4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ./hackbench 10
13 Performance counter stats for './hackbench 10':
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
27 0.154822978 seconds time elapsed
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 * Improvements and fixes by:
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
41 * Released under the GPL v2. (and only v2, not any later version)
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include <subcmd/parse-options.h>
49 #include "util/parse-events.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/color.h"
56 #include "util/stat.h"
57 #include "util/header.h"
58 #include "util/cpumap.h"
59 #include "util/thread.h"
60 #include "util/thread_map.h"
61 #include "util/counts.h"
62 #include "util/session.h"
63 #include "util/tool.h"
67 #include <sys/prctl.h>
70 #define DEFAULT_SEPARATOR " "
71 #define CNTR_NOT_SUPPORTED "<not supported>"
72 #define CNTR_NOT_COUNTED "<not counted>"
74 static void print_counters(struct timespec *ts, int argc, const char **argv);
76 /* Default events used for perf stat -T */
77 static const char *transaction_attrs = {
89 /* More limited version when the CPU does not have all events. */
90 static const char * transaction_limited_attrs = {
100 static struct perf_evlist *evsel_list;
102 static struct target target = {
106 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
108 static int run_count = 1;
109 static bool no_inherit = false;
110 static volatile pid_t child_pid = -1;
111 static bool null_run = false;
112 static int detailed_run = 0;
113 static bool transaction_run;
114 static bool big_num = true;
115 static int big_num_opt = -1;
116 static const char *csv_sep = NULL;
117 static bool csv_output = false;
118 static bool group = false;
119 static const char *pre_cmd = NULL;
120 static const char *post_cmd = NULL;
121 static bool sync_run = false;
122 static unsigned int initial_delay = 0;
123 static unsigned int unit_width = 4; /* strlen("unit") */
124 static bool forever = false;
125 static struct timespec ref_time;
126 static struct cpu_map *aggr_map;
127 static aggr_get_id_t aggr_get_id;
128 static bool append_file;
129 static const char *output_name;
130 static int output_fd;
134 struct perf_data_file file;
135 struct perf_session *session;
137 struct perf_tool tool;
139 struct cpu_map *cpus;
140 struct thread_map *threads;
141 enum aggr_mode aggr_mode;
144 static struct perf_stat perf_stat;
145 #define STAT_RECORD perf_stat.record
147 static volatile int done = 0;
149 static struct perf_stat_config stat_config = {
150 .aggr_mode = AGGR_GLOBAL,
154 static inline void diff_timespec(struct timespec *r, struct timespec *a,
157 r->tv_sec = a->tv_sec - b->tv_sec;
158 if (a->tv_nsec < b->tv_nsec) {
159 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
162 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
166 static void perf_stat__reset_stats(void)
168 perf_evlist__reset_stats(evsel_list);
169 perf_stat__reset_shadow_stats();
172 static int create_perf_stat_counter(struct perf_evsel *evsel)
174 struct perf_event_attr *attr = &evsel->attr;
176 if (stat_config.scale)
177 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
178 PERF_FORMAT_TOTAL_TIME_RUNNING;
180 attr->inherit = !no_inherit;
183 * Some events get initialized with sample_(period/type) set,
184 * like tracepoints. Clear it up for counting.
186 attr->sample_period = 0;
189 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
190 * while avoiding that older tools show confusing messages.
192 * However for pipe sessions we need to keep it zero,
193 * because script's perf_evsel__check_attr is triggered
194 * by attr->sample_type != 0, and we can't run it on
197 if (!(STAT_RECORD && perf_stat.file.is_pipe))
198 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
201 * Disabling all counters initially, they will be enabled
202 * either manually by us or by kernel via enable_on_exec
205 if (perf_evsel__is_group_leader(evsel)) {
209 * In case of initial_delay we enable tracee
212 if (target__none(&target) && !initial_delay)
213 attr->enable_on_exec = 1;
216 if (target__has_cpu(&target))
217 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
219 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
223 * Does the counter have nsecs as a unit?
225 static inline int nsec_counter(struct perf_evsel *evsel)
227 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
228 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
234 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
235 union perf_event *event,
236 struct perf_sample *sample __maybe_unused,
237 struct machine *machine __maybe_unused)
239 if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
240 pr_err("failed to write perf data, error: %m\n");
244 perf_stat.bytes_written += event->header.size;
248 static int write_stat_round_event(u64 tm, u64 type)
250 return perf_event__synthesize_stat_round(NULL, tm, type,
251 process_synthesized_event,
255 #define WRITE_STAT_ROUND_EVENT(time, interval) \
256 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
258 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
261 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
262 struct perf_counts_values *count)
264 struct perf_sample_id *sid = SID(counter, cpu, thread);
266 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
267 process_synthesized_event, NULL);
271 * Read out the results of a single counter:
272 * do not aggregate counts across CPUs in system-wide mode
274 static int read_counter(struct perf_evsel *counter)
276 int nthreads = thread_map__nr(evsel_list->threads);
277 int ncpus = perf_evsel__nr_cpus(counter);
280 if (!counter->supported)
283 if (counter->system_wide)
286 for (thread = 0; thread < nthreads; thread++) {
287 for (cpu = 0; cpu < ncpus; cpu++) {
288 struct perf_counts_values *count;
290 count = perf_counts(counter->counts, cpu, thread);
291 if (perf_evsel__read(counter, cpu, thread, count))
295 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
296 pr_err("failed to write stat event\n");
306 static void read_counters(bool close_counters)
308 struct perf_evsel *counter;
310 evlist__for_each(evsel_list, counter) {
311 if (read_counter(counter))
312 pr_debug("failed to read counter %s\n", counter->name);
314 if (perf_stat_process_counter(&stat_config, counter))
315 pr_warning("failed to process counter %s\n", counter->name);
317 if (close_counters) {
318 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
319 thread_map__nr(evsel_list->threads));
324 static void process_interval(void)
326 struct timespec ts, rs;
328 read_counters(false);
330 clock_gettime(CLOCK_MONOTONIC, &ts);
331 diff_timespec(&rs, &ts, &ref_time);
334 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSECS_PER_SEC + rs.tv_nsec, INTERVAL))
335 pr_err("failed to write stat round event\n");
338 print_counters(&rs, 0, NULL);
341 static void enable_counters(void)
344 usleep(initial_delay * 1000);
347 * We need to enable counters only if:
348 * - we don't have tracee (attaching to task or cpu)
349 * - we have initial delay configured
351 if (!target__none(&target) || initial_delay)
352 perf_evlist__enable(evsel_list);
355 static volatile int workload_exec_errno;
358 * perf_evlist__prepare_workload will send a SIGUSR1
359 * if the fork fails, since we asked by setting its
360 * want_signal to true.
362 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
363 void *ucontext __maybe_unused)
365 workload_exec_errno = info->si_value.sival_int;
368 static bool has_unit(struct perf_evsel *counter)
370 return counter->unit && *counter->unit;
373 static bool has_scale(struct perf_evsel *counter)
375 return counter->scale != 1;
378 static int perf_stat_synthesize_config(bool is_pipe)
380 struct perf_evsel *counter;
384 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
385 process_synthesized_event);
387 pr_err("Couldn't synthesize attrs.\n");
393 * Synthesize other events stuff not carried within
394 * attr event - unit, scale, name
396 evlist__for_each(evsel_list, counter) {
397 if (!counter->supported)
401 * Synthesize unit and scale only if it's defined.
403 if (has_unit(counter)) {
404 err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
406 pr_err("Couldn't synthesize evsel unit.\n");
411 if (has_scale(counter)) {
412 err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
414 pr_err("Couldn't synthesize evsel scale.\n");
419 if (counter->own_cpus) {
420 err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
422 pr_err("Couldn't synthesize evsel scale.\n");
428 * Name is needed only for pipe output,
429 * perf.data carries event names.
432 err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
434 pr_err("Couldn't synthesize evsel name.\n");
440 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
441 process_synthesized_event,
444 pr_err("Couldn't synthesize thread map.\n");
448 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
449 process_synthesized_event, NULL);
451 pr_err("Couldn't synthesize thread map.\n");
455 err = perf_event__synthesize_stat_config(NULL, &stat_config,
456 process_synthesized_event, NULL);
458 pr_err("Couldn't synthesize config.\n");
465 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
467 static int __store_counter_ids(struct perf_evsel *counter,
468 struct cpu_map *cpus,
469 struct thread_map *threads)
473 for (cpu = 0; cpu < cpus->nr; cpu++) {
474 for (thread = 0; thread < threads->nr; thread++) {
475 int fd = FD(counter, cpu, thread);
477 if (perf_evlist__id_add_fd(evsel_list, counter,
478 cpu, thread, fd) < 0)
486 static int store_counter_ids(struct perf_evsel *counter)
488 struct cpu_map *cpus = counter->cpus;
489 struct thread_map *threads = counter->threads;
491 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
494 return __store_counter_ids(counter, cpus, threads);
497 static int __run_perf_stat(int argc, const char **argv)
499 int interval = stat_config.interval;
501 unsigned long long t0, t1;
502 struct perf_evsel *counter;
506 const bool forks = (argc > 0);
507 bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
510 ts.tv_sec = interval / 1000;
511 ts.tv_nsec = (interval % 1000) * 1000000;
518 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
519 workload_exec_failed_signal) < 0) {
520 perror("failed to prepare workload");
523 child_pid = evsel_list->workload.pid;
527 perf_evlist__set_leader(evsel_list);
529 evlist__for_each(evsel_list, counter) {
530 if (create_perf_stat_counter(counter) < 0) {
532 * PPC returns ENXIO for HW counters until 2.6.37
533 * (behavior changed with commit b0a873e).
535 if (errno == EINVAL || errno == ENOSYS ||
536 errno == ENOENT || errno == EOPNOTSUPP ||
539 ui__warning("%s event is not supported by the kernel.\n",
540 perf_evsel__name(counter));
541 counter->supported = false;
543 if ((counter->leader != counter) ||
544 !(counter->leader->nr_members > 1))
548 perf_evsel__open_strerror(counter, &target,
549 errno, msg, sizeof(msg));
550 ui__error("%s\n", msg);
553 kill(child_pid, SIGTERM);
557 counter->supported = true;
559 l = strlen(counter->unit);
563 if (STAT_RECORD && store_counter_ids(counter))
567 if (perf_evlist__apply_filters(evsel_list, &counter)) {
568 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
569 counter->filter, perf_evsel__name(counter), errno,
570 strerror_r(errno, msg, sizeof(msg)));
575 int err, fd = perf_data_file__fd(&perf_stat.file);
578 err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
580 err = perf_session__write_header(perf_stat.session, evsel_list,
587 err = perf_stat_synthesize_config(is_pipe);
593 * Enable counters and exec the command:
596 clock_gettime(CLOCK_MONOTONIC, &ref_time);
599 perf_evlist__start_workload(evsel_list);
603 while (!waitpid(child_pid, &status, WNOHANG)) {
604 nanosleep(&ts, NULL);
610 if (workload_exec_errno) {
611 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
612 pr_err("Workload failed: %s\n", emsg);
616 if (WIFSIGNALED(status))
617 psignal(WTERMSIG(status), argv[0]);
621 nanosleep(&ts, NULL);
629 update_stats(&walltime_nsecs_stats, t1 - t0);
633 return WEXITSTATUS(status);
636 static int run_perf_stat(int argc, const char **argv)
641 ret = system(pre_cmd);
649 ret = __run_perf_stat(argc, argv);
654 ret = system(post_cmd);
662 static void print_running(u64 run, u64 ena)
665 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
669 ena ? 100.0 * run / ena : 100.0);
670 } else if (run != ena) {
671 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
675 static void print_noise_pct(double total, double avg)
677 double pct = rel_stddev_stats(total, avg);
680 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
682 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
685 static void print_noise(struct perf_evsel *evsel, double avg)
687 struct perf_stat_evsel *ps;
693 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
696 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
698 switch (stat_config.aggr_mode) {
700 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
701 cpu_map__id_to_socket(id),
703 cpu_map__id_to_cpu(id),
710 fprintf(stat_config.output, "S%*d%s%*d%s",
719 fprintf(stat_config.output, "CPU%*d%s",
721 perf_evsel__cpus(evsel)->map[id], csv_sep);
724 fprintf(stat_config.output, "%*s-%*d%s",
726 thread_map__comm(evsel->threads, id),
728 thread_map__pid(evsel->threads, id),
744 #define METRIC_LEN 35
746 static void new_line_std(void *ctx)
748 struct outstate *os = ctx;
753 static void do_new_line_std(struct outstate *os)
756 fputs(os->prefix, os->fh);
757 if (stat_config.aggr_mode == AGGR_NONE)
758 fprintf(os->fh, " ");
759 if (stat_config.aggr_mode == AGGR_CORE)
760 fprintf(os->fh, " ");
761 if (stat_config.aggr_mode == AGGR_SOCKET)
762 fprintf(os->fh, " ");
763 fprintf(os->fh, " ");
766 static void print_metric_std(void *ctx, const char *color, const char *fmt,
767 const char *unit, double val)
769 struct outstate *os = ctx;
772 bool newline = os->newline;
776 if (unit == NULL || fmt == NULL) {
777 fprintf(out, "%-*s", METRIC_LEN, "");
784 n = fprintf(out, " # ");
786 n += color_fprintf(out, color, fmt, val);
788 n += fprintf(out, fmt, val);
789 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
792 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
794 FILE *output = stat_config.output;
795 double msecs = avg / 1e6;
796 const char *fmt_v, *fmt_n;
799 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
800 fmt_n = csv_output ? "%s" : "%-25s";
802 aggr_printout(evsel, id, nr);
804 scnprintf(name, sizeof(name), "%s%s",
805 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
807 fprintf(output, fmt_v, msecs, csv_sep);
810 fprintf(output, "%s%s", evsel->unit, csv_sep);
812 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
814 fprintf(output, fmt_n, name);
817 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
820 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
822 FILE *output = stat_config.output;
823 double sc = evsel->scale;
827 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
830 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
832 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
835 aggr_printout(evsel, id, nr);
837 fprintf(output, fmt, avg, csv_sep);
840 fprintf(output, "%-*s%s",
841 csv_output ? 0 : unit_width,
842 evsel->unit, csv_sep);
844 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
847 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
850 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
851 char *prefix, u64 run, u64 ena, double noise)
853 struct perf_stat_output_ctx out;
854 struct outstate os = {
855 .fh = stat_config.output,
856 .prefix = prefix ? prefix : ""
858 print_metric_t pm = print_metric_std;
863 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
864 aggr_printout(counter, id, nr);
866 fprintf(stat_config.output, "%*s%s",
868 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
871 fprintf(stat_config.output, "%-*s%s",
872 csv_output ? 0 : unit_width,
873 counter->unit, csv_sep);
875 fprintf(stat_config.output, "%*s",
876 csv_output ? 0 : -25,
877 perf_evsel__name(counter));
880 fprintf(stat_config.output, "%s%s",
881 csv_sep, counter->cgrp->name);
883 print_running(run, ena);
887 if (nsec_counter(counter))
888 nsec_printout(id, nr, counter, uval);
890 abs_printout(id, nr, counter, uval);
892 out.print_metric = pm;
897 perf_stat__print_shadow_stats(counter, uval,
898 stat_config.aggr_mode == AGGR_GLOBAL ? 0 :
899 cpu_map__id_to_cpu(id),
902 print_noise(counter, noise);
903 print_running(run, ena);
906 static void print_aggr(char *prefix)
908 FILE *output = stat_config.output;
909 struct perf_evsel *counter;
910 int cpu, s, s2, id, nr;
914 if (!(aggr_map || aggr_get_id))
917 for (s = 0; s < aggr_map->nr; s++) {
918 id = aggr_map->map[s];
919 evlist__for_each(evsel_list, counter) {
922 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
923 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
926 val += perf_counts(counter->counts, cpu, 0)->val;
927 ena += perf_counts(counter->counts, cpu, 0)->ena;
928 run += perf_counts(counter->counts, cpu, 0)->run;
932 fprintf(output, "%s", prefix);
934 uval = val * counter->scale;
935 printout(id, nr, counter, uval, prefix, run, ena, 1.0);
941 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
943 FILE *output = stat_config.output;
944 int nthreads = thread_map__nr(counter->threads);
945 int ncpus = cpu_map__nr(counter->cpus);
949 for (thread = 0; thread < nthreads; thread++) {
950 u64 ena = 0, run = 0, val = 0;
952 for (cpu = 0; cpu < ncpus; cpu++) {
953 val += perf_counts(counter->counts, cpu, thread)->val;
954 ena += perf_counts(counter->counts, cpu, thread)->ena;
955 run += perf_counts(counter->counts, cpu, thread)->run;
959 fprintf(output, "%s", prefix);
961 uval = val * counter->scale;
962 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
968 * Print out the results of a single counter:
969 * aggregated counts in system-wide mode
971 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
973 FILE *output = stat_config.output;
974 struct perf_stat_evsel *ps = counter->priv;
975 double avg = avg_stats(&ps->res_stats[0]);
977 double avg_enabled, avg_running;
979 avg_enabled = avg_stats(&ps->res_stats[1]);
980 avg_running = avg_stats(&ps->res_stats[2]);
983 fprintf(output, "%s", prefix);
985 uval = avg * counter->scale;
986 printout(-1, 0, counter, uval, prefix, avg_running, avg_enabled, avg);
987 fprintf(output, "\n");
991 * Print out the results of a single counter:
992 * does not use aggregated count in system-wide
994 static void print_counter(struct perf_evsel *counter, char *prefix)
996 FILE *output = stat_config.output;
1001 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1002 val = perf_counts(counter->counts, cpu, 0)->val;
1003 ena = perf_counts(counter->counts, cpu, 0)->ena;
1004 run = perf_counts(counter->counts, cpu, 0)->run;
1007 fprintf(output, "%s", prefix);
1009 uval = val * counter->scale;
1010 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1012 fputc('\n', output);
1016 static void print_interval(char *prefix, struct timespec *ts)
1018 FILE *output = stat_config.output;
1019 static int num_print_interval;
1021 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1023 if (num_print_interval == 0 && !csv_output) {
1024 switch (stat_config.aggr_mode) {
1026 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
1029 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
1032 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
1035 fprintf(output, "# time comm-pid counts %*s events\n", unit_width, "unit");
1039 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
1045 if (++num_print_interval == 25)
1046 num_print_interval = 0;
1049 static void print_header(int argc, const char **argv)
1051 FILE *output = stat_config.output;
1057 fprintf(output, "\n");
1058 fprintf(output, " Performance counter stats for ");
1059 if (target.system_wide)
1060 fprintf(output, "\'system wide");
1061 else if (target.cpu_list)
1062 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1063 else if (!target__has_task(&target)) {
1064 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1065 for (i = 1; argv && (i < argc); i++)
1066 fprintf(output, " %s", argv[i]);
1067 } else if (target.pid)
1068 fprintf(output, "process id \'%s", target.pid);
1070 fprintf(output, "thread id \'%s", target.tid);
1072 fprintf(output, "\'");
1074 fprintf(output, " (%d runs)", run_count);
1075 fprintf(output, ":\n\n");
1079 static void print_footer(void)
1081 FILE *output = stat_config.output;
1084 fprintf(output, "\n");
1085 fprintf(output, " %17.9f seconds time elapsed",
1086 avg_stats(&walltime_nsecs_stats)/1e9);
1087 if (run_count > 1) {
1088 fprintf(output, " ");
1089 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1090 avg_stats(&walltime_nsecs_stats));
1092 fprintf(output, "\n\n");
1095 static void print_counters(struct timespec *ts, int argc, const char **argv)
1097 int interval = stat_config.interval;
1098 struct perf_evsel *counter;
1099 char buf[64], *prefix = NULL;
1101 /* Do not print anything if we record to the pipe. */
1102 if (STAT_RECORD && perf_stat.file.is_pipe)
1106 print_interval(prefix = buf, ts);
1108 print_header(argc, argv);
1110 switch (stat_config.aggr_mode) {
1116 evlist__for_each(evsel_list, counter)
1117 print_aggr_thread(counter, prefix);
1120 evlist__for_each(evsel_list, counter)
1121 print_counter_aggr(counter, prefix);
1124 evlist__for_each(evsel_list, counter)
1125 print_counter(counter, prefix);
1132 if (!interval && !csv_output)
1135 fflush(stat_config.output);
1138 static volatile int signr = -1;
1140 static void skip_signal(int signo)
1142 if ((child_pid == -1) || stat_config.interval)
1147 * render child_pid harmless
1148 * won't send SIGTERM to a random
1149 * process in case of race condition
1150 * and fast PID recycling
1155 static void sig_atexit(void)
1160 * avoid race condition with SIGCHLD handler
1161 * in skip_signal() which is modifying child_pid
1162 * goal is to avoid send SIGTERM to a random
1166 sigaddset(&set, SIGCHLD);
1167 sigprocmask(SIG_BLOCK, &set, &oset);
1169 if (child_pid != -1)
1170 kill(child_pid, SIGTERM);
1172 sigprocmask(SIG_SETMASK, &oset, NULL);
1177 signal(signr, SIG_DFL);
1178 kill(getpid(), signr);
1181 static int stat__set_big_num(const struct option *opt __maybe_unused,
1182 const char *s __maybe_unused, int unset)
1184 big_num_opt = unset ? 0 : 1;
1188 static const struct option stat_options[] = {
1189 OPT_BOOLEAN('T', "transaction", &transaction_run,
1190 "hardware transaction statistics"),
1191 OPT_CALLBACK('e', "event", &evsel_list, "event",
1192 "event selector. use 'perf list' to list available events",
1193 parse_events_option),
1194 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1195 "event filter", parse_filter),
1196 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1197 "child tasks do not inherit counters"),
1198 OPT_STRING('p', "pid", &target.pid, "pid",
1199 "stat events on existing process id"),
1200 OPT_STRING('t', "tid", &target.tid, "tid",
1201 "stat events on existing thread id"),
1202 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1203 "system-wide collection from all CPUs"),
1204 OPT_BOOLEAN('g', "group", &group,
1205 "put the counters into a counter group"),
1206 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1207 OPT_INCR('v', "verbose", &verbose,
1208 "be more verbose (show counter open errors, etc)"),
1209 OPT_INTEGER('r', "repeat", &run_count,
1210 "repeat command and print average + stddev (max: 100, forever: 0)"),
1211 OPT_BOOLEAN('n', "null", &null_run,
1212 "null run - dont start any counters"),
1213 OPT_INCR('d', "detailed", &detailed_run,
1214 "detailed run - start a lot of events"),
1215 OPT_BOOLEAN('S', "sync", &sync_run,
1216 "call sync() before starting a run"),
1217 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1218 "print large numbers with thousands\' separators",
1220 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1221 "list of cpus to monitor in system-wide"),
1222 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1223 "disable CPU count aggregation", AGGR_NONE),
1224 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1225 "print counts with custom separator"),
1226 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1227 "monitor event in cgroup name only", parse_cgroups),
1228 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1229 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1230 OPT_INTEGER(0, "log-fd", &output_fd,
1231 "log output to fd, instead of stderr"),
1232 OPT_STRING(0, "pre", &pre_cmd, "command",
1233 "command to run prior to the measured command"),
1234 OPT_STRING(0, "post", &post_cmd, "command",
1235 "command to run after to the measured command"),
1236 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1237 "print counts at regular interval in ms (>= 10)"),
1238 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1239 "aggregate counts per processor socket", AGGR_SOCKET),
1240 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1241 "aggregate counts per physical processor core", AGGR_CORE),
1242 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1243 "aggregate counts per thread", AGGR_THREAD),
1244 OPT_UINTEGER('D', "delay", &initial_delay,
1245 "ms to wait before starting measurement after program start"),
1249 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1251 return cpu_map__get_socket(map, cpu, NULL);
1254 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1256 return cpu_map__get_core(map, cpu, NULL);
1259 static int cpu_map__get_max(struct cpu_map *map)
1263 for (i = 0; i < map->nr; i++) {
1264 if (map->map[i] > max)
1271 static struct cpu_map *cpus_aggr_map;
1273 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1280 cpu = map->map[idx];
1282 if (cpus_aggr_map->map[cpu] == -1)
1283 cpus_aggr_map->map[cpu] = get_id(map, idx);
1285 return cpus_aggr_map->map[cpu];
1288 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1290 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1293 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1295 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1298 static int perf_stat_init_aggr_mode(void)
1302 switch (stat_config.aggr_mode) {
1304 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1305 perror("cannot build socket map");
1308 aggr_get_id = perf_stat__get_socket_cached;
1311 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1312 perror("cannot build core map");
1315 aggr_get_id = perf_stat__get_core_cached;
1326 * The evsel_list->cpus is the base we operate on,
1327 * taking the highest cpu number to be the size of
1328 * the aggregation translate cpumap.
1330 nr = cpu_map__get_max(evsel_list->cpus);
1331 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1332 return cpus_aggr_map ? 0 : -ENOMEM;
1335 static void perf_stat__exit_aggr_mode(void)
1337 cpu_map__put(aggr_map);
1338 cpu_map__put(cpus_aggr_map);
1340 cpus_aggr_map = NULL;
1343 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1350 cpu = map->map[idx];
1352 if (cpu >= env->nr_cpus_online)
1358 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1360 struct perf_env *env = data;
1361 int cpu = perf_env__get_cpu(env, map, idx);
1363 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1366 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1368 struct perf_env *env = data;
1369 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1372 int socket_id = env->cpu[cpu].socket_id;
1375 * Encode socket in upper 16 bits
1376 * core_id is relative to socket, and
1377 * we need a global id. So we combine
1380 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1386 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1387 struct cpu_map **sockp)
1389 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1392 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1393 struct cpu_map **corep)
1395 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1398 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1400 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1403 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1405 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1408 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1410 struct perf_env *env = &st->session->header.env;
1412 switch (stat_config.aggr_mode) {
1414 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1415 perror("cannot build socket map");
1418 aggr_get_id = perf_stat__get_socket_file;
1421 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1422 perror("cannot build core map");
1425 aggr_get_id = perf_stat__get_core_file;
1439 * Add default attributes, if there were no attributes specified or
1440 * if -d/--detailed, -d -d or -d -d -d is used:
1442 static int add_default_attributes(void)
1444 struct perf_event_attr default_attrs[] = {
1446 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1447 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1448 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1449 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1451 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1452 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1453 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1454 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1455 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1456 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1461 * Detailed stats (-d), covering the L1 and last level data caches:
1463 struct perf_event_attr detailed_attrs[] = {
1465 { .type = PERF_TYPE_HW_CACHE,
1467 PERF_COUNT_HW_CACHE_L1D << 0 |
1468 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1469 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1471 { .type = PERF_TYPE_HW_CACHE,
1473 PERF_COUNT_HW_CACHE_L1D << 0 |
1474 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1475 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1477 { .type = PERF_TYPE_HW_CACHE,
1479 PERF_COUNT_HW_CACHE_LL << 0 |
1480 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1481 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1483 { .type = PERF_TYPE_HW_CACHE,
1485 PERF_COUNT_HW_CACHE_LL << 0 |
1486 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1487 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1491 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1493 struct perf_event_attr very_detailed_attrs[] = {
1495 { .type = PERF_TYPE_HW_CACHE,
1497 PERF_COUNT_HW_CACHE_L1I << 0 |
1498 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1499 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1501 { .type = PERF_TYPE_HW_CACHE,
1503 PERF_COUNT_HW_CACHE_L1I << 0 |
1504 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1505 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1507 { .type = PERF_TYPE_HW_CACHE,
1509 PERF_COUNT_HW_CACHE_DTLB << 0 |
1510 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1511 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1513 { .type = PERF_TYPE_HW_CACHE,
1515 PERF_COUNT_HW_CACHE_DTLB << 0 |
1516 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1517 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1519 { .type = PERF_TYPE_HW_CACHE,
1521 PERF_COUNT_HW_CACHE_ITLB << 0 |
1522 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1523 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1525 { .type = PERF_TYPE_HW_CACHE,
1527 PERF_COUNT_HW_CACHE_ITLB << 0 |
1528 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1529 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1534 * Very, very detailed stats (-d -d -d), adding prefetch events:
1536 struct perf_event_attr very_very_detailed_attrs[] = {
1538 { .type = PERF_TYPE_HW_CACHE,
1540 PERF_COUNT_HW_CACHE_L1D << 0 |
1541 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1542 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1544 { .type = PERF_TYPE_HW_CACHE,
1546 PERF_COUNT_HW_CACHE_L1D << 0 |
1547 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1548 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1551 /* Set attrs if no event is selected and !null_run: */
1555 if (transaction_run) {
1557 if (pmu_have_event("cpu", "cycles-ct") &&
1558 pmu_have_event("cpu", "el-start"))
1559 err = parse_events(evsel_list, transaction_attrs, NULL);
1561 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1563 fprintf(stderr, "Cannot set up transaction events\n");
1569 if (!evsel_list->nr_entries) {
1570 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1574 /* Detailed events get appended to the event list: */
1576 if (detailed_run < 1)
1579 /* Append detailed run extra attributes: */
1580 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1583 if (detailed_run < 2)
1586 /* Append very detailed run extra attributes: */
1587 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1590 if (detailed_run < 3)
1593 /* Append very, very detailed run extra attributes: */
1594 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1597 static const char * const stat_record_usage[] = {
1598 "perf stat record [<options>]",
1602 static void init_features(struct perf_session *session)
1606 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1607 perf_header__set_feat(&session->header, feat);
1609 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1610 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1611 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1612 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1615 static int __cmd_record(int argc, const char **argv)
1617 struct perf_session *session;
1618 struct perf_data_file *file = &perf_stat.file;
1620 argc = parse_options(argc, argv, stat_options, stat_record_usage,
1621 PARSE_OPT_STOP_AT_NON_OPTION);
1624 file->path = output_name;
1626 if (run_count != 1 || forever) {
1627 pr_err("Cannot use -r option with perf stat record.\n");
1631 session = perf_session__new(file, false, NULL);
1632 if (session == NULL) {
1633 pr_err("Perf session creation failed.\n");
1637 init_features(session);
1639 session->evlist = evsel_list;
1640 perf_stat.session = session;
1641 perf_stat.record = true;
1645 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
1646 union perf_event *event,
1647 struct perf_session *session)
1649 struct stat_round_event *round = &event->stat_round;
1650 struct perf_evsel *counter;
1651 struct timespec tsh, *ts = NULL;
1652 const char **argv = session->header.env.cmdline_argv;
1653 int argc = session->header.env.nr_cmdline;
1655 evlist__for_each(evsel_list, counter)
1656 perf_stat_process_counter(&stat_config, counter);
1658 if (round->type == PERF_STAT_ROUND_TYPE__FINAL)
1659 update_stats(&walltime_nsecs_stats, round->time);
1661 if (stat_config.interval && round->time) {
1662 tsh.tv_sec = round->time / NSECS_PER_SEC;
1663 tsh.tv_nsec = round->time % NSECS_PER_SEC;
1667 print_counters(ts, argc, argv);
1672 int process_stat_config_event(struct perf_tool *tool __maybe_unused,
1673 union perf_event *event,
1674 struct perf_session *session __maybe_unused)
1676 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1678 perf_event__read_stat_config(&stat_config, &event->stat_config);
1680 if (cpu_map__empty(st->cpus)) {
1681 if (st->aggr_mode != AGGR_UNSET)
1682 pr_warning("warning: processing task data, aggregation mode not set\n");
1686 if (st->aggr_mode != AGGR_UNSET)
1687 stat_config.aggr_mode = st->aggr_mode;
1689 if (perf_stat.file.is_pipe)
1690 perf_stat_init_aggr_mode();
1692 perf_stat_init_aggr_mode_file(st);
1697 static int set_maps(struct perf_stat *st)
1699 if (!st->cpus || !st->threads)
1702 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1705 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
1707 if (perf_evlist__alloc_stats(evsel_list, true))
1710 st->maps_allocated = true;
1715 int process_thread_map_event(struct perf_tool *tool __maybe_unused,
1716 union perf_event *event,
1717 struct perf_session *session __maybe_unused)
1719 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1722 pr_warning("Extra thread map event, ignoring.\n");
1726 st->threads = thread_map__new_event(&event->thread_map);
1730 return set_maps(st);
1734 int process_cpu_map_event(struct perf_tool *tool __maybe_unused,
1735 union perf_event *event,
1736 struct perf_session *session __maybe_unused)
1738 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1739 struct cpu_map *cpus;
1742 pr_warning("Extra cpu map event, ignoring.\n");
1746 cpus = cpu_map__new_data(&event->cpu_map.data);
1751 return set_maps(st);
1754 static const char * const stat_report_usage[] = {
1755 "perf stat report [<options>]",
1759 static struct perf_stat perf_stat = {
1761 .attr = perf_event__process_attr,
1762 .event_update = perf_event__process_event_update,
1763 .thread_map = process_thread_map_event,
1764 .cpu_map = process_cpu_map_event,
1765 .stat_config = process_stat_config_event,
1766 .stat = perf_event__process_stat_event,
1767 .stat_round = process_stat_round_event,
1769 .aggr_mode = AGGR_UNSET,
1772 static int __cmd_report(int argc, const char **argv)
1774 struct perf_session *session;
1775 const struct option options[] = {
1776 OPT_STRING('i', "input", &input_name, "file", "input file name"),
1777 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1778 "aggregate counts per processor socket", AGGR_SOCKET),
1779 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1780 "aggregate counts per physical processor core", AGGR_CORE),
1781 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1782 "disable CPU count aggregation", AGGR_NONE),
1788 argc = parse_options(argc, argv, options, stat_report_usage, 0);
1790 if (!input_name || !strlen(input_name)) {
1791 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1794 input_name = "perf.data";
1797 perf_stat.file.path = input_name;
1798 perf_stat.file.mode = PERF_DATA_MODE_READ;
1800 session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
1801 if (session == NULL)
1804 perf_stat.session = session;
1805 stat_config.output = stderr;
1806 evsel_list = session->evlist;
1808 ret = perf_session__process_events(session);
1812 perf_session__delete(session);
1816 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1818 const char * const stat_usage[] = {
1819 "perf stat [<options>] [<command>]",
1822 int status = -EINVAL, run_idx;
1824 FILE *output = stderr;
1825 unsigned int interval;
1826 const char * const stat_subcommands[] = { "record", "report" };
1828 setlocale(LC_ALL, "");
1830 evsel_list = perf_evlist__new();
1831 if (evsel_list == NULL)
1834 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1835 (const char **) stat_usage,
1836 PARSE_OPT_STOP_AT_NON_OPTION);
1840 if (!strcmp(csv_sep, "\\t"))
1843 csv_sep = DEFAULT_SEPARATOR;
1845 if (argc && !strncmp(argv[0], "rec", 3)) {
1846 argc = __cmd_record(argc, argv);
1849 } else if (argc && !strncmp(argv[0], "rep", 3))
1850 return __cmd_report(argc, argv);
1852 interval = stat_config.interval;
1855 * For record command the -o is already taken care of.
1857 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1860 if (output_name && output_fd) {
1861 fprintf(stderr, "cannot use both --output and --log-fd\n");
1862 parse_options_usage(stat_usage, stat_options, "o", 1);
1863 parse_options_usage(NULL, stat_options, "log-fd", 0);
1867 if (output_fd < 0) {
1868 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1869 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1875 mode = append_file ? "a" : "w";
1877 output = fopen(output_name, mode);
1879 perror("failed to create output file");
1882 clock_gettime(CLOCK_REALTIME, &tm);
1883 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1884 } else if (output_fd > 0) {
1885 mode = append_file ? "a" : "w";
1886 output = fdopen(output_fd, mode);
1888 perror("Failed opening logfd");
1893 stat_config.output = output;
1896 * let the spreadsheet do the pretty-printing
1899 /* User explicitly passed -B? */
1900 if (big_num_opt == 1) {
1901 fprintf(stderr, "-B option not supported with -x\n");
1902 parse_options_usage(stat_usage, stat_options, "B", 1);
1903 parse_options_usage(NULL, stat_options, "x", 1);
1905 } else /* Nope, so disable big number formatting */
1907 } else if (big_num_opt == 0) /* User passed --no-big-num */
1910 if (!argc && target__none(&target))
1911 usage_with_options(stat_usage, stat_options);
1913 if (run_count < 0) {
1914 pr_err("Run count must be a positive number\n");
1915 parse_options_usage(stat_usage, stat_options, "r", 1);
1917 } else if (run_count == 0) {
1922 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
1923 fprintf(stderr, "The --per-thread option is only available "
1924 "when monitoring via -p -t options.\n");
1925 parse_options_usage(NULL, stat_options, "p", 1);
1926 parse_options_usage(NULL, stat_options, "t", 1);
1931 * no_aggr, cgroup are for system-wide only
1932 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1934 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1935 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1936 !target__has_cpu(&target)) {
1937 fprintf(stderr, "both cgroup and no-aggregation "
1938 "modes only available in system-wide mode\n");
1940 parse_options_usage(stat_usage, stat_options, "G", 1);
1941 parse_options_usage(NULL, stat_options, "A", 1);
1942 parse_options_usage(NULL, stat_options, "a", 1);
1946 if (add_default_attributes())
1949 target__validate(&target);
1951 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1952 if (target__has_task(&target)) {
1953 pr_err("Problems finding threads of monitor\n");
1954 parse_options_usage(stat_usage, stat_options, "p", 1);
1955 parse_options_usage(NULL, stat_options, "t", 1);
1956 } else if (target__has_cpu(&target)) {
1957 perror("failed to parse CPUs map");
1958 parse_options_usage(stat_usage, stat_options, "C", 1);
1959 parse_options_usage(NULL, stat_options, "a", 1);
1965 * Initialize thread_map with comm names,
1966 * so we could print it out on output.
1968 if (stat_config.aggr_mode == AGGR_THREAD)
1969 thread_map__read_comms(evsel_list->threads);
1971 if (interval && interval < 100) {
1972 if (interval < 10) {
1973 pr_err("print interval must be >= 10ms\n");
1974 parse_options_usage(stat_usage, stat_options, "I", 1);
1977 pr_warning("print interval < 100ms. "
1978 "The overhead percentage could be high in some cases. "
1979 "Please proceed with caution.\n");
1982 if (perf_evlist__alloc_stats(evsel_list, interval))
1985 if (perf_stat_init_aggr_mode())
1989 * We dont want to block the signals - that would cause
1990 * child tasks to inherit that and Ctrl-C would not work.
1991 * What we want is for Ctrl-C to work in the exec()-ed
1992 * task, but being ignored by perf stat itself:
1996 signal(SIGINT, skip_signal);
1997 signal(SIGCHLD, skip_signal);
1998 signal(SIGALRM, skip_signal);
1999 signal(SIGABRT, skip_signal);
2002 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2003 if (run_count != 1 && verbose)
2004 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2007 status = run_perf_stat(argc, argv);
2008 if (forever && status != -1) {
2009 print_counters(NULL, argc, argv);
2010 perf_stat__reset_stats();
2014 if (!forever && status != -1 && !interval)
2015 print_counters(NULL, argc, argv);
2019 * We synthesize the kernel mmap record just so that older tools
2020 * don't emit warnings about not being able to resolve symbols
2021 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2022 * a saner message about no samples being in the perf.data file.
2024 * This also serves to suppress a warning about f_header.data.size == 0
2025 * in header.c at the moment 'perf stat record' gets introduced, which
2026 * is not really needed once we start adding the stat specific PERF_RECORD_
2027 * records, but the need to suppress the kptr_restrict messages in older
2028 * tools remain -acme
2030 int fd = perf_data_file__fd(&perf_stat.file);
2031 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2032 process_synthesized_event,
2033 &perf_stat.session->machines.host);
2035 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2036 "older tools may produce warnings about this file\n.");
2040 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2041 pr_err("failed to write stat round event\n");
2044 if (!perf_stat.file.is_pipe) {
2045 perf_stat.session->header.data_size += perf_stat.bytes_written;
2046 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2049 perf_session__delete(perf_stat.session);
2052 perf_stat__exit_aggr_mode();
2053 perf_evlist__free_stats(evsel_list);
2055 perf_evlist__delete(evsel_list);