4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ./hackbench 10
13 Performance counter stats for './hackbench 10':
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
27 0.154822978 seconds time elapsed
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 * Improvements and fixes by:
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
41 * Released under the GPL v2. (and only v2, not any later version)
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include <subcmd/parse-options.h>
49 #include "util/parse-events.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/drv_configs.h"
56 #include "util/color.h"
57 #include "util/stat.h"
58 #include "util/header.h"
59 #include "util/cpumap.h"
60 #include "util/thread.h"
61 #include "util/thread_map.h"
62 #include "util/counts.h"
63 #include "util/group.h"
64 #include "util/session.h"
65 #include "util/tool.h"
66 #include "util/group.h"
69 #include <linux/time64.h>
70 #include <api/fs/fs.h>
72 #include <sys/prctl.h>
76 #define DEFAULT_SEPARATOR " "
77 #define CNTR_NOT_SUPPORTED "<not supported>"
78 #define CNTR_NOT_COUNTED "<not counted>"
80 static void print_counters(struct timespec *ts, int argc, const char **argv);
82 /* Default events used for perf stat -T */
83 static const char *transaction_attrs = {
95 /* More limited version when the CPU does not have all events. */
96 static const char * transaction_limited_attrs = {
106 static const char * topdown_attrs[] = {
107 "topdown-total-slots",
108 "topdown-slots-retired",
109 "topdown-recovery-bubbles",
110 "topdown-fetch-bubbles",
111 "topdown-slots-issued",
115 static struct perf_evlist *evsel_list;
117 static struct target target = {
121 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
123 static int run_count = 1;
124 static bool no_inherit = false;
125 static volatile pid_t child_pid = -1;
126 static bool null_run = false;
127 static int detailed_run = 0;
128 static bool transaction_run;
129 static bool topdown_run = false;
130 static bool big_num = true;
131 static int big_num_opt = -1;
132 static const char *csv_sep = NULL;
133 static bool csv_output = false;
134 static bool group = false;
135 static const char *pre_cmd = NULL;
136 static const char *post_cmd = NULL;
137 static bool sync_run = false;
138 static unsigned int initial_delay = 0;
139 static unsigned int unit_width = 4; /* strlen("unit") */
140 static bool forever = false;
141 static bool metric_only = false;
142 static bool force_metric_only = false;
143 static bool no_merge = false;
144 static struct timespec ref_time;
145 static struct cpu_map *aggr_map;
146 static aggr_get_id_t aggr_get_id;
147 static bool append_file;
148 static const char *output_name;
149 static int output_fd;
150 static int print_free_counters_hint;
154 struct perf_data_file file;
155 struct perf_session *session;
157 struct perf_tool tool;
159 struct cpu_map *cpus;
160 struct thread_map *threads;
161 enum aggr_mode aggr_mode;
164 static struct perf_stat perf_stat;
165 #define STAT_RECORD perf_stat.record
167 static volatile int done = 0;
169 static struct perf_stat_config stat_config = {
170 .aggr_mode = AGGR_GLOBAL,
174 static inline void diff_timespec(struct timespec *r, struct timespec *a,
177 r->tv_sec = a->tv_sec - b->tv_sec;
178 if (a->tv_nsec < b->tv_nsec) {
179 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
182 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
186 static void perf_stat__reset_stats(void)
188 perf_evlist__reset_stats(evsel_list);
189 perf_stat__reset_shadow_stats();
192 static int create_perf_stat_counter(struct perf_evsel *evsel)
194 struct perf_event_attr *attr = &evsel->attr;
196 if (stat_config.scale)
197 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
198 PERF_FORMAT_TOTAL_TIME_RUNNING;
200 attr->inherit = !no_inherit;
203 * Some events get initialized with sample_(period/type) set,
204 * like tracepoints. Clear it up for counting.
206 attr->sample_period = 0;
209 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
210 * while avoiding that older tools show confusing messages.
212 * However for pipe sessions we need to keep it zero,
213 * because script's perf_evsel__check_attr is triggered
214 * by attr->sample_type != 0, and we can't run it on
217 if (!(STAT_RECORD && perf_stat.file.is_pipe))
218 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
221 * Disabling all counters initially, they will be enabled
222 * either manually by us or by kernel via enable_on_exec
225 if (perf_evsel__is_group_leader(evsel)) {
229 * In case of initial_delay we enable tracee
232 if (target__none(&target) && !initial_delay)
233 attr->enable_on_exec = 1;
236 if (target__has_cpu(&target))
237 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
239 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
243 * Does the counter have nsecs as a unit?
245 static inline int nsec_counter(struct perf_evsel *evsel)
247 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
248 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
254 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
255 union perf_event *event,
256 struct perf_sample *sample __maybe_unused,
257 struct machine *machine __maybe_unused)
259 if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
260 pr_err("failed to write perf data, error: %m\n");
264 perf_stat.bytes_written += event->header.size;
268 static int write_stat_round_event(u64 tm, u64 type)
270 return perf_event__synthesize_stat_round(NULL, tm, type,
271 process_synthesized_event,
275 #define WRITE_STAT_ROUND_EVENT(time, interval) \
276 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
278 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
281 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
282 struct perf_counts_values *count)
284 struct perf_sample_id *sid = SID(counter, cpu, thread);
286 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
287 process_synthesized_event, NULL);
291 * Read out the results of a single counter:
292 * do not aggregate counts across CPUs in system-wide mode
294 static int read_counter(struct perf_evsel *counter)
296 int nthreads = thread_map__nr(evsel_list->threads);
297 int ncpus, cpu, thread;
299 if (target__has_cpu(&target))
300 ncpus = perf_evsel__nr_cpus(counter);
304 if (!counter->supported)
307 if (counter->system_wide)
310 for (thread = 0; thread < nthreads; thread++) {
311 for (cpu = 0; cpu < ncpus; cpu++) {
312 struct perf_counts_values *count;
314 count = perf_counts(counter->counts, cpu, thread);
315 if (perf_evsel__read(counter, cpu, thread, count)) {
316 counter->counts->scaled = -1;
317 perf_counts(counter->counts, cpu, thread)->ena = 0;
318 perf_counts(counter->counts, cpu, thread)->run = 0;
323 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
324 pr_err("failed to write stat event\n");
330 fprintf(stat_config.output,
331 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
332 perf_evsel__name(counter),
334 count->val, count->ena, count->run);
342 static void read_counters(void)
344 struct perf_evsel *counter;
347 evlist__for_each_entry(evsel_list, counter) {
348 ret = read_counter(counter);
350 pr_debug("failed to read counter %s\n", counter->name);
352 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
353 pr_warning("failed to process counter %s\n", counter->name);
357 static void process_interval(void)
359 struct timespec ts, rs;
363 clock_gettime(CLOCK_MONOTONIC, &ts);
364 diff_timespec(&rs, &ts, &ref_time);
367 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
368 pr_err("failed to write stat round event\n");
371 print_counters(&rs, 0, NULL);
374 static void enable_counters(void)
377 usleep(initial_delay * USEC_PER_MSEC);
380 * We need to enable counters only if:
381 * - we don't have tracee (attaching to task or cpu)
382 * - we have initial delay configured
384 if (!target__none(&target) || initial_delay)
385 perf_evlist__enable(evsel_list);
388 static void disable_counters(void)
391 * If we don't have tracee (attaching to task or cpu), counters may
392 * still be running. To get accurate group ratios, we must stop groups
393 * from counting before reading their constituent counters.
395 if (!target__none(&target))
396 perf_evlist__disable(evsel_list);
399 static volatile int workload_exec_errno;
402 * perf_evlist__prepare_workload will send a SIGUSR1
403 * if the fork fails, since we asked by setting its
404 * want_signal to true.
406 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
407 void *ucontext __maybe_unused)
409 workload_exec_errno = info->si_value.sival_int;
412 static bool has_unit(struct perf_evsel *counter)
414 return counter->unit && *counter->unit;
417 static bool has_scale(struct perf_evsel *counter)
419 return counter->scale != 1;
422 static int perf_stat_synthesize_config(bool is_pipe)
424 struct perf_evsel *counter;
428 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
429 process_synthesized_event);
431 pr_err("Couldn't synthesize attrs.\n");
437 * Synthesize other events stuff not carried within
438 * attr event - unit, scale, name
440 evlist__for_each_entry(evsel_list, counter) {
441 if (!counter->supported)
445 * Synthesize unit and scale only if it's defined.
447 if (has_unit(counter)) {
448 err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
450 pr_err("Couldn't synthesize evsel unit.\n");
455 if (has_scale(counter)) {
456 err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
458 pr_err("Couldn't synthesize evsel scale.\n");
463 if (counter->own_cpus) {
464 err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
466 pr_err("Couldn't synthesize evsel scale.\n");
472 * Name is needed only for pipe output,
473 * perf.data carries event names.
476 err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
478 pr_err("Couldn't synthesize evsel name.\n");
484 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
485 process_synthesized_event,
488 pr_err("Couldn't synthesize thread map.\n");
492 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
493 process_synthesized_event, NULL);
495 pr_err("Couldn't synthesize thread map.\n");
499 err = perf_event__synthesize_stat_config(NULL, &stat_config,
500 process_synthesized_event, NULL);
502 pr_err("Couldn't synthesize config.\n");
509 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
511 static int __store_counter_ids(struct perf_evsel *counter,
512 struct cpu_map *cpus,
513 struct thread_map *threads)
517 for (cpu = 0; cpu < cpus->nr; cpu++) {
518 for (thread = 0; thread < threads->nr; thread++) {
519 int fd = FD(counter, cpu, thread);
521 if (perf_evlist__id_add_fd(evsel_list, counter,
522 cpu, thread, fd) < 0)
530 static int store_counter_ids(struct perf_evsel *counter)
532 struct cpu_map *cpus = counter->cpus;
533 struct thread_map *threads = counter->threads;
535 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
538 return __store_counter_ids(counter, cpus, threads);
541 static int __run_perf_stat(int argc, const char **argv)
543 int interval = stat_config.interval;
545 unsigned long long t0, t1;
546 struct perf_evsel *counter;
550 const bool forks = (argc > 0);
551 bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
552 struct perf_evsel_config_term *err_term;
555 ts.tv_sec = interval / USEC_PER_MSEC;
556 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
563 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
564 workload_exec_failed_signal) < 0) {
565 perror("failed to prepare workload");
568 child_pid = evsel_list->workload.pid;
572 perf_evlist__set_leader(evsel_list);
574 evlist__for_each_entry(evsel_list, counter) {
576 if (create_perf_stat_counter(counter) < 0) {
578 * PPC returns ENXIO for HW counters until 2.6.37
579 * (behavior changed with commit b0a873e).
581 if (errno == EINVAL || errno == ENOSYS ||
582 errno == ENOENT || errno == EOPNOTSUPP ||
585 ui__warning("%s event is not supported by the kernel.\n",
586 perf_evsel__name(counter));
587 counter->supported = false;
589 if ((counter->leader != counter) ||
590 !(counter->leader->nr_members > 1))
592 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
594 ui__warning("%s\n", msg);
598 perf_evsel__open_strerror(counter, &target,
599 errno, msg, sizeof(msg));
600 ui__error("%s\n", msg);
603 kill(child_pid, SIGTERM);
607 counter->supported = true;
609 l = strlen(counter->unit);
613 if (STAT_RECORD && store_counter_ids(counter))
617 if (perf_evlist__apply_filters(evsel_list, &counter)) {
618 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
619 counter->filter, perf_evsel__name(counter), errno,
620 str_error_r(errno, msg, sizeof(msg)));
624 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
625 error("failed to set config \"%s\" on event %s with %d (%s)\n",
626 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
627 str_error_r(errno, msg, sizeof(msg)));
632 int err, fd = perf_data_file__fd(&perf_stat.file);
635 err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
637 err = perf_session__write_header(perf_stat.session, evsel_list,
644 err = perf_stat_synthesize_config(is_pipe);
650 * Enable counters and exec the command:
653 clock_gettime(CLOCK_MONOTONIC, &ref_time);
656 perf_evlist__start_workload(evsel_list);
660 while (!waitpid(child_pid, &status, WNOHANG)) {
661 nanosleep(&ts, NULL);
667 if (workload_exec_errno) {
668 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
669 pr_err("Workload failed: %s\n", emsg);
673 if (WIFSIGNALED(status))
674 psignal(WTERMSIG(status), argv[0]);
678 nanosleep(&ts, NULL);
688 update_stats(&walltime_nsecs_stats, t1 - t0);
691 * Closing a group leader splits the group, and as we only disable
692 * group leaders, results in remaining events becoming enabled. To
693 * avoid arbitrary skew, we must read all counters before closing any
697 perf_evlist__close(evsel_list);
699 return WEXITSTATUS(status);
702 static int run_perf_stat(int argc, const char **argv)
707 ret = system(pre_cmd);
715 ret = __run_perf_stat(argc, argv);
720 ret = system(post_cmd);
728 static void print_running(u64 run, u64 ena)
731 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
735 ena ? 100.0 * run / ena : 100.0);
736 } else if (run != ena) {
737 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
741 static void print_noise_pct(double total, double avg)
743 double pct = rel_stddev_stats(total, avg);
746 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
748 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
751 static void print_noise(struct perf_evsel *evsel, double avg)
753 struct perf_stat_evsel *ps;
759 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
762 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
764 switch (stat_config.aggr_mode) {
766 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
767 cpu_map__id_to_socket(id),
769 cpu_map__id_to_cpu(id),
776 fprintf(stat_config.output, "S%*d%s%*d%s",
785 fprintf(stat_config.output, "CPU%*d%s",
787 perf_evsel__cpus(evsel)->map[id], csv_sep);
790 fprintf(stat_config.output, "%*s-%*d%s",
792 thread_map__comm(evsel->threads, id),
794 thread_map__pid(evsel->threads, id),
810 struct perf_evsel *evsel;
813 #define METRIC_LEN 35
815 static void new_line_std(void *ctx)
817 struct outstate *os = ctx;
822 static void do_new_line_std(struct outstate *os)
825 fputs(os->prefix, os->fh);
826 aggr_printout(os->evsel, os->id, os->nr);
827 if (stat_config.aggr_mode == AGGR_NONE)
828 fprintf(os->fh, " ");
829 fprintf(os->fh, " ");
832 static void print_metric_std(void *ctx, const char *color, const char *fmt,
833 const char *unit, double val)
835 struct outstate *os = ctx;
838 bool newline = os->newline;
842 if (unit == NULL || fmt == NULL) {
843 fprintf(out, "%-*s", METRIC_LEN, "");
850 n = fprintf(out, " # ");
852 n += color_fprintf(out, color, fmt, val);
854 n += fprintf(out, fmt, val);
855 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
858 static void new_line_csv(void *ctx)
860 struct outstate *os = ctx;
865 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
866 aggr_printout(os->evsel, os->id, os->nr);
867 for (i = 0; i < os->nfields; i++)
868 fputs(csv_sep, os->fh);
871 static void print_metric_csv(void *ctx,
872 const char *color __maybe_unused,
873 const char *fmt, const char *unit, double val)
875 struct outstate *os = ctx;
877 char buf[64], *vals, *ends;
879 if (unit == NULL || fmt == NULL) {
880 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
883 snprintf(buf, sizeof(buf), fmt, val);
884 ends = vals = ltrim(buf);
885 while (isdigit(*ends) || *ends == '.')
888 while (isspace(*unit))
890 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
893 #define METRIC_ONLY_LEN 20
895 /* Filter out some columns that don't work well in metrics only mode */
897 static bool valid_only_metric(const char *unit)
901 if (strstr(unit, "/sec") ||
902 strstr(unit, "hz") ||
903 strstr(unit, "Hz") ||
904 strstr(unit, "CPUs utilized"))
909 static const char *fixunit(char *buf, struct perf_evsel *evsel,
912 if (!strncmp(unit, "of all", 6)) {
913 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
920 static void print_metric_only(void *ctx, const char *color, const char *fmt,
921 const char *unit, double val)
923 struct outstate *os = ctx;
927 unsigned mlen = METRIC_ONLY_LEN;
929 if (!valid_only_metric(unit))
931 unit = fixunit(buf, os->evsel, unit);
933 n = color_fprintf(out, color, fmt, val);
935 n = fprintf(out, fmt, val);
936 if (n > METRIC_ONLY_LEN)
938 if (mlen < strlen(unit))
939 mlen = strlen(unit) + 1;
940 fprintf(out, "%*s", mlen - n, "");
943 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
945 const char *unit, double val)
947 struct outstate *os = ctx;
949 char buf[64], *vals, *ends;
952 if (!valid_only_metric(unit))
954 unit = fixunit(tbuf, os->evsel, unit);
955 snprintf(buf, sizeof buf, fmt, val);
956 ends = vals = ltrim(buf);
957 while (isdigit(*ends) || *ends == '.')
960 fprintf(out, "%s%s", vals, csv_sep);
963 static void new_line_metric(void *ctx __maybe_unused)
967 static void print_metric_header(void *ctx, const char *color __maybe_unused,
968 const char *fmt __maybe_unused,
969 const char *unit, double val __maybe_unused)
971 struct outstate *os = ctx;
974 if (!valid_only_metric(unit))
976 unit = fixunit(tbuf, os->evsel, unit);
978 fprintf(os->fh, "%s%s", unit, csv_sep);
980 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
983 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
985 FILE *output = stat_config.output;
986 double msecs = avg / NSEC_PER_MSEC;
987 const char *fmt_v, *fmt_n;
990 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
991 fmt_n = csv_output ? "%s" : "%-25s";
993 aggr_printout(evsel, id, nr);
995 scnprintf(name, sizeof(name), "%s%s",
996 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
998 fprintf(output, fmt_v, msecs, csv_sep);
1001 fprintf(output, "%s%s", evsel->unit, csv_sep);
1003 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
1005 fprintf(output, fmt_n, name);
1008 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1011 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1018 if (stat_config.aggr_mode == AGGR_NONE)
1021 if (stat_config.aggr_mode == AGGR_GLOBAL)
1024 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1025 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1027 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1033 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1035 FILE *output = stat_config.output;
1036 double sc = evsel->scale;
1040 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
1043 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1045 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1048 aggr_printout(evsel, id, nr);
1050 fprintf(output, fmt, avg, csv_sep);
1053 fprintf(output, "%-*s%s",
1054 csv_output ? 0 : unit_width,
1055 evsel->unit, csv_sep);
1057 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1060 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1063 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1064 char *prefix, u64 run, u64 ena, double noise)
1066 struct perf_stat_output_ctx out;
1067 struct outstate os = {
1068 .fh = stat_config.output,
1069 .prefix = prefix ? prefix : "",
1074 print_metric_t pm = print_metric_std;
1078 nl = new_line_metric;
1080 pm = print_metric_only_csv;
1082 pm = print_metric_only;
1086 if (csv_output && !metric_only) {
1087 static int aggr_fields[] = {
1095 pm = print_metric_csv;
1098 os.nfields += aggr_fields[stat_config.aggr_mode];
1102 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1104 pm(&os, NULL, "", "", 0);
1107 aggr_printout(counter, id, nr);
1109 fprintf(stat_config.output, "%*s%s",
1110 csv_output ? 0 : 18,
1111 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1114 if (counter->supported)
1115 print_free_counters_hint = 1;
1117 fprintf(stat_config.output, "%-*s%s",
1118 csv_output ? 0 : unit_width,
1119 counter->unit, csv_sep);
1121 fprintf(stat_config.output, "%*s",
1122 csv_output ? 0 : -25,
1123 perf_evsel__name(counter));
1126 fprintf(stat_config.output, "%s%s",
1127 csv_sep, counter->cgrp->name);
1130 pm(&os, NULL, NULL, "", 0);
1131 print_noise(counter, noise);
1132 print_running(run, ena);
1134 pm(&os, NULL, NULL, "", 0);
1140 else if (nsec_counter(counter))
1141 nsec_printout(id, nr, counter, uval);
1143 abs_printout(id, nr, counter, uval);
1145 out.print_metric = pm;
1148 out.force_header = false;
1150 if (csv_output && !metric_only) {
1151 print_noise(counter, noise);
1152 print_running(run, ena);
1155 perf_stat__print_shadow_stats(counter, uval,
1156 first_shadow_cpu(counter, id),
1158 if (!csv_output && !metric_only) {
1159 print_noise(counter, noise);
1160 print_running(run, ena);
1164 static void aggr_update_shadow(void)
1168 struct perf_evsel *counter;
1170 for (s = 0; s < aggr_map->nr; s++) {
1171 id = aggr_map->map[s];
1172 evlist__for_each_entry(evsel_list, counter) {
1174 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1175 s2 = aggr_get_id(evsel_list->cpus, cpu);
1178 val += perf_counts(counter->counts, cpu, 0)->val;
1180 val = val * counter->scale;
1181 perf_stat__update_shadow_stats(counter, &val,
1182 first_shadow_cpu(counter, id));
1187 static void collect_all_aliases(struct perf_evsel *counter,
1188 void (*cb)(struct perf_evsel *counter, void *data,
1192 struct perf_evsel *alias;
1194 alias = list_prepare_entry(counter, &(evsel_list->entries), node);
1195 list_for_each_entry_continue (alias, &evsel_list->entries, node) {
1196 if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) ||
1197 alias->scale != counter->scale ||
1198 alias->cgrp != counter->cgrp ||
1199 strcmp(alias->unit, counter->unit) ||
1200 nsec_counter(alias) != nsec_counter(counter))
1202 alias->merged_stat = true;
1203 cb(alias, data, false);
1207 static bool collect_data(struct perf_evsel *counter,
1208 void (*cb)(struct perf_evsel *counter, void *data,
1212 if (counter->merged_stat)
1214 cb(counter, data, true);
1216 collect_all_aliases(counter, cb, data);
1227 static void aggr_cb(struct perf_evsel *counter, void *data, bool first)
1229 struct aggr_data *ad = data;
1232 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1233 struct perf_counts_values *counts;
1235 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1240 counts = perf_counts(counter->counts, cpu, 0);
1242 * When any result is bad, make them all to give
1243 * consistent output in interval mode.
1245 if (counts->ena == 0 || counts->run == 0 ||
1246 counter->counts->scaled == -1) {
1251 ad->val += counts->val;
1252 ad->ena += counts->ena;
1253 ad->run += counts->run;
1257 static void print_aggr(char *prefix)
1259 FILE *output = stat_config.output;
1260 struct perf_evsel *counter;
1266 if (!(aggr_map || aggr_get_id))
1269 aggr_update_shadow();
1272 * With metric_only everything is on a single line.
1273 * Without each counter has its own line.
1275 for (s = 0; s < aggr_map->nr; s++) {
1276 struct aggr_data ad;
1277 if (prefix && metric_only)
1278 fprintf(output, "%s", prefix);
1280 ad.id = id = aggr_map->map[s];
1282 evlist__for_each_entry(evsel_list, counter) {
1283 ad.val = ad.ena = ad.run = 0;
1285 if (!collect_data(counter, aggr_cb, &ad))
1291 if (first && metric_only) {
1293 aggr_printout(counter, id, nr);
1295 if (prefix && !metric_only)
1296 fprintf(output, "%s", prefix);
1298 uval = val * counter->scale;
1299 printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1301 fputc('\n', output);
1304 fputc('\n', output);
1308 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1310 FILE *output = stat_config.output;
1311 int nthreads = thread_map__nr(counter->threads);
1312 int ncpus = cpu_map__nr(counter->cpus);
1316 for (thread = 0; thread < nthreads; thread++) {
1317 u64 ena = 0, run = 0, val = 0;
1319 for (cpu = 0; cpu < ncpus; cpu++) {
1320 val += perf_counts(counter->counts, cpu, thread)->val;
1321 ena += perf_counts(counter->counts, cpu, thread)->ena;
1322 run += perf_counts(counter->counts, cpu, thread)->run;
1326 fprintf(output, "%s", prefix);
1328 uval = val * counter->scale;
1329 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1330 fputc('\n', output);
1335 double avg, avg_enabled, avg_running;
1338 static void counter_aggr_cb(struct perf_evsel *counter, void *data,
1339 bool first __maybe_unused)
1341 struct caggr_data *cd = data;
1342 struct perf_stat_evsel *ps = counter->priv;
1344 cd->avg += avg_stats(&ps->res_stats[0]);
1345 cd->avg_enabled += avg_stats(&ps->res_stats[1]);
1346 cd->avg_running += avg_stats(&ps->res_stats[2]);
1350 * Print out the results of a single counter:
1351 * aggregated counts in system-wide mode
1353 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1355 FILE *output = stat_config.output;
1357 struct caggr_data cd = { .avg = 0.0 };
1359 if (!collect_data(counter, counter_aggr_cb, &cd))
1362 if (prefix && !metric_only)
1363 fprintf(output, "%s", prefix);
1365 uval = cd.avg * counter->scale;
1366 printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled, cd.avg);
1368 fprintf(output, "\n");
1371 static void counter_cb(struct perf_evsel *counter, void *data,
1372 bool first __maybe_unused)
1374 struct aggr_data *ad = data;
1376 ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
1377 ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
1378 ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
1382 * Print out the results of a single counter:
1383 * does not use aggregated count in system-wide
1385 static void print_counter(struct perf_evsel *counter, char *prefix)
1387 FILE *output = stat_config.output;
1392 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1393 struct aggr_data ad = { .cpu = cpu };
1395 if (!collect_data(counter, counter_cb, &ad))
1402 fprintf(output, "%s", prefix);
1404 uval = val * counter->scale;
1405 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1407 fputc('\n', output);
1411 static void print_no_aggr_metric(char *prefix)
1415 struct perf_evsel *counter;
1419 nrcpus = evsel_list->cpus->nr;
1420 for (cpu = 0; cpu < nrcpus; cpu++) {
1424 fputs(prefix, stat_config.output);
1425 evlist__for_each_entry(evsel_list, counter) {
1427 aggr_printout(counter, cpu, 0);
1430 val = perf_counts(counter->counts, cpu, 0)->val;
1431 ena = perf_counts(counter->counts, cpu, 0)->ena;
1432 run = perf_counts(counter->counts, cpu, 0)->run;
1434 uval = val * counter->scale;
1435 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1437 fputc('\n', stat_config.output);
1441 static int aggr_header_lens[] = {
1449 static const char *aggr_header_csv[] = {
1450 [AGGR_CORE] = "core,cpus,",
1451 [AGGR_SOCKET] = "socket,cpus",
1452 [AGGR_NONE] = "cpu,",
1453 [AGGR_THREAD] = "comm-pid,",
1457 static void print_metric_headers(const char *prefix, bool no_indent)
1459 struct perf_stat_output_ctx out;
1460 struct perf_evsel *counter;
1461 struct outstate os = {
1462 .fh = stat_config.output
1466 fprintf(stat_config.output, "%s", prefix);
1468 if (!csv_output && !no_indent)
1469 fprintf(stat_config.output, "%*s",
1470 aggr_header_lens[stat_config.aggr_mode], "");
1472 if (stat_config.interval)
1473 fputs("time,", stat_config.output);
1474 fputs(aggr_header_csv[stat_config.aggr_mode],
1475 stat_config.output);
1478 /* Print metrics headers only */
1479 evlist__for_each_entry(evsel_list, counter) {
1482 out.print_metric = print_metric_header;
1483 out.new_line = new_line_metric;
1484 out.force_header = true;
1486 perf_stat__print_shadow_stats(counter, 0,
1490 fputc('\n', stat_config.output);
1493 static void print_interval(char *prefix, struct timespec *ts)
1495 FILE *output = stat_config.output;
1496 static int num_print_interval;
1498 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1500 if (num_print_interval == 0 && !csv_output) {
1501 switch (stat_config.aggr_mode) {
1503 fprintf(output, "# time socket cpus");
1505 fprintf(output, " counts %*s events\n", unit_width, "unit");
1508 fprintf(output, "# time core cpus");
1510 fprintf(output, " counts %*s events\n", unit_width, "unit");
1513 fprintf(output, "# time CPU");
1515 fprintf(output, " counts %*s events\n", unit_width, "unit");
1518 fprintf(output, "# time comm-pid");
1520 fprintf(output, " counts %*s events\n", unit_width, "unit");
1524 fprintf(output, "# time");
1526 fprintf(output, " counts %*s events\n", unit_width, "unit");
1532 if (num_print_interval == 0 && metric_only)
1533 print_metric_headers(" ", true);
1534 if (++num_print_interval == 25)
1535 num_print_interval = 0;
1538 static void print_header(int argc, const char **argv)
1540 FILE *output = stat_config.output;
1546 fprintf(output, "\n");
1547 fprintf(output, " Performance counter stats for ");
1548 if (target.system_wide)
1549 fprintf(output, "\'system wide");
1550 else if (target.cpu_list)
1551 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1552 else if (!target__has_task(&target)) {
1553 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1554 for (i = 1; argv && (i < argc); i++)
1555 fprintf(output, " %s", argv[i]);
1556 } else if (target.pid)
1557 fprintf(output, "process id \'%s", target.pid);
1559 fprintf(output, "thread id \'%s", target.tid);
1561 fprintf(output, "\'");
1563 fprintf(output, " (%d runs)", run_count);
1564 fprintf(output, ":\n\n");
1568 static void print_footer(void)
1570 FILE *output = stat_config.output;
1573 fprintf(output, "\n");
1574 fprintf(output, " %17.9f seconds time elapsed",
1575 avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC);
1576 if (run_count > 1) {
1577 fprintf(output, " ");
1578 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1579 avg_stats(&walltime_nsecs_stats));
1581 fprintf(output, "\n\n");
1583 if (print_free_counters_hint)
1585 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1586 " echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1588 " echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1591 static void print_counters(struct timespec *ts, int argc, const char **argv)
1593 int interval = stat_config.interval;
1594 struct perf_evsel *counter;
1595 char buf[64], *prefix = NULL;
1597 /* Do not print anything if we record to the pipe. */
1598 if (STAT_RECORD && perf_stat.file.is_pipe)
1602 print_interval(prefix = buf, ts);
1604 print_header(argc, argv);
1607 static int num_print_iv;
1609 if (num_print_iv == 0 && !interval)
1610 print_metric_headers(prefix, false);
1611 if (num_print_iv++ == 25)
1613 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1614 fprintf(stat_config.output, "%s", prefix);
1617 switch (stat_config.aggr_mode) {
1623 evlist__for_each_entry(evsel_list, counter)
1624 print_aggr_thread(counter, prefix);
1627 evlist__for_each_entry(evsel_list, counter)
1628 print_counter_aggr(counter, prefix);
1630 fputc('\n', stat_config.output);
1634 print_no_aggr_metric(prefix);
1636 evlist__for_each_entry(evsel_list, counter)
1637 print_counter(counter, prefix);
1645 if (!interval && !csv_output)
1648 fflush(stat_config.output);
1651 static volatile int signr = -1;
1653 static void skip_signal(int signo)
1655 if ((child_pid == -1) || stat_config.interval)
1660 * render child_pid harmless
1661 * won't send SIGTERM to a random
1662 * process in case of race condition
1663 * and fast PID recycling
1668 static void sig_atexit(void)
1673 * avoid race condition with SIGCHLD handler
1674 * in skip_signal() which is modifying child_pid
1675 * goal is to avoid send SIGTERM to a random
1679 sigaddset(&set, SIGCHLD);
1680 sigprocmask(SIG_BLOCK, &set, &oset);
1682 if (child_pid != -1)
1683 kill(child_pid, SIGTERM);
1685 sigprocmask(SIG_SETMASK, &oset, NULL);
1690 signal(signr, SIG_DFL);
1691 kill(getpid(), signr);
1694 static int stat__set_big_num(const struct option *opt __maybe_unused,
1695 const char *s __maybe_unused, int unset)
1697 big_num_opt = unset ? 0 : 1;
1701 static int enable_metric_only(const struct option *opt __maybe_unused,
1702 const char *s __maybe_unused, int unset)
1704 force_metric_only = true;
1705 metric_only = !unset;
1709 static const struct option stat_options[] = {
1710 OPT_BOOLEAN('T', "transaction", &transaction_run,
1711 "hardware transaction statistics"),
1712 OPT_CALLBACK('e', "event", &evsel_list, "event",
1713 "event selector. use 'perf list' to list available events",
1714 parse_events_option),
1715 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1716 "event filter", parse_filter),
1717 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1718 "child tasks do not inherit counters"),
1719 OPT_STRING('p', "pid", &target.pid, "pid",
1720 "stat events on existing process id"),
1721 OPT_STRING('t', "tid", &target.tid, "tid",
1722 "stat events on existing thread id"),
1723 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1724 "system-wide collection from all CPUs"),
1725 OPT_BOOLEAN('g', "group", &group,
1726 "put the counters into a counter group"),
1727 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1728 OPT_INCR('v', "verbose", &verbose,
1729 "be more verbose (show counter open errors, etc)"),
1730 OPT_INTEGER('r', "repeat", &run_count,
1731 "repeat command and print average + stddev (max: 100, forever: 0)"),
1732 OPT_BOOLEAN('n', "null", &null_run,
1733 "null run - dont start any counters"),
1734 OPT_INCR('d', "detailed", &detailed_run,
1735 "detailed run - start a lot of events"),
1736 OPT_BOOLEAN('S', "sync", &sync_run,
1737 "call sync() before starting a run"),
1738 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1739 "print large numbers with thousands\' separators",
1741 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1742 "list of cpus to monitor in system-wide"),
1743 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1744 "disable CPU count aggregation", AGGR_NONE),
1745 OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"),
1746 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1747 "print counts with custom separator"),
1748 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1749 "monitor event in cgroup name only", parse_cgroups),
1750 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1751 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1752 OPT_INTEGER(0, "log-fd", &output_fd,
1753 "log output to fd, instead of stderr"),
1754 OPT_STRING(0, "pre", &pre_cmd, "command",
1755 "command to run prior to the measured command"),
1756 OPT_STRING(0, "post", &post_cmd, "command",
1757 "command to run after to the measured command"),
1758 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1759 "print counts at regular interval in ms (>= 10)"),
1760 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1761 "aggregate counts per processor socket", AGGR_SOCKET),
1762 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1763 "aggregate counts per physical processor core", AGGR_CORE),
1764 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1765 "aggregate counts per thread", AGGR_THREAD),
1766 OPT_UINTEGER('D', "delay", &initial_delay,
1767 "ms to wait before starting measurement after program start"),
1768 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1769 "Only print computed metrics. No raw values", enable_metric_only),
1770 OPT_BOOLEAN(0, "topdown", &topdown_run,
1771 "measure topdown level 1 statistics"),
1775 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1777 return cpu_map__get_socket(map, cpu, NULL);
1780 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1782 return cpu_map__get_core(map, cpu, NULL);
1785 static int cpu_map__get_max(struct cpu_map *map)
1789 for (i = 0; i < map->nr; i++) {
1790 if (map->map[i] > max)
1797 static struct cpu_map *cpus_aggr_map;
1799 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1806 cpu = map->map[idx];
1808 if (cpus_aggr_map->map[cpu] == -1)
1809 cpus_aggr_map->map[cpu] = get_id(map, idx);
1811 return cpus_aggr_map->map[cpu];
1814 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1816 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1819 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1821 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1824 static int perf_stat_init_aggr_mode(void)
1828 switch (stat_config.aggr_mode) {
1830 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1831 perror("cannot build socket map");
1834 aggr_get_id = perf_stat__get_socket_cached;
1837 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1838 perror("cannot build core map");
1841 aggr_get_id = perf_stat__get_core_cached;
1852 * The evsel_list->cpus is the base we operate on,
1853 * taking the highest cpu number to be the size of
1854 * the aggregation translate cpumap.
1856 nr = cpu_map__get_max(evsel_list->cpus);
1857 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1858 return cpus_aggr_map ? 0 : -ENOMEM;
1861 static void perf_stat__exit_aggr_mode(void)
1863 cpu_map__put(aggr_map);
1864 cpu_map__put(cpus_aggr_map);
1866 cpus_aggr_map = NULL;
1869 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1876 cpu = map->map[idx];
1878 if (cpu >= env->nr_cpus_avail)
1884 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1886 struct perf_env *env = data;
1887 int cpu = perf_env__get_cpu(env, map, idx);
1889 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1892 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1894 struct perf_env *env = data;
1895 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1898 int socket_id = env->cpu[cpu].socket_id;
1901 * Encode socket in upper 16 bits
1902 * core_id is relative to socket, and
1903 * we need a global id. So we combine
1906 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1912 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1913 struct cpu_map **sockp)
1915 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1918 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1919 struct cpu_map **corep)
1921 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1924 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1926 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1929 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1931 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1934 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1936 struct perf_env *env = &st->session->header.env;
1938 switch (stat_config.aggr_mode) {
1940 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1941 perror("cannot build socket map");
1944 aggr_get_id = perf_stat__get_socket_file;
1947 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1948 perror("cannot build core map");
1951 aggr_get_id = perf_stat__get_core_file;
1964 static int topdown_filter_events(const char **attr, char **str, bool use_group)
1971 for (i = 0; attr[i]; i++) {
1972 if (pmu_have_event("cpu", attr[i])) {
1973 len += strlen(attr[i]) + 1;
1974 attr[i - off] = attr[i];
1978 attr[i - off] = NULL;
1980 *str = malloc(len + 1 + 2);
1990 for (i = 0; attr[i]; i++) {
2003 __weak bool arch_topdown_check_group(bool *warn)
2009 __weak void arch_topdown_group_warn(void)
2014 * Add default attributes, if there were no attributes specified or
2015 * if -d/--detailed, -d -d or -d -d -d is used:
2017 static int add_default_attributes(void)
2020 struct perf_event_attr default_attrs0[] = {
2022 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
2023 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
2024 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
2025 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
2027 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
2029 struct perf_event_attr frontend_attrs[] = {
2030 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
2032 struct perf_event_attr backend_attrs[] = {
2033 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
2035 struct perf_event_attr default_attrs1[] = {
2036 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
2037 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
2038 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
2043 * Detailed stats (-d), covering the L1 and last level data caches:
2045 struct perf_event_attr detailed_attrs[] = {
2047 { .type = PERF_TYPE_HW_CACHE,
2049 PERF_COUNT_HW_CACHE_L1D << 0 |
2050 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2051 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2053 { .type = PERF_TYPE_HW_CACHE,
2055 PERF_COUNT_HW_CACHE_L1D << 0 |
2056 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2057 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2059 { .type = PERF_TYPE_HW_CACHE,
2061 PERF_COUNT_HW_CACHE_LL << 0 |
2062 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2063 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2065 { .type = PERF_TYPE_HW_CACHE,
2067 PERF_COUNT_HW_CACHE_LL << 0 |
2068 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2069 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2073 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2075 struct perf_event_attr very_detailed_attrs[] = {
2077 { .type = PERF_TYPE_HW_CACHE,
2079 PERF_COUNT_HW_CACHE_L1I << 0 |
2080 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2081 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2083 { .type = PERF_TYPE_HW_CACHE,
2085 PERF_COUNT_HW_CACHE_L1I << 0 |
2086 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2087 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2089 { .type = PERF_TYPE_HW_CACHE,
2091 PERF_COUNT_HW_CACHE_DTLB << 0 |
2092 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2093 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2095 { .type = PERF_TYPE_HW_CACHE,
2097 PERF_COUNT_HW_CACHE_DTLB << 0 |
2098 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2099 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2101 { .type = PERF_TYPE_HW_CACHE,
2103 PERF_COUNT_HW_CACHE_ITLB << 0 |
2104 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2105 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2107 { .type = PERF_TYPE_HW_CACHE,
2109 PERF_COUNT_HW_CACHE_ITLB << 0 |
2110 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2111 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2116 * Very, very detailed stats (-d -d -d), adding prefetch events:
2118 struct perf_event_attr very_very_detailed_attrs[] = {
2120 { .type = PERF_TYPE_HW_CACHE,
2122 PERF_COUNT_HW_CACHE_L1D << 0 |
2123 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2124 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2126 { .type = PERF_TYPE_HW_CACHE,
2128 PERF_COUNT_HW_CACHE_L1D << 0 |
2129 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2130 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2133 /* Set attrs if no event is selected and !null_run: */
2137 if (transaction_run) {
2138 if (pmu_have_event("cpu", "cycles-ct") &&
2139 pmu_have_event("cpu", "el-start"))
2140 err = parse_events(evsel_list, transaction_attrs, NULL);
2142 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
2144 fprintf(stderr, "Cannot set up transaction events\n");
2154 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2155 stat_config.aggr_mode != AGGR_CORE) {
2156 pr_err("top down event configuration requires --per-core mode\n");
2159 stat_config.aggr_mode = AGGR_CORE;
2160 if (nr_cgroups || !target__has_cpu(&target)) {
2161 pr_err("top down event configuration requires system-wide mode (-a)\n");
2165 if (!force_metric_only)
2167 if (topdown_filter_events(topdown_attrs, &str,
2168 arch_topdown_check_group(&warn)) < 0) {
2169 pr_err("Out of memory\n");
2172 if (topdown_attrs[0] && str) {
2174 arch_topdown_group_warn();
2175 err = parse_events(evsel_list, str, NULL);
2178 "Cannot set up top down events %s: %d\n",
2184 fprintf(stderr, "System does not support topdown\n");
2190 if (!evsel_list->nr_entries) {
2191 if (target__has_cpu(&target))
2192 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2194 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2196 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2197 if (perf_evlist__add_default_attrs(evsel_list,
2198 frontend_attrs) < 0)
2201 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2202 if (perf_evlist__add_default_attrs(evsel_list,
2206 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2210 /* Detailed events get appended to the event list: */
2212 if (detailed_run < 1)
2215 /* Append detailed run extra attributes: */
2216 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2219 if (detailed_run < 2)
2222 /* Append very detailed run extra attributes: */
2223 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2226 if (detailed_run < 3)
2229 /* Append very, very detailed run extra attributes: */
2230 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2233 static const char * const stat_record_usage[] = {
2234 "perf stat record [<options>]",
2238 static void init_features(struct perf_session *session)
2242 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2243 perf_header__set_feat(&session->header, feat);
2245 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2246 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2247 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2248 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2251 static int __cmd_record(int argc, const char **argv)
2253 struct perf_session *session;
2254 struct perf_data_file *file = &perf_stat.file;
2256 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2257 PARSE_OPT_STOP_AT_NON_OPTION);
2260 file->path = output_name;
2262 if (run_count != 1 || forever) {
2263 pr_err("Cannot use -r option with perf stat record.\n");
2267 session = perf_session__new(file, false, NULL);
2268 if (session == NULL) {
2269 pr_err("Perf session creation failed.\n");
2273 init_features(session);
2275 session->evlist = evsel_list;
2276 perf_stat.session = session;
2277 perf_stat.record = true;
2281 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2282 union perf_event *event,
2283 struct perf_session *session)
2285 struct stat_round_event *stat_round = &event->stat_round;
2286 struct perf_evsel *counter;
2287 struct timespec tsh, *ts = NULL;
2288 const char **argv = session->header.env.cmdline_argv;
2289 int argc = session->header.env.nr_cmdline;
2291 evlist__for_each_entry(evsel_list, counter)
2292 perf_stat_process_counter(&stat_config, counter);
2294 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2295 update_stats(&walltime_nsecs_stats, stat_round->time);
2297 if (stat_config.interval && stat_round->time) {
2298 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2299 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2303 print_counters(ts, argc, argv);
2308 int process_stat_config_event(struct perf_tool *tool,
2309 union perf_event *event,
2310 struct perf_session *session __maybe_unused)
2312 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2314 perf_event__read_stat_config(&stat_config, &event->stat_config);
2316 if (cpu_map__empty(st->cpus)) {
2317 if (st->aggr_mode != AGGR_UNSET)
2318 pr_warning("warning: processing task data, aggregation mode not set\n");
2322 if (st->aggr_mode != AGGR_UNSET)
2323 stat_config.aggr_mode = st->aggr_mode;
2325 if (perf_stat.file.is_pipe)
2326 perf_stat_init_aggr_mode();
2328 perf_stat_init_aggr_mode_file(st);
2333 static int set_maps(struct perf_stat *st)
2335 if (!st->cpus || !st->threads)
2338 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2341 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2343 if (perf_evlist__alloc_stats(evsel_list, true))
2346 st->maps_allocated = true;
2351 int process_thread_map_event(struct perf_tool *tool,
2352 union perf_event *event,
2353 struct perf_session *session __maybe_unused)
2355 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2358 pr_warning("Extra thread map event, ignoring.\n");
2362 st->threads = thread_map__new_event(&event->thread_map);
2366 return set_maps(st);
2370 int process_cpu_map_event(struct perf_tool *tool,
2371 union perf_event *event,
2372 struct perf_session *session __maybe_unused)
2374 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2375 struct cpu_map *cpus;
2378 pr_warning("Extra cpu map event, ignoring.\n");
2382 cpus = cpu_map__new_data(&event->cpu_map.data);
2387 return set_maps(st);
2390 static const char * const stat_report_usage[] = {
2391 "perf stat report [<options>]",
2395 static struct perf_stat perf_stat = {
2397 .attr = perf_event__process_attr,
2398 .event_update = perf_event__process_event_update,
2399 .thread_map = process_thread_map_event,
2400 .cpu_map = process_cpu_map_event,
2401 .stat_config = process_stat_config_event,
2402 .stat = perf_event__process_stat_event,
2403 .stat_round = process_stat_round_event,
2405 .aggr_mode = AGGR_UNSET,
2408 static int __cmd_report(int argc, const char **argv)
2410 struct perf_session *session;
2411 const struct option options[] = {
2412 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2413 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2414 "aggregate counts per processor socket", AGGR_SOCKET),
2415 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2416 "aggregate counts per physical processor core", AGGR_CORE),
2417 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2418 "disable CPU count aggregation", AGGR_NONE),
2424 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2426 if (!input_name || !strlen(input_name)) {
2427 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2430 input_name = "perf.data";
2433 perf_stat.file.path = input_name;
2434 perf_stat.file.mode = PERF_DATA_MODE_READ;
2436 session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
2437 if (session == NULL)
2440 perf_stat.session = session;
2441 stat_config.output = stderr;
2442 evsel_list = session->evlist;
2444 ret = perf_session__process_events(session);
2448 perf_session__delete(session);
2452 static void setup_system_wide(int forks)
2455 * Make system wide (-a) the default target if
2456 * no target was specified and one of following
2457 * conditions is met:
2459 * - there's no workload specified
2460 * - there is workload specified but all requested
2461 * events are system wide events
2463 if (!target__none(&target))
2467 target.system_wide = true;
2469 struct perf_evsel *counter;
2471 evlist__for_each_entry(evsel_list, counter) {
2472 if (!counter->system_wide)
2476 if (evsel_list->nr_entries)
2477 target.system_wide = true;
2481 int cmd_stat(int argc, const char **argv)
2483 const char * const stat_usage[] = {
2484 "perf stat [<options>] [<command>]",
2487 int status = -EINVAL, run_idx;
2489 FILE *output = stderr;
2490 unsigned int interval;
2491 const char * const stat_subcommands[] = { "record", "report" };
2493 setlocale(LC_ALL, "");
2495 evsel_list = perf_evlist__new();
2496 if (evsel_list == NULL)
2499 parse_events__shrink_config_terms();
2500 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2501 (const char **) stat_usage,
2502 PARSE_OPT_STOP_AT_NON_OPTION);
2503 perf_stat__collect_metric_expr(evsel_list);
2504 perf_stat__init_shadow_stats();
2508 if (!strcmp(csv_sep, "\\t"))
2511 csv_sep = DEFAULT_SEPARATOR;
2513 if (argc && !strncmp(argv[0], "rec", 3)) {
2514 argc = __cmd_record(argc, argv);
2517 } else if (argc && !strncmp(argv[0], "rep", 3))
2518 return __cmd_report(argc, argv);
2520 interval = stat_config.interval;
2523 * For record command the -o is already taken care of.
2525 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2528 if (output_name && output_fd) {
2529 fprintf(stderr, "cannot use both --output and --log-fd\n");
2530 parse_options_usage(stat_usage, stat_options, "o", 1);
2531 parse_options_usage(NULL, stat_options, "log-fd", 0);
2535 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2536 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2540 if (metric_only && run_count > 1) {
2541 fprintf(stderr, "--metric-only is not supported with -r\n");
2545 if (output_fd < 0) {
2546 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2547 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2553 mode = append_file ? "a" : "w";
2555 output = fopen(output_name, mode);
2557 perror("failed to create output file");
2560 clock_gettime(CLOCK_REALTIME, &tm);
2561 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2562 } else if (output_fd > 0) {
2563 mode = append_file ? "a" : "w";
2564 output = fdopen(output_fd, mode);
2566 perror("Failed opening logfd");
2571 stat_config.output = output;
2574 * let the spreadsheet do the pretty-printing
2577 /* User explicitly passed -B? */
2578 if (big_num_opt == 1) {
2579 fprintf(stderr, "-B option not supported with -x\n");
2580 parse_options_usage(stat_usage, stat_options, "B", 1);
2581 parse_options_usage(NULL, stat_options, "x", 1);
2583 } else /* Nope, so disable big number formatting */
2585 } else if (big_num_opt == 0) /* User passed --no-big-num */
2588 setup_system_wide(argc);
2590 if (run_count < 0) {
2591 pr_err("Run count must be a positive number\n");
2592 parse_options_usage(stat_usage, stat_options, "r", 1);
2594 } else if (run_count == 0) {
2599 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2600 fprintf(stderr, "The --per-thread option is only available "
2601 "when monitoring via -p -t options.\n");
2602 parse_options_usage(NULL, stat_options, "p", 1);
2603 parse_options_usage(NULL, stat_options, "t", 1);
2608 * no_aggr, cgroup are for system-wide only
2609 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2611 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2612 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2613 !target__has_cpu(&target)) {
2614 fprintf(stderr, "both cgroup and no-aggregation "
2615 "modes only available in system-wide mode\n");
2617 parse_options_usage(stat_usage, stat_options, "G", 1);
2618 parse_options_usage(NULL, stat_options, "A", 1);
2619 parse_options_usage(NULL, stat_options, "a", 1);
2623 if (add_default_attributes())
2626 target__validate(&target);
2628 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2629 if (target__has_task(&target)) {
2630 pr_err("Problems finding threads of monitor\n");
2631 parse_options_usage(stat_usage, stat_options, "p", 1);
2632 parse_options_usage(NULL, stat_options, "t", 1);
2633 } else if (target__has_cpu(&target)) {
2634 perror("failed to parse CPUs map");
2635 parse_options_usage(stat_usage, stat_options, "C", 1);
2636 parse_options_usage(NULL, stat_options, "a", 1);
2642 * Initialize thread_map with comm names,
2643 * so we could print it out on output.
2645 if (stat_config.aggr_mode == AGGR_THREAD)
2646 thread_map__read_comms(evsel_list->threads);
2648 if (interval && interval < 100) {
2649 if (interval < 10) {
2650 pr_err("print interval must be >= 10ms\n");
2651 parse_options_usage(stat_usage, stat_options, "I", 1);
2654 pr_warning("print interval < 100ms. "
2655 "The overhead percentage could be high in some cases. "
2656 "Please proceed with caution.\n");
2659 if (perf_evlist__alloc_stats(evsel_list, interval))
2662 if (perf_stat_init_aggr_mode())
2666 * We dont want to block the signals - that would cause
2667 * child tasks to inherit that and Ctrl-C would not work.
2668 * What we want is for Ctrl-C to work in the exec()-ed
2669 * task, but being ignored by perf stat itself:
2673 signal(SIGINT, skip_signal);
2674 signal(SIGCHLD, skip_signal);
2675 signal(SIGALRM, skip_signal);
2676 signal(SIGABRT, skip_signal);
2679 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2680 if (run_count != 1 && verbose > 0)
2681 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2684 status = run_perf_stat(argc, argv);
2685 if (forever && status != -1) {
2686 print_counters(NULL, argc, argv);
2687 perf_stat__reset_stats();
2691 if (!forever && status != -1 && !interval)
2692 print_counters(NULL, argc, argv);
2696 * We synthesize the kernel mmap record just so that older tools
2697 * don't emit warnings about not being able to resolve symbols
2698 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2699 * a saner message about no samples being in the perf.data file.
2701 * This also serves to suppress a warning about f_header.data.size == 0
2702 * in header.c at the moment 'perf stat record' gets introduced, which
2703 * is not really needed once we start adding the stat specific PERF_RECORD_
2704 * records, but the need to suppress the kptr_restrict messages in older
2705 * tools remain -acme
2707 int fd = perf_data_file__fd(&perf_stat.file);
2708 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2709 process_synthesized_event,
2710 &perf_stat.session->machines.host);
2712 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2713 "older tools may produce warnings about this file\n.");
2717 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2718 pr_err("failed to write stat round event\n");
2721 if (!perf_stat.file.is_pipe) {
2722 perf_stat.session->header.data_size += perf_stat.bytes_written;
2723 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2726 perf_session__delete(perf_stat.session);
2729 perf_stat__exit_aggr_mode();
2730 perf_evlist__free_stats(evsel_list);
2732 perf_evlist__delete(evsel_list);