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"
63 #include <sys/prctl.h>
66 #define DEFAULT_SEPARATOR " "
67 #define CNTR_NOT_SUPPORTED "<not supported>"
68 #define CNTR_NOT_COUNTED "<not counted>"
70 static void print_stat(int argc, const char **argv);
71 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
72 static void print_counter(struct perf_evsel *counter, char *prefix);
73 static void print_aggr(char *prefix);
75 /* Default events used for perf stat -T */
76 static const char *transaction_attrs = {
88 /* More limited version when the CPU does not have all events. */
89 static const char * transaction_limited_attrs = {
99 static struct perf_evlist *evsel_list;
101 static struct target target = {
105 static int run_count = 1;
106 static bool no_inherit = false;
107 static bool scale = true;
108 static enum aggr_mode aggr_mode = AGGR_GLOBAL;
109 static volatile pid_t child_pid = -1;
110 static bool null_run = false;
111 static int detailed_run = 0;
112 static bool transaction_run;
113 static bool big_num = true;
114 static int big_num_opt = -1;
115 static const char *csv_sep = NULL;
116 static bool csv_output = false;
117 static bool group = false;
118 static FILE *output = NULL;
119 static const char *pre_cmd = NULL;
120 static const char *post_cmd = NULL;
121 static bool sync_run = false;
122 static unsigned int interval = 0;
123 static unsigned int initial_delay = 0;
124 static unsigned int unit_width = 4; /* strlen("unit") */
125 static bool forever = false;
126 static struct timespec ref_time;
127 static struct cpu_map *aggr_map;
128 static int (*aggr_get_id)(struct cpu_map *m, int cpu);
130 static volatile int done = 0;
132 static inline void diff_timespec(struct timespec *r, struct timespec *a,
135 r->tv_sec = a->tv_sec - b->tv_sec;
136 if (a->tv_nsec < b->tv_nsec) {
137 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
140 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
144 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
147 struct perf_stat *ps = evsel->priv;
149 for (i = 0; i < 3; i++)
150 init_stats(&ps->res_stats[i]);
152 perf_stat_evsel_id_init(evsel);
155 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
157 evsel->priv = zalloc(sizeof(struct perf_stat));
158 if (evsel->priv == NULL)
160 perf_evsel__reset_stat_priv(evsel);
164 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
169 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
171 struct perf_counts *counts;
173 counts = perf_counts__new(perf_evsel__nr_cpus(evsel));
175 evsel->prev_raw_counts = counts;
177 return counts ? 0 : -ENOMEM;
180 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
182 perf_counts__delete(evsel->prev_raw_counts);
183 evsel->prev_raw_counts = NULL;
186 static void perf_evlist__free_stats(struct perf_evlist *evlist)
188 struct perf_evsel *evsel;
190 evlist__for_each(evlist, evsel) {
191 perf_evsel__free_stat_priv(evsel);
192 perf_evsel__free_counts(evsel);
193 perf_evsel__free_prev_raw_counts(evsel);
197 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
199 struct perf_evsel *evsel;
201 evlist__for_each(evlist, evsel) {
202 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
203 perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
204 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
211 perf_evlist__free_stats(evlist);
215 static void perf_stat__reset_stats(struct perf_evlist *evlist)
217 struct perf_evsel *evsel;
219 evlist__for_each(evlist, evsel) {
220 perf_evsel__reset_stat_priv(evsel);
221 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
224 perf_stat__reset_shadow_stats();
227 static int create_perf_stat_counter(struct perf_evsel *evsel)
229 struct perf_event_attr *attr = &evsel->attr;
232 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
233 PERF_FORMAT_TOTAL_TIME_RUNNING;
235 attr->inherit = !no_inherit;
237 if (target__has_cpu(&target))
238 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
240 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
243 attr->enable_on_exec = 1;
246 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
250 * Does the counter have nsecs as a unit?
252 static inline int nsec_counter(struct perf_evsel *evsel)
254 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
255 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
261 static void zero_per_pkg(struct perf_evsel *counter)
263 if (counter->per_pkg_mask)
264 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
267 static int check_per_pkg(struct perf_evsel *counter, int cpu, bool *skip)
269 unsigned long *mask = counter->per_pkg_mask;
270 struct cpu_map *cpus = perf_evsel__cpus(counter);
275 if (!counter->per_pkg)
278 if (cpu_map__empty(cpus))
282 mask = zalloc(MAX_NR_CPUS);
286 counter->per_pkg_mask = mask;
289 s = cpu_map__get_socket(cpus, cpu);
293 *skip = test_and_set_bit(s, mask) == 1;
297 static int read_cb(struct perf_evsel *evsel, int cpu, int thread __maybe_unused,
298 struct perf_counts_values *count)
300 struct perf_counts_values *aggr = &evsel->counts->aggr;
301 static struct perf_counts_values zero;
304 if (check_per_pkg(evsel, cpu, &skip)) {
305 pr_err("failed to read per-pkg counter\n");
316 if (!evsel->snapshot)
317 perf_evsel__compute_deltas(evsel, cpu, count);
318 perf_counts_values__scale(count, scale, NULL);
319 *perf_counts(evsel->counts, cpu) = *count;
320 if (aggr_mode == AGGR_NONE)
321 perf_stat__update_shadow_stats(evsel, count->values, cpu);
324 aggr->val += count->val;
326 aggr->ena += count->ena;
327 aggr->run += count->run;
336 static int read_counter(struct perf_evsel *counter);
339 * Read out the results of a single counter:
340 * aggregate counts across CPUs in system-wide mode
342 static int read_counter_aggr(struct perf_evsel *counter)
344 struct perf_counts_values *aggr = &counter->counts->aggr;
345 struct perf_stat *ps = counter->priv;
346 u64 *count = counter->counts->aggr.values;
349 aggr->val = aggr->ena = aggr->run = 0;
351 if (read_counter(counter))
354 if (!counter->snapshot)
355 perf_evsel__compute_deltas(counter, -1, aggr);
356 perf_counts_values__scale(aggr, scale, &counter->counts->scaled);
358 for (i = 0; i < 3; i++)
359 update_stats(&ps->res_stats[i], count[i]);
362 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
363 perf_evsel__name(counter), count[0], count[1], count[2]);
367 * Save the full runtime - to allow normalization during printout:
369 perf_stat__update_shadow_stats(counter, count, 0);
375 * Read out the results of a single counter:
376 * do not aggregate counts across CPUs in system-wide mode
378 static int read_counter(struct perf_evsel *counter)
380 int nthreads = thread_map__nr(evsel_list->threads);
381 int ncpus = perf_evsel__nr_cpus(counter);
384 if (!counter->supported)
387 if (counter->system_wide)
390 if (counter->per_pkg)
391 zero_per_pkg(counter);
393 for (thread = 0; thread < nthreads; thread++) {
394 for (cpu = 0; cpu < ncpus; cpu++) {
395 if (perf_evsel__read_cb(counter, cpu, thread, read_cb))
403 static void print_interval(void)
405 static int num_print_interval;
406 struct perf_evsel *counter;
407 struct perf_stat *ps;
408 struct timespec ts, rs;
411 if (aggr_mode == AGGR_GLOBAL) {
412 evlist__for_each(evsel_list, counter) {
414 memset(ps->res_stats, 0, sizeof(ps->res_stats));
415 read_counter_aggr(counter);
418 evlist__for_each(evsel_list, counter) {
420 memset(ps->res_stats, 0, sizeof(ps->res_stats));
421 read_counter(counter);
425 clock_gettime(CLOCK_MONOTONIC, &ts);
426 diff_timespec(&rs, &ts, &ref_time);
427 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
429 if (num_print_interval == 0 && !csv_output) {
432 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
435 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
438 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
442 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
446 if (++num_print_interval == 25)
447 num_print_interval = 0;
455 evlist__for_each(evsel_list, counter)
456 print_counter(counter, prefix);
460 evlist__for_each(evsel_list, counter)
461 print_counter_aggr(counter, prefix);
467 static void handle_initial_delay(void)
469 struct perf_evsel *counter;
472 const int ncpus = cpu_map__nr(evsel_list->cpus),
473 nthreads = thread_map__nr(evsel_list->threads);
475 usleep(initial_delay * 1000);
476 evlist__for_each(evsel_list, counter)
477 perf_evsel__enable(counter, ncpus, nthreads);
481 static volatile int workload_exec_errno;
484 * perf_evlist__prepare_workload will send a SIGUSR1
485 * if the fork fails, since we asked by setting its
486 * want_signal to true.
488 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
489 void *ucontext __maybe_unused)
491 workload_exec_errno = info->si_value.sival_int;
494 static int __run_perf_stat(int argc, const char **argv)
497 unsigned long long t0, t1;
498 struct perf_evsel *counter;
502 const bool forks = (argc > 0);
505 ts.tv_sec = interval / 1000;
506 ts.tv_nsec = (interval % 1000) * 1000000;
513 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
514 workload_exec_failed_signal) < 0) {
515 perror("failed to prepare workload");
518 child_pid = evsel_list->workload.pid;
522 perf_evlist__set_leader(evsel_list);
524 evlist__for_each(evsel_list, counter) {
525 if (create_perf_stat_counter(counter) < 0) {
527 * PPC returns ENXIO for HW counters until 2.6.37
528 * (behavior changed with commit b0a873e).
530 if (errno == EINVAL || errno == ENOSYS ||
531 errno == ENOENT || errno == EOPNOTSUPP ||
534 ui__warning("%s event is not supported by the kernel.\n",
535 perf_evsel__name(counter));
536 counter->supported = false;
538 if ((counter->leader != counter) ||
539 !(counter->leader->nr_members > 1))
543 perf_evsel__open_strerror(counter, &target,
544 errno, msg, sizeof(msg));
545 ui__error("%s\n", msg);
548 kill(child_pid, SIGTERM);
552 counter->supported = true;
554 l = strlen(counter->unit);
559 if (perf_evlist__apply_filters(evsel_list, &counter)) {
560 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
561 counter->filter, perf_evsel__name(counter), errno,
562 strerror_r(errno, msg, sizeof(msg)));
567 * Enable counters and exec the command:
570 clock_gettime(CLOCK_MONOTONIC, &ref_time);
573 perf_evlist__start_workload(evsel_list);
574 handle_initial_delay();
577 while (!waitpid(child_pid, &status, WNOHANG)) {
578 nanosleep(&ts, NULL);
584 if (workload_exec_errno) {
585 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
586 pr_err("Workload failed: %s\n", emsg);
590 if (WIFSIGNALED(status))
591 psignal(WTERMSIG(status), argv[0]);
593 handle_initial_delay();
595 nanosleep(&ts, NULL);
603 update_stats(&walltime_nsecs_stats, t1 - t0);
605 if (aggr_mode == AGGR_GLOBAL) {
606 evlist__for_each(evsel_list, counter) {
607 read_counter_aggr(counter);
608 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
609 thread_map__nr(evsel_list->threads));
612 evlist__for_each(evsel_list, counter) {
613 read_counter(counter);
614 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
618 return WEXITSTATUS(status);
621 static int run_perf_stat(int argc, const char **argv)
626 ret = system(pre_cmd);
634 ret = __run_perf_stat(argc, argv);
639 ret = system(post_cmd);
647 static void print_running(u64 run, u64 ena)
650 fprintf(output, "%s%" PRIu64 "%s%.2f",
654 ena ? 100.0 * run / ena : 100.0);
655 } else if (run != ena) {
656 fprintf(output, " (%.2f%%)", 100.0 * run / ena);
660 static void print_noise_pct(double total, double avg)
662 double pct = rel_stddev_stats(total, avg);
665 fprintf(output, "%s%.2f%%", csv_sep, pct);
667 fprintf(output, " ( +-%6.2f%% )", pct);
670 static void print_noise(struct perf_evsel *evsel, double avg)
672 struct perf_stat *ps;
678 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
681 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
685 fprintf(output, "S%d-C%*d%s%*d%s",
686 cpu_map__id_to_socket(id),
688 cpu_map__id_to_cpu(id),
695 fprintf(output, "S%*d%s%*d%s",
704 fprintf(output, "CPU%*d%s",
706 perf_evsel__cpus(evsel)->map[id], csv_sep);
714 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
716 double msecs = avg / 1e6;
717 const char *fmt_v, *fmt_n;
720 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
721 fmt_n = csv_output ? "%s" : "%-25s";
723 aggr_printout(evsel, id, nr);
725 scnprintf(name, sizeof(name), "%s%s",
726 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
728 fprintf(output, fmt_v, msecs, csv_sep);
731 fprintf(output, "%s%s", evsel->unit, csv_sep);
733 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
735 fprintf(output, fmt_n, name);
738 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
740 if (csv_output || interval)
743 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
744 fprintf(output, " # %8.3f CPUs utilized ",
745 avg / avg_stats(&walltime_nsecs_stats));
747 fprintf(output, " ");
750 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
752 double sc = evsel->scale;
754 int cpu = cpu_map__id_to_cpu(id);
757 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
760 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
762 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
765 aggr_printout(evsel, id, nr);
767 if (aggr_mode == AGGR_GLOBAL)
770 fprintf(output, fmt, avg, csv_sep);
773 fprintf(output, "%-*s%s",
774 csv_output ? 0 : unit_width,
775 evsel->unit, csv_sep);
777 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
780 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
782 if (csv_output || interval)
785 perf_stat__print_shadow_stats(output, evsel, avg, cpu, aggr_mode);
788 static void print_aggr(char *prefix)
790 struct perf_evsel *counter;
791 int cpu, cpu2, s, s2, id, nr;
795 if (!(aggr_map || aggr_get_id))
798 for (s = 0; s < aggr_map->nr; s++) {
799 id = aggr_map->map[s];
800 evlist__for_each(evsel_list, counter) {
803 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
804 cpu2 = perf_evsel__cpus(counter)->map[cpu];
805 s2 = aggr_get_id(evsel_list->cpus, cpu2);
808 val += perf_counts(counter->counts, cpu)->val;
809 ena += perf_counts(counter->counts, cpu)->ena;
810 run += perf_counts(counter->counts, cpu)->run;
814 fprintf(output, "%s", prefix);
816 if (run == 0 || ena == 0) {
817 aggr_printout(counter, id, nr);
819 fprintf(output, "%*s%s",
821 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
824 fprintf(output, "%-*s%s",
825 csv_output ? 0 : unit_width,
826 counter->unit, csv_sep);
828 fprintf(output, "%*s",
829 csv_output ? 0 : -25,
830 perf_evsel__name(counter));
833 fprintf(output, "%s%s",
834 csv_sep, counter->cgrp->name);
836 print_running(run, ena);
840 uval = val * counter->scale;
842 if (nsec_counter(counter))
843 nsec_printout(id, nr, counter, uval);
845 abs_printout(id, nr, counter, uval);
848 print_noise(counter, 1.0);
850 print_running(run, ena);
857 * Print out the results of a single counter:
858 * aggregated counts in system-wide mode
860 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
862 struct perf_stat *ps = counter->priv;
863 double avg = avg_stats(&ps->res_stats[0]);
864 int scaled = counter->counts->scaled;
866 double avg_enabled, avg_running;
868 avg_enabled = avg_stats(&ps->res_stats[1]);
869 avg_running = avg_stats(&ps->res_stats[2]);
872 fprintf(output, "%s", prefix);
874 if (scaled == -1 || !counter->supported) {
875 fprintf(output, "%*s%s",
877 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
879 fprintf(output, "%-*s%s",
880 csv_output ? 0 : unit_width,
881 counter->unit, csv_sep);
882 fprintf(output, "%*s",
883 csv_output ? 0 : -25,
884 perf_evsel__name(counter));
887 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
889 print_running(avg_running, avg_enabled);
894 uval = avg * counter->scale;
896 if (nsec_counter(counter))
897 nsec_printout(-1, 0, counter, uval);
899 abs_printout(-1, 0, counter, uval);
901 print_noise(counter, avg);
903 print_running(avg_running, avg_enabled);
904 fprintf(output, "\n");
908 * Print out the results of a single counter:
909 * does not use aggregated count in system-wide
911 static void print_counter(struct perf_evsel *counter, char *prefix)
917 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
918 val = perf_counts(counter->counts, cpu)->val;
919 ena = perf_counts(counter->counts, cpu)->ena;
920 run = perf_counts(counter->counts, cpu)->run;
923 fprintf(output, "%s", prefix);
925 if (run == 0 || ena == 0) {
926 fprintf(output, "CPU%*d%s%*s%s",
928 perf_evsel__cpus(counter)->map[cpu], csv_sep,
930 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
933 fprintf(output, "%-*s%s",
934 csv_output ? 0 : unit_width,
935 counter->unit, csv_sep);
937 fprintf(output, "%*s",
938 csv_output ? 0 : -25,
939 perf_evsel__name(counter));
942 fprintf(output, "%s%s",
943 csv_sep, counter->cgrp->name);
945 print_running(run, ena);
950 uval = val * counter->scale;
952 if (nsec_counter(counter))
953 nsec_printout(cpu, 0, counter, uval);
955 abs_printout(cpu, 0, counter, uval);
958 print_noise(counter, 1.0);
959 print_running(run, ena);
965 static void print_stat(int argc, const char **argv)
967 struct perf_evsel *counter;
973 fprintf(output, "\n");
974 fprintf(output, " Performance counter stats for ");
975 if (target.system_wide)
976 fprintf(output, "\'system wide");
977 else if (target.cpu_list)
978 fprintf(output, "\'CPU(s) %s", target.cpu_list);
979 else if (!target__has_task(&target)) {
980 fprintf(output, "\'%s", argv[0]);
981 for (i = 1; i < argc; i++)
982 fprintf(output, " %s", argv[i]);
983 } else if (target.pid)
984 fprintf(output, "process id \'%s", target.pid);
986 fprintf(output, "thread id \'%s", target.tid);
988 fprintf(output, "\'");
990 fprintf(output, " (%d runs)", run_count);
991 fprintf(output, ":\n\n");
1000 evlist__for_each(evsel_list, counter)
1001 print_counter_aggr(counter, NULL);
1004 evlist__for_each(evsel_list, counter)
1005 print_counter(counter, NULL);
1013 fprintf(output, "\n");
1014 fprintf(output, " %17.9f seconds time elapsed",
1015 avg_stats(&walltime_nsecs_stats)/1e9);
1016 if (run_count > 1) {
1017 fprintf(output, " ");
1018 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1019 avg_stats(&walltime_nsecs_stats));
1021 fprintf(output, "\n\n");
1025 static volatile int signr = -1;
1027 static void skip_signal(int signo)
1029 if ((child_pid == -1) || interval)
1034 * render child_pid harmless
1035 * won't send SIGTERM to a random
1036 * process in case of race condition
1037 * and fast PID recycling
1042 static void sig_atexit(void)
1047 * avoid race condition with SIGCHLD handler
1048 * in skip_signal() which is modifying child_pid
1049 * goal is to avoid send SIGTERM to a random
1053 sigaddset(&set, SIGCHLD);
1054 sigprocmask(SIG_BLOCK, &set, &oset);
1056 if (child_pid != -1)
1057 kill(child_pid, SIGTERM);
1059 sigprocmask(SIG_SETMASK, &oset, NULL);
1064 signal(signr, SIG_DFL);
1065 kill(getpid(), signr);
1068 static int stat__set_big_num(const struct option *opt __maybe_unused,
1069 const char *s __maybe_unused, int unset)
1071 big_num_opt = unset ? 0 : 1;
1075 static int perf_stat_init_aggr_mode(void)
1077 switch (aggr_mode) {
1079 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1080 perror("cannot build socket map");
1083 aggr_get_id = cpu_map__get_socket;
1086 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1087 perror("cannot build core map");
1090 aggr_get_id = cpu_map__get_core;
1101 * Add default attributes, if there were no attributes specified or
1102 * if -d/--detailed, -d -d or -d -d -d is used:
1104 static int add_default_attributes(void)
1106 struct perf_event_attr default_attrs[] = {
1108 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1109 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1110 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1111 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1113 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1114 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1115 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1116 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1117 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1118 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1123 * Detailed stats (-d), covering the L1 and last level data caches:
1125 struct perf_event_attr detailed_attrs[] = {
1127 { .type = PERF_TYPE_HW_CACHE,
1129 PERF_COUNT_HW_CACHE_L1D << 0 |
1130 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1131 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1133 { .type = PERF_TYPE_HW_CACHE,
1135 PERF_COUNT_HW_CACHE_L1D << 0 |
1136 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1137 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1139 { .type = PERF_TYPE_HW_CACHE,
1141 PERF_COUNT_HW_CACHE_LL << 0 |
1142 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1143 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1145 { .type = PERF_TYPE_HW_CACHE,
1147 PERF_COUNT_HW_CACHE_LL << 0 |
1148 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1149 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1153 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1155 struct perf_event_attr very_detailed_attrs[] = {
1157 { .type = PERF_TYPE_HW_CACHE,
1159 PERF_COUNT_HW_CACHE_L1I << 0 |
1160 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1161 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1163 { .type = PERF_TYPE_HW_CACHE,
1165 PERF_COUNT_HW_CACHE_L1I << 0 |
1166 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1167 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1169 { .type = PERF_TYPE_HW_CACHE,
1171 PERF_COUNT_HW_CACHE_DTLB << 0 |
1172 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1173 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1175 { .type = PERF_TYPE_HW_CACHE,
1177 PERF_COUNT_HW_CACHE_DTLB << 0 |
1178 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1179 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1181 { .type = PERF_TYPE_HW_CACHE,
1183 PERF_COUNT_HW_CACHE_ITLB << 0 |
1184 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1185 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1187 { .type = PERF_TYPE_HW_CACHE,
1189 PERF_COUNT_HW_CACHE_ITLB << 0 |
1190 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1191 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1196 * Very, very detailed stats (-d -d -d), adding prefetch events:
1198 struct perf_event_attr very_very_detailed_attrs[] = {
1200 { .type = PERF_TYPE_HW_CACHE,
1202 PERF_COUNT_HW_CACHE_L1D << 0 |
1203 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1204 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1206 { .type = PERF_TYPE_HW_CACHE,
1208 PERF_COUNT_HW_CACHE_L1D << 0 |
1209 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1210 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1213 /* Set attrs if no event is selected and !null_run: */
1217 if (transaction_run) {
1219 if (pmu_have_event("cpu", "cycles-ct") &&
1220 pmu_have_event("cpu", "el-start"))
1221 err = parse_events(evsel_list, transaction_attrs, NULL);
1223 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1225 fprintf(stderr, "Cannot set up transaction events\n");
1231 if (!evsel_list->nr_entries) {
1232 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1236 /* Detailed events get appended to the event list: */
1238 if (detailed_run < 1)
1241 /* Append detailed run extra attributes: */
1242 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1245 if (detailed_run < 2)
1248 /* Append very detailed run extra attributes: */
1249 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1252 if (detailed_run < 3)
1255 /* Append very, very detailed run extra attributes: */
1256 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1259 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1261 bool append_file = false;
1263 const char *output_name = NULL;
1264 const struct option options[] = {
1265 OPT_BOOLEAN('T', "transaction", &transaction_run,
1266 "hardware transaction statistics"),
1267 OPT_CALLBACK('e', "event", &evsel_list, "event",
1268 "event selector. use 'perf list' to list available events",
1269 parse_events_option),
1270 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1271 "event filter", parse_filter),
1272 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1273 "child tasks do not inherit counters"),
1274 OPT_STRING('p', "pid", &target.pid, "pid",
1275 "stat events on existing process id"),
1276 OPT_STRING('t', "tid", &target.tid, "tid",
1277 "stat events on existing thread id"),
1278 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1279 "system-wide collection from all CPUs"),
1280 OPT_BOOLEAN('g', "group", &group,
1281 "put the counters into a counter group"),
1282 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1283 OPT_INCR('v', "verbose", &verbose,
1284 "be more verbose (show counter open errors, etc)"),
1285 OPT_INTEGER('r', "repeat", &run_count,
1286 "repeat command and print average + stddev (max: 100, forever: 0)"),
1287 OPT_BOOLEAN('n', "null", &null_run,
1288 "null run - dont start any counters"),
1289 OPT_INCR('d', "detailed", &detailed_run,
1290 "detailed run - start a lot of events"),
1291 OPT_BOOLEAN('S', "sync", &sync_run,
1292 "call sync() before starting a run"),
1293 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1294 "print large numbers with thousands\' separators",
1296 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1297 "list of cpus to monitor in system-wide"),
1298 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1299 "disable CPU count aggregation", AGGR_NONE),
1300 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1301 "print counts with custom separator"),
1302 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1303 "monitor event in cgroup name only", parse_cgroups),
1304 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1305 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1306 OPT_INTEGER(0, "log-fd", &output_fd,
1307 "log output to fd, instead of stderr"),
1308 OPT_STRING(0, "pre", &pre_cmd, "command",
1309 "command to run prior to the measured command"),
1310 OPT_STRING(0, "post", &post_cmd, "command",
1311 "command to run after to the measured command"),
1312 OPT_UINTEGER('I', "interval-print", &interval,
1313 "print counts at regular interval in ms (>= 100)"),
1314 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1315 "aggregate counts per processor socket", AGGR_SOCKET),
1316 OPT_SET_UINT(0, "per-core", &aggr_mode,
1317 "aggregate counts per physical processor core", AGGR_CORE),
1318 OPT_UINTEGER('D', "delay", &initial_delay,
1319 "ms to wait before starting measurement after program start"),
1322 const char * const stat_usage[] = {
1323 "perf stat [<options>] [<command>]",
1326 int status = -EINVAL, run_idx;
1329 setlocale(LC_ALL, "");
1331 evsel_list = perf_evlist__new();
1332 if (evsel_list == NULL)
1335 argc = parse_options(argc, argv, options, stat_usage,
1336 PARSE_OPT_STOP_AT_NON_OPTION);
1339 if (output_name && strcmp(output_name, "-"))
1342 if (output_name && output_fd) {
1343 fprintf(stderr, "cannot use both --output and --log-fd\n");
1344 parse_options_usage(stat_usage, options, "o", 1);
1345 parse_options_usage(NULL, options, "log-fd", 0);
1349 if (output_fd < 0) {
1350 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1351 parse_options_usage(stat_usage, options, "log-fd", 0);
1357 mode = append_file ? "a" : "w";
1359 output = fopen(output_name, mode);
1361 perror("failed to create output file");
1364 clock_gettime(CLOCK_REALTIME, &tm);
1365 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1366 } else if (output_fd > 0) {
1367 mode = append_file ? "a" : "w";
1368 output = fdopen(output_fd, mode);
1370 perror("Failed opening logfd");
1377 if (!strcmp(csv_sep, "\\t"))
1380 csv_sep = DEFAULT_SEPARATOR;
1383 * let the spreadsheet do the pretty-printing
1386 /* User explicitly passed -B? */
1387 if (big_num_opt == 1) {
1388 fprintf(stderr, "-B option not supported with -x\n");
1389 parse_options_usage(stat_usage, options, "B", 1);
1390 parse_options_usage(NULL, options, "x", 1);
1392 } else /* Nope, so disable big number formatting */
1394 } else if (big_num_opt == 0) /* User passed --no-big-num */
1397 if (!argc && target__none(&target))
1398 usage_with_options(stat_usage, options);
1400 if (run_count < 0) {
1401 pr_err("Run count must be a positive number\n");
1402 parse_options_usage(stat_usage, options, "r", 1);
1404 } else if (run_count == 0) {
1409 /* no_aggr, cgroup are for system-wide only */
1410 if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1411 !target__has_cpu(&target)) {
1412 fprintf(stderr, "both cgroup and no-aggregation "
1413 "modes only available in system-wide mode\n");
1415 parse_options_usage(stat_usage, options, "G", 1);
1416 parse_options_usage(NULL, options, "A", 1);
1417 parse_options_usage(NULL, options, "a", 1);
1421 if (add_default_attributes())
1424 target__validate(&target);
1426 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1427 if (target__has_task(&target)) {
1428 pr_err("Problems finding threads of monitor\n");
1429 parse_options_usage(stat_usage, options, "p", 1);
1430 parse_options_usage(NULL, options, "t", 1);
1431 } else if (target__has_cpu(&target)) {
1432 perror("failed to parse CPUs map");
1433 parse_options_usage(stat_usage, options, "C", 1);
1434 parse_options_usage(NULL, options, "a", 1);
1438 if (interval && interval < 100) {
1439 pr_err("print interval must be >= 100ms\n");
1440 parse_options_usage(stat_usage, options, "I", 1);
1444 if (perf_evlist__alloc_stats(evsel_list, interval))
1447 if (perf_stat_init_aggr_mode())
1451 * We dont want to block the signals - that would cause
1452 * child tasks to inherit that and Ctrl-C would not work.
1453 * What we want is for Ctrl-C to work in the exec()-ed
1454 * task, but being ignored by perf stat itself:
1458 signal(SIGINT, skip_signal);
1459 signal(SIGCHLD, skip_signal);
1460 signal(SIGALRM, skip_signal);
1461 signal(SIGABRT, skip_signal);
1464 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1465 if (run_count != 1 && verbose)
1466 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1469 status = run_perf_stat(argc, argv);
1470 if (forever && status != -1) {
1471 print_stat(argc, argv);
1472 perf_stat__reset_stats(evsel_list);
1476 if (!forever && status != -1 && !interval)
1477 print_stat(argc, argv);
1479 perf_evlist__free_stats(evsel_list);
1481 perf_evlist__delete(evsel_list);