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_map.h"
58 #include "util/counts.h"
59 #include "util/group.h"
60 #include "util/session.h"
61 #include "util/tool.h"
62 #include "util/string2.h"
63 #include "util/metricgroup.h"
64 #include "util/target.h"
65 #include "util/time-utils.h"
69 #include <linux/time64.h>
70 #include <linux/zalloc.h>
71 #include <api/fs/fs.h>
75 #include <sys/prctl.h>
79 #include <sys/types.h>
84 #include <sys/resource.h>
86 #include <linux/ctype.h>
87 #include <perf/evlist.h>
89 #define DEFAULT_SEPARATOR " "
90 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
92 static void print_counters(struct timespec *ts, int argc, const char **argv);
94 /* Default events used for perf stat -T */
95 static const char *transaction_attrs = {
107 /* More limited version when the CPU does not have all events. */
108 static const char * transaction_limited_attrs = {
118 static const char * topdown_attrs[] = {
119 "topdown-total-slots",
120 "topdown-slots-retired",
121 "topdown-recovery-bubbles",
122 "topdown-fetch-bubbles",
123 "topdown-slots-issued",
127 static const char *smi_cost_attrs = {
135 static struct evlist *evsel_list;
137 static struct target target = {
141 #define METRIC_ONLY_LEN 20
143 static volatile pid_t child_pid = -1;
144 static int detailed_run = 0;
145 static bool transaction_run;
146 static bool topdown_run = false;
147 static bool smi_cost = false;
148 static bool smi_reset = false;
149 static int big_num_opt = -1;
150 static bool group = false;
151 static const char *pre_cmd = NULL;
152 static const char *post_cmd = NULL;
153 static bool sync_run = false;
154 static bool forever = false;
155 static bool force_metric_only = false;
156 static struct timespec ref_time;
157 static bool append_file;
158 static bool interval_count;
159 static const char *output_name;
160 static int output_fd;
164 struct perf_data data;
165 struct perf_session *session;
167 struct perf_tool tool;
169 struct perf_cpu_map *cpus;
170 struct perf_thread_map *threads;
171 enum aggr_mode aggr_mode;
174 static struct perf_stat perf_stat;
175 #define STAT_RECORD perf_stat.record
177 static volatile int done = 0;
179 static struct perf_stat_config stat_config = {
180 .aggr_mode = AGGR_GLOBAL,
182 .unit_width = 4, /* strlen("unit") */
184 .metric_only_len = METRIC_ONLY_LEN,
185 .walltime_nsecs_stats = &walltime_nsecs_stats,
189 static inline void diff_timespec(struct timespec *r, struct timespec *a,
192 r->tv_sec = a->tv_sec - b->tv_sec;
193 if (a->tv_nsec < b->tv_nsec) {
194 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
197 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
201 static void perf_stat__reset_stats(void)
205 perf_evlist__reset_stats(evsel_list);
206 perf_stat__reset_shadow_stats();
208 for (i = 0; i < stat_config.stats_num; i++)
209 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
212 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
213 union perf_event *event,
214 struct perf_sample *sample __maybe_unused,
215 struct machine *machine __maybe_unused)
217 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
218 pr_err("failed to write perf data, error: %m\n");
222 perf_stat.bytes_written += event->header.size;
226 static int write_stat_round_event(u64 tm, u64 type)
228 return perf_event__synthesize_stat_round(NULL, tm, type,
229 process_synthesized_event,
233 #define WRITE_STAT_ROUND_EVENT(time, interval) \
234 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
236 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
239 perf_evsel__write_stat_event(struct evsel *counter, u32 cpu, u32 thread,
240 struct perf_counts_values *count)
242 struct perf_sample_id *sid = SID(counter, cpu, thread);
244 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
245 process_synthesized_event, NULL);
248 static int read_single_counter(struct evsel *counter, int cpu,
249 int thread, struct timespec *rs)
251 if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
252 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
253 struct perf_counts_values *count =
254 perf_counts(counter->counts, cpu, thread);
255 count->ena = count->run = val;
259 return perf_evsel__read_counter(counter, cpu, thread);
263 * Read out the results of a single counter:
264 * do not aggregate counts across CPUs in system-wide mode
266 static int read_counter(struct evsel *counter, struct timespec *rs)
268 int nthreads = perf_thread_map__nr(evsel_list->core.threads);
269 int ncpus, cpu, thread;
271 if (target__has_cpu(&target) && !target__has_per_thread(&target))
272 ncpus = perf_evsel__nr_cpus(counter);
276 if (!counter->supported)
279 if (counter->system_wide)
282 for (thread = 0; thread < nthreads; thread++) {
283 for (cpu = 0; cpu < ncpus; cpu++) {
284 struct perf_counts_values *count;
286 count = perf_counts(counter->counts, cpu, thread);
289 * The leader's group read loads data into its group members
290 * (via perf_evsel__read_counter) and sets threir count->loaded.
292 if (!perf_counts__is_loaded(counter->counts, cpu, thread) &&
293 read_single_counter(counter, cpu, thread, rs)) {
294 counter->counts->scaled = -1;
295 perf_counts(counter->counts, cpu, thread)->ena = 0;
296 perf_counts(counter->counts, cpu, thread)->run = 0;
300 perf_counts__set_loaded(counter->counts, cpu, thread, false);
303 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
304 pr_err("failed to write stat event\n");
310 fprintf(stat_config.output,
311 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
312 perf_evsel__name(counter),
314 count->val, count->ena, count->run);
322 static void read_counters(struct timespec *rs)
324 struct evsel *counter;
327 evlist__for_each_entry(evsel_list, counter) {
328 ret = read_counter(counter, rs);
330 pr_debug("failed to read counter %s\n", counter->name);
332 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
333 pr_warning("failed to process counter %s\n", counter->name);
337 static void process_interval(void)
339 struct timespec ts, rs;
341 clock_gettime(CLOCK_MONOTONIC, &ts);
342 diff_timespec(&rs, &ts, &ref_time);
347 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
348 pr_err("failed to write stat round event\n");
351 init_stats(&walltime_nsecs_stats);
352 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
353 print_counters(&rs, 0, NULL);
356 static void enable_counters(void)
358 if (stat_config.initial_delay)
359 usleep(stat_config.initial_delay * USEC_PER_MSEC);
362 * We need to enable counters only if:
363 * - we don't have tracee (attaching to task or cpu)
364 * - we have initial delay configured
366 if (!target__none(&target) || stat_config.initial_delay)
367 evlist__enable(evsel_list);
370 static void disable_counters(void)
373 * If we don't have tracee (attaching to task or cpu), counters may
374 * still be running. To get accurate group ratios, we must stop groups
375 * from counting before reading their constituent counters.
377 if (!target__none(&target))
378 evlist__disable(evsel_list);
381 static volatile int workload_exec_errno;
384 * perf_evlist__prepare_workload will send a SIGUSR1
385 * if the fork fails, since we asked by setting its
386 * want_signal to true.
388 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
389 void *ucontext __maybe_unused)
391 workload_exec_errno = info->si_value.sival_int;
394 static bool perf_evsel__should_store_id(struct evsel *counter)
396 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
399 static bool is_target_alive(struct target *_target,
400 struct perf_thread_map *threads)
405 if (!target__has_task(_target))
408 for (i = 0; i < threads->nr; i++) {
411 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
412 threads->map[i].pid);
414 if (!stat(path, &st))
421 static int __run_perf_stat(int argc, const char **argv, int run_idx)
423 int interval = stat_config.interval;
424 int times = stat_config.times;
425 int timeout = stat_config.timeout;
427 unsigned long long t0, t1;
428 struct evsel *counter;
432 const bool forks = (argc > 0);
433 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
436 ts.tv_sec = interval / USEC_PER_MSEC;
437 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
438 } else if (timeout) {
439 ts.tv_sec = timeout / USEC_PER_MSEC;
440 ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
447 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
448 workload_exec_failed_signal) < 0) {
449 perror("failed to prepare workload");
452 child_pid = evsel_list->workload.pid;
456 perf_evlist__set_leader(evsel_list);
458 evlist__for_each_entry(evsel_list, counter) {
460 if (create_perf_stat_counter(counter, &stat_config, &target) < 0) {
462 /* Weak group failed. Reset the group. */
463 if ((errno == EINVAL || errno == EBADF) &&
464 counter->leader != counter &&
465 counter->weak_group) {
466 counter = perf_evlist__reset_weak_group(evsel_list, counter);
471 * PPC returns ENXIO for HW counters until 2.6.37
472 * (behavior changed with commit b0a873e).
474 if (errno == EINVAL || errno == ENOSYS ||
475 errno == ENOENT || errno == EOPNOTSUPP ||
478 ui__warning("%s event is not supported by the kernel.\n",
479 perf_evsel__name(counter));
480 counter->supported = false;
482 if ((counter->leader != counter) ||
483 !(counter->leader->core.nr_members > 1))
485 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
487 ui__warning("%s\n", msg);
489 } else if (target__has_per_thread(&target) &&
490 evsel_list->core.threads &&
491 evsel_list->core.threads->err_thread != -1) {
493 * For global --per-thread case, skip current
496 if (!thread_map__remove(evsel_list->core.threads,
497 evsel_list->core.threads->err_thread)) {
498 evsel_list->core.threads->err_thread = -1;
503 perf_evsel__open_strerror(counter, &target,
504 errno, msg, sizeof(msg));
505 ui__error("%s\n", msg);
508 kill(child_pid, SIGTERM);
512 counter->supported = true;
514 l = strlen(counter->unit);
515 if (l > stat_config.unit_width)
516 stat_config.unit_width = l;
518 if (perf_evsel__should_store_id(counter) &&
519 perf_evsel__store_ids(counter, evsel_list))
523 if (perf_evlist__apply_filters(evsel_list, &counter)) {
524 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
525 counter->filter, perf_evsel__name(counter), errno,
526 str_error_r(errno, msg, sizeof(msg)));
531 int err, fd = perf_data__fd(&perf_stat.data);
534 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
536 err = perf_session__write_header(perf_stat.session, evsel_list,
543 err = perf_stat_synthesize_config(&stat_config, NULL, evsel_list,
544 process_synthesized_event, is_pipe);
550 * Enable counters and exec the command:
553 clock_gettime(CLOCK_MONOTONIC, &ref_time);
556 perf_evlist__start_workload(evsel_list);
559 if (interval || timeout) {
560 while (!waitpid(child_pid, &status, WNOHANG)) {
561 nanosleep(&ts, NULL);
565 if (interval_count && !(--times))
570 wait4(child_pid, &status, 0, &stat_config.ru_data);
572 if (workload_exec_errno) {
573 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
574 pr_err("Workload failed: %s\n", emsg);
578 if (WIFSIGNALED(status))
579 psignal(WTERMSIG(status), argv[0]);
583 nanosleep(&ts, NULL);
584 if (!is_target_alive(&target, evsel_list->core.threads))
590 if (interval_count && !(--times))
600 if (stat_config.walltime_run_table)
601 stat_config.walltime_run[run_idx] = t1 - t0;
603 update_stats(&walltime_nsecs_stats, t1 - t0);
606 * Closing a group leader splits the group, and as we only disable
607 * group leaders, results in remaining events becoming enabled. To
608 * avoid arbitrary skew, we must read all counters before closing any
611 read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
614 * We need to keep evsel_list alive, because it's processed
615 * later the evsel_list will be closed after.
618 evlist__close(evsel_list);
620 return WEXITSTATUS(status);
623 static int run_perf_stat(int argc, const char **argv, int run_idx)
628 ret = system(pre_cmd);
636 ret = __run_perf_stat(argc, argv, run_idx);
641 ret = system(post_cmd);
649 static void print_counters(struct timespec *ts, int argc, const char **argv)
651 /* Do not print anything if we record to the pipe. */
652 if (STAT_RECORD && perf_stat.data.is_pipe)
655 perf_evlist__print_counters(evsel_list, &stat_config, &target,
659 static volatile int signr = -1;
661 static void skip_signal(int signo)
663 if ((child_pid == -1) || stat_config.interval)
668 * render child_pid harmless
669 * won't send SIGTERM to a random
670 * process in case of race condition
671 * and fast PID recycling
676 static void sig_atexit(void)
681 * avoid race condition with SIGCHLD handler
682 * in skip_signal() which is modifying child_pid
683 * goal is to avoid send SIGTERM to a random
687 sigaddset(&set, SIGCHLD);
688 sigprocmask(SIG_BLOCK, &set, &oset);
691 kill(child_pid, SIGTERM);
693 sigprocmask(SIG_SETMASK, &oset, NULL);
698 signal(signr, SIG_DFL);
699 kill(getpid(), signr);
702 static int stat__set_big_num(const struct option *opt __maybe_unused,
703 const char *s __maybe_unused, int unset)
705 big_num_opt = unset ? 0 : 1;
709 static int enable_metric_only(const struct option *opt __maybe_unused,
710 const char *s __maybe_unused, int unset)
712 force_metric_only = true;
713 stat_config.metric_only = !unset;
717 static int parse_metric_groups(const struct option *opt,
719 int unset __maybe_unused)
721 return metricgroup__parse_groups(opt, str, &stat_config.metric_events);
724 static struct option stat_options[] = {
725 OPT_BOOLEAN('T', "transaction", &transaction_run,
726 "hardware transaction statistics"),
727 OPT_CALLBACK('e', "event", &evsel_list, "event",
728 "event selector. use 'perf list' to list available events",
729 parse_events_option),
730 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
731 "event filter", parse_filter),
732 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
733 "child tasks do not inherit counters"),
734 OPT_STRING('p', "pid", &target.pid, "pid",
735 "stat events on existing process id"),
736 OPT_STRING('t', "tid", &target.tid, "tid",
737 "stat events on existing thread id"),
738 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
739 "system-wide collection from all CPUs"),
740 OPT_BOOLEAN('g', "group", &group,
741 "put the counters into a counter group"),
742 OPT_BOOLEAN(0, "scale", &stat_config.scale,
743 "Use --no-scale to disable counter scaling for multiplexing"),
744 OPT_INCR('v', "verbose", &verbose,
745 "be more verbose (show counter open errors, etc)"),
746 OPT_INTEGER('r', "repeat", &stat_config.run_count,
747 "repeat command and print average + stddev (max: 100, forever: 0)"),
748 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
749 "display details about each run (only with -r option)"),
750 OPT_BOOLEAN('n', "null", &stat_config.null_run,
751 "null run - dont start any counters"),
752 OPT_INCR('d', "detailed", &detailed_run,
753 "detailed run - start a lot of events"),
754 OPT_BOOLEAN('S', "sync", &sync_run,
755 "call sync() before starting a run"),
756 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
757 "print large numbers with thousands\' separators",
759 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
760 "list of cpus to monitor in system-wide"),
761 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
762 "disable CPU count aggregation", AGGR_NONE),
763 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
764 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
765 "print counts with custom separator"),
766 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
767 "monitor event in cgroup name only", parse_cgroups),
768 OPT_STRING('o', "output", &output_name, "file", "output file name"),
769 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
770 OPT_INTEGER(0, "log-fd", &output_fd,
771 "log output to fd, instead of stderr"),
772 OPT_STRING(0, "pre", &pre_cmd, "command",
773 "command to run prior to the measured command"),
774 OPT_STRING(0, "post", &post_cmd, "command",
775 "command to run after to the measured command"),
776 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
777 "print counts at regular interval in ms "
778 "(overhead is possible for values <= 100ms)"),
779 OPT_INTEGER(0, "interval-count", &stat_config.times,
780 "print counts for fixed number of times"),
781 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
782 "clear screen in between new interval"),
783 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
784 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
785 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
786 "aggregate counts per processor socket", AGGR_SOCKET),
787 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
788 "aggregate counts per processor die", AGGR_DIE),
789 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
790 "aggregate counts per physical processor core", AGGR_CORE),
791 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
792 "aggregate counts per thread", AGGR_THREAD),
793 OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
794 "ms to wait before starting measurement after program start"),
795 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
796 "Only print computed metrics. No raw values", enable_metric_only),
797 OPT_BOOLEAN(0, "topdown", &topdown_run,
798 "measure topdown level 1 statistics"),
799 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
801 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
802 "monitor specified metrics or metric groups (separated by ,)",
803 parse_metric_groups),
807 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
808 struct perf_cpu_map *map, int cpu)
810 return cpu_map__get_socket(map, cpu, NULL);
813 static int perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
814 struct perf_cpu_map *map, int cpu)
816 return cpu_map__get_die(map, cpu, NULL);
819 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
820 struct perf_cpu_map *map, int cpu)
822 return cpu_map__get_core(map, cpu, NULL);
825 static int cpu_map__get_max(struct perf_cpu_map *map)
829 for (i = 0; i < map->nr; i++) {
830 if (map->map[i] > max)
837 static int perf_stat__get_aggr(struct perf_stat_config *config,
838 aggr_get_id_t get_id, struct perf_cpu_map *map, int idx)
847 if (config->cpus_aggr_map->map[cpu] == -1)
848 config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
850 return config->cpus_aggr_map->map[cpu];
853 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
854 struct perf_cpu_map *map, int idx)
856 return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
859 static int perf_stat__get_die_cached(struct perf_stat_config *config,
860 struct perf_cpu_map *map, int idx)
862 return perf_stat__get_aggr(config, perf_stat__get_die, map, idx);
865 static int perf_stat__get_core_cached(struct perf_stat_config *config,
866 struct perf_cpu_map *map, int idx)
868 return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
871 static bool term_percore_set(void)
873 struct evsel *counter;
875 evlist__for_each_entry(evsel_list, counter) {
876 if (counter->percore)
883 static int perf_stat_init_aggr_mode(void)
887 switch (stat_config.aggr_mode) {
889 if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
890 perror("cannot build socket map");
893 stat_config.aggr_get_id = perf_stat__get_socket_cached;
896 if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
897 perror("cannot build die map");
900 stat_config.aggr_get_id = perf_stat__get_die_cached;
903 if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
904 perror("cannot build core map");
907 stat_config.aggr_get_id = perf_stat__get_core_cached;
910 if (term_percore_set()) {
911 if (cpu_map__build_core_map(evsel_list->core.cpus,
912 &stat_config.aggr_map)) {
913 perror("cannot build core map");
916 stat_config.aggr_get_id = perf_stat__get_core_cached;
927 * The evsel_list->cpus is the base we operate on,
928 * taking the highest cpu number to be the size of
929 * the aggregation translate cpumap.
931 nr = cpu_map__get_max(evsel_list->core.cpus);
932 stat_config.cpus_aggr_map = perf_cpu_map__empty_new(nr + 1);
933 return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
936 static void perf_stat__exit_aggr_mode(void)
938 perf_cpu_map__put(stat_config.aggr_map);
939 perf_cpu_map__put(stat_config.cpus_aggr_map);
940 stat_config.aggr_map = NULL;
941 stat_config.cpus_aggr_map = NULL;
944 static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx)
953 if (cpu >= env->nr_cpus_avail)
959 static int perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data)
961 struct perf_env *env = data;
962 int cpu = perf_env__get_cpu(env, map, idx);
964 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
967 static int perf_env__get_die(struct perf_cpu_map *map, int idx, void *data)
969 struct perf_env *env = data;
970 int die_id = -1, cpu = perf_env__get_cpu(env, map, idx);
974 * Encode socket in bit range 15:8
975 * die_id is relative to socket,
976 * we need a global id. So we combine
979 if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
982 if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
985 die_id = (env->cpu[cpu].socket_id << 8) | (env->cpu[cpu].die_id & 0xff);
991 static int perf_env__get_core(struct perf_cpu_map *map, int idx, void *data)
993 struct perf_env *env = data;
994 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
998 * Encode socket in bit range 31:24
999 * encode die id in bit range 23:16
1000 * core_id is relative to socket and die,
1001 * we need a global id. So we combine
1002 * socket + die id + core id
1004 if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
1007 if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
1010 if (WARN_ONCE(env->cpu[cpu].core_id >> 16, "The core id number is too big.\n"))
1013 core = (env->cpu[cpu].socket_id << 24) |
1014 (env->cpu[cpu].die_id << 16) |
1015 (env->cpu[cpu].core_id & 0xffff);
1021 static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus,
1022 struct perf_cpu_map **sockp)
1024 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1027 static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus,
1028 struct perf_cpu_map **diep)
1030 return cpu_map__build_map(cpus, diep, perf_env__get_die, env);
1033 static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus,
1034 struct perf_cpu_map **corep)
1036 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1039 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1040 struct perf_cpu_map *map, int idx)
1042 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1044 static int perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1045 struct perf_cpu_map *map, int idx)
1047 return perf_env__get_die(map, idx, &perf_stat.session->header.env);
1050 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1051 struct perf_cpu_map *map, int idx)
1053 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1056 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1058 struct perf_env *env = &st->session->header.env;
1060 switch (stat_config.aggr_mode) {
1062 if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1063 perror("cannot build socket map");
1066 stat_config.aggr_get_id = perf_stat__get_socket_file;
1069 if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1070 perror("cannot build die map");
1073 stat_config.aggr_get_id = perf_stat__get_die_file;
1076 if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1077 perror("cannot build core map");
1080 stat_config.aggr_get_id = perf_stat__get_core_file;
1093 static int topdown_filter_events(const char **attr, char **str, bool use_group)
1100 for (i = 0; attr[i]; i++) {
1101 if (pmu_have_event("cpu", attr[i])) {
1102 len += strlen(attr[i]) + 1;
1103 attr[i - off] = attr[i];
1107 attr[i - off] = NULL;
1109 *str = malloc(len + 1 + 2);
1119 for (i = 0; attr[i]; i++) {
1132 __weak bool arch_topdown_check_group(bool *warn)
1138 __weak void arch_topdown_group_warn(void)
1143 * Add default attributes, if there were no attributes specified or
1144 * if -d/--detailed, -d -d or -d -d -d is used:
1146 static int add_default_attributes(void)
1149 struct perf_event_attr default_attrs0[] = {
1151 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1152 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1153 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1154 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1156 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1158 struct perf_event_attr frontend_attrs[] = {
1159 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1161 struct perf_event_attr backend_attrs[] = {
1162 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1164 struct perf_event_attr default_attrs1[] = {
1165 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1166 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1167 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1172 * Detailed stats (-d), covering the L1 and last level data caches:
1174 struct perf_event_attr detailed_attrs[] = {
1176 { .type = PERF_TYPE_HW_CACHE,
1178 PERF_COUNT_HW_CACHE_L1D << 0 |
1179 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1180 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1182 { .type = PERF_TYPE_HW_CACHE,
1184 PERF_COUNT_HW_CACHE_L1D << 0 |
1185 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1186 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1188 { .type = PERF_TYPE_HW_CACHE,
1190 PERF_COUNT_HW_CACHE_LL << 0 |
1191 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1192 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1194 { .type = PERF_TYPE_HW_CACHE,
1196 PERF_COUNT_HW_CACHE_LL << 0 |
1197 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1198 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1202 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1204 struct perf_event_attr very_detailed_attrs[] = {
1206 { .type = PERF_TYPE_HW_CACHE,
1208 PERF_COUNT_HW_CACHE_L1I << 0 |
1209 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1210 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1212 { .type = PERF_TYPE_HW_CACHE,
1214 PERF_COUNT_HW_CACHE_L1I << 0 |
1215 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1216 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1218 { .type = PERF_TYPE_HW_CACHE,
1220 PERF_COUNT_HW_CACHE_DTLB << 0 |
1221 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1222 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1224 { .type = PERF_TYPE_HW_CACHE,
1226 PERF_COUNT_HW_CACHE_DTLB << 0 |
1227 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1228 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1230 { .type = PERF_TYPE_HW_CACHE,
1232 PERF_COUNT_HW_CACHE_ITLB << 0 |
1233 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1234 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1236 { .type = PERF_TYPE_HW_CACHE,
1238 PERF_COUNT_HW_CACHE_ITLB << 0 |
1239 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1240 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1245 * Very, very detailed stats (-d -d -d), adding prefetch events:
1247 struct perf_event_attr very_very_detailed_attrs[] = {
1249 { .type = PERF_TYPE_HW_CACHE,
1251 PERF_COUNT_HW_CACHE_L1D << 0 |
1252 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1253 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1255 { .type = PERF_TYPE_HW_CACHE,
1257 PERF_COUNT_HW_CACHE_L1D << 0 |
1258 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1259 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1261 struct parse_events_error errinfo;
1263 /* Set attrs if no event is selected and !null_run: */
1264 if (stat_config.null_run)
1267 if (transaction_run) {
1268 /* Handle -T as -M transaction. Once platform specific metrics
1269 * support has been added to the json files, all archictures
1270 * will use this approach. To determine transaction support
1271 * on an architecture test for such a metric name.
1273 if (metricgroup__has_metric("transaction")) {
1274 struct option opt = { .value = &evsel_list };
1276 return metricgroup__parse_groups(&opt, "transaction",
1277 &stat_config.metric_events);
1280 if (pmu_have_event("cpu", "cycles-ct") &&
1281 pmu_have_event("cpu", "el-start"))
1282 err = parse_events(evsel_list, transaction_attrs,
1285 err = parse_events(evsel_list,
1286 transaction_limited_attrs,
1289 fprintf(stderr, "Cannot set up transaction events\n");
1290 parse_events_print_error(&errinfo, transaction_attrs);
1299 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1300 fprintf(stderr, "freeze_on_smi is not supported.\n");
1305 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1306 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1312 if (pmu_have_event("msr", "aperf") &&
1313 pmu_have_event("msr", "smi")) {
1314 if (!force_metric_only)
1315 stat_config.metric_only = true;
1316 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1318 fprintf(stderr, "To measure SMI cost, it needs "
1319 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1320 parse_events_print_error(&errinfo, smi_cost_attrs);
1324 fprintf(stderr, "Cannot set up SMI cost events\n");
1334 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1335 stat_config.aggr_mode != AGGR_CORE) {
1336 pr_err("top down event configuration requires --per-core mode\n");
1339 stat_config.aggr_mode = AGGR_CORE;
1340 if (nr_cgroups || !target__has_cpu(&target)) {
1341 pr_err("top down event configuration requires system-wide mode (-a)\n");
1345 if (!force_metric_only)
1346 stat_config.metric_only = true;
1347 if (topdown_filter_events(topdown_attrs, &str,
1348 arch_topdown_check_group(&warn)) < 0) {
1349 pr_err("Out of memory\n");
1352 if (topdown_attrs[0] && str) {
1354 arch_topdown_group_warn();
1355 err = parse_events(evsel_list, str, &errinfo);
1358 "Cannot set up top down events %s: %d\n",
1360 parse_events_print_error(&errinfo, str);
1365 fprintf(stderr, "System does not support topdown\n");
1371 if (!evsel_list->core.nr_entries) {
1372 if (target__has_cpu(&target))
1373 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1375 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1377 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1378 if (perf_evlist__add_default_attrs(evsel_list,
1379 frontend_attrs) < 0)
1382 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1383 if (perf_evlist__add_default_attrs(evsel_list,
1387 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1391 /* Detailed events get appended to the event list: */
1393 if (detailed_run < 1)
1396 /* Append detailed run extra attributes: */
1397 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1400 if (detailed_run < 2)
1403 /* Append very detailed run extra attributes: */
1404 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1407 if (detailed_run < 3)
1410 /* Append very, very detailed run extra attributes: */
1411 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1414 static const char * const stat_record_usage[] = {
1415 "perf stat record [<options>]",
1419 static void init_features(struct perf_session *session)
1423 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1424 perf_header__set_feat(&session->header, feat);
1426 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1427 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1428 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1429 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1430 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1433 static int __cmd_record(int argc, const char **argv)
1435 struct perf_session *session;
1436 struct perf_data *data = &perf_stat.data;
1438 argc = parse_options(argc, argv, stat_options, stat_record_usage,
1439 PARSE_OPT_STOP_AT_NON_OPTION);
1442 data->path = output_name;
1444 if (stat_config.run_count != 1 || forever) {
1445 pr_err("Cannot use -r option with perf stat record.\n");
1449 session = perf_session__new(data, false, NULL);
1450 if (session == NULL) {
1451 pr_err("Perf session creation failed.\n");
1455 init_features(session);
1457 session->evlist = evsel_list;
1458 perf_stat.session = session;
1459 perf_stat.record = true;
1463 static int process_stat_round_event(struct perf_session *session,
1464 union perf_event *event)
1466 struct perf_record_stat_round *stat_round = &event->stat_round;
1467 struct evsel *counter;
1468 struct timespec tsh, *ts = NULL;
1469 const char **argv = session->header.env.cmdline_argv;
1470 int argc = session->header.env.nr_cmdline;
1472 evlist__for_each_entry(evsel_list, counter)
1473 perf_stat_process_counter(&stat_config, counter);
1475 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1476 update_stats(&walltime_nsecs_stats, stat_round->time);
1478 if (stat_config.interval && stat_round->time) {
1479 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
1480 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1484 print_counters(ts, argc, argv);
1489 int process_stat_config_event(struct perf_session *session,
1490 union perf_event *event)
1492 struct perf_tool *tool = session->tool;
1493 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1495 perf_event__read_stat_config(&stat_config, &event->stat_config);
1497 if (perf_cpu_map__empty(st->cpus)) {
1498 if (st->aggr_mode != AGGR_UNSET)
1499 pr_warning("warning: processing task data, aggregation mode not set\n");
1503 if (st->aggr_mode != AGGR_UNSET)
1504 stat_config.aggr_mode = st->aggr_mode;
1506 if (perf_stat.data.is_pipe)
1507 perf_stat_init_aggr_mode();
1509 perf_stat_init_aggr_mode_file(st);
1514 static int set_maps(struct perf_stat *st)
1516 if (!st->cpus || !st->threads)
1519 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1522 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
1524 if (perf_evlist__alloc_stats(evsel_list, true))
1527 st->maps_allocated = true;
1532 int process_thread_map_event(struct perf_session *session,
1533 union perf_event *event)
1535 struct perf_tool *tool = session->tool;
1536 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1539 pr_warning("Extra thread map event, ignoring.\n");
1543 st->threads = thread_map__new_event(&event->thread_map);
1547 return set_maps(st);
1551 int process_cpu_map_event(struct perf_session *session,
1552 union perf_event *event)
1554 struct perf_tool *tool = session->tool;
1555 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1556 struct perf_cpu_map *cpus;
1559 pr_warning("Extra cpu map event, ignoring.\n");
1563 cpus = cpu_map__new_data(&event->cpu_map.data);
1568 return set_maps(st);
1571 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
1575 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
1579 config->stats_num = nthreads;
1581 for (i = 0; i < nthreads; i++)
1582 runtime_stat__init(&config->stats[i]);
1587 static void runtime_stat_delete(struct perf_stat_config *config)
1594 for (i = 0; i < config->stats_num; i++)
1595 runtime_stat__exit(&config->stats[i]);
1597 zfree(&config->stats);
1600 static const char * const stat_report_usage[] = {
1601 "perf stat report [<options>]",
1605 static struct perf_stat perf_stat = {
1607 .attr = perf_event__process_attr,
1608 .event_update = perf_event__process_event_update,
1609 .thread_map = process_thread_map_event,
1610 .cpu_map = process_cpu_map_event,
1611 .stat_config = process_stat_config_event,
1612 .stat = perf_event__process_stat_event,
1613 .stat_round = process_stat_round_event,
1615 .aggr_mode = AGGR_UNSET,
1618 static int __cmd_report(int argc, const char **argv)
1620 struct perf_session *session;
1621 const struct option options[] = {
1622 OPT_STRING('i', "input", &input_name, "file", "input file name"),
1623 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1624 "aggregate counts per processor socket", AGGR_SOCKET),
1625 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
1626 "aggregate counts per processor die", AGGR_DIE),
1627 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1628 "aggregate counts per physical processor core", AGGR_CORE),
1629 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1630 "disable CPU count aggregation", AGGR_NONE),
1636 argc = parse_options(argc, argv, options, stat_report_usage, 0);
1638 if (!input_name || !strlen(input_name)) {
1639 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1642 input_name = "perf.data";
1645 perf_stat.data.path = input_name;
1646 perf_stat.data.mode = PERF_DATA_MODE_READ;
1648 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
1649 if (session == NULL)
1652 perf_stat.session = session;
1653 stat_config.output = stderr;
1654 evsel_list = session->evlist;
1656 ret = perf_session__process_events(session);
1660 perf_session__delete(session);
1664 static void setup_system_wide(int forks)
1667 * Make system wide (-a) the default target if
1668 * no target was specified and one of following
1669 * conditions is met:
1671 * - there's no workload specified
1672 * - there is workload specified but all requested
1673 * events are system wide events
1675 if (!target__none(&target))
1679 target.system_wide = true;
1681 struct evsel *counter;
1683 evlist__for_each_entry(evsel_list, counter) {
1684 if (!counter->system_wide)
1688 if (evsel_list->core.nr_entries)
1689 target.system_wide = true;
1693 int cmd_stat(int argc, const char **argv)
1695 const char * const stat_usage[] = {
1696 "perf stat [<options>] [<command>]",
1699 int status = -EINVAL, run_idx;
1701 FILE *output = stderr;
1702 unsigned int interval, timeout;
1703 const char * const stat_subcommands[] = { "record", "report" };
1705 setlocale(LC_ALL, "");
1707 evsel_list = evlist__new();
1708 if (evsel_list == NULL)
1711 parse_events__shrink_config_terms();
1713 /* String-parsing callback-based options would segfault when negated */
1714 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
1715 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
1716 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
1718 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1719 (const char **) stat_usage,
1720 PARSE_OPT_STOP_AT_NON_OPTION);
1721 perf_stat__collect_metric_expr(evsel_list);
1722 perf_stat__init_shadow_stats();
1724 if (stat_config.csv_sep) {
1725 stat_config.csv_output = true;
1726 if (!strcmp(stat_config.csv_sep, "\\t"))
1727 stat_config.csv_sep = "\t";
1729 stat_config.csv_sep = DEFAULT_SEPARATOR;
1731 if (argc && !strncmp(argv[0], "rec", 3)) {
1732 argc = __cmd_record(argc, argv);
1735 } else if (argc && !strncmp(argv[0], "rep", 3))
1736 return __cmd_report(argc, argv);
1738 interval = stat_config.interval;
1739 timeout = stat_config.timeout;
1742 * For record command the -o is already taken care of.
1744 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1747 if (output_name && output_fd) {
1748 fprintf(stderr, "cannot use both --output and --log-fd\n");
1749 parse_options_usage(stat_usage, stat_options, "o", 1);
1750 parse_options_usage(NULL, stat_options, "log-fd", 0);
1754 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
1755 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
1759 if (stat_config.metric_only && stat_config.run_count > 1) {
1760 fprintf(stderr, "--metric-only is not supported with -r\n");
1764 if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
1765 fprintf(stderr, "--table is only supported with -r\n");
1766 parse_options_usage(stat_usage, stat_options, "r", 1);
1767 parse_options_usage(NULL, stat_options, "table", 0);
1771 if (output_fd < 0) {
1772 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1773 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1779 mode = append_file ? "a" : "w";
1781 output = fopen(output_name, mode);
1783 perror("failed to create output file");
1786 clock_gettime(CLOCK_REALTIME, &tm);
1787 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1788 } else if (output_fd > 0) {
1789 mode = append_file ? "a" : "w";
1790 output = fdopen(output_fd, mode);
1792 perror("Failed opening logfd");
1797 stat_config.output = output;
1800 * let the spreadsheet do the pretty-printing
1802 if (stat_config.csv_output) {
1803 /* User explicitly passed -B? */
1804 if (big_num_opt == 1) {
1805 fprintf(stderr, "-B option not supported with -x\n");
1806 parse_options_usage(stat_usage, stat_options, "B", 1);
1807 parse_options_usage(NULL, stat_options, "x", 1);
1809 } else /* Nope, so disable big number formatting */
1810 stat_config.big_num = false;
1811 } else if (big_num_opt == 0) /* User passed --no-big-num */
1812 stat_config.big_num = false;
1814 setup_system_wide(argc);
1817 * Display user/system times only for single
1818 * run and when there's specified tracee.
1820 if ((stat_config.run_count == 1) && target__none(&target))
1821 stat_config.ru_display = true;
1823 if (stat_config.run_count < 0) {
1824 pr_err("Run count must be a positive number\n");
1825 parse_options_usage(stat_usage, stat_options, "r", 1);
1827 } else if (stat_config.run_count == 0) {
1829 stat_config.run_count = 1;
1832 if (stat_config.walltime_run_table) {
1833 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
1834 if (!stat_config.walltime_run) {
1835 pr_err("failed to setup -r option");
1840 if ((stat_config.aggr_mode == AGGR_THREAD) &&
1841 !target__has_task(&target)) {
1842 if (!target.system_wide || target.cpu_list) {
1843 fprintf(stderr, "The --per-thread option is only "
1844 "available when monitoring via -p -t -a "
1845 "options or only --per-thread.\n");
1846 parse_options_usage(NULL, stat_options, "p", 1);
1847 parse_options_usage(NULL, stat_options, "t", 1);
1853 * no_aggr, cgroup are for system-wide only
1854 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1856 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1857 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1858 !target__has_cpu(&target)) {
1859 fprintf(stderr, "both cgroup and no-aggregation "
1860 "modes only available in system-wide mode\n");
1862 parse_options_usage(stat_usage, stat_options, "G", 1);
1863 parse_options_usage(NULL, stat_options, "A", 1);
1864 parse_options_usage(NULL, stat_options, "a", 1);
1868 if (add_default_attributes())
1871 target__validate(&target);
1873 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
1874 target.per_thread = true;
1876 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1877 if (target__has_task(&target)) {
1878 pr_err("Problems finding threads of monitor\n");
1879 parse_options_usage(stat_usage, stat_options, "p", 1);
1880 parse_options_usage(NULL, stat_options, "t", 1);
1881 } else if (target__has_cpu(&target)) {
1882 perror("failed to parse CPUs map");
1883 parse_options_usage(stat_usage, stat_options, "C", 1);
1884 parse_options_usage(NULL, stat_options, "a", 1);
1890 * Initialize thread_map with comm names,
1891 * so we could print it out on output.
1893 if (stat_config.aggr_mode == AGGR_THREAD) {
1894 thread_map__read_comms(evsel_list->core.threads);
1895 if (target.system_wide) {
1896 if (runtime_stat_new(&stat_config,
1897 perf_thread_map__nr(evsel_list->core.threads))) {
1903 if (stat_config.times && interval)
1904 interval_count = true;
1905 else if (stat_config.times && !interval) {
1906 pr_err("interval-count option should be used together with "
1907 "interval-print.\n");
1908 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
1909 parse_options_usage(stat_usage, stat_options, "I", 1);
1913 if (timeout && timeout < 100) {
1915 pr_err("timeout must be >= 10ms.\n");
1916 parse_options_usage(stat_usage, stat_options, "timeout", 0);
1919 pr_warning("timeout < 100ms. "
1920 "The overhead percentage could be high in some cases. "
1921 "Please proceed with caution.\n");
1923 if (timeout && interval) {
1924 pr_err("timeout option is not supported with interval-print.\n");
1925 parse_options_usage(stat_usage, stat_options, "timeout", 0);
1926 parse_options_usage(stat_usage, stat_options, "I", 1);
1930 if (perf_evlist__alloc_stats(evsel_list, interval))
1933 if (perf_stat_init_aggr_mode())
1937 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
1938 * while avoiding that older tools show confusing messages.
1940 * However for pipe sessions we need to keep it zero,
1941 * because script's perf_evsel__check_attr is triggered
1942 * by attr->sample_type != 0, and we can't run it on
1945 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
1948 * We dont want to block the signals - that would cause
1949 * child tasks to inherit that and Ctrl-C would not work.
1950 * What we want is for Ctrl-C to work in the exec()-ed
1951 * task, but being ignored by perf stat itself:
1955 signal(SIGINT, skip_signal);
1956 signal(SIGCHLD, skip_signal);
1957 signal(SIGALRM, skip_signal);
1958 signal(SIGABRT, skip_signal);
1961 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
1962 if (stat_config.run_count != 1 && verbose > 0)
1963 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1966 status = run_perf_stat(argc, argv, run_idx);
1967 if (forever && status != -1) {
1968 print_counters(NULL, argc, argv);
1969 perf_stat__reset_stats();
1973 if (!forever && status != -1 && !interval)
1974 print_counters(NULL, argc, argv);
1978 * We synthesize the kernel mmap record just so that older tools
1979 * don't emit warnings about not being able to resolve symbols
1980 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
1981 * a saner message about no samples being in the perf.data file.
1983 * This also serves to suppress a warning about f_header.data.size == 0
1984 * in header.c at the moment 'perf stat record' gets introduced, which
1985 * is not really needed once we start adding the stat specific PERF_RECORD_
1986 * records, but the need to suppress the kptr_restrict messages in older
1987 * tools remain -acme
1989 int fd = perf_data__fd(&perf_stat.data);
1990 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
1991 process_synthesized_event,
1992 &perf_stat.session->machines.host);
1994 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
1995 "older tools may produce warnings about this file\n.");
1999 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2000 pr_err("failed to write stat round event\n");
2003 if (!perf_stat.data.is_pipe) {
2004 perf_stat.session->header.data_size += perf_stat.bytes_written;
2005 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2008 evlist__close(evsel_list);
2009 perf_session__delete(perf_stat.session);
2012 perf_stat__exit_aggr_mode();
2013 perf_evlist__free_stats(evsel_list);
2015 zfree(&stat_config.walltime_run);
2017 if (smi_cost && smi_reset)
2018 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2020 evlist__delete(evsel_list);
2022 runtime_stat_delete(&stat_config);