1 // SPDX-License-Identifier: GPL-2.0-only
5 * Builtin stat command: Give a precise performance counters summary
6 * overview about any workload, CPU or specific PID.
10 $ perf stat ./hackbench 10
14 Performance counter stats for './hackbench 10':
16 1708.761321 task-clock # 11.037 CPUs utilized
17 41,190 context-switches # 0.024 M/sec
18 6,735 CPU-migrations # 0.004 M/sec
19 17,318 page-faults # 0.010 M/sec
20 5,205,202,243 cycles # 3.046 GHz
21 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
22 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
23 2,603,501,247 instructions # 0.50 insns per cycle
24 # 1.48 stalled cycles per insn
25 484,357,498 branches # 283.455 M/sec
26 6,388,934 branch-misses # 1.32% of all branches
28 0.154822978 seconds time elapsed
31 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
33 * Improvements and fixes by:
35 * Arjan van de Ven <arjan@linux.intel.com>
36 * Yanmin Zhang <yanmin.zhang@intel.com>
37 * Wu Fengguang <fengguang.wu@intel.com>
38 * Mike Galbraith <efault@gmx.de>
39 * Paul Mackerras <paulus@samba.org>
40 * Jaswinder Singh Rajput <jaswinder@kernel.org>
45 #include "util/cgroup.h"
46 #include <subcmd/parse-options.h>
47 #include "util/parse-events.h"
49 #include "util/event.h"
50 #include "util/evlist.h"
51 #include "util/evsel.h"
52 #include "util/debug.h"
53 #include "util/color.h"
54 #include "util/stat.h"
55 #include "util/header.h"
56 #include "util/cpumap.h"
57 #include "util/thread.h"
58 #include "util/thread_map.h"
59 #include "util/counts.h"
60 #include "util/group.h"
61 #include "util/session.h"
62 #include "util/tool.h"
63 #include "util/string2.h"
64 #include "util/metricgroup.h"
68 #include <linux/time64.h>
69 #include <linux/zalloc.h>
70 #include <api/fs/fs.h>
74 #include <sys/prctl.h>
78 #include <sys/types.h>
83 #include <sys/resource.h>
85 #include <linux/ctype.h>
87 #define DEFAULT_SEPARATOR " "
88 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
90 static void print_counters(struct timespec *ts, int argc, const char **argv);
92 /* Default events used for perf stat -T */
93 static const char *transaction_attrs = {
105 /* More limited version when the CPU does not have all events. */
106 static const char * transaction_limited_attrs = {
116 static const char * topdown_attrs[] = {
117 "topdown-total-slots",
118 "topdown-slots-retired",
119 "topdown-recovery-bubbles",
120 "topdown-fetch-bubbles",
121 "topdown-slots-issued",
125 static const char *smi_cost_attrs = {
133 static struct perf_evlist *evsel_list;
135 static struct target target = {
139 #define METRIC_ONLY_LEN 20
141 static volatile pid_t child_pid = -1;
142 static int detailed_run = 0;
143 static bool transaction_run;
144 static bool topdown_run = false;
145 static bool smi_cost = false;
146 static bool smi_reset = false;
147 static int big_num_opt = -1;
148 static bool group = false;
149 static const char *pre_cmd = NULL;
150 static const char *post_cmd = NULL;
151 static bool sync_run = false;
152 static bool forever = false;
153 static bool force_metric_only = false;
154 static struct timespec ref_time;
155 static bool append_file;
156 static bool interval_count;
157 static const char *output_name;
158 static int output_fd;
162 struct perf_data data;
163 struct perf_session *session;
165 struct perf_tool tool;
167 struct cpu_map *cpus;
168 struct thread_map *threads;
169 enum aggr_mode aggr_mode;
172 static struct perf_stat perf_stat;
173 #define STAT_RECORD perf_stat.record
175 static volatile int done = 0;
177 static struct perf_stat_config stat_config = {
178 .aggr_mode = AGGR_GLOBAL,
180 .unit_width = 4, /* strlen("unit") */
182 .metric_only_len = METRIC_ONLY_LEN,
183 .walltime_nsecs_stats = &walltime_nsecs_stats,
187 static inline void diff_timespec(struct timespec *r, struct timespec *a,
190 r->tv_sec = a->tv_sec - b->tv_sec;
191 if (a->tv_nsec < b->tv_nsec) {
192 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
195 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
199 static void perf_stat__reset_stats(void)
203 perf_evlist__reset_stats(evsel_list);
204 perf_stat__reset_shadow_stats();
206 for (i = 0; i < stat_config.stats_num; i++)
207 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
210 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
211 union perf_event *event,
212 struct perf_sample *sample __maybe_unused,
213 struct machine *machine __maybe_unused)
215 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
216 pr_err("failed to write perf data, error: %m\n");
220 perf_stat.bytes_written += event->header.size;
224 static int write_stat_round_event(u64 tm, u64 type)
226 return perf_event__synthesize_stat_round(NULL, tm, type,
227 process_synthesized_event,
231 #define WRITE_STAT_ROUND_EVENT(time, interval) \
232 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
234 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
237 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
238 struct perf_counts_values *count)
240 struct perf_sample_id *sid = SID(counter, cpu, thread);
242 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
243 process_synthesized_event, NULL);
246 static int read_single_counter(struct perf_evsel *counter, int cpu,
247 int thread, struct timespec *rs)
249 if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
250 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
251 struct perf_counts_values *count =
252 perf_counts(counter->counts, cpu, thread);
253 count->ena = count->run = val;
257 return perf_evsel__read_counter(counter, cpu, thread);
261 * Read out the results of a single counter:
262 * do not aggregate counts across CPUs in system-wide mode
264 static int read_counter(struct perf_evsel *counter, struct timespec *rs)
266 int nthreads = thread_map__nr(evsel_list->threads);
267 int ncpus, cpu, thread;
269 if (target__has_cpu(&target) && !target__has_per_thread(&target))
270 ncpus = perf_evsel__nr_cpus(counter);
274 if (!counter->supported)
277 if (counter->system_wide)
280 for (thread = 0; thread < nthreads; thread++) {
281 for (cpu = 0; cpu < ncpus; cpu++) {
282 struct perf_counts_values *count;
284 count = perf_counts(counter->counts, cpu, thread);
287 * The leader's group read loads data into its group members
288 * (via perf_evsel__read_counter) and sets threir count->loaded.
290 if (!count->loaded &&
291 read_single_counter(counter, cpu, thread, rs)) {
292 counter->counts->scaled = -1;
293 perf_counts(counter->counts, cpu, thread)->ena = 0;
294 perf_counts(counter->counts, cpu, thread)->run = 0;
298 count->loaded = false;
301 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
302 pr_err("failed to write stat event\n");
308 fprintf(stat_config.output,
309 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
310 perf_evsel__name(counter),
312 count->val, count->ena, count->run);
320 static void read_counters(struct timespec *rs)
322 struct perf_evsel *counter;
325 evlist__for_each_entry(evsel_list, counter) {
326 ret = read_counter(counter, rs);
328 pr_debug("failed to read counter %s\n", counter->name);
330 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
331 pr_warning("failed to process counter %s\n", counter->name);
335 static void process_interval(void)
337 struct timespec ts, rs;
339 clock_gettime(CLOCK_MONOTONIC, &ts);
340 diff_timespec(&rs, &ts, &ref_time);
345 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
346 pr_err("failed to write stat round event\n");
349 init_stats(&walltime_nsecs_stats);
350 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
351 print_counters(&rs, 0, NULL);
354 static void enable_counters(void)
356 if (stat_config.initial_delay)
357 usleep(stat_config.initial_delay * USEC_PER_MSEC);
360 * We need to enable counters only if:
361 * - we don't have tracee (attaching to task or cpu)
362 * - we have initial delay configured
364 if (!target__none(&target) || stat_config.initial_delay)
365 perf_evlist__enable(evsel_list);
368 static void disable_counters(void)
371 * If we don't have tracee (attaching to task or cpu), counters may
372 * still be running. To get accurate group ratios, we must stop groups
373 * from counting before reading their constituent counters.
375 if (!target__none(&target))
376 perf_evlist__disable(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 perf_evsel__should_store_id(struct perf_evsel *counter)
394 return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
397 static bool is_target_alive(struct target *_target,
398 struct thread_map *threads)
403 if (!target__has_task(_target))
406 for (i = 0; i < threads->nr; i++) {
409 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
410 threads->map[i].pid);
412 if (!stat(path, &st))
419 static int __run_perf_stat(int argc, const char **argv, int run_idx)
421 int interval = stat_config.interval;
422 int times = stat_config.times;
423 int timeout = stat_config.timeout;
425 unsigned long long t0, t1;
426 struct perf_evsel *counter;
430 const bool forks = (argc > 0);
431 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
434 ts.tv_sec = interval / USEC_PER_MSEC;
435 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
436 } else if (timeout) {
437 ts.tv_sec = timeout / USEC_PER_MSEC;
438 ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
445 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
446 workload_exec_failed_signal) < 0) {
447 perror("failed to prepare workload");
450 child_pid = evsel_list->workload.pid;
454 perf_evlist__set_leader(evsel_list);
456 evlist__for_each_entry(evsel_list, counter) {
458 if (create_perf_stat_counter(counter, &stat_config, &target) < 0) {
460 /* Weak group failed. Reset the group. */
461 if ((errno == EINVAL || errno == EBADF) &&
462 counter->leader != counter &&
463 counter->weak_group) {
464 counter = perf_evlist__reset_weak_group(evsel_list, counter);
469 * PPC returns ENXIO for HW counters until 2.6.37
470 * (behavior changed with commit b0a873e).
472 if (errno == EINVAL || errno == ENOSYS ||
473 errno == ENOENT || errno == EOPNOTSUPP ||
476 ui__warning("%s event is not supported by the kernel.\n",
477 perf_evsel__name(counter));
478 counter->supported = false;
480 if ((counter->leader != counter) ||
481 !(counter->leader->nr_members > 1))
483 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
485 ui__warning("%s\n", msg);
487 } else if (target__has_per_thread(&target) &&
488 evsel_list->threads &&
489 evsel_list->threads->err_thread != -1) {
491 * For global --per-thread case, skip current
494 if (!thread_map__remove(evsel_list->threads,
495 evsel_list->threads->err_thread)) {
496 evsel_list->threads->err_thread = -1;
501 perf_evsel__open_strerror(counter, &target,
502 errno, msg, sizeof(msg));
503 ui__error("%s\n", msg);
506 kill(child_pid, SIGTERM);
510 counter->supported = true;
512 l = strlen(counter->unit);
513 if (l > stat_config.unit_width)
514 stat_config.unit_width = l;
516 if (perf_evsel__should_store_id(counter) &&
517 perf_evsel__store_ids(counter, evsel_list))
521 if (perf_evlist__apply_filters(evsel_list, &counter)) {
522 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
523 counter->filter, perf_evsel__name(counter), errno,
524 str_error_r(errno, msg, sizeof(msg)));
529 int err, fd = perf_data__fd(&perf_stat.data);
532 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
534 err = perf_session__write_header(perf_stat.session, evsel_list,
541 err = perf_stat_synthesize_config(&stat_config, NULL, evsel_list,
542 process_synthesized_event, is_pipe);
548 * Enable counters and exec the command:
551 clock_gettime(CLOCK_MONOTONIC, &ref_time);
554 perf_evlist__start_workload(evsel_list);
557 if (interval || timeout) {
558 while (!waitpid(child_pid, &status, WNOHANG)) {
559 nanosleep(&ts, NULL);
563 if (interval_count && !(--times))
568 wait4(child_pid, &status, 0, &stat_config.ru_data);
570 if (workload_exec_errno) {
571 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
572 pr_err("Workload failed: %s\n", emsg);
576 if (WIFSIGNALED(status))
577 psignal(WTERMSIG(status), argv[0]);
581 nanosleep(&ts, NULL);
582 if (!is_target_alive(&target, evsel_list->threads))
588 if (interval_count && !(--times))
598 if (stat_config.walltime_run_table)
599 stat_config.walltime_run[run_idx] = t1 - t0;
601 update_stats(&walltime_nsecs_stats, t1 - t0);
604 * Closing a group leader splits the group, and as we only disable
605 * group leaders, results in remaining events becoming enabled. To
606 * avoid arbitrary skew, we must read all counters before closing any
609 read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
612 * We need to keep evsel_list alive, because it's processed
613 * later the evsel_list will be closed after.
616 perf_evlist__close(evsel_list);
618 return WEXITSTATUS(status);
621 static int run_perf_stat(int argc, const char **argv, int run_idx)
626 ret = system(pre_cmd);
634 ret = __run_perf_stat(argc, argv, run_idx);
639 ret = system(post_cmd);
647 static void print_counters(struct timespec *ts, int argc, const char **argv)
649 /* Do not print anything if we record to the pipe. */
650 if (STAT_RECORD && perf_stat.data.is_pipe)
653 perf_evlist__print_counters(evsel_list, &stat_config, &target,
657 static volatile int signr = -1;
659 static void skip_signal(int signo)
661 if ((child_pid == -1) || stat_config.interval)
666 * render child_pid harmless
667 * won't send SIGTERM to a random
668 * process in case of race condition
669 * and fast PID recycling
674 static void sig_atexit(void)
679 * avoid race condition with SIGCHLD handler
680 * in skip_signal() which is modifying child_pid
681 * goal is to avoid send SIGTERM to a random
685 sigaddset(&set, SIGCHLD);
686 sigprocmask(SIG_BLOCK, &set, &oset);
689 kill(child_pid, SIGTERM);
691 sigprocmask(SIG_SETMASK, &oset, NULL);
696 signal(signr, SIG_DFL);
697 kill(getpid(), signr);
700 static int stat__set_big_num(const struct option *opt __maybe_unused,
701 const char *s __maybe_unused, int unset)
703 big_num_opt = unset ? 0 : 1;
707 static int enable_metric_only(const struct option *opt __maybe_unused,
708 const char *s __maybe_unused, int unset)
710 force_metric_only = true;
711 stat_config.metric_only = !unset;
715 static int parse_metric_groups(const struct option *opt,
717 int unset __maybe_unused)
719 return metricgroup__parse_groups(opt, str, &stat_config.metric_events);
722 static struct option stat_options[] = {
723 OPT_BOOLEAN('T', "transaction", &transaction_run,
724 "hardware transaction statistics"),
725 OPT_CALLBACK('e', "event", &evsel_list, "event",
726 "event selector. use 'perf list' to list available events",
727 parse_events_option),
728 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
729 "event filter", parse_filter),
730 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
731 "child tasks do not inherit counters"),
732 OPT_STRING('p', "pid", &target.pid, "pid",
733 "stat events on existing process id"),
734 OPT_STRING('t', "tid", &target.tid, "tid",
735 "stat events on existing thread id"),
736 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
737 "system-wide collection from all CPUs"),
738 OPT_BOOLEAN('g', "group", &group,
739 "put the counters into a counter group"),
740 OPT_BOOLEAN(0, "scale", &stat_config.scale,
741 "Use --no-scale to disable counter scaling for multiplexing"),
742 OPT_INCR('v', "verbose", &verbose,
743 "be more verbose (show counter open errors, etc)"),
744 OPT_INTEGER('r', "repeat", &stat_config.run_count,
745 "repeat command and print average + stddev (max: 100, forever: 0)"),
746 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
747 "display details about each run (only with -r option)"),
748 OPT_BOOLEAN('n', "null", &stat_config.null_run,
749 "null run - dont start any counters"),
750 OPT_INCR('d', "detailed", &detailed_run,
751 "detailed run - start a lot of events"),
752 OPT_BOOLEAN('S', "sync", &sync_run,
753 "call sync() before starting a run"),
754 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
755 "print large numbers with thousands\' separators",
757 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
758 "list of cpus to monitor in system-wide"),
759 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
760 "disable CPU count aggregation", AGGR_NONE),
761 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
762 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
763 "print counts with custom separator"),
764 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
765 "monitor event in cgroup name only", parse_cgroups),
766 OPT_STRING('o', "output", &output_name, "file", "output file name"),
767 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
768 OPT_INTEGER(0, "log-fd", &output_fd,
769 "log output to fd, instead of stderr"),
770 OPT_STRING(0, "pre", &pre_cmd, "command",
771 "command to run prior to the measured command"),
772 OPT_STRING(0, "post", &post_cmd, "command",
773 "command to run after to the measured command"),
774 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
775 "print counts at regular interval in ms "
776 "(overhead is possible for values <= 100ms)"),
777 OPT_INTEGER(0, "interval-count", &stat_config.times,
778 "print counts for fixed number of times"),
779 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
780 "clear screen in between new interval"),
781 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
782 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
783 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
784 "aggregate counts per processor socket", AGGR_SOCKET),
785 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
786 "aggregate counts per processor die", AGGR_DIE),
787 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
788 "aggregate counts per physical processor core", AGGR_CORE),
789 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
790 "aggregate counts per thread", AGGR_THREAD),
791 OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
792 "ms to wait before starting measurement after program start"),
793 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
794 "Only print computed metrics. No raw values", enable_metric_only),
795 OPT_BOOLEAN(0, "topdown", &topdown_run,
796 "measure topdown level 1 statistics"),
797 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
799 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
800 "monitor specified metrics or metric groups (separated by ,)",
801 parse_metric_groups),
805 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
806 struct cpu_map *map, int cpu)
808 return cpu_map__get_socket(map, cpu, NULL);
811 static int perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
812 struct cpu_map *map, int cpu)
814 return cpu_map__get_die(map, cpu, NULL);
817 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
818 struct cpu_map *map, int cpu)
820 return cpu_map__get_core(map, cpu, NULL);
823 static int cpu_map__get_max(struct cpu_map *map)
827 for (i = 0; i < map->nr; i++) {
828 if (map->map[i] > max)
835 static int perf_stat__get_aggr(struct perf_stat_config *config,
836 aggr_get_id_t get_id, struct cpu_map *map, int idx)
845 if (config->cpus_aggr_map->map[cpu] == -1)
846 config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
848 return config->cpus_aggr_map->map[cpu];
851 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
852 struct cpu_map *map, int idx)
854 return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
857 static int perf_stat__get_die_cached(struct perf_stat_config *config,
858 struct cpu_map *map, int idx)
860 return perf_stat__get_aggr(config, perf_stat__get_die, map, idx);
863 static int perf_stat__get_core_cached(struct perf_stat_config *config,
864 struct cpu_map *map, int idx)
866 return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
869 static bool term_percore_set(void)
871 struct perf_evsel *counter;
873 evlist__for_each_entry(evsel_list, counter) {
874 if (counter->percore)
881 static int perf_stat_init_aggr_mode(void)
885 switch (stat_config.aggr_mode) {
887 if (cpu_map__build_socket_map(evsel_list->cpus, &stat_config.aggr_map)) {
888 perror("cannot build socket map");
891 stat_config.aggr_get_id = perf_stat__get_socket_cached;
894 if (cpu_map__build_die_map(evsel_list->cpus, &stat_config.aggr_map)) {
895 perror("cannot build die map");
898 stat_config.aggr_get_id = perf_stat__get_die_cached;
901 if (cpu_map__build_core_map(evsel_list->cpus, &stat_config.aggr_map)) {
902 perror("cannot build core map");
905 stat_config.aggr_get_id = perf_stat__get_core_cached;
908 if (term_percore_set()) {
909 if (cpu_map__build_core_map(evsel_list->cpus,
910 &stat_config.aggr_map)) {
911 perror("cannot build core map");
914 stat_config.aggr_get_id = perf_stat__get_core_cached;
925 * The evsel_list->cpus is the base we operate on,
926 * taking the highest cpu number to be the size of
927 * the aggregation translate cpumap.
929 nr = cpu_map__get_max(evsel_list->cpus);
930 stat_config.cpus_aggr_map = cpu_map__empty_new(nr + 1);
931 return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
934 static void perf_stat__exit_aggr_mode(void)
936 cpu_map__put(stat_config.aggr_map);
937 cpu_map__put(stat_config.cpus_aggr_map);
938 stat_config.aggr_map = NULL;
939 stat_config.cpus_aggr_map = NULL;
942 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
951 if (cpu >= env->nr_cpus_avail)
957 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
959 struct perf_env *env = data;
960 int cpu = perf_env__get_cpu(env, map, idx);
962 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
965 static int perf_env__get_die(struct cpu_map *map, int idx, void *data)
967 struct perf_env *env = data;
968 int die_id = -1, cpu = perf_env__get_cpu(env, map, idx);
972 * Encode socket in bit range 15:8
973 * die_id is relative to socket,
974 * we need a global id. So we combine
977 if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
980 if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
983 die_id = (env->cpu[cpu].socket_id << 8) | (env->cpu[cpu].die_id & 0xff);
989 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
991 struct perf_env *env = data;
992 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
996 * Encode socket in bit range 31:24
997 * encode die id in bit range 23:16
998 * core_id is relative to socket and die,
999 * we need a global id. So we combine
1000 * socket + die id + core id
1002 if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
1005 if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
1008 if (WARN_ONCE(env->cpu[cpu].core_id >> 16, "The core id number is too big.\n"))
1011 core = (env->cpu[cpu].socket_id << 24) |
1012 (env->cpu[cpu].die_id << 16) |
1013 (env->cpu[cpu].core_id & 0xffff);
1019 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1020 struct cpu_map **sockp)
1022 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1025 static int perf_env__build_die_map(struct perf_env *env, struct cpu_map *cpus,
1026 struct cpu_map **diep)
1028 return cpu_map__build_map(cpus, diep, perf_env__get_die, env);
1031 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1032 struct cpu_map **corep)
1034 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1037 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1038 struct cpu_map *map, int idx)
1040 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1042 static int perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1043 struct cpu_map *map, int idx)
1045 return perf_env__get_die(map, idx, &perf_stat.session->header.env);
1048 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1049 struct cpu_map *map, int idx)
1051 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1054 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1056 struct perf_env *env = &st->session->header.env;
1058 switch (stat_config.aggr_mode) {
1060 if (perf_env__build_socket_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
1061 perror("cannot build socket map");
1064 stat_config.aggr_get_id = perf_stat__get_socket_file;
1067 if (perf_env__build_die_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
1068 perror("cannot build die map");
1071 stat_config.aggr_get_id = perf_stat__get_die_file;
1074 if (perf_env__build_core_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
1075 perror("cannot build core map");
1078 stat_config.aggr_get_id = perf_stat__get_core_file;
1091 static int topdown_filter_events(const char **attr, char **str, bool use_group)
1098 for (i = 0; attr[i]; i++) {
1099 if (pmu_have_event("cpu", attr[i])) {
1100 len += strlen(attr[i]) + 1;
1101 attr[i - off] = attr[i];
1105 attr[i - off] = NULL;
1107 *str = malloc(len + 1 + 2);
1117 for (i = 0; attr[i]; i++) {
1130 __weak bool arch_topdown_check_group(bool *warn)
1136 __weak void arch_topdown_group_warn(void)
1141 * Add default attributes, if there were no attributes specified or
1142 * if -d/--detailed, -d -d or -d -d -d is used:
1144 static int add_default_attributes(void)
1147 struct perf_event_attr default_attrs0[] = {
1149 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1150 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1151 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1152 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1154 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1156 struct perf_event_attr frontend_attrs[] = {
1157 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1159 struct perf_event_attr backend_attrs[] = {
1160 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1162 struct perf_event_attr default_attrs1[] = {
1163 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1164 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1165 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1170 * Detailed stats (-d), covering the L1 and last level data caches:
1172 struct perf_event_attr detailed_attrs[] = {
1174 { .type = PERF_TYPE_HW_CACHE,
1176 PERF_COUNT_HW_CACHE_L1D << 0 |
1177 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1178 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1180 { .type = PERF_TYPE_HW_CACHE,
1182 PERF_COUNT_HW_CACHE_L1D << 0 |
1183 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1184 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1186 { .type = PERF_TYPE_HW_CACHE,
1188 PERF_COUNT_HW_CACHE_LL << 0 |
1189 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1190 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1192 { .type = PERF_TYPE_HW_CACHE,
1194 PERF_COUNT_HW_CACHE_LL << 0 |
1195 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1196 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1200 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1202 struct perf_event_attr very_detailed_attrs[] = {
1204 { .type = PERF_TYPE_HW_CACHE,
1206 PERF_COUNT_HW_CACHE_L1I << 0 |
1207 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1208 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1210 { .type = PERF_TYPE_HW_CACHE,
1212 PERF_COUNT_HW_CACHE_L1I << 0 |
1213 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1214 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1216 { .type = PERF_TYPE_HW_CACHE,
1218 PERF_COUNT_HW_CACHE_DTLB << 0 |
1219 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1220 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1222 { .type = PERF_TYPE_HW_CACHE,
1224 PERF_COUNT_HW_CACHE_DTLB << 0 |
1225 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1226 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1228 { .type = PERF_TYPE_HW_CACHE,
1230 PERF_COUNT_HW_CACHE_ITLB << 0 |
1231 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1232 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1234 { .type = PERF_TYPE_HW_CACHE,
1236 PERF_COUNT_HW_CACHE_ITLB << 0 |
1237 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1238 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1243 * Very, very detailed stats (-d -d -d), adding prefetch events:
1245 struct perf_event_attr very_very_detailed_attrs[] = {
1247 { .type = PERF_TYPE_HW_CACHE,
1249 PERF_COUNT_HW_CACHE_L1D << 0 |
1250 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1251 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1253 { .type = PERF_TYPE_HW_CACHE,
1255 PERF_COUNT_HW_CACHE_L1D << 0 |
1256 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1257 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1259 struct parse_events_error errinfo;
1261 /* Set attrs if no event is selected and !null_run: */
1262 if (stat_config.null_run)
1265 if (transaction_run) {
1266 /* Handle -T as -M transaction. Once platform specific metrics
1267 * support has been added to the json files, all archictures
1268 * will use this approach. To determine transaction support
1269 * on an architecture test for such a metric name.
1271 if (metricgroup__has_metric("transaction")) {
1272 struct option opt = { .value = &evsel_list };
1274 return metricgroup__parse_groups(&opt, "transaction",
1275 &stat_config.metric_events);
1278 if (pmu_have_event("cpu", "cycles-ct") &&
1279 pmu_have_event("cpu", "el-start"))
1280 err = parse_events(evsel_list, transaction_attrs,
1283 err = parse_events(evsel_list,
1284 transaction_limited_attrs,
1287 fprintf(stderr, "Cannot set up transaction events\n");
1288 parse_events_print_error(&errinfo, transaction_attrs);
1297 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1298 fprintf(stderr, "freeze_on_smi is not supported.\n");
1303 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1304 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1310 if (pmu_have_event("msr", "aperf") &&
1311 pmu_have_event("msr", "smi")) {
1312 if (!force_metric_only)
1313 stat_config.metric_only = true;
1314 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1316 fprintf(stderr, "To measure SMI cost, it needs "
1317 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1318 parse_events_print_error(&errinfo, smi_cost_attrs);
1322 fprintf(stderr, "Cannot set up SMI cost events\n");
1332 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1333 stat_config.aggr_mode != AGGR_CORE) {
1334 pr_err("top down event configuration requires --per-core mode\n");
1337 stat_config.aggr_mode = AGGR_CORE;
1338 if (nr_cgroups || !target__has_cpu(&target)) {
1339 pr_err("top down event configuration requires system-wide mode (-a)\n");
1343 if (!force_metric_only)
1344 stat_config.metric_only = true;
1345 if (topdown_filter_events(topdown_attrs, &str,
1346 arch_topdown_check_group(&warn)) < 0) {
1347 pr_err("Out of memory\n");
1350 if (topdown_attrs[0] && str) {
1352 arch_topdown_group_warn();
1353 err = parse_events(evsel_list, str, &errinfo);
1356 "Cannot set up top down events %s: %d\n",
1358 parse_events_print_error(&errinfo, str);
1363 fprintf(stderr, "System does not support topdown\n");
1369 if (!evsel_list->nr_entries) {
1370 if (target__has_cpu(&target))
1371 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1373 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1375 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1376 if (perf_evlist__add_default_attrs(evsel_list,
1377 frontend_attrs) < 0)
1380 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1381 if (perf_evlist__add_default_attrs(evsel_list,
1385 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1389 /* Detailed events get appended to the event list: */
1391 if (detailed_run < 1)
1394 /* Append detailed run extra attributes: */
1395 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1398 if (detailed_run < 2)
1401 /* Append very detailed run extra attributes: */
1402 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1405 if (detailed_run < 3)
1408 /* Append very, very detailed run extra attributes: */
1409 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1412 static const char * const stat_record_usage[] = {
1413 "perf stat record [<options>]",
1417 static void init_features(struct perf_session *session)
1421 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1422 perf_header__set_feat(&session->header, feat);
1424 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1425 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1426 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1427 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1428 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1431 static int __cmd_record(int argc, const char **argv)
1433 struct perf_session *session;
1434 struct perf_data *data = &perf_stat.data;
1436 argc = parse_options(argc, argv, stat_options, stat_record_usage,
1437 PARSE_OPT_STOP_AT_NON_OPTION);
1440 data->path = output_name;
1442 if (stat_config.run_count != 1 || forever) {
1443 pr_err("Cannot use -r option with perf stat record.\n");
1447 session = perf_session__new(data, false, NULL);
1448 if (session == NULL) {
1449 pr_err("Perf session creation failed.\n");
1453 init_features(session);
1455 session->evlist = evsel_list;
1456 perf_stat.session = session;
1457 perf_stat.record = true;
1461 static int process_stat_round_event(struct perf_session *session,
1462 union perf_event *event)
1464 struct stat_round_event *stat_round = &event->stat_round;
1465 struct perf_evsel *counter;
1466 struct timespec tsh, *ts = NULL;
1467 const char **argv = session->header.env.cmdline_argv;
1468 int argc = session->header.env.nr_cmdline;
1470 evlist__for_each_entry(evsel_list, counter)
1471 perf_stat_process_counter(&stat_config, counter);
1473 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1474 update_stats(&walltime_nsecs_stats, stat_round->time);
1476 if (stat_config.interval && stat_round->time) {
1477 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
1478 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1482 print_counters(ts, argc, argv);
1487 int process_stat_config_event(struct perf_session *session,
1488 union perf_event *event)
1490 struct perf_tool *tool = session->tool;
1491 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1493 perf_event__read_stat_config(&stat_config, &event->stat_config);
1495 if (cpu_map__empty(st->cpus)) {
1496 if (st->aggr_mode != AGGR_UNSET)
1497 pr_warning("warning: processing task data, aggregation mode not set\n");
1501 if (st->aggr_mode != AGGR_UNSET)
1502 stat_config.aggr_mode = st->aggr_mode;
1504 if (perf_stat.data.is_pipe)
1505 perf_stat_init_aggr_mode();
1507 perf_stat_init_aggr_mode_file(st);
1512 static int set_maps(struct perf_stat *st)
1514 if (!st->cpus || !st->threads)
1517 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1520 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
1522 if (perf_evlist__alloc_stats(evsel_list, true))
1525 st->maps_allocated = true;
1530 int process_thread_map_event(struct perf_session *session,
1531 union perf_event *event)
1533 struct perf_tool *tool = session->tool;
1534 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1537 pr_warning("Extra thread map event, ignoring.\n");
1541 st->threads = thread_map__new_event(&event->thread_map);
1545 return set_maps(st);
1549 int process_cpu_map_event(struct perf_session *session,
1550 union perf_event *event)
1552 struct perf_tool *tool = session->tool;
1553 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1554 struct cpu_map *cpus;
1557 pr_warning("Extra cpu map event, ignoring.\n");
1561 cpus = cpu_map__new_data(&event->cpu_map.data);
1566 return set_maps(st);
1569 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
1573 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
1577 config->stats_num = nthreads;
1579 for (i = 0; i < nthreads; i++)
1580 runtime_stat__init(&config->stats[i]);
1585 static void runtime_stat_delete(struct perf_stat_config *config)
1592 for (i = 0; i < config->stats_num; i++)
1593 runtime_stat__exit(&config->stats[i]);
1595 zfree(&config->stats);
1598 static const char * const stat_report_usage[] = {
1599 "perf stat report [<options>]",
1603 static struct perf_stat perf_stat = {
1605 .attr = perf_event__process_attr,
1606 .event_update = perf_event__process_event_update,
1607 .thread_map = process_thread_map_event,
1608 .cpu_map = process_cpu_map_event,
1609 .stat_config = process_stat_config_event,
1610 .stat = perf_event__process_stat_event,
1611 .stat_round = process_stat_round_event,
1613 .aggr_mode = AGGR_UNSET,
1616 static int __cmd_report(int argc, const char **argv)
1618 struct perf_session *session;
1619 const struct option options[] = {
1620 OPT_STRING('i', "input", &input_name, "file", "input file name"),
1621 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1622 "aggregate counts per processor socket", AGGR_SOCKET),
1623 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
1624 "aggregate counts per processor die", AGGR_DIE),
1625 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1626 "aggregate counts per physical processor core", AGGR_CORE),
1627 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1628 "disable CPU count aggregation", AGGR_NONE),
1634 argc = parse_options(argc, argv, options, stat_report_usage, 0);
1636 if (!input_name || !strlen(input_name)) {
1637 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1640 input_name = "perf.data";
1643 perf_stat.data.path = input_name;
1644 perf_stat.data.mode = PERF_DATA_MODE_READ;
1646 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
1647 if (session == NULL)
1650 perf_stat.session = session;
1651 stat_config.output = stderr;
1652 evsel_list = session->evlist;
1654 ret = perf_session__process_events(session);
1658 perf_session__delete(session);
1662 static void setup_system_wide(int forks)
1665 * Make system wide (-a) the default target if
1666 * no target was specified and one of following
1667 * conditions is met:
1669 * - there's no workload specified
1670 * - there is workload specified but all requested
1671 * events are system wide events
1673 if (!target__none(&target))
1677 target.system_wide = true;
1679 struct perf_evsel *counter;
1681 evlist__for_each_entry(evsel_list, counter) {
1682 if (!counter->system_wide)
1686 if (evsel_list->nr_entries)
1687 target.system_wide = true;
1691 int cmd_stat(int argc, const char **argv)
1693 const char * const stat_usage[] = {
1694 "perf stat [<options>] [<command>]",
1697 int status = -EINVAL, run_idx;
1699 FILE *output = stderr;
1700 unsigned int interval, timeout;
1701 const char * const stat_subcommands[] = { "record", "report" };
1703 setlocale(LC_ALL, "");
1705 evsel_list = perf_evlist__new();
1706 if (evsel_list == NULL)
1709 parse_events__shrink_config_terms();
1711 /* String-parsing callback-based options would segfault when negated */
1712 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
1713 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
1714 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
1716 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1717 (const char **) stat_usage,
1718 PARSE_OPT_STOP_AT_NON_OPTION);
1719 perf_stat__collect_metric_expr(evsel_list);
1720 perf_stat__init_shadow_stats();
1722 if (stat_config.csv_sep) {
1723 stat_config.csv_output = true;
1724 if (!strcmp(stat_config.csv_sep, "\\t"))
1725 stat_config.csv_sep = "\t";
1727 stat_config.csv_sep = DEFAULT_SEPARATOR;
1729 if (argc && !strncmp(argv[0], "rec", 3)) {
1730 argc = __cmd_record(argc, argv);
1733 } else if (argc && !strncmp(argv[0], "rep", 3))
1734 return __cmd_report(argc, argv);
1736 interval = stat_config.interval;
1737 timeout = stat_config.timeout;
1740 * For record command the -o is already taken care of.
1742 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1745 if (output_name && output_fd) {
1746 fprintf(stderr, "cannot use both --output and --log-fd\n");
1747 parse_options_usage(stat_usage, stat_options, "o", 1);
1748 parse_options_usage(NULL, stat_options, "log-fd", 0);
1752 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
1753 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
1757 if (stat_config.metric_only && stat_config.run_count > 1) {
1758 fprintf(stderr, "--metric-only is not supported with -r\n");
1762 if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
1763 fprintf(stderr, "--table is only supported with -r\n");
1764 parse_options_usage(stat_usage, stat_options, "r", 1);
1765 parse_options_usage(NULL, stat_options, "table", 0);
1769 if (output_fd < 0) {
1770 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1771 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1777 mode = append_file ? "a" : "w";
1779 output = fopen(output_name, mode);
1781 perror("failed to create output file");
1784 clock_gettime(CLOCK_REALTIME, &tm);
1785 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1786 } else if (output_fd > 0) {
1787 mode = append_file ? "a" : "w";
1788 output = fdopen(output_fd, mode);
1790 perror("Failed opening logfd");
1795 stat_config.output = output;
1798 * let the spreadsheet do the pretty-printing
1800 if (stat_config.csv_output) {
1801 /* User explicitly passed -B? */
1802 if (big_num_opt == 1) {
1803 fprintf(stderr, "-B option not supported with -x\n");
1804 parse_options_usage(stat_usage, stat_options, "B", 1);
1805 parse_options_usage(NULL, stat_options, "x", 1);
1807 } else /* Nope, so disable big number formatting */
1808 stat_config.big_num = false;
1809 } else if (big_num_opt == 0) /* User passed --no-big-num */
1810 stat_config.big_num = false;
1812 setup_system_wide(argc);
1815 * Display user/system times only for single
1816 * run and when there's specified tracee.
1818 if ((stat_config.run_count == 1) && target__none(&target))
1819 stat_config.ru_display = true;
1821 if (stat_config.run_count < 0) {
1822 pr_err("Run count must be a positive number\n");
1823 parse_options_usage(stat_usage, stat_options, "r", 1);
1825 } else if (stat_config.run_count == 0) {
1827 stat_config.run_count = 1;
1830 if (stat_config.walltime_run_table) {
1831 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
1832 if (!stat_config.walltime_run) {
1833 pr_err("failed to setup -r option");
1838 if ((stat_config.aggr_mode == AGGR_THREAD) &&
1839 !target__has_task(&target)) {
1840 if (!target.system_wide || target.cpu_list) {
1841 fprintf(stderr, "The --per-thread option is only "
1842 "available when monitoring via -p -t -a "
1843 "options or only --per-thread.\n");
1844 parse_options_usage(NULL, stat_options, "p", 1);
1845 parse_options_usage(NULL, stat_options, "t", 1);
1851 * no_aggr, cgroup are for system-wide only
1852 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1854 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1855 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1856 !target__has_cpu(&target)) {
1857 fprintf(stderr, "both cgroup and no-aggregation "
1858 "modes only available in system-wide mode\n");
1860 parse_options_usage(stat_usage, stat_options, "G", 1);
1861 parse_options_usage(NULL, stat_options, "A", 1);
1862 parse_options_usage(NULL, stat_options, "a", 1);
1866 if (add_default_attributes())
1869 target__validate(&target);
1871 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
1872 target.per_thread = true;
1874 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1875 if (target__has_task(&target)) {
1876 pr_err("Problems finding threads of monitor\n");
1877 parse_options_usage(stat_usage, stat_options, "p", 1);
1878 parse_options_usage(NULL, stat_options, "t", 1);
1879 } else if (target__has_cpu(&target)) {
1880 perror("failed to parse CPUs map");
1881 parse_options_usage(stat_usage, stat_options, "C", 1);
1882 parse_options_usage(NULL, stat_options, "a", 1);
1888 * Initialize thread_map with comm names,
1889 * so we could print it out on output.
1891 if (stat_config.aggr_mode == AGGR_THREAD) {
1892 thread_map__read_comms(evsel_list->threads);
1893 if (target.system_wide) {
1894 if (runtime_stat_new(&stat_config,
1895 thread_map__nr(evsel_list->threads))) {
1901 if (stat_config.times && interval)
1902 interval_count = true;
1903 else if (stat_config.times && !interval) {
1904 pr_err("interval-count option should be used together with "
1905 "interval-print.\n");
1906 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
1907 parse_options_usage(stat_usage, stat_options, "I", 1);
1911 if (timeout && timeout < 100) {
1913 pr_err("timeout must be >= 10ms.\n");
1914 parse_options_usage(stat_usage, stat_options, "timeout", 0);
1917 pr_warning("timeout < 100ms. "
1918 "The overhead percentage could be high in some cases. "
1919 "Please proceed with caution.\n");
1921 if (timeout && interval) {
1922 pr_err("timeout option is not supported with interval-print.\n");
1923 parse_options_usage(stat_usage, stat_options, "timeout", 0);
1924 parse_options_usage(stat_usage, stat_options, "I", 1);
1928 if (perf_evlist__alloc_stats(evsel_list, interval))
1931 if (perf_stat_init_aggr_mode())
1935 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
1936 * while avoiding that older tools show confusing messages.
1938 * However for pipe sessions we need to keep it zero,
1939 * because script's perf_evsel__check_attr is triggered
1940 * by attr->sample_type != 0, and we can't run it on
1943 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
1946 * We dont want to block the signals - that would cause
1947 * child tasks to inherit that and Ctrl-C would not work.
1948 * What we want is for Ctrl-C to work in the exec()-ed
1949 * task, but being ignored by perf stat itself:
1953 signal(SIGINT, skip_signal);
1954 signal(SIGCHLD, skip_signal);
1955 signal(SIGALRM, skip_signal);
1956 signal(SIGABRT, skip_signal);
1959 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
1960 if (stat_config.run_count != 1 && verbose > 0)
1961 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1964 status = run_perf_stat(argc, argv, run_idx);
1965 if (forever && status != -1) {
1966 print_counters(NULL, argc, argv);
1967 perf_stat__reset_stats();
1971 if (!forever && status != -1 && !interval)
1972 print_counters(NULL, argc, argv);
1976 * We synthesize the kernel mmap record just so that older tools
1977 * don't emit warnings about not being able to resolve symbols
1978 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
1979 * a saner message about no samples being in the perf.data file.
1981 * This also serves to suppress a warning about f_header.data.size == 0
1982 * in header.c at the moment 'perf stat record' gets introduced, which
1983 * is not really needed once we start adding the stat specific PERF_RECORD_
1984 * records, but the need to suppress the kptr_restrict messages in older
1985 * tools remain -acme
1987 int fd = perf_data__fd(&perf_stat.data);
1988 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
1989 process_synthesized_event,
1990 &perf_stat.session->machines.host);
1992 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
1993 "older tools may produce warnings about this file\n.");
1997 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
1998 pr_err("failed to write stat round event\n");
2001 if (!perf_stat.data.is_pipe) {
2002 perf_stat.session->header.data_size += perf_stat.bytes_written;
2003 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2006 perf_evlist__close(evsel_list);
2007 perf_session__delete(perf_stat.session);
2010 perf_stat__exit_aggr_mode();
2011 perf_evlist__free_stats(evsel_list);
2013 zfree(&stat_config.walltime_run);
2015 if (smi_cost && smi_reset)
2016 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2018 perf_evlist__delete(evsel_list);
2020 runtime_stat_delete(&stat_config);