4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ./hackbench 10
13 Performance counter stats for './hackbench 10':
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
27 0.154822978 seconds time elapsed
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 * Improvements and fixes by:
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
41 * Released under the GPL v2. (and only v2, not any later version)
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include <subcmd/parse-options.h>
49 #include "util/parse-events.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/color.h"
56 #include "util/stat.h"
57 #include "util/header.h"
58 #include "util/cpumap.h"
59 #include "util/thread.h"
60 #include "util/thread_map.h"
61 #include "util/counts.h"
62 #include "util/group.h"
63 #include "util/session.h"
64 #include "util/tool.h"
65 #include "util/group.h"
68 #include <api/fs/fs.h>
70 #include <sys/prctl.h>
74 #define DEFAULT_SEPARATOR " "
75 #define CNTR_NOT_SUPPORTED "<not supported>"
76 #define CNTR_NOT_COUNTED "<not counted>"
78 static void print_counters(struct timespec *ts, int argc, const char **argv);
80 /* Default events used for perf stat -T */
81 static const char *transaction_attrs = {
93 /* More limited version when the CPU does not have all events. */
94 static const char * transaction_limited_attrs = {
104 static const char * topdown_attrs[] = {
105 "topdown-total-slots",
106 "topdown-slots-retired",
107 "topdown-recovery-bubbles",
108 "topdown-fetch-bubbles",
109 "topdown-slots-issued",
113 static struct perf_evlist *evsel_list;
115 static struct target target = {
119 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
121 static int run_count = 1;
122 static bool no_inherit = false;
123 static volatile pid_t child_pid = -1;
124 static bool null_run = false;
125 static int detailed_run = 0;
126 static bool transaction_run;
127 static bool topdown_run = false;
128 static bool big_num = true;
129 static int big_num_opt = -1;
130 static const char *csv_sep = NULL;
131 static bool csv_output = false;
132 static bool group = false;
133 static const char *pre_cmd = NULL;
134 static const char *post_cmd = NULL;
135 static bool sync_run = false;
136 static unsigned int initial_delay = 0;
137 static unsigned int unit_width = 4; /* strlen("unit") */
138 static bool forever = false;
139 static bool metric_only = false;
140 static bool force_metric_only = false;
141 static struct timespec ref_time;
142 static struct cpu_map *aggr_map;
143 static aggr_get_id_t aggr_get_id;
144 static bool append_file;
145 static const char *output_name;
146 static int output_fd;
150 struct perf_data_file file;
151 struct perf_session *session;
153 struct perf_tool tool;
155 struct cpu_map *cpus;
156 struct thread_map *threads;
157 enum aggr_mode aggr_mode;
160 static struct perf_stat perf_stat;
161 #define STAT_RECORD perf_stat.record
163 static volatile int done = 0;
165 static struct perf_stat_config stat_config = {
166 .aggr_mode = AGGR_GLOBAL,
170 static inline void diff_timespec(struct timespec *r, struct timespec *a,
173 r->tv_sec = a->tv_sec - b->tv_sec;
174 if (a->tv_nsec < b->tv_nsec) {
175 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
178 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
182 static void perf_stat__reset_stats(void)
184 perf_evlist__reset_stats(evsel_list);
185 perf_stat__reset_shadow_stats();
188 static int create_perf_stat_counter(struct perf_evsel *evsel)
190 struct perf_event_attr *attr = &evsel->attr;
192 if (stat_config.scale)
193 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
194 PERF_FORMAT_TOTAL_TIME_RUNNING;
196 attr->inherit = !no_inherit;
199 * Some events get initialized with sample_(period/type) set,
200 * like tracepoints. Clear it up for counting.
202 attr->sample_period = 0;
205 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
206 * while avoiding that older tools show confusing messages.
208 * However for pipe sessions we need to keep it zero,
209 * because script's perf_evsel__check_attr is triggered
210 * by attr->sample_type != 0, and we can't run it on
213 if (!(STAT_RECORD && perf_stat.file.is_pipe))
214 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
217 * Disabling all counters initially, they will be enabled
218 * either manually by us or by kernel via enable_on_exec
221 if (perf_evsel__is_group_leader(evsel)) {
225 * In case of initial_delay we enable tracee
228 if (target__none(&target) && !initial_delay)
229 attr->enable_on_exec = 1;
232 if (target__has_cpu(&target))
233 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
235 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
239 * Does the counter have nsecs as a unit?
241 static inline int nsec_counter(struct perf_evsel *evsel)
243 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
244 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
250 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
251 union perf_event *event,
252 struct perf_sample *sample __maybe_unused,
253 struct machine *machine __maybe_unused)
255 if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
256 pr_err("failed to write perf data, error: %m\n");
260 perf_stat.bytes_written += event->header.size;
264 static int write_stat_round_event(u64 tm, u64 type)
266 return perf_event__synthesize_stat_round(NULL, tm, type,
267 process_synthesized_event,
271 #define WRITE_STAT_ROUND_EVENT(time, interval) \
272 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
274 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
277 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
278 struct perf_counts_values *count)
280 struct perf_sample_id *sid = SID(counter, cpu, thread);
282 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
283 process_synthesized_event, NULL);
287 * Read out the results of a single counter:
288 * do not aggregate counts across CPUs in system-wide mode
290 static int read_counter(struct perf_evsel *counter)
292 int nthreads = thread_map__nr(evsel_list->threads);
293 int ncpus = perf_evsel__nr_cpus(counter);
296 if (!counter->supported)
299 if (counter->system_wide)
302 for (thread = 0; thread < nthreads; thread++) {
303 for (cpu = 0; cpu < ncpus; cpu++) {
304 struct perf_counts_values *count;
306 count = perf_counts(counter->counts, cpu, thread);
307 if (perf_evsel__read(counter, cpu, thread, count))
311 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
312 pr_err("failed to write stat event\n");
318 fprintf(stat_config.output,
319 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
320 perf_evsel__name(counter),
322 count->val, count->ena, count->run);
330 static void read_counters(bool close_counters)
332 struct perf_evsel *counter;
334 evlist__for_each(evsel_list, counter) {
335 if (read_counter(counter))
336 pr_debug("failed to read counter %s\n", counter->name);
338 if (perf_stat_process_counter(&stat_config, counter))
339 pr_warning("failed to process counter %s\n", counter->name);
341 if (close_counters) {
342 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
343 thread_map__nr(evsel_list->threads));
348 static void process_interval(void)
350 struct timespec ts, rs;
352 read_counters(false);
354 clock_gettime(CLOCK_MONOTONIC, &ts);
355 diff_timespec(&rs, &ts, &ref_time);
358 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSECS_PER_SEC + rs.tv_nsec, INTERVAL))
359 pr_err("failed to write stat round event\n");
362 print_counters(&rs, 0, NULL);
365 static void enable_counters(void)
368 usleep(initial_delay * 1000);
371 * We need to enable counters only if:
372 * - we don't have tracee (attaching to task or cpu)
373 * - we have initial delay configured
375 if (!target__none(&target) || initial_delay)
376 perf_evlist__enable(evsel_list);
379 static volatile int workload_exec_errno;
382 * perf_evlist__prepare_workload will send a SIGUSR1
383 * if the fork fails, since we asked by setting its
384 * want_signal to true.
386 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
387 void *ucontext __maybe_unused)
389 workload_exec_errno = info->si_value.sival_int;
392 static bool has_unit(struct perf_evsel *counter)
394 return counter->unit && *counter->unit;
397 static bool has_scale(struct perf_evsel *counter)
399 return counter->scale != 1;
402 static int perf_stat_synthesize_config(bool is_pipe)
404 struct perf_evsel *counter;
408 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
409 process_synthesized_event);
411 pr_err("Couldn't synthesize attrs.\n");
417 * Synthesize other events stuff not carried within
418 * attr event - unit, scale, name
420 evlist__for_each(evsel_list, counter) {
421 if (!counter->supported)
425 * Synthesize unit and scale only if it's defined.
427 if (has_unit(counter)) {
428 err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
430 pr_err("Couldn't synthesize evsel unit.\n");
435 if (has_scale(counter)) {
436 err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
438 pr_err("Couldn't synthesize evsel scale.\n");
443 if (counter->own_cpus) {
444 err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
446 pr_err("Couldn't synthesize evsel scale.\n");
452 * Name is needed only for pipe output,
453 * perf.data carries event names.
456 err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
458 pr_err("Couldn't synthesize evsel name.\n");
464 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
465 process_synthesized_event,
468 pr_err("Couldn't synthesize thread map.\n");
472 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
473 process_synthesized_event, NULL);
475 pr_err("Couldn't synthesize thread map.\n");
479 err = perf_event__synthesize_stat_config(NULL, &stat_config,
480 process_synthesized_event, NULL);
482 pr_err("Couldn't synthesize config.\n");
489 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
491 static int __store_counter_ids(struct perf_evsel *counter,
492 struct cpu_map *cpus,
493 struct thread_map *threads)
497 for (cpu = 0; cpu < cpus->nr; cpu++) {
498 for (thread = 0; thread < threads->nr; thread++) {
499 int fd = FD(counter, cpu, thread);
501 if (perf_evlist__id_add_fd(evsel_list, counter,
502 cpu, thread, fd) < 0)
510 static int store_counter_ids(struct perf_evsel *counter)
512 struct cpu_map *cpus = counter->cpus;
513 struct thread_map *threads = counter->threads;
515 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
518 return __store_counter_ids(counter, cpus, threads);
521 static int __run_perf_stat(int argc, const char **argv)
523 int interval = stat_config.interval;
525 unsigned long long t0, t1;
526 struct perf_evsel *counter;
530 const bool forks = (argc > 0);
531 bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
534 ts.tv_sec = interval / 1000;
535 ts.tv_nsec = (interval % 1000) * 1000000;
542 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
543 workload_exec_failed_signal) < 0) {
544 perror("failed to prepare workload");
547 child_pid = evsel_list->workload.pid;
551 perf_evlist__set_leader(evsel_list);
553 evlist__for_each(evsel_list, counter) {
555 if (create_perf_stat_counter(counter) < 0) {
557 * PPC returns ENXIO for HW counters until 2.6.37
558 * (behavior changed with commit b0a873e).
560 if (errno == EINVAL || errno == ENOSYS ||
561 errno == ENOENT || errno == EOPNOTSUPP ||
564 ui__warning("%s event is not supported by the kernel.\n",
565 perf_evsel__name(counter));
566 counter->supported = false;
568 if ((counter->leader != counter) ||
569 !(counter->leader->nr_members > 1))
571 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
573 ui__warning("%s\n", msg);
577 perf_evsel__open_strerror(counter, &target,
578 errno, msg, sizeof(msg));
579 ui__error("%s\n", msg);
582 kill(child_pid, SIGTERM);
586 counter->supported = true;
588 l = strlen(counter->unit);
592 if (STAT_RECORD && store_counter_ids(counter))
596 if (perf_evlist__apply_filters(evsel_list, &counter)) {
597 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
598 counter->filter, perf_evsel__name(counter), errno,
599 strerror_r(errno, msg, sizeof(msg)));
604 int err, fd = perf_data_file__fd(&perf_stat.file);
607 err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
609 err = perf_session__write_header(perf_stat.session, evsel_list,
616 err = perf_stat_synthesize_config(is_pipe);
622 * Enable counters and exec the command:
625 clock_gettime(CLOCK_MONOTONIC, &ref_time);
628 perf_evlist__start_workload(evsel_list);
632 while (!waitpid(child_pid, &status, WNOHANG)) {
633 nanosleep(&ts, NULL);
639 if (workload_exec_errno) {
640 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
641 pr_err("Workload failed: %s\n", emsg);
645 if (WIFSIGNALED(status))
646 psignal(WTERMSIG(status), argv[0]);
650 nanosleep(&ts, NULL);
658 update_stats(&walltime_nsecs_stats, t1 - t0);
662 return WEXITSTATUS(status);
665 static int run_perf_stat(int argc, const char **argv)
670 ret = system(pre_cmd);
678 ret = __run_perf_stat(argc, argv);
683 ret = system(post_cmd);
691 static void print_running(u64 run, u64 ena)
694 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
698 ena ? 100.0 * run / ena : 100.0);
699 } else if (run != ena) {
700 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
704 static void print_noise_pct(double total, double avg)
706 double pct = rel_stddev_stats(total, avg);
709 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
711 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
714 static void print_noise(struct perf_evsel *evsel, double avg)
716 struct perf_stat_evsel *ps;
722 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
725 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
727 switch (stat_config.aggr_mode) {
729 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
730 cpu_map__id_to_socket(id),
732 cpu_map__id_to_cpu(id),
739 fprintf(stat_config.output, "S%*d%s%*d%s",
748 fprintf(stat_config.output, "CPU%*d%s",
750 perf_evsel__cpus(evsel)->map[id], csv_sep);
753 fprintf(stat_config.output, "%*s-%*d%s",
755 thread_map__comm(evsel->threads, id),
757 thread_map__pid(evsel->threads, id),
773 struct perf_evsel *evsel;
776 #define METRIC_LEN 35
778 static void new_line_std(void *ctx)
780 struct outstate *os = ctx;
785 static void do_new_line_std(struct outstate *os)
788 fputs(os->prefix, os->fh);
789 aggr_printout(os->evsel, os->id, os->nr);
790 if (stat_config.aggr_mode == AGGR_NONE)
791 fprintf(os->fh, " ");
792 fprintf(os->fh, " ");
795 static void print_metric_std(void *ctx, const char *color, const char *fmt,
796 const char *unit, double val)
798 struct outstate *os = ctx;
801 bool newline = os->newline;
805 if (unit == NULL || fmt == NULL) {
806 fprintf(out, "%-*s", METRIC_LEN, "");
813 n = fprintf(out, " # ");
815 n += color_fprintf(out, color, fmt, val);
817 n += fprintf(out, fmt, val);
818 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
821 static void new_line_csv(void *ctx)
823 struct outstate *os = ctx;
828 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
829 aggr_printout(os->evsel, os->id, os->nr);
830 for (i = 0; i < os->nfields; i++)
831 fputs(csv_sep, os->fh);
834 static void print_metric_csv(void *ctx,
835 const char *color __maybe_unused,
836 const char *fmt, const char *unit, double val)
838 struct outstate *os = ctx;
840 char buf[64], *vals, *ends;
842 if (unit == NULL || fmt == NULL) {
843 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
846 snprintf(buf, sizeof(buf), fmt, val);
848 while (isspace(*vals))
851 while (isdigit(*ends) || *ends == '.')
854 while (isspace(*unit))
856 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
859 #define METRIC_ONLY_LEN 20
861 /* Filter out some columns that don't work well in metrics only mode */
863 static bool valid_only_metric(const char *unit)
867 if (strstr(unit, "/sec") ||
868 strstr(unit, "hz") ||
869 strstr(unit, "Hz") ||
870 strstr(unit, "CPUs utilized"))
875 static const char *fixunit(char *buf, struct perf_evsel *evsel,
878 if (!strncmp(unit, "of all", 6)) {
879 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
886 static void print_metric_only(void *ctx, const char *color, const char *fmt,
887 const char *unit, double val)
889 struct outstate *os = ctx;
893 unsigned mlen = METRIC_ONLY_LEN;
895 if (!valid_only_metric(unit))
897 unit = fixunit(buf, os->evsel, unit);
899 n = color_fprintf(out, color, fmt, val);
901 n = fprintf(out, fmt, val);
902 if (n > METRIC_ONLY_LEN)
904 if (mlen < strlen(unit))
905 mlen = strlen(unit) + 1;
906 fprintf(out, "%*s", mlen - n, "");
909 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
911 const char *unit, double val)
913 struct outstate *os = ctx;
915 char buf[64], *vals, *ends;
918 if (!valid_only_metric(unit))
920 unit = fixunit(tbuf, os->evsel, unit);
921 snprintf(buf, sizeof buf, fmt, val);
923 while (isspace(*vals))
926 while (isdigit(*ends) || *ends == '.')
929 fprintf(out, "%s%s", vals, csv_sep);
932 static void new_line_metric(void *ctx __maybe_unused)
936 static void print_metric_header(void *ctx, const char *color __maybe_unused,
937 const char *fmt __maybe_unused,
938 const char *unit, double val __maybe_unused)
940 struct outstate *os = ctx;
943 if (!valid_only_metric(unit))
945 unit = fixunit(tbuf, os->evsel, unit);
947 fprintf(os->fh, "%s%s", unit, csv_sep);
949 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
952 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
954 FILE *output = stat_config.output;
955 double msecs = avg / 1e6;
956 const char *fmt_v, *fmt_n;
959 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
960 fmt_n = csv_output ? "%s" : "%-25s";
962 aggr_printout(evsel, id, nr);
964 scnprintf(name, sizeof(name), "%s%s",
965 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
967 fprintf(output, fmt_v, msecs, csv_sep);
970 fprintf(output, "%s%s", evsel->unit, csv_sep);
972 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
974 fprintf(output, fmt_n, name);
977 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
980 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
987 if (stat_config.aggr_mode == AGGR_NONE)
990 if (stat_config.aggr_mode == AGGR_GLOBAL)
993 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
994 int cpu2 = perf_evsel__cpus(evsel)->map[i];
996 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1002 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1004 FILE *output = stat_config.output;
1005 double sc = evsel->scale;
1009 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
1012 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1014 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1017 aggr_printout(evsel, id, nr);
1019 fprintf(output, fmt, avg, csv_sep);
1022 fprintf(output, "%-*s%s",
1023 csv_output ? 0 : unit_width,
1024 evsel->unit, csv_sep);
1026 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1029 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1032 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1033 char *prefix, u64 run, u64 ena, double noise)
1035 struct perf_stat_output_ctx out;
1036 struct outstate os = {
1037 .fh = stat_config.output,
1038 .prefix = prefix ? prefix : "",
1043 print_metric_t pm = print_metric_std;
1047 nl = new_line_metric;
1049 pm = print_metric_only_csv;
1051 pm = print_metric_only;
1055 if (csv_output && !metric_only) {
1056 static int aggr_fields[] = {
1064 pm = print_metric_csv;
1067 os.nfields += aggr_fields[stat_config.aggr_mode];
1071 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1073 pm(&os, NULL, "", "", 0);
1076 aggr_printout(counter, id, nr);
1078 fprintf(stat_config.output, "%*s%s",
1079 csv_output ? 0 : 18,
1080 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1083 fprintf(stat_config.output, "%-*s%s",
1084 csv_output ? 0 : unit_width,
1085 counter->unit, csv_sep);
1087 fprintf(stat_config.output, "%*s",
1088 csv_output ? 0 : -25,
1089 perf_evsel__name(counter));
1092 fprintf(stat_config.output, "%s%s",
1093 csv_sep, counter->cgrp->name);
1096 pm(&os, NULL, NULL, "", 0);
1097 print_noise(counter, noise);
1098 print_running(run, ena);
1100 pm(&os, NULL, NULL, "", 0);
1106 else if (nsec_counter(counter))
1107 nsec_printout(id, nr, counter, uval);
1109 abs_printout(id, nr, counter, uval);
1111 out.print_metric = pm;
1115 if (csv_output && !metric_only) {
1116 print_noise(counter, noise);
1117 print_running(run, ena);
1120 perf_stat__print_shadow_stats(counter, uval,
1121 first_shadow_cpu(counter, id),
1123 if (!csv_output && !metric_only) {
1124 print_noise(counter, noise);
1125 print_running(run, ena);
1129 static void aggr_update_shadow(void)
1133 struct perf_evsel *counter;
1135 for (s = 0; s < aggr_map->nr; s++) {
1136 id = aggr_map->map[s];
1137 evlist__for_each(evsel_list, counter) {
1139 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1140 s2 = aggr_get_id(evsel_list->cpus, cpu);
1143 val += perf_counts(counter->counts, cpu, 0)->val;
1145 val = val * counter->scale;
1146 perf_stat__update_shadow_stats(counter, &val,
1147 first_shadow_cpu(counter, id));
1152 static void print_aggr(char *prefix)
1154 FILE *output = stat_config.output;
1155 struct perf_evsel *counter;
1156 int cpu, s, s2, id, nr;
1161 if (!(aggr_map || aggr_get_id))
1164 aggr_update_shadow();
1167 * With metric_only everything is on a single line.
1168 * Without each counter has its own line.
1170 for (s = 0; s < aggr_map->nr; s++) {
1171 if (prefix && metric_only)
1172 fprintf(output, "%s", prefix);
1174 id = aggr_map->map[s];
1176 evlist__for_each(evsel_list, counter) {
1177 val = ena = run = 0;
1179 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1180 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1183 val += perf_counts(counter->counts, cpu, 0)->val;
1184 ena += perf_counts(counter->counts, cpu, 0)->ena;
1185 run += perf_counts(counter->counts, cpu, 0)->run;
1188 if (first && metric_only) {
1190 aggr_printout(counter, id, nr);
1192 if (prefix && !metric_only)
1193 fprintf(output, "%s", prefix);
1195 uval = val * counter->scale;
1196 printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1198 fputc('\n', output);
1201 fputc('\n', output);
1205 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1207 FILE *output = stat_config.output;
1208 int nthreads = thread_map__nr(counter->threads);
1209 int ncpus = cpu_map__nr(counter->cpus);
1213 for (thread = 0; thread < nthreads; thread++) {
1214 u64 ena = 0, run = 0, val = 0;
1216 for (cpu = 0; cpu < ncpus; cpu++) {
1217 val += perf_counts(counter->counts, cpu, thread)->val;
1218 ena += perf_counts(counter->counts, cpu, thread)->ena;
1219 run += perf_counts(counter->counts, cpu, thread)->run;
1223 fprintf(output, "%s", prefix);
1225 uval = val * counter->scale;
1226 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1227 fputc('\n', output);
1232 * Print out the results of a single counter:
1233 * aggregated counts in system-wide mode
1235 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1237 FILE *output = stat_config.output;
1238 struct perf_stat_evsel *ps = counter->priv;
1239 double avg = avg_stats(&ps->res_stats[0]);
1241 double avg_enabled, avg_running;
1243 avg_enabled = avg_stats(&ps->res_stats[1]);
1244 avg_running = avg_stats(&ps->res_stats[2]);
1246 if (prefix && !metric_only)
1247 fprintf(output, "%s", prefix);
1249 uval = avg * counter->scale;
1250 printout(-1, 0, counter, uval, prefix, avg_running, avg_enabled, avg);
1252 fprintf(output, "\n");
1256 * Print out the results of a single counter:
1257 * does not use aggregated count in system-wide
1259 static void print_counter(struct perf_evsel *counter, char *prefix)
1261 FILE *output = stat_config.output;
1266 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1267 val = perf_counts(counter->counts, cpu, 0)->val;
1268 ena = perf_counts(counter->counts, cpu, 0)->ena;
1269 run = perf_counts(counter->counts, cpu, 0)->run;
1272 fprintf(output, "%s", prefix);
1274 uval = val * counter->scale;
1275 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1277 fputc('\n', output);
1281 static void print_no_aggr_metric(char *prefix)
1285 struct perf_evsel *counter;
1289 nrcpus = evsel_list->cpus->nr;
1290 for (cpu = 0; cpu < nrcpus; cpu++) {
1294 fputs(prefix, stat_config.output);
1295 evlist__for_each(evsel_list, counter) {
1297 aggr_printout(counter, cpu, 0);
1300 val = perf_counts(counter->counts, cpu, 0)->val;
1301 ena = perf_counts(counter->counts, cpu, 0)->ena;
1302 run = perf_counts(counter->counts, cpu, 0)->run;
1304 uval = val * counter->scale;
1305 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1307 fputc('\n', stat_config.output);
1311 static int aggr_header_lens[] = {
1319 static const char *aggr_header_csv[] = {
1320 [AGGR_CORE] = "core,cpus,",
1321 [AGGR_SOCKET] = "socket,cpus",
1322 [AGGR_NONE] = "cpu,",
1323 [AGGR_THREAD] = "comm-pid,",
1327 static void print_metric_headers(const char *prefix, bool no_indent)
1329 struct perf_stat_output_ctx out;
1330 struct perf_evsel *counter;
1331 struct outstate os = {
1332 .fh = stat_config.output
1336 fprintf(stat_config.output, "%s", prefix);
1338 if (!csv_output && !no_indent)
1339 fprintf(stat_config.output, "%*s",
1340 aggr_header_lens[stat_config.aggr_mode], "");
1342 if (stat_config.interval)
1343 fputs("time,", stat_config.output);
1344 fputs(aggr_header_csv[stat_config.aggr_mode],
1345 stat_config.output);
1348 /* Print metrics headers only */
1349 evlist__for_each(evsel_list, counter) {
1352 out.print_metric = print_metric_header;
1353 out.new_line = new_line_metric;
1355 perf_stat__print_shadow_stats(counter, 0,
1359 fputc('\n', stat_config.output);
1362 static void print_interval(char *prefix, struct timespec *ts)
1364 FILE *output = stat_config.output;
1365 static int num_print_interval;
1367 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1369 if (num_print_interval == 0 && !csv_output) {
1370 switch (stat_config.aggr_mode) {
1372 fprintf(output, "# time socket cpus");
1374 fprintf(output, " counts %*s events\n", unit_width, "unit");
1377 fprintf(output, "# time core cpus");
1379 fprintf(output, " counts %*s events\n", unit_width, "unit");
1382 fprintf(output, "# time CPU");
1384 fprintf(output, " counts %*s events\n", unit_width, "unit");
1387 fprintf(output, "# time comm-pid");
1389 fprintf(output, " counts %*s events\n", unit_width, "unit");
1393 fprintf(output, "# time");
1395 fprintf(output, " counts %*s events\n", unit_width, "unit");
1401 if (num_print_interval == 0 && metric_only)
1402 print_metric_headers(" ", true);
1403 if (++num_print_interval == 25)
1404 num_print_interval = 0;
1407 static void print_header(int argc, const char **argv)
1409 FILE *output = stat_config.output;
1415 fprintf(output, "\n");
1416 fprintf(output, " Performance counter stats for ");
1417 if (target.system_wide)
1418 fprintf(output, "\'system wide");
1419 else if (target.cpu_list)
1420 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1421 else if (!target__has_task(&target)) {
1422 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1423 for (i = 1; argv && (i < argc); i++)
1424 fprintf(output, " %s", argv[i]);
1425 } else if (target.pid)
1426 fprintf(output, "process id \'%s", target.pid);
1428 fprintf(output, "thread id \'%s", target.tid);
1430 fprintf(output, "\'");
1432 fprintf(output, " (%d runs)", run_count);
1433 fprintf(output, ":\n\n");
1437 static void print_footer(void)
1439 FILE *output = stat_config.output;
1442 fprintf(output, "\n");
1443 fprintf(output, " %17.9f seconds time elapsed",
1444 avg_stats(&walltime_nsecs_stats)/1e9);
1445 if (run_count > 1) {
1446 fprintf(output, " ");
1447 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1448 avg_stats(&walltime_nsecs_stats));
1450 fprintf(output, "\n\n");
1453 static void print_counters(struct timespec *ts, int argc, const char **argv)
1455 int interval = stat_config.interval;
1456 struct perf_evsel *counter;
1457 char buf[64], *prefix = NULL;
1459 /* Do not print anything if we record to the pipe. */
1460 if (STAT_RECORD && perf_stat.file.is_pipe)
1464 print_interval(prefix = buf, ts);
1466 print_header(argc, argv);
1469 static int num_print_iv;
1471 if (num_print_iv == 0 && !interval)
1472 print_metric_headers(prefix, false);
1473 if (num_print_iv++ == 25)
1475 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1476 fprintf(stat_config.output, "%s", prefix);
1479 switch (stat_config.aggr_mode) {
1485 evlist__for_each(evsel_list, counter)
1486 print_aggr_thread(counter, prefix);
1489 evlist__for_each(evsel_list, counter)
1490 print_counter_aggr(counter, prefix);
1492 fputc('\n', stat_config.output);
1496 print_no_aggr_metric(prefix);
1498 evlist__for_each(evsel_list, counter)
1499 print_counter(counter, prefix);
1507 if (!interval && !csv_output)
1510 fflush(stat_config.output);
1513 static volatile int signr = -1;
1515 static void skip_signal(int signo)
1517 if ((child_pid == -1) || stat_config.interval)
1522 * render child_pid harmless
1523 * won't send SIGTERM to a random
1524 * process in case of race condition
1525 * and fast PID recycling
1530 static void sig_atexit(void)
1535 * avoid race condition with SIGCHLD handler
1536 * in skip_signal() which is modifying child_pid
1537 * goal is to avoid send SIGTERM to a random
1541 sigaddset(&set, SIGCHLD);
1542 sigprocmask(SIG_BLOCK, &set, &oset);
1544 if (child_pid != -1)
1545 kill(child_pid, SIGTERM);
1547 sigprocmask(SIG_SETMASK, &oset, NULL);
1552 signal(signr, SIG_DFL);
1553 kill(getpid(), signr);
1556 static int stat__set_big_num(const struct option *opt __maybe_unused,
1557 const char *s __maybe_unused, int unset)
1559 big_num_opt = unset ? 0 : 1;
1563 static int enable_metric_only(const struct option *opt __maybe_unused,
1564 const char *s __maybe_unused, int unset)
1566 force_metric_only = true;
1567 metric_only = !unset;
1571 static const struct option stat_options[] = {
1572 OPT_BOOLEAN('T', "transaction", &transaction_run,
1573 "hardware transaction statistics"),
1574 OPT_CALLBACK('e', "event", &evsel_list, "event",
1575 "event selector. use 'perf list' to list available events",
1576 parse_events_option),
1577 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1578 "event filter", parse_filter),
1579 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1580 "child tasks do not inherit counters"),
1581 OPT_STRING('p', "pid", &target.pid, "pid",
1582 "stat events on existing process id"),
1583 OPT_STRING('t', "tid", &target.tid, "tid",
1584 "stat events on existing thread id"),
1585 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1586 "system-wide collection from all CPUs"),
1587 OPT_BOOLEAN('g', "group", &group,
1588 "put the counters into a counter group"),
1589 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1590 OPT_INCR('v', "verbose", &verbose,
1591 "be more verbose (show counter open errors, etc)"),
1592 OPT_INTEGER('r', "repeat", &run_count,
1593 "repeat command and print average + stddev (max: 100, forever: 0)"),
1594 OPT_BOOLEAN('n', "null", &null_run,
1595 "null run - dont start any counters"),
1596 OPT_INCR('d', "detailed", &detailed_run,
1597 "detailed run - start a lot of events"),
1598 OPT_BOOLEAN('S', "sync", &sync_run,
1599 "call sync() before starting a run"),
1600 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1601 "print large numbers with thousands\' separators",
1603 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1604 "list of cpus to monitor in system-wide"),
1605 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1606 "disable CPU count aggregation", AGGR_NONE),
1607 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1608 "print counts with custom separator"),
1609 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1610 "monitor event in cgroup name only", parse_cgroups),
1611 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1612 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1613 OPT_INTEGER(0, "log-fd", &output_fd,
1614 "log output to fd, instead of stderr"),
1615 OPT_STRING(0, "pre", &pre_cmd, "command",
1616 "command to run prior to the measured command"),
1617 OPT_STRING(0, "post", &post_cmd, "command",
1618 "command to run after to the measured command"),
1619 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1620 "print counts at regular interval in ms (>= 10)"),
1621 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1622 "aggregate counts per processor socket", AGGR_SOCKET),
1623 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1624 "aggregate counts per physical processor core", AGGR_CORE),
1625 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1626 "aggregate counts per thread", AGGR_THREAD),
1627 OPT_UINTEGER('D', "delay", &initial_delay,
1628 "ms to wait before starting measurement after program start"),
1629 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1630 "Only print computed metrics. No raw values", enable_metric_only),
1631 OPT_BOOLEAN(0, "topdown", &topdown_run,
1632 "measure topdown level 1 statistics"),
1636 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1638 return cpu_map__get_socket(map, cpu, NULL);
1641 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1643 return cpu_map__get_core(map, cpu, NULL);
1646 static int cpu_map__get_max(struct cpu_map *map)
1650 for (i = 0; i < map->nr; i++) {
1651 if (map->map[i] > max)
1658 static struct cpu_map *cpus_aggr_map;
1660 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1667 cpu = map->map[idx];
1669 if (cpus_aggr_map->map[cpu] == -1)
1670 cpus_aggr_map->map[cpu] = get_id(map, idx);
1672 return cpus_aggr_map->map[cpu];
1675 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1677 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1680 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1682 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1685 static int perf_stat_init_aggr_mode(void)
1689 switch (stat_config.aggr_mode) {
1691 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1692 perror("cannot build socket map");
1695 aggr_get_id = perf_stat__get_socket_cached;
1698 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1699 perror("cannot build core map");
1702 aggr_get_id = perf_stat__get_core_cached;
1713 * The evsel_list->cpus is the base we operate on,
1714 * taking the highest cpu number to be the size of
1715 * the aggregation translate cpumap.
1717 nr = cpu_map__get_max(evsel_list->cpus);
1718 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1719 return cpus_aggr_map ? 0 : -ENOMEM;
1722 static void perf_stat__exit_aggr_mode(void)
1724 cpu_map__put(aggr_map);
1725 cpu_map__put(cpus_aggr_map);
1727 cpus_aggr_map = NULL;
1730 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1737 cpu = map->map[idx];
1739 if (cpu >= env->nr_cpus_online)
1745 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1747 struct perf_env *env = data;
1748 int cpu = perf_env__get_cpu(env, map, idx);
1750 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1753 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1755 struct perf_env *env = data;
1756 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1759 int socket_id = env->cpu[cpu].socket_id;
1762 * Encode socket in upper 16 bits
1763 * core_id is relative to socket, and
1764 * we need a global id. So we combine
1767 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1773 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1774 struct cpu_map **sockp)
1776 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1779 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1780 struct cpu_map **corep)
1782 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1785 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1787 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1790 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1792 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1795 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1797 struct perf_env *env = &st->session->header.env;
1799 switch (stat_config.aggr_mode) {
1801 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1802 perror("cannot build socket map");
1805 aggr_get_id = perf_stat__get_socket_file;
1808 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1809 perror("cannot build core map");
1812 aggr_get_id = perf_stat__get_core_file;
1825 static int topdown_filter_events(const char **attr, char **str, bool use_group)
1832 for (i = 0; attr[i]; i++) {
1833 if (pmu_have_event("cpu", attr[i])) {
1834 len += strlen(attr[i]) + 1;
1835 attr[i - off] = attr[i];
1839 attr[i - off] = NULL;
1841 *str = malloc(len + 1 + 2);
1851 for (i = 0; attr[i]; i++) {
1864 __weak bool arch_topdown_check_group(bool *warn)
1870 __weak void arch_topdown_group_warn(void)
1875 * Add default attributes, if there were no attributes specified or
1876 * if -d/--detailed, -d -d or -d -d -d is used:
1878 static int add_default_attributes(void)
1881 struct perf_event_attr default_attrs0[] = {
1883 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1884 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1885 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1886 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1888 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1890 struct perf_event_attr frontend_attrs[] = {
1891 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1893 struct perf_event_attr backend_attrs[] = {
1894 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1896 struct perf_event_attr default_attrs1[] = {
1897 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1898 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1899 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1904 * Detailed stats (-d), covering the L1 and last level data caches:
1906 struct perf_event_attr detailed_attrs[] = {
1908 { .type = PERF_TYPE_HW_CACHE,
1910 PERF_COUNT_HW_CACHE_L1D << 0 |
1911 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1912 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1914 { .type = PERF_TYPE_HW_CACHE,
1916 PERF_COUNT_HW_CACHE_L1D << 0 |
1917 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1918 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1920 { .type = PERF_TYPE_HW_CACHE,
1922 PERF_COUNT_HW_CACHE_LL << 0 |
1923 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1924 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1926 { .type = PERF_TYPE_HW_CACHE,
1928 PERF_COUNT_HW_CACHE_LL << 0 |
1929 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1930 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1934 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1936 struct perf_event_attr very_detailed_attrs[] = {
1938 { .type = PERF_TYPE_HW_CACHE,
1940 PERF_COUNT_HW_CACHE_L1I << 0 |
1941 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1942 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1944 { .type = PERF_TYPE_HW_CACHE,
1946 PERF_COUNT_HW_CACHE_L1I << 0 |
1947 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1948 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1950 { .type = PERF_TYPE_HW_CACHE,
1952 PERF_COUNT_HW_CACHE_DTLB << 0 |
1953 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1954 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1956 { .type = PERF_TYPE_HW_CACHE,
1958 PERF_COUNT_HW_CACHE_DTLB << 0 |
1959 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1960 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1962 { .type = PERF_TYPE_HW_CACHE,
1964 PERF_COUNT_HW_CACHE_ITLB << 0 |
1965 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1966 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1968 { .type = PERF_TYPE_HW_CACHE,
1970 PERF_COUNT_HW_CACHE_ITLB << 0 |
1971 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1972 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1977 * Very, very detailed stats (-d -d -d), adding prefetch events:
1979 struct perf_event_attr very_very_detailed_attrs[] = {
1981 { .type = PERF_TYPE_HW_CACHE,
1983 PERF_COUNT_HW_CACHE_L1D << 0 |
1984 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1985 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1987 { .type = PERF_TYPE_HW_CACHE,
1989 PERF_COUNT_HW_CACHE_L1D << 0 |
1990 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1991 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1994 /* Set attrs if no event is selected and !null_run: */
1998 if (transaction_run) {
1999 if (pmu_have_event("cpu", "cycles-ct") &&
2000 pmu_have_event("cpu", "el-start"))
2001 err = parse_events(evsel_list, transaction_attrs, NULL);
2003 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
2005 fprintf(stderr, "Cannot set up transaction events\n");
2015 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2016 stat_config.aggr_mode != AGGR_CORE) {
2017 pr_err("top down event configuration requires --per-core mode\n");
2020 stat_config.aggr_mode = AGGR_CORE;
2021 if (nr_cgroups || !target__has_cpu(&target)) {
2022 pr_err("top down event configuration requires system-wide mode (-a)\n");
2026 if (!force_metric_only)
2028 if (topdown_filter_events(topdown_attrs, &str,
2029 arch_topdown_check_group(&warn)) < 0) {
2030 pr_err("Out of memory\n");
2033 if (topdown_attrs[0] && str) {
2035 arch_topdown_group_warn();
2036 err = parse_events(evsel_list, str, NULL);
2039 "Cannot set up top down events %s: %d\n",
2045 fprintf(stderr, "System does not support topdown\n");
2051 if (!evsel_list->nr_entries) {
2052 if (target__has_cpu(&target))
2053 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2055 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2057 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2058 if (perf_evlist__add_default_attrs(evsel_list,
2059 frontend_attrs) < 0)
2062 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2063 if (perf_evlist__add_default_attrs(evsel_list,
2067 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2071 /* Detailed events get appended to the event list: */
2073 if (detailed_run < 1)
2076 /* Append detailed run extra attributes: */
2077 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2080 if (detailed_run < 2)
2083 /* Append very detailed run extra attributes: */
2084 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2087 if (detailed_run < 3)
2090 /* Append very, very detailed run extra attributes: */
2091 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2094 static const char * const stat_record_usage[] = {
2095 "perf stat record [<options>]",
2099 static void init_features(struct perf_session *session)
2103 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2104 perf_header__set_feat(&session->header, feat);
2106 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2107 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2108 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2109 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2112 static int __cmd_record(int argc, const char **argv)
2114 struct perf_session *session;
2115 struct perf_data_file *file = &perf_stat.file;
2117 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2118 PARSE_OPT_STOP_AT_NON_OPTION);
2121 file->path = output_name;
2123 if (run_count != 1 || forever) {
2124 pr_err("Cannot use -r option with perf stat record.\n");
2128 session = perf_session__new(file, false, NULL);
2129 if (session == NULL) {
2130 pr_err("Perf session creation failed.\n");
2134 init_features(session);
2136 session->evlist = evsel_list;
2137 perf_stat.session = session;
2138 perf_stat.record = true;
2142 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2143 union perf_event *event,
2144 struct perf_session *session)
2146 struct stat_round_event *stat_round = &event->stat_round;
2147 struct perf_evsel *counter;
2148 struct timespec tsh, *ts = NULL;
2149 const char **argv = session->header.env.cmdline_argv;
2150 int argc = session->header.env.nr_cmdline;
2152 evlist__for_each(evsel_list, counter)
2153 perf_stat_process_counter(&stat_config, counter);
2155 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2156 update_stats(&walltime_nsecs_stats, stat_round->time);
2158 if (stat_config.interval && stat_round->time) {
2159 tsh.tv_sec = stat_round->time / NSECS_PER_SEC;
2160 tsh.tv_nsec = stat_round->time % NSECS_PER_SEC;
2164 print_counters(ts, argc, argv);
2169 int process_stat_config_event(struct perf_tool *tool __maybe_unused,
2170 union perf_event *event,
2171 struct perf_session *session __maybe_unused)
2173 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2175 perf_event__read_stat_config(&stat_config, &event->stat_config);
2177 if (cpu_map__empty(st->cpus)) {
2178 if (st->aggr_mode != AGGR_UNSET)
2179 pr_warning("warning: processing task data, aggregation mode not set\n");
2183 if (st->aggr_mode != AGGR_UNSET)
2184 stat_config.aggr_mode = st->aggr_mode;
2186 if (perf_stat.file.is_pipe)
2187 perf_stat_init_aggr_mode();
2189 perf_stat_init_aggr_mode_file(st);
2194 static int set_maps(struct perf_stat *st)
2196 if (!st->cpus || !st->threads)
2199 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2202 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2204 if (perf_evlist__alloc_stats(evsel_list, true))
2207 st->maps_allocated = true;
2212 int process_thread_map_event(struct perf_tool *tool __maybe_unused,
2213 union perf_event *event,
2214 struct perf_session *session __maybe_unused)
2216 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2219 pr_warning("Extra thread map event, ignoring.\n");
2223 st->threads = thread_map__new_event(&event->thread_map);
2227 return set_maps(st);
2231 int process_cpu_map_event(struct perf_tool *tool __maybe_unused,
2232 union perf_event *event,
2233 struct perf_session *session __maybe_unused)
2235 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2236 struct cpu_map *cpus;
2239 pr_warning("Extra cpu map event, ignoring.\n");
2243 cpus = cpu_map__new_data(&event->cpu_map.data);
2248 return set_maps(st);
2251 static const char * const stat_report_usage[] = {
2252 "perf stat report [<options>]",
2256 static struct perf_stat perf_stat = {
2258 .attr = perf_event__process_attr,
2259 .event_update = perf_event__process_event_update,
2260 .thread_map = process_thread_map_event,
2261 .cpu_map = process_cpu_map_event,
2262 .stat_config = process_stat_config_event,
2263 .stat = perf_event__process_stat_event,
2264 .stat_round = process_stat_round_event,
2266 .aggr_mode = AGGR_UNSET,
2269 static int __cmd_report(int argc, const char **argv)
2271 struct perf_session *session;
2272 const struct option options[] = {
2273 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2274 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2275 "aggregate counts per processor socket", AGGR_SOCKET),
2276 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2277 "aggregate counts per physical processor core", AGGR_CORE),
2278 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2279 "disable CPU count aggregation", AGGR_NONE),
2285 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2287 if (!input_name || !strlen(input_name)) {
2288 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2291 input_name = "perf.data";
2294 perf_stat.file.path = input_name;
2295 perf_stat.file.mode = PERF_DATA_MODE_READ;
2297 session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
2298 if (session == NULL)
2301 perf_stat.session = session;
2302 stat_config.output = stderr;
2303 evsel_list = session->evlist;
2305 ret = perf_session__process_events(session);
2309 perf_session__delete(session);
2313 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
2315 const char * const stat_usage[] = {
2316 "perf stat [<options>] [<command>]",
2319 int status = -EINVAL, run_idx;
2321 FILE *output = stderr;
2322 unsigned int interval;
2323 const char * const stat_subcommands[] = { "record", "report" };
2325 setlocale(LC_ALL, "");
2327 evsel_list = perf_evlist__new();
2328 if (evsel_list == NULL)
2331 parse_events__shrink_config_terms();
2332 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2333 (const char **) stat_usage,
2334 PARSE_OPT_STOP_AT_NON_OPTION);
2335 perf_stat__init_shadow_stats();
2339 if (!strcmp(csv_sep, "\\t"))
2342 csv_sep = DEFAULT_SEPARATOR;
2344 if (argc && !strncmp(argv[0], "rec", 3)) {
2345 argc = __cmd_record(argc, argv);
2348 } else if (argc && !strncmp(argv[0], "rep", 3))
2349 return __cmd_report(argc, argv);
2351 interval = stat_config.interval;
2354 * For record command the -o is already taken care of.
2356 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2359 if (output_name && output_fd) {
2360 fprintf(stderr, "cannot use both --output and --log-fd\n");
2361 parse_options_usage(stat_usage, stat_options, "o", 1);
2362 parse_options_usage(NULL, stat_options, "log-fd", 0);
2366 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2367 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2371 if (metric_only && run_count > 1) {
2372 fprintf(stderr, "--metric-only is not supported with -r\n");
2376 if (output_fd < 0) {
2377 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2378 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2384 mode = append_file ? "a" : "w";
2386 output = fopen(output_name, mode);
2388 perror("failed to create output file");
2391 clock_gettime(CLOCK_REALTIME, &tm);
2392 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2393 } else if (output_fd > 0) {
2394 mode = append_file ? "a" : "w";
2395 output = fdopen(output_fd, mode);
2397 perror("Failed opening logfd");
2402 stat_config.output = output;
2405 * let the spreadsheet do the pretty-printing
2408 /* User explicitly passed -B? */
2409 if (big_num_opt == 1) {
2410 fprintf(stderr, "-B option not supported with -x\n");
2411 parse_options_usage(stat_usage, stat_options, "B", 1);
2412 parse_options_usage(NULL, stat_options, "x", 1);
2414 } else /* Nope, so disable big number formatting */
2416 } else if (big_num_opt == 0) /* User passed --no-big-num */
2419 if (!argc && target__none(&target))
2420 usage_with_options(stat_usage, stat_options);
2422 if (run_count < 0) {
2423 pr_err("Run count must be a positive number\n");
2424 parse_options_usage(stat_usage, stat_options, "r", 1);
2426 } else if (run_count == 0) {
2431 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2432 fprintf(stderr, "The --per-thread option is only available "
2433 "when monitoring via -p -t options.\n");
2434 parse_options_usage(NULL, stat_options, "p", 1);
2435 parse_options_usage(NULL, stat_options, "t", 1);
2440 * no_aggr, cgroup are for system-wide only
2441 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2443 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2444 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2445 !target__has_cpu(&target)) {
2446 fprintf(stderr, "both cgroup and no-aggregation "
2447 "modes only available in system-wide mode\n");
2449 parse_options_usage(stat_usage, stat_options, "G", 1);
2450 parse_options_usage(NULL, stat_options, "A", 1);
2451 parse_options_usage(NULL, stat_options, "a", 1);
2455 if (add_default_attributes())
2458 target__validate(&target);
2460 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2461 if (target__has_task(&target)) {
2462 pr_err("Problems finding threads of monitor\n");
2463 parse_options_usage(stat_usage, stat_options, "p", 1);
2464 parse_options_usage(NULL, stat_options, "t", 1);
2465 } else if (target__has_cpu(&target)) {
2466 perror("failed to parse CPUs map");
2467 parse_options_usage(stat_usage, stat_options, "C", 1);
2468 parse_options_usage(NULL, stat_options, "a", 1);
2474 * Initialize thread_map with comm names,
2475 * so we could print it out on output.
2477 if (stat_config.aggr_mode == AGGR_THREAD)
2478 thread_map__read_comms(evsel_list->threads);
2480 if (interval && interval < 100) {
2481 if (interval < 10) {
2482 pr_err("print interval must be >= 10ms\n");
2483 parse_options_usage(stat_usage, stat_options, "I", 1);
2486 pr_warning("print interval < 100ms. "
2487 "The overhead percentage could be high in some cases. "
2488 "Please proceed with caution.\n");
2491 if (perf_evlist__alloc_stats(evsel_list, interval))
2494 if (perf_stat_init_aggr_mode())
2498 * We dont want to block the signals - that would cause
2499 * child tasks to inherit that and Ctrl-C would not work.
2500 * What we want is for Ctrl-C to work in the exec()-ed
2501 * task, but being ignored by perf stat itself:
2505 signal(SIGINT, skip_signal);
2506 signal(SIGCHLD, skip_signal);
2507 signal(SIGALRM, skip_signal);
2508 signal(SIGABRT, skip_signal);
2511 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2512 if (run_count != 1 && verbose)
2513 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2516 status = run_perf_stat(argc, argv);
2517 if (forever && status != -1) {
2518 print_counters(NULL, argc, argv);
2519 perf_stat__reset_stats();
2523 if (!forever && status != -1 && !interval)
2524 print_counters(NULL, argc, argv);
2528 * We synthesize the kernel mmap record just so that older tools
2529 * don't emit warnings about not being able to resolve symbols
2530 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2531 * a saner message about no samples being in the perf.data file.
2533 * This also serves to suppress a warning about f_header.data.size == 0
2534 * in header.c at the moment 'perf stat record' gets introduced, which
2535 * is not really needed once we start adding the stat specific PERF_RECORD_
2536 * records, but the need to suppress the kptr_restrict messages in older
2537 * tools remain -acme
2539 int fd = perf_data_file__fd(&perf_stat.file);
2540 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2541 process_synthesized_event,
2542 &perf_stat.session->machines.host);
2544 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2545 "older tools may produce warnings about this file\n.");
2549 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2550 pr_err("failed to write stat round event\n");
2553 if (!perf_stat.file.is_pipe) {
2554 perf_stat.session->header.data_size += perf_stat.bytes_written;
2555 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2558 perf_session__delete(perf_stat.session);
2561 perf_stat__exit_aggr_mode();
2562 perf_evlist__free_stats(evsel_list);
2564 perf_evlist__delete(evsel_list);