print_running(avg_running, avg_enabled);
703 uval = avg * counter->scale;
704 printout(-1, 0, counter, uval);
706 print_noise(counter, avg);
708 print_running(avg_running, avg_enabled);
709 fprintf(output, "\n");
713 * Print out the results of a single counter:
714 * does not use aggregated count in system-wide
716 static void print_counter(struct perf_evsel *counter, char *prefix)
718 FILE *output = stat_config.output;
723 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
724 val = perf_counts(counter->counts, cpu, 0)->val;
725 ena = perf_counts(counter->counts, cpu, 0)->ena;
726 run = perf_counts(counter->counts, cpu, 0)->run;
729 fprintf(output, "%s", prefix);
731 if (run == 0 || ena == 0) {
732 fprintf(output, "CPU%*d%s%*s%s",
734 perf_evsel__cpus(counter)->map[cpu], csv_sep,
736 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
739 fprintf(output, "%-*s%s",
740 csv_output ? 0 : unit_width,
741 counter->unit, csv_sep);
743 fprintf(output, "%*s",
744 csv_output ? 0 : -25,
745 perf_evsel__name(counter));
748 fprintf(output, "%s%s",
749 csv_sep, counter->cgrp->name);
751 print_running(run, ena);
756 uval = val * counter->scale;
757 printout(cpu, 0, counter, uval);
759 print_noise(counter, 1.0);
760 print_running(run, ena);
766 static void print_interval(char *prefix, struct timespec *ts)
768 FILE *output = stat_config.output;
769 static int num_print_interval;
771 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
773 if (num_print_interval == 0 && !csv_output) {
774 switch (stat_config.aggr_mode) {
776 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
779 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
782 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
785 fprintf(output, "# time comm-pid counts %*s events\n", unit_width, "unit");
789 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
795 if (++num_print_interval == 25)
796 num_print_interval = 0;
799 static void print_header(int argc, const char **argv)
801 FILE *output = stat_config.output;
807 fprintf(output, "\n");
808 fprintf(output, " Performance counter stats for ");
809 if (target.system_wide)
810 fprintf(output, "\'system wide");
811 else if (target.cpu_list)
812 fprintf(output, "\'CPU(s) %s", target.cpu_list);
813 else if (!target__has_task(&target)) {
814 fprintf(output, "\'%s", argv[0]);
815 for (i = 1; i < argc; i++)
816 fprintf(output, " %s", argv[i]);
817 } else if (target.pid)
818 fprintf(output, "process id \'%s", target.pid);
820 fprintf(output, "thread id \'%s", target.tid);
822 fprintf(output, "\'");
824 fprintf(output, " (%d runs)", run_count);
825 fprintf(output, ":\n\n");
829 static void print_footer(void)
831 FILE *output = stat_config.output;
834 fprintf(output, "\n");
835 fprintf(output, " %17.9f seconds time elapsed",
836 avg_stats(&walltime_nsecs_stats)/1e9);
838 fprintf(output, " ");
839 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
840 avg_stats(&walltime_nsecs_stats));
842 fprintf(output, "\n\n");
845 static void print_counters(struct timespec *ts, int argc, const char **argv)
847 int interval = stat_config.interval;
848 struct perf_evsel *counter;
849 char buf[64], *prefix = NULL;
852 print_interval(prefix = buf, ts);
854 print_header(argc, argv);
856 switch (stat_config.aggr_mode) {
862 evlist__for_each(evsel_list, counter)
863 print_aggr_thread(counter, prefix);
866 evlist__for_each(evsel_list, counter)
867 print_counter_aggr(counter, prefix);
870 evlist__for_each(evsel_list, counter)
871 print_counter(counter, prefix);
878 if (!interval && !csv_output)
881 fflush(stat_config.output);
884 static volatile int signr = -1;
886 static void skip_signal(int signo)
888 if ((child_pid == -1) || stat_config.interval)
893 * render child_pid harmless
894 * won't send SIGTERM to a random
895 * process in case of race condition
896 * and fast PID recycling
901 static void sig_atexit(void)
906 * avoid race condition with SIGCHLD handler
907 * in skip_signal() which is modifying child_pid
908 * goal is to avoid send SIGTERM to a random
912 sigaddset(&set, SIGCHLD);
913 sigprocmask(SIG_BLOCK, &set, &oset);
916 kill(child_pid, SIGTERM);
918 sigprocmask(SIG_SETMASK, &oset, NULL);
923 signal(signr, SIG_DFL);
924 kill(getpid(), signr);
927 static int stat__set_big_num(const struct option *opt __maybe_unused,
928 const char *s __maybe_unused, int unset)
930 big_num_opt = unset ? 0 : 1;
934 static const struct option stat_options[] = {
935 OPT_BOOLEAN('T', "transaction", &transaction_run,
936 "hardware transaction statistics"),
937 OPT_CALLBACK('e', "event", &evsel_list, "event",
938 "event selector. use 'perf list' to list available events",
939 parse_events_option),
940 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
941 "event filter", parse_filter),
942 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
943 "child tasks do not inherit counters"),
944 OPT_STRING('p', "pid", &target.pid, "pid",
945 "stat events on existing process id"),
946 OPT_STRING('t', "tid", &target.tid, "tid",
947 "stat events on existing thread id"),
948 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
949 "system-wide collection from all CPUs"),
950 OPT_BOOLEAN('g', "group", &group,
951 "put the counters into a counter group"),
952 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
953 OPT_INCR('v', "verbose", &verbose,
954 "be more verbose (show counter open errors, etc)"),
955 OPT_INTEGER('r', "repeat", &run_count,
956 "repeat command and print average + stddev (max: 100, forever: 0)"),
957 OPT_BOOLEAN('n', "null", &null_run,
958 "null run - dont start any counters"),
959 OPT_INCR('d', "detailed", &detailed_run,
960 "detailed run - start a lot of events"),
961 OPT_BOOLEAN('S', "sync", &sync_run,
962 "call sync() before starting a run"),
963 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
964 "print large numbers with thousands\' separators",
966 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
967 "list of cpus to monitor in system-wide"),
968 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
969 "disable CPU count aggregation", AGGR_NONE),
970 OPT_STRING('x', "field-separator", &csv_sep, "separator",
971 "print counts with custom separator"),
972 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
973 "monitor event in cgroup name only", parse_cgroups),
974 OPT_STRING('o', "output", &output_name, "file", "output file name"),
975 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
976 OPT_INTEGER(0, "log-fd", &output_fd,
977 "log output to fd, instead of stderr"),
978 OPT_STRING(0, "pre", &pre_cmd, "command",
979 "command to run prior to the measured command"),
980 OPT_STRING(0, "post", &post_cmd, "command",
981 "command to run after to the measured command"),
982 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
983 "print counts at regular interval in ms (>= 10)"),
984 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
985 "aggregate counts per processor socket", AGGR_SOCKET),
986 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
987 "aggregate counts per physical processor core", AGGR_CORE),
988 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
989 "aggregate counts per thread", AGGR_THREAD),
990 OPT_UINTEGER('D', "delay", &initial_delay,
991 "ms to wait before starting measurement after program start"),
995 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
997 return cpu_map__get_socket(map, cpu, NULL);
1000 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1002 return cpu_map__get_core(map, cpu, NULL);
1005 static int cpu_map__get_max(struct cpu_map *map)
1009 for (i = 0; i < map->nr; i++) {
1010 if (map->map[i] > max)
1017 static struct cpu_map *cpus_aggr_map;
1019 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1026 cpu = map->map[idx];
1028 if (cpus_aggr_map->map[cpu] == -1)
1029 cpus_aggr_map->map[cpu] = get_id(map, idx);
1031 return cpus_aggr_map->map[cpu];
1034 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1036 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1039 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1041 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1044 static int perf_stat_init_aggr_mode(void)
1048 switch (stat_config.aggr_mode) {
1050 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1051 perror("cannot build socket map");
1054 aggr_get_id = perf_stat__get_socket_cached;
1057 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1058 perror("cannot build core map");
1061 aggr_get_id = perf_stat__get_core_cached;
1072 * The evsel_list->cpus is the base we operate on,
1073 * taking the highest cpu number to be the size of
1074 * the aggregation translate cpumap.
1076 nr = cpu_map__get_max(evsel_list->cpus);
1077 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1078 return cpus_aggr_map ? 0 : -ENOMEM;
1082 * Add default attributes, if there were no attributes specified or
1083 * if -d/--detailed, -d -d or -d -d -d is used:
1085 static int add_default_attributes(void)
1087 struct perf_event_attr default_attrs[] = {
1089 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1090 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1091 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1092 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1094 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1095 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1096 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1097 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1098 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1099 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1104 * Detailed stats (-d), covering the L1 and last level data caches:
1106 struct perf_event_attr detailed_attrs[] = {
1108 { .type = PERF_TYPE_HW_CACHE,
1110 PERF_COUNT_HW_CACHE_L1D << 0 |
1111 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1112 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1114 { .type = PERF_TYPE_HW_CACHE,
1116 PERF_COUNT_HW_CACHE_L1D << 0 |
1117 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1118 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1120 { .type = PERF_TYPE_HW_CACHE,
1122 PERF_COUNT_HW_CACHE_LL << 0 |
1123 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1124 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1126 { .type = PERF_TYPE_HW_CACHE,
1128 PERF_COUNT_HW_CACHE_LL << 0 |
1129 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1130 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1134 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1136 struct perf_event_attr very_detailed_attrs[] = {
1138 { .type = PERF_TYPE_HW_CACHE,
1140 PERF_COUNT_HW_CACHE_L1I << 0 |
1141 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1142 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1144 { .type = PERF_TYPE_HW_CACHE,
1146 PERF_COUNT_HW_CACHE_L1I << 0 |
1147 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1148 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1150 { .type = PERF_TYPE_HW_CACHE,
1152 PERF_COUNT_HW_CACHE_DTLB << 0 |
1153 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1154 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1156 { .type = PERF_TYPE_HW_CACHE,
1158 PERF_COUNT_HW_CACHE_DTLB << 0 |
1159 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1160 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1162 { .type = PERF_TYPE_HW_CACHE,
1164 PERF_COUNT_HW_CACHE_ITLB << 0 |
1165 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1166 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1168 { .type = PERF_TYPE_HW_CACHE,
1170 PERF_COUNT_HW_CACHE_ITLB << 0 |
1171 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1172 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1177 * Very, very detailed stats (-d -d -d), adding prefetch events:
1179 struct perf_event_attr very_very_detailed_attrs[] = {
1181 { .type = PERF_TYPE_HW_CACHE,
1183 PERF_COUNT_HW_CACHE_L1D << 0 |
1184 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1185 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1187 { .type = PERF_TYPE_HW_CACHE,
1189 PERF_COUNT_HW_CACHE_L1D << 0 |
1190 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1191 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1194 /* Set attrs if no event is selected and !null_run: */
1198 if (transaction_run) {
1200 if (pmu_have_event("cpu", "cycles-ct") &&
1201 pmu_have_event("cpu", "el-start"))
1202 err = parse_events(evsel_list, transaction_attrs, NULL);
1204 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1206 fprintf(stderr, "Cannot set up transaction events\n");
1212 if (!evsel_list->nr_entries) {
1213 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1217 /* Detailed events get appended to the event list: */
1219 if (detailed_run < 1)
1222 /* Append detailed run extra attributes: */
1223 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1226 if (detailed_run < 2)
1229 /* Append very detailed run extra attributes: */
1230 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1233 if (detailed_run < 3)
1236 /* Append very, very detailed run extra attributes: */
1237 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1240 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1242 const char * const stat_usage[] = {
1243 "perf stat [<options>] [<command>]",
1246 int status = -EINVAL, run_idx;
1248 FILE *output = stderr;
1249 unsigned int interval;
1251 setlocale(LC_ALL, "");
1253 evsel_list = perf_evlist__new();
1254 if (evsel_list == NULL)
1257 argc = parse_options(argc, argv, stat_options, stat_usage,
1258 PARSE_OPT_STOP_AT_NON_OPTION);
1260 interval = stat_config.interval;
1262 if (output_name && strcmp(output_name, "-"))
1265 if (output_name && output_fd) {
1266 fprintf(stderr, "cannot use both --output and --log-fd\n");
1267 parse_options_usage(stat_usage, stat_options, "o", 1);
1268 parse_options_usage(NULL, stat_options, "log-fd", 0);
1272 if (output_fd < 0) {
1273 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1274 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1280 mode = append_file ? "a" : "w";
1282 output = fopen(output_name, mode);
1284 perror("failed to create output file");
1287 clock_gettime(CLOCK_REALTIME, &tm);
1288 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1289 } else if (output_fd > 0) {
1290 mode = append_file ? "a" : "w";
1291 output = fdopen(output_fd, mode);
1293 perror("Failed opening logfd");
1298 stat_config.output = output;
1302 if (!strcmp(csv_sep, "\\t"))
1305 csv_sep = DEFAULT_SEPARATOR;
1308 * let the spreadsheet do the pretty-printing
1311 /* User explicitly passed -B? */
1312 if (big_num_opt == 1) {
1313 fprintf(stderr, "-B option not supported with -x\n");
1314 parse_options_usage(stat_usage, stat_options, "B", 1);
1315 parse_options_usage(NULL, stat_options, "x", 1);
1317 } else /* Nope, so disable big number formatting */
1319 } else if (big_num_opt == 0) /* User passed --no-big-num */
1322 if (!argc && target__none(&target))
1323 usage_with_options(stat_usage, stat_options);
1325 if (run_count < 0) {
1326 pr_err("Run count must be a positive number\n");
1327 parse_options_usage(stat_usage, stat_options, "r", 1);
1329 } else if (run_count == 0) {
1334 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
1335 fprintf(stderr, "The --per-thread option is only available "
1336 "when monitoring via -p -t options.\n");
1337 parse_options_usage(NULL, stat_options, "p", 1);
1338 parse_options_usage(NULL, stat_options, "t", 1);
1343 * no_aggr, cgroup are for system-wide only
1344 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1346 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1347 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1348 !target__has_cpu(&target)) {
1349 fprintf(stderr, "both cgroup and no-aggregation "
1350 "modes only available in system-wide mode\n");
1352 parse_options_usage(stat_usage, stat_options, "G", 1);
1353 parse_options_usage(NULL, stat_options, "A", 1);
1354 parse_options_usage(NULL, stat_options, "a", 1);
1358 if (add_default_attributes())
1361 target__validate(&target);
1363 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1364 if (target__has_task(&target)) {
1365 pr_err("Problems finding threads of monitor\n");
1366 parse_options_usage(stat_usage, stat_options, "p", 1);
1367 parse_options_usage(NULL, stat_options, "t", 1);
1368 } else if (target__has_cpu(&target)) {
1369 perror("failed to parse CPUs map");
1370 parse_options_usage(stat_usage, stat_options, "C", 1);
1371 parse_options_usage(NULL, stat_options, "a", 1);
1377 * Initialize thread_map with comm names,
1378 * so we could print it out on output.
1380 if (stat_config.aggr_mode == AGGR_THREAD)
1381 thread_map__read_comms(evsel_list->threads);
1383 if (interval && interval < 100) {
1384 if (interval < 10) {
1385 pr_err("print interval must be >= 10ms\n");
1386 parse_options_usage(stat_usage, stat_options, "I", 1);
1389 pr_warning("print interval < 100ms. "
1390 "The overhead percentage could be high in some cases. "
1391 "Please proceed with caution.\n");
1394 if (perf_evlist__alloc_stats(evsel_list, interval))
1397 if (perf_stat_init_aggr_mode())
1401 * We dont want to block the signals - that would cause
1402 * child tasks to inherit that and Ctrl-C would not work.
1403 * What we want is for Ctrl-C to work in the exec()-ed
1404 * task, but being ignored by perf stat itself:
1408 signal(SIGINT, skip_signal);
1409 signal(SIGCHLD, skip_signal);
1410 signal(SIGALRM, skip_signal);
1411 signal(SIGABRT, skip_signal);
1414 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1415 if (run_count != 1 && verbose)
1416 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1419 status = run_perf_stat(argc, argv);
1420 if (forever && status != -1) {
1421 print_counters(NULL, argc, argv);
1422 perf_stat__reset_stats();
1426 if (!forever && status != -1 && !interval)
1427 print_counters(NULL, argc, argv);
1429 perf_evlist__free_stats(evsel_list);
1431 perf_evlist__delete(evsel_list);