4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ./hackbench 10
13 Performance counter stats for './hackbench 10':
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
27 0.154822978 seconds time elapsed
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 * Improvements and fixes by:
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
41 * Released under the GPL v2. (and only v2, not any later version)
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include <subcmd/parse-options.h>
49 #include "util/parse-events.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/color.h"
56 #include "util/stat.h"
57 #include "util/header.h"
58 #include "util/cpumap.h"
59 #include "util/thread.h"
60 #include "util/thread_map.h"
61 #include "util/counts.h"
62 #include "util/group.h"
63 #include "util/session.h"
64 #include "util/tool.h"
65 #include "util/string2.h"
66 #include "util/metricgroup.h"
70 #include <linux/time64.h>
71 #include <api/fs/fs.h>
75 #include <sys/prctl.h>
79 #include <sys/types.h>
84 #include <sys/resource.h>
86 #include "sane_ctype.h"
88 #define DEFAULT_SEPARATOR " "
89 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
91 static void print_counters(struct timespec *ts, int argc, const char **argv);
93 /* Default events used for perf stat -T */
94 static const char *transaction_attrs = {
106 /* More limited version when the CPU does not have all events. */
107 static const char * transaction_limited_attrs = {
117 static const char * topdown_attrs[] = {
118 "topdown-total-slots",
119 "topdown-slots-retired",
120 "topdown-recovery-bubbles",
121 "topdown-fetch-bubbles",
122 "topdown-slots-issued",
126 static const char *smi_cost_attrs = {
134 static struct perf_evlist *evsel_list;
136 static struct target target = {
140 #define METRIC_ONLY_LEN 20
142 static volatile pid_t child_pid = -1;
143 static int detailed_run = 0;
144 static bool transaction_run;
145 static bool topdown_run = false;
146 static bool smi_cost = false;
147 static bool smi_reset = false;
148 static int big_num_opt = -1;
149 static bool group = false;
150 static const char *pre_cmd = NULL;
151 static const char *post_cmd = NULL;
152 static bool sync_run = false;
153 static bool forever = false;
154 static bool force_metric_only = false;
155 static struct timespec ref_time;
156 static bool append_file;
157 static bool interval_count;
158 static const char *output_name;
159 static int output_fd;
163 struct perf_data data;
164 struct perf_session *session;
166 struct perf_tool tool;
168 struct cpu_map *cpus;
169 struct thread_map *threads;
170 enum aggr_mode aggr_mode;
173 static struct perf_stat perf_stat;
174 #define STAT_RECORD perf_stat.record
176 static volatile int done = 0;
178 static struct perf_stat_config stat_config = {
179 .aggr_mode = AGGR_GLOBAL,
181 .unit_width = 4, /* strlen("unit") */
183 .metric_only_len = METRIC_ONLY_LEN,
184 .walltime_nsecs_stats = &walltime_nsecs_stats,
188 static inline void diff_timespec(struct timespec *r, struct timespec *a,
191 r->tv_sec = a->tv_sec - b->tv_sec;
192 if (a->tv_nsec < b->tv_nsec) {
193 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
196 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
200 static void perf_stat__reset_stats(void)
204 perf_evlist__reset_stats(evsel_list);
205 perf_stat__reset_shadow_stats();
207 for (i = 0; i < stat_config.stats_num; i++)
208 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
211 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
212 union perf_event *event,
213 struct perf_sample *sample __maybe_unused,
214 struct machine *machine __maybe_unused)
216 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
217 pr_err("failed to write perf data, error: %m\n");
221 perf_stat.bytes_written += event->header.size;
225 static int write_stat_round_event(u64 tm, u64 type)
227 return perf_event__synthesize_stat_round(NULL, tm, type,
228 process_synthesized_event,
232 #define WRITE_STAT_ROUND_EVENT(time, interval) \
233 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
235 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
238 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
239 struct perf_counts_values *count)
241 struct perf_sample_id *sid = SID(counter, cpu, thread);
243 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
244 process_synthesized_event, NULL);
247 static int read_single_counter(struct perf_evsel *counter, int cpu,
248 int thread, struct timespec *rs)
250 if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
251 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
252 struct perf_counts_values *count =
253 perf_counts(counter->counts, cpu, thread);
254 count->ena = count->run = val;
258 return perf_evsel__read_counter(counter, cpu, thread);
262 * Read out the results of a single counter:
263 * do not aggregate counts across CPUs in system-wide mode
265 static int read_counter(struct perf_evsel *counter, struct timespec *rs)
267 int nthreads = thread_map__nr(evsel_list->threads);
268 int ncpus, cpu, thread;
270 if (target__has_cpu(&target) && !target__has_per_thread(&target))
271 ncpus = perf_evsel__nr_cpus(counter);
275 if (!counter->supported)
278 if (counter->system_wide)
281 for (thread = 0; thread < nthreads; thread++) {
282 for (cpu = 0; cpu < ncpus; cpu++) {
283 struct perf_counts_values *count;
285 count = perf_counts(counter->counts, cpu, thread);
288 * The leader's group read loads data into its group members
289 * (via perf_evsel__read_counter) and sets threir count->loaded.
291 if (!count->loaded &&
292 read_single_counter(counter, cpu, thread, rs)) {
293 counter->counts->scaled = -1;
294 perf_counts(counter->counts, cpu, thread)->ena = 0;
295 perf_counts(counter->counts, cpu, thread)->run = 0;
299 count->loaded = false;
302 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
303 pr_err("failed to write stat event\n");
309 fprintf(stat_config.output,
310 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
311 perf_evsel__name(counter),
313 count->val, count->ena, count->run);
321 static void read_counters(struct timespec *rs)
323 struct perf_evsel *counter;
326 evlist__for_each_entry(evsel_list, counter) {
327 ret = read_counter(counter, rs);
329 pr_debug("failed to read counter %s\n", counter->name);
331 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
332 pr_warning("failed to process counter %s\n", counter->name);
336 static void process_interval(void)
338 struct timespec ts, rs;
340 clock_gettime(CLOCK_MONOTONIC, &ts);
341 diff_timespec(&rs, &ts, &ref_time);
346 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
347 pr_err("failed to write stat round event\n");
350 init_stats(&walltime_nsecs_stats);
351 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
352 print_counters(&rs, 0, NULL);
355 static void enable_counters(void)
357 if (stat_config.initial_delay)
358 usleep(stat_config.initial_delay * USEC_PER_MSEC);
361 * We need to enable counters only if:
362 * - we don't have tracee (attaching to task or cpu)
363 * - we have initial delay configured
365 if (!target__none(&target) || stat_config.initial_delay)
366 perf_evlist__enable(evsel_list);
369 static void disable_counters(void)
372 * If we don't have tracee (attaching to task or cpu), counters may
373 * still be running. To get accurate group ratios, we must stop groups
374 * from counting before reading their constituent counters.
376 if (!target__none(&target))
377 perf_evlist__disable(evsel_list);
380 static volatile int workload_exec_errno;
383 * perf_evlist__prepare_workload will send a SIGUSR1
384 * if the fork fails, since we asked by setting its
385 * want_signal to true.
387 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
388 void *ucontext __maybe_unused)
390 workload_exec_errno = info->si_value.sival_int;
393 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
395 return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
398 static bool is_target_alive(struct target *_target,
399 struct thread_map *threads)
404 if (!target__has_task(_target))
407 for (i = 0; i < threads->nr; i++) {
410 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
411 threads->map[i].pid);
413 if (!stat(path, &st))
420 static int __run_perf_stat(int argc, const char **argv, int run_idx)
422 int interval = stat_config.interval;
423 int times = stat_config.times;
424 int timeout = stat_config.timeout;
426 unsigned long long t0, t1;
427 struct perf_evsel *counter;
431 const bool forks = (argc > 0);
432 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
435 ts.tv_sec = interval / USEC_PER_MSEC;
436 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
437 } else if (timeout) {
438 ts.tv_sec = timeout / USEC_PER_MSEC;
439 ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
446 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
447 workload_exec_failed_signal) < 0) {
448 perror("failed to prepare workload");
451 child_pid = evsel_list->workload.pid;
455 perf_evlist__set_leader(evsel_list);
457 evlist__for_each_entry(evsel_list, counter) {
459 if (create_perf_stat_counter(counter, &stat_config, &target) < 0) {
461 /* Weak group failed. Reset the group. */
462 if ((errno == EINVAL || errno == EBADF) &&
463 counter->leader != counter &&
464 counter->weak_group) {
465 counter = perf_evlist__reset_weak_group(evsel_list, counter);
470 * PPC returns ENXIO for HW counters until 2.6.37
471 * (behavior changed with commit b0a873e).
473 if (errno == EINVAL || errno == ENOSYS ||
474 errno == ENOENT || errno == EOPNOTSUPP ||
477 ui__warning("%s event is not supported by the kernel.\n",
478 perf_evsel__name(counter));
479 counter->supported = false;
481 if ((counter->leader != counter) ||
482 !(counter->leader->nr_members > 1))
484 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
486 ui__warning("%s\n", msg);
488 } else if (target__has_per_thread(&target) &&
489 evsel_list->threads &&
490 evsel_list->threads->err_thread != -1) {
492 * For global --per-thread case, skip current
495 if (!thread_map__remove(evsel_list->threads,
496 evsel_list->threads->err_thread)) {
497 evsel_list->threads->err_thread = -1;
502 perf_evsel__open_strerror(counter, &target,
503 errno, msg, sizeof(msg));
504 ui__error("%s\n", msg);
507 kill(child_pid, SIGTERM);
511 counter->supported = true;
513 l = strlen(counter->unit);
514 if (l > stat_config.unit_width)
515 stat_config.unit_width = l;
517 if (perf_evsel__should_store_id(counter) &&
518 perf_evsel__store_ids(counter, evsel_list))
522 if (perf_evlist__apply_filters(evsel_list, &counter)) {
523 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
524 counter->filter, perf_evsel__name(counter), errno,
525 str_error_r(errno, msg, sizeof(msg)));
530 int err, fd = perf_data__fd(&perf_stat.data);
533 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
535 err = perf_session__write_header(perf_stat.session, evsel_list,
542 err = perf_stat_synthesize_config(&stat_config, NULL, evsel_list,
543 process_synthesized_event, is_pipe);
549 * Enable counters and exec the command:
552 clock_gettime(CLOCK_MONOTONIC, &ref_time);
555 perf_evlist__start_workload(evsel_list);
558 if (interval || timeout) {
559 while (!waitpid(child_pid, &status, WNOHANG)) {
560 nanosleep(&ts, NULL);
564 if (interval_count && !(--times))
569 wait4(child_pid, &status, 0, &stat_config.ru_data);
571 if (workload_exec_errno) {
572 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
573 pr_err("Workload failed: %s\n", emsg);
577 if (WIFSIGNALED(status))
578 psignal(WTERMSIG(status), argv[0]);
582 nanosleep(&ts, NULL);
583 if (!is_target_alive(&target, evsel_list->threads))
589 if (interval_count && !(--times))
599 if (stat_config.walltime_run_table)
600 stat_config.walltime_run[run_idx] = t1 - t0;
602 update_stats(&walltime_nsecs_stats, t1 - t0);
605 * Closing a group leader splits the group, and as we only disable
606 * group leaders, results in remaining events becoming enabled. To
607 * avoid arbitrary skew, we must read all counters before closing any
610 read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
611 perf_evlist__close(evsel_list);
613 return WEXITSTATUS(status);
616 static int run_perf_stat(int argc, const char **argv, int run_idx)
621 ret = system(pre_cmd);
629 ret = __run_perf_stat(argc, argv, run_idx);
634 ret = system(post_cmd);
642 static void print_counters(struct timespec *ts, int argc, const char **argv)
644 /* Do not print anything if we record to the pipe. */
645 if (STAT_RECORD && perf_stat.data.is_pipe)
648 perf_evlist__print_counters(evsel_list, &stat_config, &target,
652 static volatile int signr = -1;
654 static void skip_signal(int signo)
656 if ((child_pid == -1) || stat_config.interval)
661 * render child_pid harmless
662 * won't send SIGTERM to a random
663 * process in case of race condition
664 * and fast PID recycling
669 static void sig_atexit(void)
674 * avoid race condition with SIGCHLD handler
675 * in skip_signal() which is modifying child_pid
676 * goal is to avoid send SIGTERM to a random
680 sigaddset(&set, SIGCHLD);
681 sigprocmask(SIG_BLOCK, &set, &oset);
684 kill(child_pid, SIGTERM);
686 sigprocmask(SIG_SETMASK, &oset, NULL);
691 signal(signr, SIG_DFL);
692 kill(getpid(), signr);
695 static int stat__set_big_num(const struct option *opt __maybe_unused,
696 const char *s __maybe_unused, int unset)
698 big_num_opt = unset ? 0 : 1;
702 static int enable_metric_only(const struct option *opt __maybe_unused,
703 const char *s __maybe_unused, int unset)
705 force_metric_only = true;
706 stat_config.metric_only = !unset;
710 static int parse_metric_groups(const struct option *opt,
712 int unset __maybe_unused)
714 return metricgroup__parse_groups(opt, str, &stat_config.metric_events);
717 static struct option stat_options[] = {
718 OPT_BOOLEAN('T', "transaction", &transaction_run,
719 "hardware transaction statistics"),
720 OPT_CALLBACK('e', "event", &evsel_list, "event",
721 "event selector. use 'perf list' to list available events",
722 parse_events_option),
723 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
724 "event filter", parse_filter),
725 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
726 "child tasks do not inherit counters"),
727 OPT_STRING('p', "pid", &target.pid, "pid",
728 "stat events on existing process id"),
729 OPT_STRING('t', "tid", &target.tid, "tid",
730 "stat events on existing thread id"),
731 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
732 "system-wide collection from all CPUs"),
733 OPT_BOOLEAN('g', "group", &group,
734 "put the counters into a counter group"),
735 OPT_BOOLEAN(0, "scale", &stat_config.scale,
736 "Use --no-scale to disable counter scaling for multiplexing"),
737 OPT_INCR('v', "verbose", &verbose,
738 "be more verbose (show counter open errors, etc)"),
739 OPT_INTEGER('r', "repeat", &stat_config.run_count,
740 "repeat command and print average + stddev (max: 100, forever: 0)"),
741 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
742 "display details about each run (only with -r option)"),
743 OPT_BOOLEAN('n', "null", &stat_config.null_run,
744 "null run - dont start any counters"),
745 OPT_INCR('d', "detailed", &detailed_run,
746 "detailed run - start a lot of events"),
747 OPT_BOOLEAN('S', "sync", &sync_run,
748 "call sync() before starting a run"),
749 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
750 "print large numbers with thousands\' separators",
752 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
753 "list of cpus to monitor in system-wide"),
754 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
755 "disable CPU count aggregation", AGGR_NONE),
756 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
757 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
758 "print counts with custom separator"),
759 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
760 "monitor event in cgroup name only", parse_cgroups),
761 OPT_STRING('o', "output", &output_name, "file", "output file name"),
762 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
763 OPT_INTEGER(0, "log-fd", &output_fd,
764 "log output to fd, instead of stderr"),
765 OPT_STRING(0, "pre", &pre_cmd, "command",
766 "command to run prior to the measured command"),
767 OPT_STRING(0, "post", &post_cmd, "command",
768 "command to run after to the measured command"),
769 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
770 "print counts at regular interval in ms "
771 "(overhead is possible for values <= 100ms)"),
772 OPT_INTEGER(0, "interval-count", &stat_config.times,
773 "print counts for fixed number of times"),
774 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
775 "clear screen in between new interval"),
776 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
777 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
778 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
779 "aggregate counts per processor socket", AGGR_SOCKET),
780 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
781 "aggregate counts per physical processor core", AGGR_CORE),
782 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
783 "aggregate counts per thread", AGGR_THREAD),
784 OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
785 "ms to wait before starting measurement after program start"),
786 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
787 "Only print computed metrics. No raw values", enable_metric_only),
788 OPT_BOOLEAN(0, "topdown", &topdown_run,
789 "measure topdown level 1 statistics"),
790 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
792 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
793 "monitor specified metrics or metric groups (separated by ,)",
794 parse_metric_groups),
798 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
799 struct cpu_map *map, int cpu)
801 return cpu_map__get_socket(map, cpu, NULL);
804 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
805 struct cpu_map *map, int cpu)
807 return cpu_map__get_core(map, cpu, NULL);
810 static int cpu_map__get_max(struct cpu_map *map)
814 for (i = 0; i < map->nr; i++) {
815 if (map->map[i] > max)
822 static int perf_stat__get_aggr(struct perf_stat_config *config,
823 aggr_get_id_t get_id, struct cpu_map *map, int idx)
832 if (config->cpus_aggr_map->map[cpu] == -1)
833 config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
835 return config->cpus_aggr_map->map[cpu];
838 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
839 struct cpu_map *map, int idx)
841 return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
844 static int perf_stat__get_core_cached(struct perf_stat_config *config,
845 struct cpu_map *map, int idx)
847 return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
850 static bool term_percore_set(void)
852 struct perf_evsel *counter;
854 evlist__for_each_entry(evsel_list, counter) {
855 if (counter->percore)
862 static int perf_stat_init_aggr_mode(void)
866 switch (stat_config.aggr_mode) {
868 if (cpu_map__build_socket_map(evsel_list->cpus, &stat_config.aggr_map)) {
869 perror("cannot build socket map");
872 stat_config.aggr_get_id = perf_stat__get_socket_cached;
875 if (cpu_map__build_core_map(evsel_list->cpus, &stat_config.aggr_map)) {
876 perror("cannot build core map");
879 stat_config.aggr_get_id = perf_stat__get_core_cached;
882 if (term_percore_set()) {
883 if (cpu_map__build_core_map(evsel_list->cpus,
884 &stat_config.aggr_map)) {
885 perror("cannot build core map");
888 stat_config.aggr_get_id = perf_stat__get_core_cached;
899 * The evsel_list->cpus is the base we operate on,
900 * taking the highest cpu number to be the size of
901 * the aggregation translate cpumap.
903 nr = cpu_map__get_max(evsel_list->cpus);
904 stat_config.cpus_aggr_map = cpu_map__empty_new(nr + 1);
905 return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
908 static void perf_stat__exit_aggr_mode(void)
910 cpu_map__put(stat_config.aggr_map);
911 cpu_map__put(stat_config.cpus_aggr_map);
912 stat_config.aggr_map = NULL;
913 stat_config.cpus_aggr_map = NULL;
916 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
925 if (cpu >= env->nr_cpus_avail)
931 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
933 struct perf_env *env = data;
934 int cpu = perf_env__get_cpu(env, map, idx);
936 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
939 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
941 struct perf_env *env = data;
942 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
945 int socket_id = env->cpu[cpu].socket_id;
948 * Encode socket in upper 16 bits
949 * core_id is relative to socket, and
950 * we need a global id. So we combine
953 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
959 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
960 struct cpu_map **sockp)
962 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
965 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
966 struct cpu_map **corep)
968 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
971 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
972 struct cpu_map *map, int idx)
974 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
977 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
978 struct cpu_map *map, int idx)
980 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
983 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
985 struct perf_env *env = &st->session->header.env;
987 switch (stat_config.aggr_mode) {
989 if (perf_env__build_socket_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
990 perror("cannot build socket map");
993 stat_config.aggr_get_id = perf_stat__get_socket_file;
996 if (perf_env__build_core_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
997 perror("cannot build core map");
1000 stat_config.aggr_get_id = perf_stat__get_core_file;
1013 static int topdown_filter_events(const char **attr, char **str, bool use_group)
1020 for (i = 0; attr[i]; i++) {
1021 if (pmu_have_event("cpu", attr[i])) {
1022 len += strlen(attr[i]) + 1;
1023 attr[i - off] = attr[i];
1027 attr[i - off] = NULL;
1029 *str = malloc(len + 1 + 2);
1039 for (i = 0; attr[i]; i++) {
1052 __weak bool arch_topdown_check_group(bool *warn)
1058 __weak void arch_topdown_group_warn(void)
1063 * Add default attributes, if there were no attributes specified or
1064 * if -d/--detailed, -d -d or -d -d -d is used:
1066 static int add_default_attributes(void)
1069 struct perf_event_attr default_attrs0[] = {
1071 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1072 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1073 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1074 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1076 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1078 struct perf_event_attr frontend_attrs[] = {
1079 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1081 struct perf_event_attr backend_attrs[] = {
1082 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1084 struct perf_event_attr default_attrs1[] = {
1085 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1086 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1087 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1092 * Detailed stats (-d), covering the L1 and last level data caches:
1094 struct perf_event_attr detailed_attrs[] = {
1096 { .type = PERF_TYPE_HW_CACHE,
1098 PERF_COUNT_HW_CACHE_L1D << 0 |
1099 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1100 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1102 { .type = PERF_TYPE_HW_CACHE,
1104 PERF_COUNT_HW_CACHE_L1D << 0 |
1105 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1106 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1108 { .type = PERF_TYPE_HW_CACHE,
1110 PERF_COUNT_HW_CACHE_LL << 0 |
1111 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1112 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1114 { .type = PERF_TYPE_HW_CACHE,
1116 PERF_COUNT_HW_CACHE_LL << 0 |
1117 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1118 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1122 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1124 struct perf_event_attr very_detailed_attrs[] = {
1126 { .type = PERF_TYPE_HW_CACHE,
1128 PERF_COUNT_HW_CACHE_L1I << 0 |
1129 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1130 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1132 { .type = PERF_TYPE_HW_CACHE,
1134 PERF_COUNT_HW_CACHE_L1I << 0 |
1135 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1136 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1138 { .type = PERF_TYPE_HW_CACHE,
1140 PERF_COUNT_HW_CACHE_DTLB << 0 |
1141 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1142 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1144 { .type = PERF_TYPE_HW_CACHE,
1146 PERF_COUNT_HW_CACHE_DTLB << 0 |
1147 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1148 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1150 { .type = PERF_TYPE_HW_CACHE,
1152 PERF_COUNT_HW_CACHE_ITLB << 0 |
1153 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1154 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1156 { .type = PERF_TYPE_HW_CACHE,
1158 PERF_COUNT_HW_CACHE_ITLB << 0 |
1159 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1160 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1165 * Very, very detailed stats (-d -d -d), adding prefetch events:
1167 struct perf_event_attr very_very_detailed_attrs[] = {
1169 { .type = PERF_TYPE_HW_CACHE,
1171 PERF_COUNT_HW_CACHE_L1D << 0 |
1172 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1173 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1175 { .type = PERF_TYPE_HW_CACHE,
1177 PERF_COUNT_HW_CACHE_L1D << 0 |
1178 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1179 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1181 struct parse_events_error errinfo;
1183 /* Set attrs if no event is selected and !null_run: */
1184 if (stat_config.null_run)
1187 if (transaction_run) {
1188 /* Handle -T as -M transaction. Once platform specific metrics
1189 * support has been added to the json files, all archictures
1190 * will use this approach. To determine transaction support
1191 * on an architecture test for such a metric name.
1193 if (metricgroup__has_metric("transaction")) {
1194 struct option opt = { .value = &evsel_list };
1196 return metricgroup__parse_groups(&opt, "transaction",
1197 &stat_config.metric_events);
1200 if (pmu_have_event("cpu", "cycles-ct") &&
1201 pmu_have_event("cpu", "el-start"))
1202 err = parse_events(evsel_list, transaction_attrs,
1205 err = parse_events(evsel_list,
1206 transaction_limited_attrs,
1209 fprintf(stderr, "Cannot set up transaction events\n");
1210 parse_events_print_error(&errinfo, transaction_attrs);
1219 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1220 fprintf(stderr, "freeze_on_smi is not supported.\n");
1225 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1226 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1232 if (pmu_have_event("msr", "aperf") &&
1233 pmu_have_event("msr", "smi")) {
1234 if (!force_metric_only)
1235 stat_config.metric_only = true;
1236 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1238 fprintf(stderr, "To measure SMI cost, it needs "
1239 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1240 parse_events_print_error(&errinfo, smi_cost_attrs);
1244 fprintf(stderr, "Cannot set up SMI cost events\n");
1254 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1255 stat_config.aggr_mode != AGGR_CORE) {
1256 pr_err("top down event configuration requires --per-core mode\n");
1259 stat_config.aggr_mode = AGGR_CORE;
1260 if (nr_cgroups || !target__has_cpu(&target)) {
1261 pr_err("top down event configuration requires system-wide mode (-a)\n");
1265 if (!force_metric_only)
1266 stat_config.metric_only = true;
1267 if (topdown_filter_events(topdown_attrs, &str,
1268 arch_topdown_check_group(&warn)) < 0) {
1269 pr_err("Out of memory\n");
1272 if (topdown_attrs[0] && str) {
1274 arch_topdown_group_warn();
1275 err = parse_events(evsel_list, str, &errinfo);
1278 "Cannot set up top down events %s: %d\n",
1281 parse_events_print_error(&errinfo, str);
1285 fprintf(stderr, "System does not support topdown\n");
1291 if (!evsel_list->nr_entries) {
1292 if (target__has_cpu(&target))
1293 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1295 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1297 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1298 if (perf_evlist__add_default_attrs(evsel_list,
1299 frontend_attrs) < 0)
1302 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1303 if (perf_evlist__add_default_attrs(evsel_list,
1307 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1311 /* Detailed events get appended to the event list: */
1313 if (detailed_run < 1)
1316 /* Append detailed run extra attributes: */
1317 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1320 if (detailed_run < 2)
1323 /* Append very detailed run extra attributes: */
1324 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1327 if (detailed_run < 3)
1330 /* Append very, very detailed run extra attributes: */
1331 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1334 static const char * const stat_record_usage[] = {
1335 "perf stat record [<options>]",
1339 static void init_features(struct perf_session *session)
1343 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1344 perf_header__set_feat(&session->header, feat);
1346 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1347 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1348 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1349 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1350 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1353 static int __cmd_record(int argc, const char **argv)
1355 struct perf_session *session;
1356 struct perf_data *data = &perf_stat.data;
1358 argc = parse_options(argc, argv, stat_options, stat_record_usage,
1359 PARSE_OPT_STOP_AT_NON_OPTION);
1362 data->path = output_name;
1364 if (stat_config.run_count != 1 || forever) {
1365 pr_err("Cannot use -r option with perf stat record.\n");
1369 session = perf_session__new(data, false, NULL);
1370 if (session == NULL) {
1371 pr_err("Perf session creation failed.\n");
1375 init_features(session);
1377 session->evlist = evsel_list;
1378 perf_stat.session = session;
1379 perf_stat.record = true;
1383 static int process_stat_round_event(struct perf_session *session,
1384 union perf_event *event)
1386 struct stat_round_event *stat_round = &event->stat_round;
1387 struct perf_evsel *counter;
1388 struct timespec tsh, *ts = NULL;
1389 const char **argv = session->header.env.cmdline_argv;
1390 int argc = session->header.env.nr_cmdline;
1392 evlist__for_each_entry(evsel_list, counter)
1393 perf_stat_process_counter(&stat_config, counter);
1395 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1396 update_stats(&walltime_nsecs_stats, stat_round->time);
1398 if (stat_config.interval && stat_round->time) {
1399 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
1400 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1404 print_counters(ts, argc, argv);
1409 int process_stat_config_event(struct perf_session *session,
1410 union perf_event *event)
1412 struct perf_tool *tool = session->tool;
1413 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1415 perf_event__read_stat_config(&stat_config, &event->stat_config);
1417 if (cpu_map__empty(st->cpus)) {
1418 if (st->aggr_mode != AGGR_UNSET)
1419 pr_warning("warning: processing task data, aggregation mode not set\n");
1423 if (st->aggr_mode != AGGR_UNSET)
1424 stat_config.aggr_mode = st->aggr_mode;
1426 if (perf_stat.data.is_pipe)
1427 perf_stat_init_aggr_mode();
1429 perf_stat_init_aggr_mode_file(st);
1434 static int set_maps(struct perf_stat *st)
1436 if (!st->cpus || !st->threads)
1439 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1442 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
1444 if (perf_evlist__alloc_stats(evsel_list, true))
1447 st->maps_allocated = true;
1452 int process_thread_map_event(struct perf_session *session,
1453 union perf_event *event)
1455 struct perf_tool *tool = session->tool;
1456 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1459 pr_warning("Extra thread map event, ignoring.\n");
1463 st->threads = thread_map__new_event(&event->thread_map);
1467 return set_maps(st);
1471 int process_cpu_map_event(struct perf_session *session,
1472 union perf_event *event)
1474 struct perf_tool *tool = session->tool;
1475 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1476 struct cpu_map *cpus;
1479 pr_warning("Extra cpu map event, ignoring.\n");
1483 cpus = cpu_map__new_data(&event->cpu_map.data);
1488 return set_maps(st);
1491 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
1495 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
1499 config->stats_num = nthreads;
1501 for (i = 0; i < nthreads; i++)
1502 runtime_stat__init(&config->stats[i]);
1507 static void runtime_stat_delete(struct perf_stat_config *config)
1514 for (i = 0; i < config->stats_num; i++)
1515 runtime_stat__exit(&config->stats[i]);
1517 free(config->stats);
1520 static const char * const stat_report_usage[] = {
1521 "perf stat report [<options>]",
1525 static struct perf_stat perf_stat = {
1527 .attr = perf_event__process_attr,
1528 .event_update = perf_event__process_event_update,
1529 .thread_map = process_thread_map_event,
1530 .cpu_map = process_cpu_map_event,
1531 .stat_config = process_stat_config_event,
1532 .stat = perf_event__process_stat_event,
1533 .stat_round = process_stat_round_event,
1535 .aggr_mode = AGGR_UNSET,
1538 static int __cmd_report(int argc, const char **argv)
1540 struct perf_session *session;
1541 const struct option options[] = {
1542 OPT_STRING('i', "input", &input_name, "file", "input file name"),
1543 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1544 "aggregate counts per processor socket", AGGR_SOCKET),
1545 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1546 "aggregate counts per physical processor core", AGGR_CORE),
1547 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1548 "disable CPU count aggregation", AGGR_NONE),
1554 argc = parse_options(argc, argv, options, stat_report_usage, 0);
1556 if (!input_name || !strlen(input_name)) {
1557 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1560 input_name = "perf.data";
1563 perf_stat.data.path = input_name;
1564 perf_stat.data.mode = PERF_DATA_MODE_READ;
1566 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
1567 if (session == NULL)
1570 perf_stat.session = session;
1571 stat_config.output = stderr;
1572 evsel_list = session->evlist;
1574 ret = perf_session__process_events(session);
1578 perf_session__delete(session);
1582 static void setup_system_wide(int forks)
1585 * Make system wide (-a) the default target if
1586 * no target was specified and one of following
1587 * conditions is met:
1589 * - there's no workload specified
1590 * - there is workload specified but all requested
1591 * events are system wide events
1593 if (!target__none(&target))
1597 target.system_wide = true;
1599 struct perf_evsel *counter;
1601 evlist__for_each_entry(evsel_list, counter) {
1602 if (!counter->system_wide)
1606 if (evsel_list->nr_entries)
1607 target.system_wide = true;
1611 int cmd_stat(int argc, const char **argv)
1613 const char * const stat_usage[] = {
1614 "perf stat [<options>] [<command>]",
1617 int status = -EINVAL, run_idx;
1619 FILE *output = stderr;
1620 unsigned int interval, timeout;
1621 const char * const stat_subcommands[] = { "record", "report" };
1623 setlocale(LC_ALL, "");
1625 evsel_list = perf_evlist__new();
1626 if (evsel_list == NULL)
1629 parse_events__shrink_config_terms();
1631 /* String-parsing callback-based options would segfault when negated */
1632 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
1633 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
1634 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
1636 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1637 (const char **) stat_usage,
1638 PARSE_OPT_STOP_AT_NON_OPTION);
1639 perf_stat__collect_metric_expr(evsel_list);
1640 perf_stat__init_shadow_stats();
1642 if (stat_config.csv_sep) {
1643 stat_config.csv_output = true;
1644 if (!strcmp(stat_config.csv_sep, "\\t"))
1645 stat_config.csv_sep = "\t";
1647 stat_config.csv_sep = DEFAULT_SEPARATOR;
1649 if (argc && !strncmp(argv[0], "rec", 3)) {
1650 argc = __cmd_record(argc, argv);
1653 } else if (argc && !strncmp(argv[0], "rep", 3))
1654 return __cmd_report(argc, argv);
1656 interval = stat_config.interval;
1657 timeout = stat_config.timeout;
1660 * For record command the -o is already taken care of.
1662 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1665 if (output_name && output_fd) {
1666 fprintf(stderr, "cannot use both --output and --log-fd\n");
1667 parse_options_usage(stat_usage, stat_options, "o", 1);
1668 parse_options_usage(NULL, stat_options, "log-fd", 0);
1672 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
1673 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
1677 if (stat_config.metric_only && stat_config.run_count > 1) {
1678 fprintf(stderr, "--metric-only is not supported with -r\n");
1682 if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
1683 fprintf(stderr, "--table is only supported with -r\n");
1684 parse_options_usage(stat_usage, stat_options, "r", 1);
1685 parse_options_usage(NULL, stat_options, "table", 0);
1689 if (output_fd < 0) {
1690 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1691 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1697 mode = append_file ? "a" : "w";
1699 output = fopen(output_name, mode);
1701 perror("failed to create output file");
1704 clock_gettime(CLOCK_REALTIME, &tm);
1705 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1706 } else if (output_fd > 0) {
1707 mode = append_file ? "a" : "w";
1708 output = fdopen(output_fd, mode);
1710 perror("Failed opening logfd");
1715 stat_config.output = output;
1718 * let the spreadsheet do the pretty-printing
1720 if (stat_config.csv_output) {
1721 /* User explicitly passed -B? */
1722 if (big_num_opt == 1) {
1723 fprintf(stderr, "-B option not supported with -x\n");
1724 parse_options_usage(stat_usage, stat_options, "B", 1);
1725 parse_options_usage(NULL, stat_options, "x", 1);
1727 } else /* Nope, so disable big number formatting */
1728 stat_config.big_num = false;
1729 } else if (big_num_opt == 0) /* User passed --no-big-num */
1730 stat_config.big_num = false;
1732 setup_system_wide(argc);
1735 * Display user/system times only for single
1736 * run and when there's specified tracee.
1738 if ((stat_config.run_count == 1) && target__none(&target))
1739 stat_config.ru_display = true;
1741 if (stat_config.run_count < 0) {
1742 pr_err("Run count must be a positive number\n");
1743 parse_options_usage(stat_usage, stat_options, "r", 1);
1745 } else if (stat_config.run_count == 0) {
1747 stat_config.run_count = 1;
1750 if (stat_config.walltime_run_table) {
1751 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
1752 if (!stat_config.walltime_run) {
1753 pr_err("failed to setup -r option");
1758 if ((stat_config.aggr_mode == AGGR_THREAD) &&
1759 !target__has_task(&target)) {
1760 if (!target.system_wide || target.cpu_list) {
1761 fprintf(stderr, "The --per-thread option is only "
1762 "available when monitoring via -p -t -a "
1763 "options or only --per-thread.\n");
1764 parse_options_usage(NULL, stat_options, "p", 1);
1765 parse_options_usage(NULL, stat_options, "t", 1);
1771 * no_aggr, cgroup are for system-wide only
1772 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1774 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1775 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1776 !target__has_cpu(&target)) {
1777 fprintf(stderr, "both cgroup and no-aggregation "
1778 "modes only available in system-wide mode\n");
1780 parse_options_usage(stat_usage, stat_options, "G", 1);
1781 parse_options_usage(NULL, stat_options, "A", 1);
1782 parse_options_usage(NULL, stat_options, "a", 1);
1786 if (add_default_attributes())
1789 target__validate(&target);
1791 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
1792 target.per_thread = true;
1794 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1795 if (target__has_task(&target)) {
1796 pr_err("Problems finding threads of monitor\n");
1797 parse_options_usage(stat_usage, stat_options, "p", 1);
1798 parse_options_usage(NULL, stat_options, "t", 1);
1799 } else if (target__has_cpu(&target)) {
1800 perror("failed to parse CPUs map");
1801 parse_options_usage(stat_usage, stat_options, "C", 1);
1802 parse_options_usage(NULL, stat_options, "a", 1);
1808 * Initialize thread_map with comm names,
1809 * so we could print it out on output.
1811 if (stat_config.aggr_mode == AGGR_THREAD) {
1812 thread_map__read_comms(evsel_list->threads);
1813 if (target.system_wide) {
1814 if (runtime_stat_new(&stat_config,
1815 thread_map__nr(evsel_list->threads))) {
1821 if (stat_config.times && interval)
1822 interval_count = true;
1823 else if (stat_config.times && !interval) {
1824 pr_err("interval-count option should be used together with "
1825 "interval-print.\n");
1826 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
1827 parse_options_usage(stat_usage, stat_options, "I", 1);
1831 if (timeout && timeout < 100) {
1833 pr_err("timeout must be >= 10ms.\n");
1834 parse_options_usage(stat_usage, stat_options, "timeout", 0);
1837 pr_warning("timeout < 100ms. "
1838 "The overhead percentage could be high in some cases. "
1839 "Please proceed with caution.\n");
1841 if (timeout && interval) {
1842 pr_err("timeout option is not supported with interval-print.\n");
1843 parse_options_usage(stat_usage, stat_options, "timeout", 0);
1844 parse_options_usage(stat_usage, stat_options, "I", 1);
1848 if (perf_evlist__alloc_stats(evsel_list, interval))
1851 if (perf_stat_init_aggr_mode())
1855 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
1856 * while avoiding that older tools show confusing messages.
1858 * However for pipe sessions we need to keep it zero,
1859 * because script's perf_evsel__check_attr is triggered
1860 * by attr->sample_type != 0, and we can't run it on
1863 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
1866 * We dont want to block the signals - that would cause
1867 * child tasks to inherit that and Ctrl-C would not work.
1868 * What we want is for Ctrl-C to work in the exec()-ed
1869 * task, but being ignored by perf stat itself:
1873 signal(SIGINT, skip_signal);
1874 signal(SIGCHLD, skip_signal);
1875 signal(SIGALRM, skip_signal);
1876 signal(SIGABRT, skip_signal);
1879 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
1880 if (stat_config.run_count != 1 && verbose > 0)
1881 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1884 status = run_perf_stat(argc, argv, run_idx);
1885 if (forever && status != -1) {
1886 print_counters(NULL, argc, argv);
1887 perf_stat__reset_stats();
1891 if (!forever && status != -1 && !interval)
1892 print_counters(NULL, argc, argv);
1896 * We synthesize the kernel mmap record just so that older tools
1897 * don't emit warnings about not being able to resolve symbols
1898 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
1899 * a saner message about no samples being in the perf.data file.
1901 * This also serves to suppress a warning about f_header.data.size == 0
1902 * in header.c at the moment 'perf stat record' gets introduced, which
1903 * is not really needed once we start adding the stat specific PERF_RECORD_
1904 * records, but the need to suppress the kptr_restrict messages in older
1905 * tools remain -acme
1907 int fd = perf_data__fd(&perf_stat.data);
1908 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
1909 process_synthesized_event,
1910 &perf_stat.session->machines.host);
1912 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
1913 "older tools may produce warnings about this file\n.");
1917 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
1918 pr_err("failed to write stat round event\n");
1921 if (!perf_stat.data.is_pipe) {
1922 perf_stat.session->header.data_size += perf_stat.bytes_written;
1923 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
1926 perf_session__delete(perf_stat.session);
1929 perf_stat__exit_aggr_mode();
1930 perf_evlist__free_stats(evsel_list);
1932 free(stat_config.walltime_run);
1934 if (smi_cost && smi_reset)
1935 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
1937 perf_evlist__delete(evsel_list);
1939 runtime_stat_delete(&stat_config);