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/util.h"
47 #include "util/parse-options.h"
48 #include "util/parse-events.h"
50 #include "util/event.h"
51 #include "util/evlist.h"
52 #include "util/evsel.h"
53 #include "util/debug.h"
54 #include "util/color.h"
55 #include "util/stat.h"
56 #include "util/header.h"
57 #include "util/cpumap.h"
58 #include "util/thread.h"
59 #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)
178 memset(evsel->priv, 0, sizeof(struct perf_stat));
181 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
183 evsel->priv = zalloc(sizeof(struct perf_stat));
184 return evsel->priv == NULL ? -ENOMEM : 0;
187 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
192 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
197 sz = sizeof(*evsel->counts) +
198 (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
204 evsel->prev_raw_counts = addr;
209 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
211 zfree(&evsel->prev_raw_counts);
214 static void perf_evlist__free_stats(struct perf_evlist *evlist)
216 struct perf_evsel *evsel;
218 list_for_each_entry(evsel, &evlist->entries, node) {
219 perf_evsel__free_stat_priv(evsel);
220 perf_evsel__free_counts(evsel);
221 perf_evsel__free_prev_raw_counts(evsel);
225 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
227 struct perf_evsel *evsel;
229 list_for_each_entry(evsel, &evlist->entries, node) {
230 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
231 perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
232 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
239 perf_evlist__free_stats(evlist);
243 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
244 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
245 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
246 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
247 static struct stats runtime_branches_stats[MAX_NR_CPUS];
248 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
249 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
250 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
251 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
252 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
253 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
254 static struct stats runtime_cycles_in_tx_stats[MAX_NR_CPUS];
255 static struct stats walltime_nsecs_stats;
256 static struct stats runtime_transaction_stats[MAX_NR_CPUS];
257 static struct stats runtime_elision_stats[MAX_NR_CPUS];
259 static void perf_stat__reset_stats(struct perf_evlist *evlist)
261 struct perf_evsel *evsel;
263 list_for_each_entry(evsel, &evlist->entries, node) {
264 perf_evsel__reset_stat_priv(evsel);
265 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
268 memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
269 memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
270 memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
271 memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
272 memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
273 memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
274 memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
275 memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
276 memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
277 memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
278 memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
279 memset(runtime_cycles_in_tx_stats, 0,
280 sizeof(runtime_cycles_in_tx_stats));
281 memset(runtime_transaction_stats, 0,
282 sizeof(runtime_transaction_stats));
283 memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
284 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
287 static int create_perf_stat_counter(struct perf_evsel *evsel)
289 struct perf_event_attr *attr = &evsel->attr;
292 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
293 PERF_FORMAT_TOTAL_TIME_RUNNING;
295 attr->inherit = !no_inherit;
297 if (target__has_cpu(&target))
298 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
300 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
303 attr->enable_on_exec = 1;
306 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
310 * Does the counter have nsecs as a unit?
312 static inline int nsec_counter(struct perf_evsel *evsel)
314 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
315 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
321 static struct perf_evsel *nth_evsel(int n)
323 static struct perf_evsel **array;
324 static int array_len;
325 struct perf_evsel *ev;
328 /* Assumes this only called when evsel_list does not change anymore. */
330 list_for_each_entry(ev, &evsel_list->entries, node)
332 array = malloc(array_len * sizeof(void *));
336 list_for_each_entry(ev, &evsel_list->entries, node)
345 * Update various tracking values we maintain to print
346 * more semantic information such as miss/hit ratios,
347 * instruction rates, etc:
349 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
351 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
352 update_stats(&runtime_nsecs_stats[0], count[0]);
353 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
354 update_stats(&runtime_cycles_stats[0], count[0]);
355 else if (transaction_run &&
356 perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX)))
357 update_stats(&runtime_cycles_in_tx_stats[0], count[0]);
358 else if (transaction_run &&
359 perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START)))
360 update_stats(&runtime_transaction_stats[0], count[0]);
361 else if (transaction_run &&
362 perf_evsel__cmp(counter, nth_evsel(T_ELISION_START)))
363 update_stats(&runtime_elision_stats[0], count[0]);
364 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
365 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
366 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
367 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
368 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
369 update_stats(&runtime_branches_stats[0], count[0]);
370 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
371 update_stats(&runtime_cacherefs_stats[0], count[0]);
372 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
373 update_stats(&runtime_l1_dcache_stats[0], count[0]);
374 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
375 update_stats(&runtime_l1_icache_stats[0], count[0]);
376 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
377 update_stats(&runtime_ll_cache_stats[0], count[0]);
378 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
379 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
380 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
381 update_stats(&runtime_itlb_cache_stats[0], count[0]);
385 * Read out the results of a single counter:
386 * aggregate counts across CPUs in system-wide mode
388 static int read_counter_aggr(struct perf_evsel *counter)
390 struct perf_stat *ps = counter->priv;
391 u64 *count = counter->counts->aggr.values;
394 if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
395 thread_map__nr(evsel_list->threads), scale) < 0)
398 for (i = 0; i < 3; i++)
399 update_stats(&ps->res_stats[i], count[i]);
402 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
403 perf_evsel__name(counter), count[0], count[1], count[2]);
407 * Save the full runtime - to allow normalization during printout:
409 update_shadow_stats(counter, count);
415 * Read out the results of a single counter:
416 * do not aggregate counts across CPUs in system-wide mode
418 static int read_counter(struct perf_evsel *counter)
423 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
424 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
427 count = counter->counts->cpu[cpu].values;
429 update_shadow_stats(counter, count);
435 static void print_interval(void)
437 static int num_print_interval;
438 struct perf_evsel *counter;
439 struct perf_stat *ps;
440 struct timespec ts, rs;
443 if (aggr_mode == AGGR_GLOBAL) {
444 list_for_each_entry(counter, &evsel_list->entries, node) {
446 memset(ps->res_stats, 0, sizeof(ps->res_stats));
447 read_counter_aggr(counter);
450 list_for_each_entry(counter, &evsel_list->entries, node) {
452 memset(ps->res_stats, 0, sizeof(ps->res_stats));
453 read_counter(counter);
457 clock_gettime(CLOCK_MONOTONIC, &ts);
458 diff_timespec(&rs, &ts, &ref_time);
459 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
461 if (num_print_interval == 0 && !csv_output) {
464 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
467 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
470 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
474 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
478 if (++num_print_interval == 25)
479 num_print_interval = 0;
487 list_for_each_entry(counter, &evsel_list->entries, node)
488 print_counter(counter, prefix);
492 list_for_each_entry(counter, &evsel_list->entries, node)
493 print_counter_aggr(counter, prefix);
499 static void handle_initial_delay(void)
501 struct perf_evsel *counter;
504 const int ncpus = cpu_map__nr(evsel_list->cpus),
505 nthreads = thread_map__nr(evsel_list->threads);
507 usleep(initial_delay * 1000);
508 list_for_each_entry(counter, &evsel_list->entries, node)
509 perf_evsel__enable(counter, ncpus, nthreads);
513 static volatile int workload_exec_errno;
516 * perf_evlist__prepare_workload will send a SIGUSR1
517 * if the fork fails, since we asked by setting its
518 * want_signal to true.
520 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
521 void *ucontext __maybe_unused)
523 workload_exec_errno = info->si_value.sival_int;
526 static int __run_perf_stat(int argc, const char **argv)
529 unsigned long long t0, t1;
530 struct perf_evsel *counter;
534 const bool forks = (argc > 0);
537 ts.tv_sec = interval / 1000;
538 ts.tv_nsec = (interval % 1000) * 1000000;
545 if (perf_evlist__prepare_workload(evsel_list, &target, argv,
547 perror("failed to prepare workload");
550 child_pid = evsel_list->workload.pid;
554 perf_evlist__set_leader(evsel_list);
556 list_for_each_entry(counter, &evsel_list->entries, node) {
557 if (create_perf_stat_counter(counter) < 0) {
559 * PPC returns ENXIO for HW counters until 2.6.37
560 * (behavior changed with commit b0a873e).
562 if (errno == EINVAL || errno == ENOSYS ||
563 errno == ENOENT || errno == EOPNOTSUPP ||
566 ui__warning("%s event is not supported by the kernel.\n",
567 perf_evsel__name(counter));
568 counter->supported = false;
572 perf_evsel__open_strerror(counter, &target,
573 errno, msg, sizeof(msg));
574 ui__error("%s\n", msg);
577 kill(child_pid, SIGTERM);
581 counter->supported = true;
583 l = strlen(counter->unit);
588 if (perf_evlist__apply_filters(evsel_list)) {
589 error("failed to set filter with %d (%s)\n", errno,
595 * Enable counters and exec the command:
598 clock_gettime(CLOCK_MONOTONIC, &ref_time);
601 struct sigaction act = {
602 .sa_flags = SA_SIGINFO,
603 .sa_sigaction = workload_exec_failed_signal,
606 * perf_evlist__prepare_workload will, after we call
607 * perf_evlist__start_Workload, send a SIGUSR1 if the exec call
608 * fails, that we will catch in workload_signal to flip
609 * workload_exec_errno.
611 sigaction(SIGUSR1, &act, NULL);
613 perf_evlist__start_workload(evsel_list);
614 handle_initial_delay();
617 while (!waitpid(child_pid, &status, WNOHANG)) {
618 nanosleep(&ts, NULL);
624 if (workload_exec_errno) {
625 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
626 pr_err("Workload failed: %s\n", emsg);
630 if (WIFSIGNALED(status))
631 psignal(WTERMSIG(status), argv[0]);
633 handle_initial_delay();
635 nanosleep(&ts, NULL);
643 update_stats(&walltime_nsecs_stats, t1 - t0);
645 if (aggr_mode == AGGR_GLOBAL) {
646 list_for_each_entry(counter, &evsel_list->entries, node) {
647 read_counter_aggr(counter);
648 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
649 thread_map__nr(evsel_list->threads));
652 list_for_each_entry(counter, &evsel_list->entries, node) {
653 read_counter(counter);
654 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
658 return WEXITSTATUS(status);
661 static int run_perf_stat(int argc __maybe_unused, const char **argv)
666 ret = system(pre_cmd);
674 ret = __run_perf_stat(argc, argv);
679 ret = system(post_cmd);
687 static void print_noise_pct(double total, double avg)
689 double pct = rel_stddev_stats(total, avg);
692 fprintf(output, "%s%.2f%%", csv_sep, pct);
694 fprintf(output, " ( +-%6.2f%% )", pct);
697 static void print_noise(struct perf_evsel *evsel, double avg)
699 struct perf_stat *ps;
705 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
708 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
712 fprintf(output, "S%d-C%*d%s%*d%s",
713 cpu_map__id_to_socket(id),
715 cpu_map__id_to_cpu(id),
722 fprintf(output, "S%*d%s%*d%s",
731 fprintf(output, "CPU%*d%s",
733 perf_evsel__cpus(evsel)->map[id], csv_sep);
741 static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
743 double msecs = avg / 1e6;
744 const char *fmt_v, *fmt_n;
747 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
748 fmt_n = csv_output ? "%s" : "%-25s";
750 aggr_printout(evsel, cpu, nr);
752 scnprintf(name, sizeof(name), "%s%s",
753 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
755 fprintf(output, fmt_v, msecs, csv_sep);
758 fprintf(output, "%s%s", evsel->unit, csv_sep);
760 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
762 fprintf(output, fmt_n, name);
765 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
767 if (csv_output || interval)
770 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
771 fprintf(output, " # %8.3f CPUs utilized ",
772 avg / avg_stats(&walltime_nsecs_stats));
774 fprintf(output, " ");
777 /* used for get_ratio_color() */
779 GRC_STALLED_CYCLES_FE,
780 GRC_STALLED_CYCLES_BE,
785 static const char *get_ratio_color(enum grc_type type, double ratio)
787 static const double grc_table[GRC_MAX_NR][3] = {
788 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
789 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
790 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
792 const char *color = PERF_COLOR_NORMAL;
794 if (ratio > grc_table[type][0])
795 color = PERF_COLOR_RED;
796 else if (ratio > grc_table[type][1])
797 color = PERF_COLOR_MAGENTA;
798 else if (ratio > grc_table[type][2])
799 color = PERF_COLOR_YELLOW;
804 static void print_stalled_cycles_frontend(int cpu,
805 struct perf_evsel *evsel
806 __maybe_unused, double avg)
808 double total, ratio = 0.0;
811 total = avg_stats(&runtime_cycles_stats[cpu]);
814 ratio = avg / total * 100.0;
816 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
818 fprintf(output, " # ");
819 color_fprintf(output, color, "%6.2f%%", ratio);
820 fprintf(output, " frontend cycles idle ");
823 static void print_stalled_cycles_backend(int cpu,
824 struct perf_evsel *evsel
825 __maybe_unused, double avg)
827 double total, ratio = 0.0;
830 total = avg_stats(&runtime_cycles_stats[cpu]);
833 ratio = avg / total * 100.0;
835 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
837 fprintf(output, " # ");
838 color_fprintf(output, color, "%6.2f%%", ratio);
839 fprintf(output, " backend cycles idle ");
842 static void print_branch_misses(int cpu,
843 struct perf_evsel *evsel __maybe_unused,
846 double total, ratio = 0.0;
849 total = avg_stats(&runtime_branches_stats[cpu]);
852 ratio = avg / total * 100.0;
854 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
856 fprintf(output, " # ");
857 color_fprintf(output, color, "%6.2f%%", ratio);
858 fprintf(output, " of all branches ");
861 static void print_l1_dcache_misses(int cpu,
862 struct perf_evsel *evsel __maybe_unused,
865 double total, ratio = 0.0;
868 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
871 ratio = avg / total * 100.0;
873 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
875 fprintf(output, " # ");
876 color_fprintf(output, color, "%6.2f%%", ratio);
877 fprintf(output, " of all L1-dcache hits ");
880 static void print_l1_icache_misses(int cpu,
881 struct perf_evsel *evsel __maybe_unused,
884 double total, ratio = 0.0;
887 total = avg_stats(&runtime_l1_icache_stats[cpu]);
890 ratio = avg / total * 100.0;
892 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
894 fprintf(output, " # ");
895 color_fprintf(output, color, "%6.2f%%", ratio);
896 fprintf(output, " of all L1-icache hits ");
899 static void print_dtlb_cache_misses(int cpu,
900 struct perf_evsel *evsel __maybe_unused,
903 double total, ratio = 0.0;
906 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
909 ratio = avg / total * 100.0;
911 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
913 fprintf(output, " # ");
914 color_fprintf(output, color, "%6.2f%%", ratio);
915 fprintf(output, " of all dTLB cache hits ");
918 static void print_itlb_cache_misses(int cpu,
919 struct perf_evsel *evsel __maybe_unused,
922 double total, ratio = 0.0;
925 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
928 ratio = avg / total * 100.0;
930 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
932 fprintf(output, " # ");
933 color_fprintf(output, color, "%6.2f%%", ratio);
934 fprintf(output, " of all iTLB cache hits ");
937 static void print_ll_cache_misses(int cpu,
938 struct perf_evsel *evsel __maybe_unused,
941 double total, ratio = 0.0;
944 total = avg_stats(&runtime_ll_cache_stats[cpu]);
947 ratio = avg / total * 100.0;
949 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
951 fprintf(output, " # ");
952 color_fprintf(output, color, "%6.2f%%", ratio);
953 fprintf(output, " of all LL-cache hits ");
956 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
958 double total, ratio = 0.0, total2;
959 double sc = evsel->scale;
963 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
966 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
968 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
971 aggr_printout(evsel, cpu, nr);
973 if (aggr_mode == AGGR_GLOBAL)
976 fprintf(output, fmt, avg, csv_sep);
979 fprintf(output, "%-*s%s",
980 csv_output ? 0 : unit_width,
981 evsel->unit, csv_sep);
983 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
986 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
988 if (csv_output || interval)
991 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
992 total = avg_stats(&runtime_cycles_stats[cpu]);
995 fprintf(output, " # %5.2f insns per cycle ", ratio);
997 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
998 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
1001 ratio = total / avg;
1002 fprintf(output, "\n");
1003 if (aggr_mode == AGGR_NONE)
1004 fprintf(output, " ");
1005 fprintf(output, " # %5.2f stalled cycles per insn", ratio);
1008 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
1009 runtime_branches_stats[cpu].n != 0) {
1010 print_branch_misses(cpu, evsel, avg);
1012 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1013 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
1014 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1015 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1016 runtime_l1_dcache_stats[cpu].n != 0) {
1017 print_l1_dcache_misses(cpu, evsel, avg);
1019 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1020 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
1021 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1022 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1023 runtime_l1_icache_stats[cpu].n != 0) {
1024 print_l1_icache_misses(cpu, evsel, avg);
1026 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1027 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
1028 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1029 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1030 runtime_dtlb_cache_stats[cpu].n != 0) {
1031 print_dtlb_cache_misses(cpu, evsel, avg);
1033 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1034 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
1035 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1036 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1037 runtime_itlb_cache_stats[cpu].n != 0) {
1038 print_itlb_cache_misses(cpu, evsel, avg);
1040 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1041 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
1042 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1043 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1044 runtime_ll_cache_stats[cpu].n != 0) {
1045 print_ll_cache_misses(cpu, evsel, avg);
1046 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
1047 runtime_cacherefs_stats[cpu].n != 0) {
1048 total = avg_stats(&runtime_cacherefs_stats[cpu]);
1051 ratio = avg * 100 / total;
1053 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
1055 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1056 print_stalled_cycles_frontend(cpu, evsel, avg);
1057 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1058 print_stalled_cycles_backend(cpu, evsel, avg);
1059 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1060 total = avg_stats(&runtime_nsecs_stats[cpu]);
1063 ratio = avg / total;
1064 fprintf(output, " # %8.3f GHz ", ratio);
1066 } else if (transaction_run &&
1067 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1068 total = avg_stats(&runtime_cycles_stats[cpu]);
1071 " # %5.2f%% transactional cycles ",
1072 100.0 * (avg / total));
1073 } else if (transaction_run &&
1074 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1075 total = avg_stats(&runtime_cycles_stats[cpu]);
1076 total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1081 " # %5.2f%% aborted cycles ",
1082 100.0 * ((total2-avg) / total));
1083 } else if (transaction_run &&
1084 perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1086 runtime_cycles_in_tx_stats[cpu].n != 0) {
1087 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1090 ratio = total / avg;
1092 fprintf(output, " # %8.0f cycles / transaction ", ratio);
1093 } else if (transaction_run &&
1094 perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1096 runtime_cycles_in_tx_stats[cpu].n != 0) {
1097 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1100 ratio = total / avg;
1102 fprintf(output, " # %8.0f cycles / elision ", ratio);
1103 } else if (runtime_nsecs_stats[cpu].n != 0) {
1106 total = avg_stats(&runtime_nsecs_stats[cpu]);
1109 ratio = 1000.0 * avg / total;
1110 if (ratio < 0.001) {
1115 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
1117 fprintf(output, " ");
1121 static void print_aggr(char *prefix)
1123 struct perf_evsel *counter;
1124 int cpu, cpu2, s, s2, id, nr;
1128 if (!(aggr_map || aggr_get_id))
1131 for (s = 0; s < aggr_map->nr; s++) {
1132 id = aggr_map->map[s];
1133 list_for_each_entry(counter, &evsel_list->entries, node) {
1134 val = ena = run = 0;
1136 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1137 cpu2 = perf_evsel__cpus(counter)->map[cpu];
1138 s2 = aggr_get_id(evsel_list->cpus, cpu2);
1141 val += counter->counts->cpu[cpu].val;
1142 ena += counter->counts->cpu[cpu].ena;
1143 run += counter->counts->cpu[cpu].run;
1147 fprintf(output, "%s", prefix);
1149 if (run == 0 || ena == 0) {
1150 aggr_printout(counter, id, nr);
1152 fprintf(output, "%*s%s",
1153 csv_output ? 0 : 18,
1154 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1157 fprintf(output, "%-*s%s",
1158 csv_output ? 0 : unit_width,
1159 counter->unit, csv_sep);
1161 fprintf(output, "%*s",
1162 csv_output ? 0 : -25,
1163 perf_evsel__name(counter));
1166 fprintf(output, "%s%s",
1167 csv_sep, counter->cgrp->name);
1169 fputc('\n', output);
1172 uval = val * counter->scale;
1174 if (nsec_counter(counter))
1175 nsec_printout(id, nr, counter, uval);
1177 abs_printout(id, nr, counter, uval);
1180 print_noise(counter, 1.0);
1183 fprintf(output, " (%.2f%%)",
1186 fputc('\n', output);
1192 * Print out the results of a single counter:
1193 * aggregated counts in system-wide mode
1195 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1197 struct perf_stat *ps = counter->priv;
1198 double avg = avg_stats(&ps->res_stats[0]);
1199 int scaled = counter->counts->scaled;
1203 fprintf(output, "%s", prefix);
1206 fprintf(output, "%*s%s",
1207 csv_output ? 0 : 18,
1208 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1210 fprintf(output, "%-*s%s",
1211 csv_output ? 0 : unit_width,
1212 counter->unit, csv_sep);
1213 fprintf(output, "%*s",
1214 csv_output ? 0 : -25,
1215 perf_evsel__name(counter));
1218 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1220 fputc('\n', output);
1224 uval = avg * counter->scale;
1226 if (nsec_counter(counter))
1227 nsec_printout(-1, 0, counter, uval);
1229 abs_printout(-1, 0, counter, uval);
1231 print_noise(counter, avg);
1234 fputc('\n', output);
1239 double avg_enabled, avg_running;
1241 avg_enabled = avg_stats(&ps->res_stats[1]);
1242 avg_running = avg_stats(&ps->res_stats[2]);
1244 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1246 fprintf(output, "\n");
1250 * Print out the results of a single counter:
1251 * does not use aggregated count in system-wide
1253 static void print_counter(struct perf_evsel *counter, char *prefix)
1259 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1260 val = counter->counts->cpu[cpu].val;
1261 ena = counter->counts->cpu[cpu].ena;
1262 run = counter->counts->cpu[cpu].run;
1265 fprintf(output, "%s", prefix);
1267 if (run == 0 || ena == 0) {
1268 fprintf(output, "CPU%*d%s%*s%s",
1269 csv_output ? 0 : -4,
1270 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1271 csv_output ? 0 : 18,
1272 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1275 fprintf(output, "%-*s%s",
1276 csv_output ? 0 : unit_width,
1277 counter->unit, csv_sep);
1279 fprintf(output, "%*s",
1280 csv_output ? 0 : -25,
1281 perf_evsel__name(counter));
1284 fprintf(output, "%s%s",
1285 csv_sep, counter->cgrp->name);
1287 fputc('\n', output);
1291 uval = val * counter->scale;
1293 if (nsec_counter(counter))
1294 nsec_printout(cpu, 0, counter, uval);
1296 abs_printout(cpu, 0, counter, uval);
1299 print_noise(counter, 1.0);
1302 fprintf(output, " (%.2f%%)",
1305 fputc('\n', output);
1309 static void print_stat(int argc, const char **argv)
1311 struct perf_evsel *counter;
1317 fprintf(output, "\n");
1318 fprintf(output, " Performance counter stats for ");
1319 if (target.system_wide)
1320 fprintf(output, "\'system wide");
1321 else if (target.cpu_list)
1322 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1323 else if (!target__has_task(&target)) {
1324 fprintf(output, "\'%s", argv[0]);
1325 for (i = 1; i < argc; i++)
1326 fprintf(output, " %s", argv[i]);
1327 } else if (target.pid)
1328 fprintf(output, "process id \'%s", target.pid);
1330 fprintf(output, "thread id \'%s", target.tid);
1332 fprintf(output, "\'");
1334 fprintf(output, " (%d runs)", run_count);
1335 fprintf(output, ":\n\n");
1338 switch (aggr_mode) {
1344 list_for_each_entry(counter, &evsel_list->entries, node)
1345 print_counter_aggr(counter, NULL);
1348 list_for_each_entry(counter, &evsel_list->entries, node)
1349 print_counter(counter, NULL);
1357 fprintf(output, "\n");
1358 fprintf(output, " %17.9f seconds time elapsed",
1359 avg_stats(&walltime_nsecs_stats)/1e9);
1360 if (run_count > 1) {
1361 fprintf(output, " ");
1362 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1363 avg_stats(&walltime_nsecs_stats));
1365 fprintf(output, "\n\n");
1369 static volatile int signr = -1;
1371 static void skip_signal(int signo)
1373 if ((child_pid == -1) || interval)
1378 * render child_pid harmless
1379 * won't send SIGTERM to a random
1380 * process in case of race condition
1381 * and fast PID recycling
1386 static void sig_atexit(void)
1391 * avoid race condition with SIGCHLD handler
1392 * in skip_signal() which is modifying child_pid
1393 * goal is to avoid send SIGTERM to a random
1397 sigaddset(&set, SIGCHLD);
1398 sigprocmask(SIG_BLOCK, &set, &oset);
1400 if (child_pid != -1)
1401 kill(child_pid, SIGTERM);
1403 sigprocmask(SIG_SETMASK, &oset, NULL);
1408 signal(signr, SIG_DFL);
1409 kill(getpid(), signr);
1412 static int stat__set_big_num(const struct option *opt __maybe_unused,
1413 const char *s __maybe_unused, int unset)
1415 big_num_opt = unset ? 0 : 1;
1419 static int perf_stat_init_aggr_mode(void)
1421 switch (aggr_mode) {
1423 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1424 perror("cannot build socket map");
1427 aggr_get_id = cpu_map__get_socket;
1430 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1431 perror("cannot build core map");
1434 aggr_get_id = cpu_map__get_core;
1444 static int setup_events(const char * const *attrs, unsigned len)
1448 for (i = 0; i < len; i++) {
1449 if (parse_events(evsel_list, attrs[i]))
1456 * Add default attributes, if there were no attributes specified or
1457 * if -d/--detailed, -d -d or -d -d -d is used:
1459 static int add_default_attributes(void)
1461 struct perf_event_attr default_attrs[] = {
1463 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1464 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1465 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1466 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1468 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1469 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1470 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1471 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1472 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1473 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1478 * Detailed stats (-d), covering the L1 and last level data caches:
1480 struct perf_event_attr detailed_attrs[] = {
1482 { .type = PERF_TYPE_HW_CACHE,
1484 PERF_COUNT_HW_CACHE_L1D << 0 |
1485 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1486 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1488 { .type = PERF_TYPE_HW_CACHE,
1490 PERF_COUNT_HW_CACHE_L1D << 0 |
1491 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1492 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1494 { .type = PERF_TYPE_HW_CACHE,
1496 PERF_COUNT_HW_CACHE_LL << 0 |
1497 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1498 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1500 { .type = PERF_TYPE_HW_CACHE,
1502 PERF_COUNT_HW_CACHE_LL << 0 |
1503 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1504 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1508 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1510 struct perf_event_attr very_detailed_attrs[] = {
1512 { .type = PERF_TYPE_HW_CACHE,
1514 PERF_COUNT_HW_CACHE_L1I << 0 |
1515 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1516 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1518 { .type = PERF_TYPE_HW_CACHE,
1520 PERF_COUNT_HW_CACHE_L1I << 0 |
1521 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1522 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1524 { .type = PERF_TYPE_HW_CACHE,
1526 PERF_COUNT_HW_CACHE_DTLB << 0 |
1527 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1528 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1530 { .type = PERF_TYPE_HW_CACHE,
1532 PERF_COUNT_HW_CACHE_DTLB << 0 |
1533 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1534 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1536 { .type = PERF_TYPE_HW_CACHE,
1538 PERF_COUNT_HW_CACHE_ITLB << 0 |
1539 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1540 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1542 { .type = PERF_TYPE_HW_CACHE,
1544 PERF_COUNT_HW_CACHE_ITLB << 0 |
1545 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1546 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1551 * Very, very detailed stats (-d -d -d), adding prefetch events:
1553 struct perf_event_attr very_very_detailed_attrs[] = {
1555 { .type = PERF_TYPE_HW_CACHE,
1557 PERF_COUNT_HW_CACHE_L1D << 0 |
1558 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1559 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1561 { .type = PERF_TYPE_HW_CACHE,
1563 PERF_COUNT_HW_CACHE_L1D << 0 |
1564 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1565 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1568 /* Set attrs if no event is selected and !null_run: */
1572 if (transaction_run) {
1574 if (pmu_have_event("cpu", "cycles-ct") &&
1575 pmu_have_event("cpu", "el-start"))
1576 err = setup_events(transaction_attrs,
1577 ARRAY_SIZE(transaction_attrs));
1579 err = setup_events(transaction_limited_attrs,
1580 ARRAY_SIZE(transaction_limited_attrs));
1582 fprintf(stderr, "Cannot set up transaction events\n");
1588 if (!evsel_list->nr_entries) {
1589 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1593 /* Detailed events get appended to the event list: */
1595 if (detailed_run < 1)
1598 /* Append detailed run extra attributes: */
1599 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1602 if (detailed_run < 2)
1605 /* Append very detailed run extra attributes: */
1606 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1609 if (detailed_run < 3)
1612 /* Append very, very detailed run extra attributes: */
1613 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1616 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1618 bool append_file = false;
1620 const char *output_name = NULL;
1621 const struct option options[] = {
1622 OPT_BOOLEAN('T', "transaction", &transaction_run,
1623 "hardware transaction statistics"),
1624 OPT_CALLBACK('e', "event", &evsel_list, "event",
1625 "event selector. use 'perf list' to list available events",
1626 parse_events_option),
1627 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1628 "event filter", parse_filter),
1629 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1630 "child tasks do not inherit counters"),
1631 OPT_STRING('p', "pid", &target.pid, "pid",
1632 "stat events on existing process id"),
1633 OPT_STRING('t', "tid", &target.tid, "tid",
1634 "stat events on existing thread id"),
1635 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1636 "system-wide collection from all CPUs"),
1637 OPT_BOOLEAN('g', "group", &group,
1638 "put the counters into a counter group"),
1639 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1640 OPT_INCR('v', "verbose", &verbose,
1641 "be more verbose (show counter open errors, etc)"),
1642 OPT_INTEGER('r', "repeat", &run_count,
1643 "repeat command and print average + stddev (max: 100, forever: 0)"),
1644 OPT_BOOLEAN('n', "null", &null_run,
1645 "null run - dont start any counters"),
1646 OPT_INCR('d', "detailed", &detailed_run,
1647 "detailed run - start a lot of events"),
1648 OPT_BOOLEAN('S', "sync", &sync_run,
1649 "call sync() before starting a run"),
1650 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1651 "print large numbers with thousands\' separators",
1653 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1654 "list of cpus to monitor in system-wide"),
1655 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1656 "disable CPU count aggregation", AGGR_NONE),
1657 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1658 "print counts with custom separator"),
1659 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1660 "monitor event in cgroup name only", parse_cgroups),
1661 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1662 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1663 OPT_INTEGER(0, "log-fd", &output_fd,
1664 "log output to fd, instead of stderr"),
1665 OPT_STRING(0, "pre", &pre_cmd, "command",
1666 "command to run prior to the measured command"),
1667 OPT_STRING(0, "post", &post_cmd, "command",
1668 "command to run after to the measured command"),
1669 OPT_UINTEGER('I', "interval-print", &interval,
1670 "print counts at regular interval in ms (>= 100)"),
1671 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1672 "aggregate counts per processor socket", AGGR_SOCKET),
1673 OPT_SET_UINT(0, "per-core", &aggr_mode,
1674 "aggregate counts per physical processor core", AGGR_CORE),
1675 OPT_UINTEGER('D', "delay", &initial_delay,
1676 "ms to wait before starting measurement after program start"),
1679 const char * const stat_usage[] = {
1680 "perf stat [<options>] [<command>]",
1683 int status = -EINVAL, run_idx;
1686 setlocale(LC_ALL, "");
1688 evsel_list = perf_evlist__new();
1689 if (evsel_list == NULL)
1692 argc = parse_options(argc, argv, options, stat_usage,
1693 PARSE_OPT_STOP_AT_NON_OPTION);
1696 if (output_name && strcmp(output_name, "-"))
1699 if (output_name && output_fd) {
1700 fprintf(stderr, "cannot use both --output and --log-fd\n");
1701 parse_options_usage(stat_usage, options, "o", 1);
1702 parse_options_usage(NULL, options, "log-fd", 0);
1706 if (output_fd < 0) {
1707 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1708 parse_options_usage(stat_usage, options, "log-fd", 0);
1714 mode = append_file ? "a" : "w";
1716 output = fopen(output_name, mode);
1718 perror("failed to create output file");
1721 clock_gettime(CLOCK_REALTIME, &tm);
1722 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1723 } else if (output_fd > 0) {
1724 mode = append_file ? "a" : "w";
1725 output = fdopen(output_fd, mode);
1727 perror("Failed opening logfd");
1734 if (!strcmp(csv_sep, "\\t"))
1737 csv_sep = DEFAULT_SEPARATOR;
1740 * let the spreadsheet do the pretty-printing
1743 /* User explicitly passed -B? */
1744 if (big_num_opt == 1) {
1745 fprintf(stderr, "-B option not supported with -x\n");
1746 parse_options_usage(stat_usage, options, "B", 1);
1747 parse_options_usage(NULL, options, "x", 1);
1749 } else /* Nope, so disable big number formatting */
1751 } else if (big_num_opt == 0) /* User passed --no-big-num */
1754 if (!argc && target__none(&target))
1755 usage_with_options(stat_usage, options);
1757 if (run_count < 0) {
1758 pr_err("Run count must be a positive number\n");
1759 parse_options_usage(stat_usage, options, "r", 1);
1761 } else if (run_count == 0) {
1766 /* no_aggr, cgroup are for system-wide only */
1767 if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1768 !target__has_cpu(&target)) {
1769 fprintf(stderr, "both cgroup and no-aggregation "
1770 "modes only available in system-wide mode\n");
1772 parse_options_usage(stat_usage, options, "G", 1);
1773 parse_options_usage(NULL, options, "A", 1);
1774 parse_options_usage(NULL, options, "a", 1);
1778 if (add_default_attributes())
1781 target__validate(&target);
1783 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1784 if (target__has_task(&target)) {
1785 pr_err("Problems finding threads of monitor\n");
1786 parse_options_usage(stat_usage, options, "p", 1);
1787 parse_options_usage(NULL, options, "t", 1);
1788 } else if (target__has_cpu(&target)) {
1789 perror("failed to parse CPUs map");
1790 parse_options_usage(stat_usage, options, "C", 1);
1791 parse_options_usage(NULL, options, "a", 1);
1795 if (interval && interval < 100) {
1796 pr_err("print interval must be >= 100ms\n");
1797 parse_options_usage(stat_usage, options, "I", 1);
1801 if (perf_evlist__alloc_stats(evsel_list, interval))
1804 if (perf_stat_init_aggr_mode())
1808 * We dont want to block the signals - that would cause
1809 * child tasks to inherit that and Ctrl-C would not work.
1810 * What we want is for Ctrl-C to work in the exec()-ed
1811 * task, but being ignored by perf stat itself:
1815 signal(SIGINT, skip_signal);
1816 signal(SIGCHLD, skip_signal);
1817 signal(SIGALRM, skip_signal);
1818 signal(SIGABRT, skip_signal);
1821 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1822 if (run_count != 1 && verbose)
1823 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1826 status = run_perf_stat(argc, argv);
1827 if (forever && status != -1) {
1828 print_stat(argc, argv);
1829 perf_stat__reset_stats(evsel_list);
1833 if (!forever && status != -1 && !interval)
1834 print_stat(argc, argv);
1836 perf_evlist__free_stats(evsel_list);
1838 perf_evlist__delete_maps(evsel_list);
1840 perf_evlist__delete(evsel_list);