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 "util/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"
63 #include <sys/prctl.h>
66 #define DEFAULT_SEPARATOR " "
67 #define CNTR_NOT_SUPPORTED "<not supported>"
68 #define CNTR_NOT_COUNTED "<not counted>"
70 static void print_stat(int argc, const char **argv);
71 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
72 static void print_counter(struct perf_evsel *counter, char *prefix);
73 static void print_aggr(char *prefix);
75 /* Default events used for perf stat -T */
76 static const char * const transaction_attrs[] = {
88 /* More limited version when the CPU does not have all events. */
89 static const char * const transaction_limited_attrs[] = {
99 /* must match transaction_attrs and the beginning limited_attrs */
110 static struct perf_evlist *evsel_list;
112 static struct target target = {
123 static int run_count = 1;
124 static bool no_inherit = false;
125 static bool scale = true;
126 static enum aggr_mode aggr_mode = AGGR_GLOBAL;
127 static volatile pid_t child_pid = -1;
128 static bool null_run = false;
129 static int detailed_run = 0;
130 static bool transaction_run;
131 static bool big_num = true;
132 static int big_num_opt = -1;
133 static const char *csv_sep = NULL;
134 static bool csv_output = false;
135 static bool group = false;
136 static FILE *output = NULL;
137 static const char *pre_cmd = NULL;
138 static const char *post_cmd = NULL;
139 static bool sync_run = false;
140 static unsigned int interval = 0;
141 static unsigned int initial_delay = 0;
142 static unsigned int unit_width = 4; /* strlen("unit") */
143 static bool forever = false;
144 static struct timespec ref_time;
145 static struct cpu_map *aggr_map;
146 static int (*aggr_get_id)(struct cpu_map *m, int cpu);
148 static volatile int done = 0;
151 struct stats res_stats[3];
154 static inline void diff_timespec(struct timespec *r, struct timespec *a,
157 r->tv_sec = a->tv_sec - b->tv_sec;
158 if (a->tv_nsec < b->tv_nsec) {
159 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
162 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
166 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
168 return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
171 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
173 return perf_evsel__cpus(evsel)->nr;
176 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
179 struct perf_stat *ps = evsel->priv;
181 for (i = 0; i < 3; i++)
182 init_stats(&ps->res_stats[i]);
185 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
187 evsel->priv = zalloc(sizeof(struct perf_stat));
188 if (evsel->priv == NULL)
190 perf_evsel__reset_stat_priv(evsel);
194 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
199 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
204 sz = sizeof(*evsel->counts) +
205 (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
211 evsel->prev_raw_counts = addr;
216 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
218 zfree(&evsel->prev_raw_counts);
221 static void perf_evlist__free_stats(struct perf_evlist *evlist)
223 struct perf_evsel *evsel;
225 evlist__for_each(evlist, evsel) {
226 perf_evsel__free_stat_priv(evsel);
227 perf_evsel__free_counts(evsel);
228 perf_evsel__free_prev_raw_counts(evsel);
232 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
234 struct perf_evsel *evsel;
236 evlist__for_each(evlist, evsel) {
237 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
238 perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
239 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
246 perf_evlist__free_stats(evlist);
250 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
251 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
252 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
253 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
254 static struct stats runtime_branches_stats[MAX_NR_CPUS];
255 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
256 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
257 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
258 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
259 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
260 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
261 static struct stats runtime_cycles_in_tx_stats[MAX_NR_CPUS];
262 static struct stats walltime_nsecs_stats;
263 static struct stats runtime_transaction_stats[MAX_NR_CPUS];
264 static struct stats runtime_elision_stats[MAX_NR_CPUS];
266 static void perf_stat__reset_stats(struct perf_evlist *evlist)
268 struct perf_evsel *evsel;
270 evlist__for_each(evlist, evsel) {
271 perf_evsel__reset_stat_priv(evsel);
272 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
275 memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
276 memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
277 memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
278 memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
279 memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
280 memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
281 memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
282 memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
283 memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
284 memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
285 memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
286 memset(runtime_cycles_in_tx_stats, 0,
287 sizeof(runtime_cycles_in_tx_stats));
288 memset(runtime_transaction_stats, 0,
289 sizeof(runtime_transaction_stats));
290 memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
291 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
294 static int create_perf_stat_counter(struct perf_evsel *evsel)
296 struct perf_event_attr *attr = &evsel->attr;
299 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
300 PERF_FORMAT_TOTAL_TIME_RUNNING;
302 attr->inherit = !no_inherit;
304 if (target__has_cpu(&target))
305 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
307 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
310 attr->enable_on_exec = 1;
313 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
317 * Does the counter have nsecs as a unit?
319 static inline int nsec_counter(struct perf_evsel *evsel)
321 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
322 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
328 static struct perf_evsel *nth_evsel(int n)
330 static struct perf_evsel **array;
331 static int array_len;
332 struct perf_evsel *ev;
335 /* Assumes this only called when evsel_list does not change anymore. */
337 evlist__for_each(evsel_list, ev)
339 array = malloc(array_len * sizeof(void *));
343 evlist__for_each(evsel_list, ev)
352 * Update various tracking values we maintain to print
353 * more semantic information such as miss/hit ratios,
354 * instruction rates, etc:
356 static void update_shadow_stats(struct perf_evsel *counter, u64 *count,
359 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
360 update_stats(&runtime_nsecs_stats[cpu], count[0]);
361 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
362 update_stats(&runtime_cycles_stats[cpu], count[0]);
363 else if (transaction_run &&
364 perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX)))
365 update_stats(&runtime_cycles_in_tx_stats[cpu], count[0]);
366 else if (transaction_run &&
367 perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START)))
368 update_stats(&runtime_transaction_stats[cpu], count[0]);
369 else if (transaction_run &&
370 perf_evsel__cmp(counter, nth_evsel(T_ELISION_START)))
371 update_stats(&runtime_elision_stats[cpu], count[0]);
372 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
373 update_stats(&runtime_stalled_cycles_front_stats[cpu], count[0]);
374 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
375 update_stats(&runtime_stalled_cycles_back_stats[cpu], count[0]);
376 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
377 update_stats(&runtime_branches_stats[cpu], count[0]);
378 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
379 update_stats(&runtime_cacherefs_stats[cpu], count[0]);
380 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
381 update_stats(&runtime_l1_dcache_stats[cpu], count[0]);
382 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
383 update_stats(&runtime_l1_icache_stats[cpu], count[0]);
384 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
385 update_stats(&runtime_ll_cache_stats[cpu], count[0]);
386 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
387 update_stats(&runtime_dtlb_cache_stats[cpu], count[0]);
388 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
389 update_stats(&runtime_itlb_cache_stats[cpu], count[0]);
392 static void zero_per_pkg(struct perf_evsel *counter)
394 if (counter->per_pkg_mask)
395 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
398 static int check_per_pkg(struct perf_evsel *counter, int cpu, bool *skip)
400 unsigned long *mask = counter->per_pkg_mask;
401 struct cpu_map *cpus = perf_evsel__cpus(counter);
406 if (!counter->per_pkg)
409 if (cpu_map__empty(cpus))
413 mask = zalloc(MAX_NR_CPUS);
417 counter->per_pkg_mask = mask;
420 s = cpu_map__get_socket(cpus, cpu);
424 *skip = test_and_set_bit(s, mask) == 1;
428 static int read_cb(struct perf_evsel *evsel, int cpu, int thread __maybe_unused,
429 struct perf_counts_values *count)
431 struct perf_counts_values *aggr = &evsel->counts->aggr;
432 static struct perf_counts_values zero;
435 if (check_per_pkg(evsel, cpu, &skip)) {
436 pr_err("failed to read per-pkg counter\n");
447 if (!evsel->snapshot)
448 perf_evsel__compute_deltas(evsel, cpu, count);
449 perf_counts_values__scale(count, scale, NULL);
450 evsel->counts->cpu[cpu] = *count;
451 if (aggr_mode == AGGR_NONE)
452 update_shadow_stats(evsel, count->values, cpu);
455 aggr->val += count->val;
457 aggr->ena += count->ena;
458 aggr->run += count->run;
467 static int read_counter(struct perf_evsel *counter);
470 * Read out the results of a single counter:
471 * aggregate counts across CPUs in system-wide mode
473 static int read_counter_aggr(struct perf_evsel *counter)
475 struct perf_counts_values *aggr = &counter->counts->aggr;
476 struct perf_stat *ps = counter->priv;
477 u64 *count = counter->counts->aggr.values;
480 aggr->val = aggr->ena = aggr->run = 0;
482 if (read_counter(counter))
485 if (!counter->snapshot)
486 perf_evsel__compute_deltas(counter, -1, aggr);
487 perf_counts_values__scale(aggr, scale, &counter->counts->scaled);
489 for (i = 0; i < 3; i++)
490 update_stats(&ps->res_stats[i], count[i]);
493 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
494 perf_evsel__name(counter), count[0], count[1], count[2]);
498 * Save the full runtime - to allow normalization during printout:
500 update_shadow_stats(counter, count, 0);
506 * Read out the results of a single counter:
507 * do not aggregate counts across CPUs in system-wide mode
509 static int read_counter(struct perf_evsel *counter)
511 int nthreads = thread_map__nr(evsel_list->threads);
512 int ncpus = perf_evsel__nr_cpus(counter);
515 if (!counter->supported)
518 if (counter->system_wide)
521 if (counter->per_pkg)
522 zero_per_pkg(counter);
524 for (thread = 0; thread < nthreads; thread++) {
525 for (cpu = 0; cpu < ncpus; cpu++) {
526 if (perf_evsel__read_cb(counter, cpu, thread, read_cb))
534 static void print_interval(void)
536 static int num_print_interval;
537 struct perf_evsel *counter;
538 struct perf_stat *ps;
539 struct timespec ts, rs;
542 if (aggr_mode == AGGR_GLOBAL) {
543 evlist__for_each(evsel_list, counter) {
545 memset(ps->res_stats, 0, sizeof(ps->res_stats));
546 read_counter_aggr(counter);
549 evlist__for_each(evsel_list, counter) {
551 memset(ps->res_stats, 0, sizeof(ps->res_stats));
552 read_counter(counter);
556 clock_gettime(CLOCK_MONOTONIC, &ts);
557 diff_timespec(&rs, &ts, &ref_time);
558 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
560 if (num_print_interval == 0 && !csv_output) {
563 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
566 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
569 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
573 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
577 if (++num_print_interval == 25)
578 num_print_interval = 0;
586 evlist__for_each(evsel_list, counter)
587 print_counter(counter, prefix);
591 evlist__for_each(evsel_list, counter)
592 print_counter_aggr(counter, prefix);
598 static void handle_initial_delay(void)
600 struct perf_evsel *counter;
603 const int ncpus = cpu_map__nr(evsel_list->cpus),
604 nthreads = thread_map__nr(evsel_list->threads);
606 usleep(initial_delay * 1000);
607 evlist__for_each(evsel_list, counter)
608 perf_evsel__enable(counter, ncpus, nthreads);
612 static volatile int workload_exec_errno;
615 * perf_evlist__prepare_workload will send a SIGUSR1
616 * if the fork fails, since we asked by setting its
617 * want_signal to true.
619 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
620 void *ucontext __maybe_unused)
622 workload_exec_errno = info->si_value.sival_int;
625 static int __run_perf_stat(int argc, const char **argv)
628 unsigned long long t0, t1;
629 struct perf_evsel *counter;
633 const bool forks = (argc > 0);
636 ts.tv_sec = interval / 1000;
637 ts.tv_nsec = (interval % 1000) * 1000000;
644 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
645 workload_exec_failed_signal) < 0) {
646 perror("failed to prepare workload");
649 child_pid = evsel_list->workload.pid;
653 perf_evlist__set_leader(evsel_list);
655 evlist__for_each(evsel_list, counter) {
656 if (create_perf_stat_counter(counter) < 0) {
658 * PPC returns ENXIO for HW counters until 2.6.37
659 * (behavior changed with commit b0a873e).
661 if (errno == EINVAL || errno == ENOSYS ||
662 errno == ENOENT || errno == EOPNOTSUPP ||
665 ui__warning("%s event is not supported by the kernel.\n",
666 perf_evsel__name(counter));
667 counter->supported = false;
671 perf_evsel__open_strerror(counter, &target,
672 errno, msg, sizeof(msg));
673 ui__error("%s\n", msg);
676 kill(child_pid, SIGTERM);
680 counter->supported = true;
682 l = strlen(counter->unit);
687 if (perf_evlist__apply_filters(evsel_list)) {
688 error("failed to set filter with %d (%s)\n", errno,
689 strerror_r(errno, msg, sizeof(msg)));
694 * Enable counters and exec the command:
697 clock_gettime(CLOCK_MONOTONIC, &ref_time);
700 perf_evlist__start_workload(evsel_list);
701 handle_initial_delay();
704 while (!waitpid(child_pid, &status, WNOHANG)) {
705 nanosleep(&ts, NULL);
711 if (workload_exec_errno) {
712 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
713 pr_err("Workload failed: %s\n", emsg);
717 if (WIFSIGNALED(status))
718 psignal(WTERMSIG(status), argv[0]);
720 handle_initial_delay();
722 nanosleep(&ts, NULL);
730 update_stats(&walltime_nsecs_stats, t1 - t0);
732 if (aggr_mode == AGGR_GLOBAL) {
733 evlist__for_each(evsel_list, counter) {
734 read_counter_aggr(counter);
735 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
736 thread_map__nr(evsel_list->threads));
739 evlist__for_each(evsel_list, counter) {
740 read_counter(counter);
741 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
745 return WEXITSTATUS(status);
748 static int run_perf_stat(int argc, const char **argv)
753 ret = system(pre_cmd);
761 ret = __run_perf_stat(argc, argv);
766 ret = system(post_cmd);
774 static void print_running(u64 run, u64 ena)
777 fprintf(output, "%s%" PRIu64 "%s%.2f",
781 ena ? 100.0 * run / ena : 100.0);
782 } else if (run != ena) {
783 fprintf(output, " (%.2f%%)", 100.0 * run / ena);
787 static void print_noise_pct(double total, double avg)
789 double pct = rel_stddev_stats(total, avg);
792 fprintf(output, "%s%.2f%%", csv_sep, pct);
794 fprintf(output, " ( +-%6.2f%% )", pct);
797 static void print_noise(struct perf_evsel *evsel, double avg)
799 struct perf_stat *ps;
805 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
808 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
812 fprintf(output, "S%d-C%*d%s%*d%s",
813 cpu_map__id_to_socket(id),
815 cpu_map__id_to_cpu(id),
822 fprintf(output, "S%*d%s%*d%s",
831 fprintf(output, "CPU%*d%s",
833 perf_evsel__cpus(evsel)->map[id], csv_sep);
841 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
843 double msecs = avg / 1e6;
844 const char *fmt_v, *fmt_n;
847 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
848 fmt_n = csv_output ? "%s" : "%-25s";
850 aggr_printout(evsel, id, nr);
852 scnprintf(name, sizeof(name), "%s%s",
853 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
855 fprintf(output, fmt_v, msecs, csv_sep);
858 fprintf(output, "%s%s", evsel->unit, csv_sep);
860 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
862 fprintf(output, fmt_n, name);
865 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
867 if (csv_output || interval)
870 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
871 fprintf(output, " # %8.3f CPUs utilized ",
872 avg / avg_stats(&walltime_nsecs_stats));
874 fprintf(output, " ");
877 /* used for get_ratio_color() */
879 GRC_STALLED_CYCLES_FE,
880 GRC_STALLED_CYCLES_BE,
885 static const char *get_ratio_color(enum grc_type type, double ratio)
887 static const double grc_table[GRC_MAX_NR][3] = {
888 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
889 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
890 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
892 const char *color = PERF_COLOR_NORMAL;
894 if (ratio > grc_table[type][0])
895 color = PERF_COLOR_RED;
896 else if (ratio > grc_table[type][1])
897 color = PERF_COLOR_MAGENTA;
898 else if (ratio > grc_table[type][2])
899 color = PERF_COLOR_YELLOW;
904 static void print_stalled_cycles_frontend(int cpu,
905 struct perf_evsel *evsel
906 __maybe_unused, double avg)
908 double total, ratio = 0.0;
911 total = avg_stats(&runtime_cycles_stats[cpu]);
914 ratio = avg / total * 100.0;
916 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
918 fprintf(output, " # ");
919 color_fprintf(output, color, "%6.2f%%", ratio);
920 fprintf(output, " frontend cycles idle ");
923 static void print_stalled_cycles_backend(int cpu,
924 struct perf_evsel *evsel
925 __maybe_unused, double avg)
927 double total, ratio = 0.0;
930 total = avg_stats(&runtime_cycles_stats[cpu]);
933 ratio = avg / total * 100.0;
935 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
937 fprintf(output, " # ");
938 color_fprintf(output, color, "%6.2f%%", ratio);
939 fprintf(output, " backend cycles idle ");
942 static void print_branch_misses(int cpu,
943 struct perf_evsel *evsel __maybe_unused,
946 double total, ratio = 0.0;
949 total = avg_stats(&runtime_branches_stats[cpu]);
952 ratio = avg / total * 100.0;
954 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
956 fprintf(output, " # ");
957 color_fprintf(output, color, "%6.2f%%", ratio);
958 fprintf(output, " of all branches ");
961 static void print_l1_dcache_misses(int cpu,
962 struct perf_evsel *evsel __maybe_unused,
965 double total, ratio = 0.0;
968 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
971 ratio = avg / total * 100.0;
973 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
975 fprintf(output, " # ");
976 color_fprintf(output, color, "%6.2f%%", ratio);
977 fprintf(output, " of all L1-dcache hits ");
980 static void print_l1_icache_misses(int cpu,
981 struct perf_evsel *evsel __maybe_unused,
984 double total, ratio = 0.0;
987 total = avg_stats(&runtime_l1_icache_stats[cpu]);
990 ratio = avg / total * 100.0;
992 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
994 fprintf(output, " # ");
995 color_fprintf(output, color, "%6.2f%%", ratio);
996 fprintf(output, " of all L1-icache hits ");
999 static void print_dtlb_cache_misses(int cpu,
1000 struct perf_evsel *evsel __maybe_unused,
1003 double total, ratio = 0.0;
1006 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
1009 ratio = avg / total * 100.0;
1011 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
1013 fprintf(output, " # ");
1014 color_fprintf(output, color, "%6.2f%%", ratio);
1015 fprintf(output, " of all dTLB cache hits ");
1018 static void print_itlb_cache_misses(int cpu,
1019 struct perf_evsel *evsel __maybe_unused,
1022 double total, ratio = 0.0;
1025 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
1028 ratio = avg / total * 100.0;
1030 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
1032 fprintf(output, " # ");
1033 color_fprintf(output, color, "%6.2f%%", ratio);
1034 fprintf(output, " of all iTLB cache hits ");
1037 static void print_ll_cache_misses(int cpu,
1038 struct perf_evsel *evsel __maybe_unused,
1041 double total, ratio = 0.0;
1044 total = avg_stats(&runtime_ll_cache_stats[cpu]);
1047 ratio = avg / total * 100.0;
1049 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
1051 fprintf(output, " # ");
1052 color_fprintf(output, color, "%6.2f%%", ratio);
1053 fprintf(output, " of all LL-cache hits ");
1056 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1058 double total, ratio = 0.0, total2;
1059 double sc = evsel->scale;
1061 int cpu = cpu_map__id_to_cpu(id);
1064 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
1067 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
1069 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
1072 aggr_printout(evsel, id, nr);
1074 if (aggr_mode == AGGR_GLOBAL)
1077 fprintf(output, fmt, avg, csv_sep);
1080 fprintf(output, "%-*s%s",
1081 csv_output ? 0 : unit_width,
1082 evsel->unit, csv_sep);
1084 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1087 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1089 if (csv_output || interval)
1092 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
1093 total = avg_stats(&runtime_cycles_stats[cpu]);
1095 ratio = avg / total;
1096 fprintf(output, " # %5.2f insns per cycle ", ratio);
1098 fprintf(output, " ");
1100 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
1101 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
1104 ratio = total / avg;
1105 fprintf(output, "\n");
1106 if (aggr_mode == AGGR_NONE)
1107 fprintf(output, " ");
1108 fprintf(output, " # %5.2f stalled cycles per insn", ratio);
1111 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
1112 runtime_branches_stats[cpu].n != 0) {
1113 print_branch_misses(cpu, evsel, avg);
1115 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1116 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
1117 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1118 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1119 runtime_l1_dcache_stats[cpu].n != 0) {
1120 print_l1_dcache_misses(cpu, evsel, avg);
1122 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1123 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
1124 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1125 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1126 runtime_l1_icache_stats[cpu].n != 0) {
1127 print_l1_icache_misses(cpu, evsel, avg);
1129 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1130 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
1131 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1132 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1133 runtime_dtlb_cache_stats[cpu].n != 0) {
1134 print_dtlb_cache_misses(cpu, evsel, avg);
1136 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1137 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
1138 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1139 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1140 runtime_itlb_cache_stats[cpu].n != 0) {
1141 print_itlb_cache_misses(cpu, evsel, avg);
1143 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1144 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
1145 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1146 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1147 runtime_ll_cache_stats[cpu].n != 0) {
1148 print_ll_cache_misses(cpu, evsel, avg);
1149 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
1150 runtime_cacherefs_stats[cpu].n != 0) {
1151 total = avg_stats(&runtime_cacherefs_stats[cpu]);
1154 ratio = avg * 100 / total;
1156 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
1158 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1159 print_stalled_cycles_frontend(cpu, evsel, avg);
1160 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1161 print_stalled_cycles_backend(cpu, evsel, avg);
1162 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1163 total = avg_stats(&runtime_nsecs_stats[cpu]);
1166 ratio = avg / total;
1167 fprintf(output, " # %8.3f GHz ", ratio);
1169 fprintf(output, " ");
1171 } else if (transaction_run &&
1172 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1173 total = avg_stats(&runtime_cycles_stats[cpu]);
1176 " # %5.2f%% transactional cycles ",
1177 100.0 * (avg / total));
1178 } else if (transaction_run &&
1179 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1180 total = avg_stats(&runtime_cycles_stats[cpu]);
1181 total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1186 " # %5.2f%% aborted cycles ",
1187 100.0 * ((total2-avg) / total));
1188 } else if (transaction_run &&
1189 perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1191 runtime_cycles_in_tx_stats[cpu].n != 0) {
1192 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1195 ratio = total / avg;
1197 fprintf(output, " # %8.0f cycles / transaction ", ratio);
1198 } else if (transaction_run &&
1199 perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1201 runtime_cycles_in_tx_stats[cpu].n != 0) {
1202 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1205 ratio = total / avg;
1207 fprintf(output, " # %8.0f cycles / elision ", ratio);
1208 } else if (runtime_nsecs_stats[cpu].n != 0) {
1211 total = avg_stats(&runtime_nsecs_stats[cpu]);
1214 ratio = 1000.0 * avg / total;
1215 if (ratio < 0.001) {
1220 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
1222 fprintf(output, " ");
1226 static void print_aggr(char *prefix)
1228 struct perf_evsel *counter;
1229 int cpu, cpu2, s, s2, id, nr;
1233 if (!(aggr_map || aggr_get_id))
1236 for (s = 0; s < aggr_map->nr; s++) {
1237 id = aggr_map->map[s];
1238 evlist__for_each(evsel_list, counter) {
1239 val = ena = run = 0;
1241 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1242 cpu2 = perf_evsel__cpus(counter)->map[cpu];
1243 s2 = aggr_get_id(evsel_list->cpus, cpu2);
1246 val += counter->counts->cpu[cpu].val;
1247 ena += counter->counts->cpu[cpu].ena;
1248 run += counter->counts->cpu[cpu].run;
1252 fprintf(output, "%s", prefix);
1254 if (run == 0 || ena == 0) {
1255 aggr_printout(counter, id, nr);
1257 fprintf(output, "%*s%s",
1258 csv_output ? 0 : 18,
1259 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1262 fprintf(output, "%-*s%s",
1263 csv_output ? 0 : unit_width,
1264 counter->unit, csv_sep);
1266 fprintf(output, "%*s",
1267 csv_output ? 0 : -25,
1268 perf_evsel__name(counter));
1271 fprintf(output, "%s%s",
1272 csv_sep, counter->cgrp->name);
1274 print_running(run, ena);
1275 fputc('\n', output);
1278 uval = val * counter->scale;
1280 if (nsec_counter(counter))
1281 nsec_printout(id, nr, counter, uval);
1283 abs_printout(id, nr, counter, uval);
1286 print_noise(counter, 1.0);
1288 print_running(run, ena);
1289 fputc('\n', output);
1295 * Print out the results of a single counter:
1296 * aggregated counts in system-wide mode
1298 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1300 struct perf_stat *ps = counter->priv;
1301 double avg = avg_stats(&ps->res_stats[0]);
1302 int scaled = counter->counts->scaled;
1304 double avg_enabled, avg_running;
1306 avg_enabled = avg_stats(&ps->res_stats[1]);
1307 avg_running = avg_stats(&ps->res_stats[2]);
1310 fprintf(output, "%s", prefix);
1312 if (scaled == -1 || !counter->supported) {
1313 fprintf(output, "%*s%s",
1314 csv_output ? 0 : 18,
1315 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1317 fprintf(output, "%-*s%s",
1318 csv_output ? 0 : unit_width,
1319 counter->unit, csv_sep);
1320 fprintf(output, "%*s",
1321 csv_output ? 0 : -25,
1322 perf_evsel__name(counter));
1325 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1327 print_running(avg_running, avg_enabled);
1328 fputc('\n', output);
1332 uval = avg * counter->scale;
1334 if (nsec_counter(counter))
1335 nsec_printout(-1, 0, counter, uval);
1337 abs_printout(-1, 0, counter, uval);
1339 print_noise(counter, avg);
1341 print_running(avg_running, avg_enabled);
1342 fprintf(output, "\n");
1346 * Print out the results of a single counter:
1347 * does not use aggregated count in system-wide
1349 static void print_counter(struct perf_evsel *counter, char *prefix)
1355 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1356 val = counter->counts->cpu[cpu].val;
1357 ena = counter->counts->cpu[cpu].ena;
1358 run = counter->counts->cpu[cpu].run;
1361 fprintf(output, "%s", prefix);
1363 if (run == 0 || ena == 0) {
1364 fprintf(output, "CPU%*d%s%*s%s",
1365 csv_output ? 0 : -4,
1366 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1367 csv_output ? 0 : 18,
1368 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1371 fprintf(output, "%-*s%s",
1372 csv_output ? 0 : unit_width,
1373 counter->unit, csv_sep);
1375 fprintf(output, "%*s",
1376 csv_output ? 0 : -25,
1377 perf_evsel__name(counter));
1380 fprintf(output, "%s%s",
1381 csv_sep, counter->cgrp->name);
1383 print_running(run, ena);
1384 fputc('\n', output);
1388 uval = val * counter->scale;
1390 if (nsec_counter(counter))
1391 nsec_printout(cpu, 0, counter, uval);
1393 abs_printout(cpu, 0, counter, uval);
1396 print_noise(counter, 1.0);
1397 print_running(run, ena);
1399 fputc('\n', output);
1403 static void print_stat(int argc, const char **argv)
1405 struct perf_evsel *counter;
1411 fprintf(output, "\n");
1412 fprintf(output, " Performance counter stats for ");
1413 if (target.system_wide)
1414 fprintf(output, "\'system wide");
1415 else if (target.cpu_list)
1416 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1417 else if (!target__has_task(&target)) {
1418 fprintf(output, "\'%s", argv[0]);
1419 for (i = 1; i < argc; i++)
1420 fprintf(output, " %s", argv[i]);
1421 } else if (target.pid)
1422 fprintf(output, "process id \'%s", target.pid);
1424 fprintf(output, "thread id \'%s", target.tid);
1426 fprintf(output, "\'");
1428 fprintf(output, " (%d runs)", run_count);
1429 fprintf(output, ":\n\n");
1432 switch (aggr_mode) {
1438 evlist__for_each(evsel_list, counter)
1439 print_counter_aggr(counter, NULL);
1442 evlist__for_each(evsel_list, counter)
1443 print_counter(counter, NULL);
1451 fprintf(output, "\n");
1452 fprintf(output, " %17.9f seconds time elapsed",
1453 avg_stats(&walltime_nsecs_stats)/1e9);
1454 if (run_count > 1) {
1455 fprintf(output, " ");
1456 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1457 avg_stats(&walltime_nsecs_stats));
1459 fprintf(output, "\n\n");
1463 static volatile int signr = -1;
1465 static void skip_signal(int signo)
1467 if ((child_pid == -1) || interval)
1472 * render child_pid harmless
1473 * won't send SIGTERM to a random
1474 * process in case of race condition
1475 * and fast PID recycling
1480 static void sig_atexit(void)
1485 * avoid race condition with SIGCHLD handler
1486 * in skip_signal() which is modifying child_pid
1487 * goal is to avoid send SIGTERM to a random
1491 sigaddset(&set, SIGCHLD);
1492 sigprocmask(SIG_BLOCK, &set, &oset);
1494 if (child_pid != -1)
1495 kill(child_pid, SIGTERM);
1497 sigprocmask(SIG_SETMASK, &oset, NULL);
1502 signal(signr, SIG_DFL);
1503 kill(getpid(), signr);
1506 static int stat__set_big_num(const struct option *opt __maybe_unused,
1507 const char *s __maybe_unused, int unset)
1509 big_num_opt = unset ? 0 : 1;
1513 static int perf_stat_init_aggr_mode(void)
1515 switch (aggr_mode) {
1517 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1518 perror("cannot build socket map");
1521 aggr_get_id = cpu_map__get_socket;
1524 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1525 perror("cannot build core map");
1528 aggr_get_id = cpu_map__get_core;
1538 static int setup_events(const char * const *attrs, unsigned len)
1542 for (i = 0; i < len; i++) {
1543 if (parse_events(evsel_list, attrs[i]))
1550 * Add default attributes, if there were no attributes specified or
1551 * if -d/--detailed, -d -d or -d -d -d is used:
1553 static int add_default_attributes(void)
1555 struct perf_event_attr default_attrs[] = {
1557 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1558 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1559 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1560 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1562 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1563 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1564 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1565 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1566 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1567 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1572 * Detailed stats (-d), covering the L1 and last level data caches:
1574 struct perf_event_attr detailed_attrs[] = {
1576 { .type = PERF_TYPE_HW_CACHE,
1578 PERF_COUNT_HW_CACHE_L1D << 0 |
1579 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1580 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1582 { .type = PERF_TYPE_HW_CACHE,
1584 PERF_COUNT_HW_CACHE_L1D << 0 |
1585 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1586 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1588 { .type = PERF_TYPE_HW_CACHE,
1590 PERF_COUNT_HW_CACHE_LL << 0 |
1591 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1592 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1594 { .type = PERF_TYPE_HW_CACHE,
1596 PERF_COUNT_HW_CACHE_LL << 0 |
1597 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1598 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1602 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1604 struct perf_event_attr very_detailed_attrs[] = {
1606 { .type = PERF_TYPE_HW_CACHE,
1608 PERF_COUNT_HW_CACHE_L1I << 0 |
1609 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1610 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1612 { .type = PERF_TYPE_HW_CACHE,
1614 PERF_COUNT_HW_CACHE_L1I << 0 |
1615 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1616 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1618 { .type = PERF_TYPE_HW_CACHE,
1620 PERF_COUNT_HW_CACHE_DTLB << 0 |
1621 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1622 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1624 { .type = PERF_TYPE_HW_CACHE,
1626 PERF_COUNT_HW_CACHE_DTLB << 0 |
1627 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1628 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1630 { .type = PERF_TYPE_HW_CACHE,
1632 PERF_COUNT_HW_CACHE_ITLB << 0 |
1633 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1634 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1636 { .type = PERF_TYPE_HW_CACHE,
1638 PERF_COUNT_HW_CACHE_ITLB << 0 |
1639 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1640 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1645 * Very, very detailed stats (-d -d -d), adding prefetch events:
1647 struct perf_event_attr very_very_detailed_attrs[] = {
1649 { .type = PERF_TYPE_HW_CACHE,
1651 PERF_COUNT_HW_CACHE_L1D << 0 |
1652 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1653 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1655 { .type = PERF_TYPE_HW_CACHE,
1657 PERF_COUNT_HW_CACHE_L1D << 0 |
1658 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1659 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1662 /* Set attrs if no event is selected and !null_run: */
1666 if (transaction_run) {
1668 if (pmu_have_event("cpu", "cycles-ct") &&
1669 pmu_have_event("cpu", "el-start"))
1670 err = setup_events(transaction_attrs,
1671 ARRAY_SIZE(transaction_attrs));
1673 err = setup_events(transaction_limited_attrs,
1674 ARRAY_SIZE(transaction_limited_attrs));
1676 fprintf(stderr, "Cannot set up transaction events\n");
1682 if (!evsel_list->nr_entries) {
1683 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1687 /* Detailed events get appended to the event list: */
1689 if (detailed_run < 1)
1692 /* Append detailed run extra attributes: */
1693 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1696 if (detailed_run < 2)
1699 /* Append very detailed run extra attributes: */
1700 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1703 if (detailed_run < 3)
1706 /* Append very, very detailed run extra attributes: */
1707 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1710 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1712 bool append_file = false;
1714 const char *output_name = NULL;
1715 const struct option options[] = {
1716 OPT_BOOLEAN('T', "transaction", &transaction_run,
1717 "hardware transaction statistics"),
1718 OPT_CALLBACK('e', "event", &evsel_list, "event",
1719 "event selector. use 'perf list' to list available events",
1720 parse_events_option),
1721 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1722 "event filter", parse_filter),
1723 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1724 "child tasks do not inherit counters"),
1725 OPT_STRING('p', "pid", &target.pid, "pid",
1726 "stat events on existing process id"),
1727 OPT_STRING('t', "tid", &target.tid, "tid",
1728 "stat events on existing thread id"),
1729 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1730 "system-wide collection from all CPUs"),
1731 OPT_BOOLEAN('g', "group", &group,
1732 "put the counters into a counter group"),
1733 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1734 OPT_INCR('v', "verbose", &verbose,
1735 "be more verbose (show counter open errors, etc)"),
1736 OPT_INTEGER('r', "repeat", &run_count,
1737 "repeat command and print average + stddev (max: 100, forever: 0)"),
1738 OPT_BOOLEAN('n', "null", &null_run,
1739 "null run - dont start any counters"),
1740 OPT_INCR('d', "detailed", &detailed_run,
1741 "detailed run - start a lot of events"),
1742 OPT_BOOLEAN('S', "sync", &sync_run,
1743 "call sync() before starting a run"),
1744 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1745 "print large numbers with thousands\' separators",
1747 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1748 "list of cpus to monitor in system-wide"),
1749 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1750 "disable CPU count aggregation", AGGR_NONE),
1751 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1752 "print counts with custom separator"),
1753 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1754 "monitor event in cgroup name only", parse_cgroups),
1755 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1756 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1757 OPT_INTEGER(0, "log-fd", &output_fd,
1758 "log output to fd, instead of stderr"),
1759 OPT_STRING(0, "pre", &pre_cmd, "command",
1760 "command to run prior to the measured command"),
1761 OPT_STRING(0, "post", &post_cmd, "command",
1762 "command to run after to the measured command"),
1763 OPT_UINTEGER('I', "interval-print", &interval,
1764 "print counts at regular interval in ms (>= 100)"),
1765 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1766 "aggregate counts per processor socket", AGGR_SOCKET),
1767 OPT_SET_UINT(0, "per-core", &aggr_mode,
1768 "aggregate counts per physical processor core", AGGR_CORE),
1769 OPT_UINTEGER('D', "delay", &initial_delay,
1770 "ms to wait before starting measurement after program start"),
1773 const char * const stat_usage[] = {
1774 "perf stat [<options>] [<command>]",
1777 int status = -EINVAL, run_idx;
1780 setlocale(LC_ALL, "");
1782 evsel_list = perf_evlist__new();
1783 if (evsel_list == NULL)
1786 argc = parse_options(argc, argv, options, stat_usage,
1787 PARSE_OPT_STOP_AT_NON_OPTION);
1790 if (output_name && strcmp(output_name, "-"))
1793 if (output_name && output_fd) {
1794 fprintf(stderr, "cannot use both --output and --log-fd\n");
1795 parse_options_usage(stat_usage, options, "o", 1);
1796 parse_options_usage(NULL, options, "log-fd", 0);
1800 if (output_fd < 0) {
1801 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1802 parse_options_usage(stat_usage, options, "log-fd", 0);
1808 mode = append_file ? "a" : "w";
1810 output = fopen(output_name, mode);
1812 perror("failed to create output file");
1815 clock_gettime(CLOCK_REALTIME, &tm);
1816 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1817 } else if (output_fd > 0) {
1818 mode = append_file ? "a" : "w";
1819 output = fdopen(output_fd, mode);
1821 perror("Failed opening logfd");
1828 if (!strcmp(csv_sep, "\\t"))
1831 csv_sep = DEFAULT_SEPARATOR;
1834 * let the spreadsheet do the pretty-printing
1837 /* User explicitly passed -B? */
1838 if (big_num_opt == 1) {
1839 fprintf(stderr, "-B option not supported with -x\n");
1840 parse_options_usage(stat_usage, options, "B", 1);
1841 parse_options_usage(NULL, options, "x", 1);
1843 } else /* Nope, so disable big number formatting */
1845 } else if (big_num_opt == 0) /* User passed --no-big-num */
1848 if (!argc && target__none(&target))
1849 usage_with_options(stat_usage, options);
1851 if (run_count < 0) {
1852 pr_err("Run count must be a positive number\n");
1853 parse_options_usage(stat_usage, options, "r", 1);
1855 } else if (run_count == 0) {
1860 /* no_aggr, cgroup are for system-wide only */
1861 if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1862 !target__has_cpu(&target)) {
1863 fprintf(stderr, "both cgroup and no-aggregation "
1864 "modes only available in system-wide mode\n");
1866 parse_options_usage(stat_usage, options, "G", 1);
1867 parse_options_usage(NULL, options, "A", 1);
1868 parse_options_usage(NULL, options, "a", 1);
1872 if (add_default_attributes())
1875 target__validate(&target);
1877 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1878 if (target__has_task(&target)) {
1879 pr_err("Problems finding threads of monitor\n");
1880 parse_options_usage(stat_usage, options, "p", 1);
1881 parse_options_usage(NULL, options, "t", 1);
1882 } else if (target__has_cpu(&target)) {
1883 perror("failed to parse CPUs map");
1884 parse_options_usage(stat_usage, options, "C", 1);
1885 parse_options_usage(NULL, options, "a", 1);
1889 if (interval && interval < 100) {
1890 pr_err("print interval must be >= 100ms\n");
1891 parse_options_usage(stat_usage, options, "I", 1);
1895 if (perf_evlist__alloc_stats(evsel_list, interval))
1898 if (perf_stat_init_aggr_mode())
1902 * We dont want to block the signals - that would cause
1903 * child tasks to inherit that and Ctrl-C would not work.
1904 * What we want is for Ctrl-C to work in the exec()-ed
1905 * task, but being ignored by perf stat itself:
1909 signal(SIGINT, skip_signal);
1910 signal(SIGCHLD, skip_signal);
1911 signal(SIGALRM, skip_signal);
1912 signal(SIGABRT, skip_signal);
1915 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1916 if (run_count != 1 && verbose)
1917 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1920 status = run_perf_stat(argc, argv);
1921 if (forever && status != -1) {
1922 print_stat(argc, argv);
1923 perf_stat__reset_stats(evsel_list);
1927 if (!forever && status != -1 && !interval)
1928 print_stat(argc, argv);
1930 perf_evlist__free_stats(evsel_list);
1932 perf_evlist__delete(evsel_list);