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"
62 #include <sys/prctl.h>
65 #define DEFAULT_SEPARATOR " "
66 #define CNTR_NOT_SUPPORTED "<not supported>"
67 #define CNTR_NOT_COUNTED "<not counted>"
69 static void print_stat(int argc, const char **argv);
70 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
71 static void print_counter(struct perf_evsel *counter, char *prefix);
72 static void print_aggr(char *prefix);
74 /* Default events used for perf stat -T */
75 static const char * const transaction_attrs[] = {
87 /* More limited version when the CPU does not have all events. */
88 static const char * const transaction_limited_attrs[] = {
98 /* must match transaction_attrs and the beginning limited_attrs */
109 static struct perf_evlist *evsel_list;
111 static struct target target = {
122 static int run_count = 1;
123 static bool no_inherit = false;
124 static bool scale = true;
125 static enum aggr_mode aggr_mode = AGGR_GLOBAL;
126 static volatile pid_t child_pid = -1;
127 static bool null_run = false;
128 static int detailed_run = 0;
129 static bool transaction_run;
130 static bool big_num = true;
131 static int big_num_opt = -1;
132 static const char *csv_sep = NULL;
133 static bool csv_output = false;
134 static bool group = false;
135 static FILE *output = NULL;
136 static const char *pre_cmd = NULL;
137 static const char *post_cmd = NULL;
138 static bool sync_run = false;
139 static unsigned int interval = 0;
140 static unsigned int initial_delay = 0;
141 static unsigned int unit_width = 4; /* strlen("unit") */
142 static bool forever = false;
143 static struct timespec ref_time;
144 static struct cpu_map *aggr_map;
145 static int (*aggr_get_id)(struct cpu_map *m, int cpu);
147 static volatile int done = 0;
150 struct stats res_stats[3];
153 static inline void diff_timespec(struct timespec *r, struct timespec *a,
156 r->tv_sec = a->tv_sec - b->tv_sec;
157 if (a->tv_nsec < b->tv_nsec) {
158 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
161 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
165 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
167 return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
170 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
172 return perf_evsel__cpus(evsel)->nr;
175 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
178 struct perf_stat *ps = evsel->priv;
180 for (i = 0; i < 3; i++)
181 init_stats(&ps->res_stats[i]);
184 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
186 evsel->priv = zalloc(sizeof(struct perf_stat));
187 if (evsel->priv == NULL)
189 perf_evsel__reset_stat_priv(evsel);
193 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
198 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
203 sz = sizeof(*evsel->counts) +
204 (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
210 evsel->prev_raw_counts = addr;
215 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
217 zfree(&evsel->prev_raw_counts);
220 static void perf_evlist__free_stats(struct perf_evlist *evlist)
222 struct perf_evsel *evsel;
224 evlist__for_each(evlist, evsel) {
225 perf_evsel__free_stat_priv(evsel);
226 perf_evsel__free_counts(evsel);
227 perf_evsel__free_prev_raw_counts(evsel);
231 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
233 struct perf_evsel *evsel;
235 evlist__for_each(evlist, evsel) {
236 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
237 perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
238 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
245 perf_evlist__free_stats(evlist);
249 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
250 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
251 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
252 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
253 static struct stats runtime_branches_stats[MAX_NR_CPUS];
254 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
255 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
256 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
257 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
258 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
259 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
260 static struct stats runtime_cycles_in_tx_stats[MAX_NR_CPUS];
261 static struct stats walltime_nsecs_stats;
262 static struct stats runtime_transaction_stats[MAX_NR_CPUS];
263 static struct stats runtime_elision_stats[MAX_NR_CPUS];
265 static void perf_stat__reset_stats(struct perf_evlist *evlist)
267 struct perf_evsel *evsel;
269 evlist__for_each(evlist, evsel) {
270 perf_evsel__reset_stat_priv(evsel);
271 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
274 memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
275 memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
276 memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
277 memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
278 memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
279 memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
280 memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
281 memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
282 memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
283 memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
284 memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
285 memset(runtime_cycles_in_tx_stats, 0,
286 sizeof(runtime_cycles_in_tx_stats));
287 memset(runtime_transaction_stats, 0,
288 sizeof(runtime_transaction_stats));
289 memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
290 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
293 static int create_perf_stat_counter(struct perf_evsel *evsel)
295 struct perf_event_attr *attr = &evsel->attr;
298 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
299 PERF_FORMAT_TOTAL_TIME_RUNNING;
301 attr->inherit = !no_inherit;
303 if (target__has_cpu(&target))
304 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
306 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
309 attr->enable_on_exec = 1;
312 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
316 * Does the counter have nsecs as a unit?
318 static inline int nsec_counter(struct perf_evsel *evsel)
320 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
321 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
327 static struct perf_evsel *nth_evsel(int n)
329 static struct perf_evsel **array;
330 static int array_len;
331 struct perf_evsel *ev;
334 /* Assumes this only called when evsel_list does not change anymore. */
336 evlist__for_each(evsel_list, ev)
338 array = malloc(array_len * sizeof(void *));
342 evlist__for_each(evsel_list, ev)
351 * Update various tracking values we maintain to print
352 * more semantic information such as miss/hit ratios,
353 * instruction rates, etc:
355 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
357 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
358 update_stats(&runtime_nsecs_stats[0], count[0]);
359 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
360 update_stats(&runtime_cycles_stats[0], count[0]);
361 else if (transaction_run &&
362 perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX)))
363 update_stats(&runtime_cycles_in_tx_stats[0], count[0]);
364 else if (transaction_run &&
365 perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START)))
366 update_stats(&runtime_transaction_stats[0], count[0]);
367 else if (transaction_run &&
368 perf_evsel__cmp(counter, nth_evsel(T_ELISION_START)))
369 update_stats(&runtime_elision_stats[0], count[0]);
370 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
371 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
372 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
373 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
374 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
375 update_stats(&runtime_branches_stats[0], count[0]);
376 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
377 update_stats(&runtime_cacherefs_stats[0], count[0]);
378 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
379 update_stats(&runtime_l1_dcache_stats[0], count[0]);
380 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
381 update_stats(&runtime_l1_icache_stats[0], count[0]);
382 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
383 update_stats(&runtime_ll_cache_stats[0], count[0]);
384 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
385 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
386 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
387 update_stats(&runtime_itlb_cache_stats[0], count[0]);
391 * Read out the results of a single counter:
392 * aggregate counts across CPUs in system-wide mode
394 static int read_counter_aggr(struct perf_evsel *counter)
396 struct perf_stat *ps = counter->priv;
397 u64 *count = counter->counts->aggr.values;
400 if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
401 thread_map__nr(evsel_list->threads), scale) < 0)
404 for (i = 0; i < 3; i++)
405 update_stats(&ps->res_stats[i], count[i]);
408 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
409 perf_evsel__name(counter), count[0], count[1], count[2]);
413 * Save the full runtime - to allow normalization during printout:
415 update_shadow_stats(counter, count);
421 * Read out the results of a single counter:
422 * do not aggregate counts across CPUs in system-wide mode
424 static int read_counter(struct perf_evsel *counter)
429 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
430 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
433 count = counter->counts->cpu[cpu].values;
435 update_shadow_stats(counter, count);
441 static void print_interval(void)
443 static int num_print_interval;
444 struct perf_evsel *counter;
445 struct perf_stat *ps;
446 struct timespec ts, rs;
449 if (aggr_mode == AGGR_GLOBAL) {
450 evlist__for_each(evsel_list, counter) {
452 memset(ps->res_stats, 0, sizeof(ps->res_stats));
453 read_counter_aggr(counter);
456 evlist__for_each(evsel_list, counter) {
458 memset(ps->res_stats, 0, sizeof(ps->res_stats));
459 read_counter(counter);
463 clock_gettime(CLOCK_MONOTONIC, &ts);
464 diff_timespec(&rs, &ts, &ref_time);
465 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
467 if (num_print_interval == 0 && !csv_output) {
470 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
473 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
476 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
480 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
484 if (++num_print_interval == 25)
485 num_print_interval = 0;
493 evlist__for_each(evsel_list, counter)
494 print_counter(counter, prefix);
498 evlist__for_each(evsel_list, counter)
499 print_counter_aggr(counter, prefix);
505 static void handle_initial_delay(void)
507 struct perf_evsel *counter;
510 const int ncpus = cpu_map__nr(evsel_list->cpus),
511 nthreads = thread_map__nr(evsel_list->threads);
513 usleep(initial_delay * 1000);
514 evlist__for_each(evsel_list, counter)
515 perf_evsel__enable(counter, ncpus, nthreads);
519 static volatile int workload_exec_errno;
522 * perf_evlist__prepare_workload will send a SIGUSR1
523 * if the fork fails, since we asked by setting its
524 * want_signal to true.
526 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
527 void *ucontext __maybe_unused)
529 workload_exec_errno = info->si_value.sival_int;
532 static int __run_perf_stat(int argc, const char **argv)
535 unsigned long long t0, t1;
536 struct perf_evsel *counter;
540 const bool forks = (argc > 0);
543 ts.tv_sec = interval / 1000;
544 ts.tv_nsec = (interval % 1000) * 1000000;
551 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
552 workload_exec_failed_signal) < 0) {
553 perror("failed to prepare workload");
556 child_pid = evsel_list->workload.pid;
560 perf_evlist__set_leader(evsel_list);
562 evlist__for_each(evsel_list, counter) {
563 if (create_perf_stat_counter(counter) < 0) {
565 * PPC returns ENXIO for HW counters until 2.6.37
566 * (behavior changed with commit b0a873e).
568 if (errno == EINVAL || errno == ENOSYS ||
569 errno == ENOENT || errno == EOPNOTSUPP ||
572 ui__warning("%s event is not supported by the kernel.\n",
573 perf_evsel__name(counter));
574 counter->supported = false;
578 perf_evsel__open_strerror(counter, &target,
579 errno, msg, sizeof(msg));
580 ui__error("%s\n", msg);
583 kill(child_pid, SIGTERM);
587 counter->supported = true;
589 l = strlen(counter->unit);
594 if (perf_evlist__apply_filters(evsel_list)) {
595 error("failed to set filter with %d (%s)\n", errno,
601 * Enable counters and exec the command:
604 clock_gettime(CLOCK_MONOTONIC, &ref_time);
607 perf_evlist__start_workload(evsel_list);
608 handle_initial_delay();
611 while (!waitpid(child_pid, &status, WNOHANG)) {
612 nanosleep(&ts, NULL);
618 if (workload_exec_errno) {
619 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
620 pr_err("Workload failed: %s\n", emsg);
624 if (WIFSIGNALED(status))
625 psignal(WTERMSIG(status), argv[0]);
627 handle_initial_delay();
629 nanosleep(&ts, NULL);
637 update_stats(&walltime_nsecs_stats, t1 - t0);
639 if (aggr_mode == AGGR_GLOBAL) {
640 evlist__for_each(evsel_list, counter) {
641 read_counter_aggr(counter);
642 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
643 thread_map__nr(evsel_list->threads));
646 evlist__for_each(evsel_list, counter) {
647 read_counter(counter);
648 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
652 return WEXITSTATUS(status);
655 static int run_perf_stat(int argc, const char **argv)
660 ret = system(pre_cmd);
668 ret = __run_perf_stat(argc, argv);
673 ret = system(post_cmd);
681 static void print_noise_pct(double total, double avg)
683 double pct = rel_stddev_stats(total, avg);
686 fprintf(output, "%s%.2f%%", csv_sep, pct);
688 fprintf(output, " ( +-%6.2f%% )", pct);
691 static void print_noise(struct perf_evsel *evsel, double avg)
693 struct perf_stat *ps;
699 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
702 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
706 fprintf(output, "S%d-C%*d%s%*d%s",
707 cpu_map__id_to_socket(id),
709 cpu_map__id_to_cpu(id),
716 fprintf(output, "S%*d%s%*d%s",
725 fprintf(output, "CPU%*d%s",
727 perf_evsel__cpus(evsel)->map[id], csv_sep);
735 static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
737 double msecs = avg / 1e6;
738 const char *fmt_v, *fmt_n;
741 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
742 fmt_n = csv_output ? "%s" : "%-25s";
744 aggr_printout(evsel, cpu, nr);
746 scnprintf(name, sizeof(name), "%s%s",
747 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
749 fprintf(output, fmt_v, msecs, csv_sep);
752 fprintf(output, "%s%s", evsel->unit, csv_sep);
754 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
756 fprintf(output, fmt_n, name);
759 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
761 if (csv_output || interval)
764 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
765 fprintf(output, " # %8.3f CPUs utilized ",
766 avg / avg_stats(&walltime_nsecs_stats));
768 fprintf(output, " ");
771 /* used for get_ratio_color() */
773 GRC_STALLED_CYCLES_FE,
774 GRC_STALLED_CYCLES_BE,
779 static const char *get_ratio_color(enum grc_type type, double ratio)
781 static const double grc_table[GRC_MAX_NR][3] = {
782 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
783 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
784 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
786 const char *color = PERF_COLOR_NORMAL;
788 if (ratio > grc_table[type][0])
789 color = PERF_COLOR_RED;
790 else if (ratio > grc_table[type][1])
791 color = PERF_COLOR_MAGENTA;
792 else if (ratio > grc_table[type][2])
793 color = PERF_COLOR_YELLOW;
798 static void print_stalled_cycles_frontend(int cpu,
799 struct perf_evsel *evsel
800 __maybe_unused, double avg)
802 double total, ratio = 0.0;
805 total = avg_stats(&runtime_cycles_stats[cpu]);
808 ratio = avg / total * 100.0;
810 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
812 fprintf(output, " # ");
813 color_fprintf(output, color, "%6.2f%%", ratio);
814 fprintf(output, " frontend cycles idle ");
817 static void print_stalled_cycles_backend(int cpu,
818 struct perf_evsel *evsel
819 __maybe_unused, double avg)
821 double total, ratio = 0.0;
824 total = avg_stats(&runtime_cycles_stats[cpu]);
827 ratio = avg / total * 100.0;
829 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
831 fprintf(output, " # ");
832 color_fprintf(output, color, "%6.2f%%", ratio);
833 fprintf(output, " backend cycles idle ");
836 static void print_branch_misses(int cpu,
837 struct perf_evsel *evsel __maybe_unused,
840 double total, ratio = 0.0;
843 total = avg_stats(&runtime_branches_stats[cpu]);
846 ratio = avg / total * 100.0;
848 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
850 fprintf(output, " # ");
851 color_fprintf(output, color, "%6.2f%%", ratio);
852 fprintf(output, " of all branches ");
855 static void print_l1_dcache_misses(int cpu,
856 struct perf_evsel *evsel __maybe_unused,
859 double total, ratio = 0.0;
862 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
865 ratio = avg / total * 100.0;
867 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
869 fprintf(output, " # ");
870 color_fprintf(output, color, "%6.2f%%", ratio);
871 fprintf(output, " of all L1-dcache hits ");
874 static void print_l1_icache_misses(int cpu,
875 struct perf_evsel *evsel __maybe_unused,
878 double total, ratio = 0.0;
881 total = avg_stats(&runtime_l1_icache_stats[cpu]);
884 ratio = avg / total * 100.0;
886 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
888 fprintf(output, " # ");
889 color_fprintf(output, color, "%6.2f%%", ratio);
890 fprintf(output, " of all L1-icache hits ");
893 static void print_dtlb_cache_misses(int cpu,
894 struct perf_evsel *evsel __maybe_unused,
897 double total, ratio = 0.0;
900 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
903 ratio = avg / total * 100.0;
905 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
907 fprintf(output, " # ");
908 color_fprintf(output, color, "%6.2f%%", ratio);
909 fprintf(output, " of all dTLB cache hits ");
912 static void print_itlb_cache_misses(int cpu,
913 struct perf_evsel *evsel __maybe_unused,
916 double total, ratio = 0.0;
919 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
922 ratio = avg / total * 100.0;
924 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
926 fprintf(output, " # ");
927 color_fprintf(output, color, "%6.2f%%", ratio);
928 fprintf(output, " of all iTLB cache hits ");
931 static void print_ll_cache_misses(int cpu,
932 struct perf_evsel *evsel __maybe_unused,
935 double total, ratio = 0.0;
938 total = avg_stats(&runtime_ll_cache_stats[cpu]);
941 ratio = avg / total * 100.0;
943 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
945 fprintf(output, " # ");
946 color_fprintf(output, color, "%6.2f%%", ratio);
947 fprintf(output, " of all LL-cache hits ");
950 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
952 double total, ratio = 0.0, total2;
953 double sc = evsel->scale;
957 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
960 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
962 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
965 aggr_printout(evsel, cpu, nr);
967 if (aggr_mode == AGGR_GLOBAL)
970 fprintf(output, fmt, avg, csv_sep);
973 fprintf(output, "%-*s%s",
974 csv_output ? 0 : unit_width,
975 evsel->unit, csv_sep);
977 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
980 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
982 if (csv_output || interval)
985 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
986 total = avg_stats(&runtime_cycles_stats[cpu]);
989 fprintf(output, " # %5.2f insns per cycle ", ratio);
991 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
992 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
996 fprintf(output, "\n");
997 if (aggr_mode == AGGR_NONE)
998 fprintf(output, " ");
999 fprintf(output, " # %5.2f stalled cycles per insn", ratio);
1002 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
1003 runtime_branches_stats[cpu].n != 0) {
1004 print_branch_misses(cpu, evsel, avg);
1006 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1007 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
1008 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1009 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1010 runtime_l1_dcache_stats[cpu].n != 0) {
1011 print_l1_dcache_misses(cpu, evsel, avg);
1013 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1014 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
1015 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1016 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1017 runtime_l1_icache_stats[cpu].n != 0) {
1018 print_l1_icache_misses(cpu, evsel, avg);
1020 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1021 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
1022 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1023 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1024 runtime_dtlb_cache_stats[cpu].n != 0) {
1025 print_dtlb_cache_misses(cpu, evsel, avg);
1027 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1028 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
1029 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1030 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1031 runtime_itlb_cache_stats[cpu].n != 0) {
1032 print_itlb_cache_misses(cpu, evsel, avg);
1034 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1035 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
1036 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1037 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1038 runtime_ll_cache_stats[cpu].n != 0) {
1039 print_ll_cache_misses(cpu, evsel, avg);
1040 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
1041 runtime_cacherefs_stats[cpu].n != 0) {
1042 total = avg_stats(&runtime_cacherefs_stats[cpu]);
1045 ratio = avg * 100 / total;
1047 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
1049 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1050 print_stalled_cycles_frontend(cpu, evsel, avg);
1051 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1052 print_stalled_cycles_backend(cpu, evsel, avg);
1053 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1054 total = avg_stats(&runtime_nsecs_stats[cpu]);
1057 ratio = avg / total;
1058 fprintf(output, " # %8.3f GHz ", ratio);
1060 } else if (transaction_run &&
1061 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1062 total = avg_stats(&runtime_cycles_stats[cpu]);
1065 " # %5.2f%% transactional cycles ",
1066 100.0 * (avg / total));
1067 } else if (transaction_run &&
1068 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1069 total = avg_stats(&runtime_cycles_stats[cpu]);
1070 total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1075 " # %5.2f%% aborted cycles ",
1076 100.0 * ((total2-avg) / total));
1077 } else if (transaction_run &&
1078 perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1080 runtime_cycles_in_tx_stats[cpu].n != 0) {
1081 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1084 ratio = total / avg;
1086 fprintf(output, " # %8.0f cycles / transaction ", ratio);
1087 } else if (transaction_run &&
1088 perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1090 runtime_cycles_in_tx_stats[cpu].n != 0) {
1091 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1094 ratio = total / avg;
1096 fprintf(output, " # %8.0f cycles / elision ", ratio);
1097 } else if (runtime_nsecs_stats[cpu].n != 0) {
1100 total = avg_stats(&runtime_nsecs_stats[cpu]);
1103 ratio = 1000.0 * avg / total;
1104 if (ratio < 0.001) {
1109 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
1111 fprintf(output, " ");
1115 static void print_aggr(char *prefix)
1117 struct perf_evsel *counter;
1118 int cpu, cpu2, s, s2, id, nr;
1122 if (!(aggr_map || aggr_get_id))
1125 for (s = 0; s < aggr_map->nr; s++) {
1126 id = aggr_map->map[s];
1127 evlist__for_each(evsel_list, counter) {
1128 val = ena = run = 0;
1130 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1131 cpu2 = perf_evsel__cpus(counter)->map[cpu];
1132 s2 = aggr_get_id(evsel_list->cpus, cpu2);
1135 val += counter->counts->cpu[cpu].val;
1136 ena += counter->counts->cpu[cpu].ena;
1137 run += counter->counts->cpu[cpu].run;
1141 fprintf(output, "%s", prefix);
1143 if (run == 0 || ena == 0) {
1144 aggr_printout(counter, id, nr);
1146 fprintf(output, "%*s%s",
1147 csv_output ? 0 : 18,
1148 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1151 fprintf(output, "%-*s%s",
1152 csv_output ? 0 : unit_width,
1153 counter->unit, csv_sep);
1155 fprintf(output, "%*s",
1156 csv_output ? 0 : -25,
1157 perf_evsel__name(counter));
1160 fprintf(output, "%s%s",
1161 csv_sep, counter->cgrp->name);
1163 fputc('\n', output);
1166 uval = val * counter->scale;
1168 if (nsec_counter(counter))
1169 nsec_printout(id, nr, counter, uval);
1171 abs_printout(id, nr, counter, uval);
1174 print_noise(counter, 1.0);
1177 fprintf(output, " (%.2f%%)",
1180 fputc('\n', output);
1186 * Print out the results of a single counter:
1187 * aggregated counts in system-wide mode
1189 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1191 struct perf_stat *ps = counter->priv;
1192 double avg = avg_stats(&ps->res_stats[0]);
1193 int scaled = counter->counts->scaled;
1197 fprintf(output, "%s", prefix);
1200 fprintf(output, "%*s%s",
1201 csv_output ? 0 : 18,
1202 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1204 fprintf(output, "%-*s%s",
1205 csv_output ? 0 : unit_width,
1206 counter->unit, csv_sep);
1207 fprintf(output, "%*s",
1208 csv_output ? 0 : -25,
1209 perf_evsel__name(counter));
1212 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1214 fputc('\n', output);
1218 uval = avg * counter->scale;
1220 if (nsec_counter(counter))
1221 nsec_printout(-1, 0, counter, uval);
1223 abs_printout(-1, 0, counter, uval);
1225 print_noise(counter, avg);
1228 fputc('\n', output);
1233 double avg_enabled, avg_running;
1235 avg_enabled = avg_stats(&ps->res_stats[1]);
1236 avg_running = avg_stats(&ps->res_stats[2]);
1238 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1240 fprintf(output, "\n");
1244 * Print out the results of a single counter:
1245 * does not use aggregated count in system-wide
1247 static void print_counter(struct perf_evsel *counter, char *prefix)
1253 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1254 val = counter->counts->cpu[cpu].val;
1255 ena = counter->counts->cpu[cpu].ena;
1256 run = counter->counts->cpu[cpu].run;
1259 fprintf(output, "%s", prefix);
1261 if (run == 0 || ena == 0) {
1262 fprintf(output, "CPU%*d%s%*s%s",
1263 csv_output ? 0 : -4,
1264 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1265 csv_output ? 0 : 18,
1266 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1269 fprintf(output, "%-*s%s",
1270 csv_output ? 0 : unit_width,
1271 counter->unit, csv_sep);
1273 fprintf(output, "%*s",
1274 csv_output ? 0 : -25,
1275 perf_evsel__name(counter));
1278 fprintf(output, "%s%s",
1279 csv_sep, counter->cgrp->name);
1281 fputc('\n', output);
1285 uval = val * counter->scale;
1287 if (nsec_counter(counter))
1288 nsec_printout(cpu, 0, counter, uval);
1290 abs_printout(cpu, 0, counter, uval);
1293 print_noise(counter, 1.0);
1296 fprintf(output, " (%.2f%%)",
1299 fputc('\n', output);
1303 static void print_stat(int argc, const char **argv)
1305 struct perf_evsel *counter;
1311 fprintf(output, "\n");
1312 fprintf(output, " Performance counter stats for ");
1313 if (target.system_wide)
1314 fprintf(output, "\'system wide");
1315 else if (target.cpu_list)
1316 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1317 else if (!target__has_task(&target)) {
1318 fprintf(output, "\'%s", argv[0]);
1319 for (i = 1; i < argc; i++)
1320 fprintf(output, " %s", argv[i]);
1321 } else if (target.pid)
1322 fprintf(output, "process id \'%s", target.pid);
1324 fprintf(output, "thread id \'%s", target.tid);
1326 fprintf(output, "\'");
1328 fprintf(output, " (%d runs)", run_count);
1329 fprintf(output, ":\n\n");
1332 switch (aggr_mode) {
1338 evlist__for_each(evsel_list, counter)
1339 print_counter_aggr(counter, NULL);
1342 evlist__for_each(evsel_list, counter)
1343 print_counter(counter, NULL);
1351 fprintf(output, "\n");
1352 fprintf(output, " %17.9f seconds time elapsed",
1353 avg_stats(&walltime_nsecs_stats)/1e9);
1354 if (run_count > 1) {
1355 fprintf(output, " ");
1356 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1357 avg_stats(&walltime_nsecs_stats));
1359 fprintf(output, "\n\n");
1363 static volatile int signr = -1;
1365 static void skip_signal(int signo)
1367 if ((child_pid == -1) || interval)
1372 * render child_pid harmless
1373 * won't send SIGTERM to a random
1374 * process in case of race condition
1375 * and fast PID recycling
1380 static void sig_atexit(void)
1385 * avoid race condition with SIGCHLD handler
1386 * in skip_signal() which is modifying child_pid
1387 * goal is to avoid send SIGTERM to a random
1391 sigaddset(&set, SIGCHLD);
1392 sigprocmask(SIG_BLOCK, &set, &oset);
1394 if (child_pid != -1)
1395 kill(child_pid, SIGTERM);
1397 sigprocmask(SIG_SETMASK, &oset, NULL);
1402 signal(signr, SIG_DFL);
1403 kill(getpid(), signr);
1406 static int stat__set_big_num(const struct option *opt __maybe_unused,
1407 const char *s __maybe_unused, int unset)
1409 big_num_opt = unset ? 0 : 1;
1413 static int perf_stat_init_aggr_mode(void)
1415 switch (aggr_mode) {
1417 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1418 perror("cannot build socket map");
1421 aggr_get_id = cpu_map__get_socket;
1424 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1425 perror("cannot build core map");
1428 aggr_get_id = cpu_map__get_core;
1438 static int setup_events(const char * const *attrs, unsigned len)
1442 for (i = 0; i < len; i++) {
1443 if (parse_events(evsel_list, attrs[i]))
1450 * Add default attributes, if there were no attributes specified or
1451 * if -d/--detailed, -d -d or -d -d -d is used:
1453 static int add_default_attributes(void)
1455 struct perf_event_attr default_attrs[] = {
1457 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1458 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1459 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1460 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1462 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1463 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1464 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1465 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1466 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1467 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1472 * Detailed stats (-d), covering the L1 and last level data caches:
1474 struct perf_event_attr detailed_attrs[] = {
1476 { .type = PERF_TYPE_HW_CACHE,
1478 PERF_COUNT_HW_CACHE_L1D << 0 |
1479 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1480 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
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_MISS << 16) },
1488 { .type = PERF_TYPE_HW_CACHE,
1490 PERF_COUNT_HW_CACHE_LL << 0 |
1491 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1492 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 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_MISS << 16) },
1502 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1504 struct perf_event_attr very_detailed_attrs[] = {
1506 { .type = PERF_TYPE_HW_CACHE,
1508 PERF_COUNT_HW_CACHE_L1I << 0 |
1509 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1510 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
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_MISS << 16) },
1518 { .type = PERF_TYPE_HW_CACHE,
1520 PERF_COUNT_HW_CACHE_DTLB << 0 |
1521 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1522 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 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_MISS << 16) },
1530 { .type = PERF_TYPE_HW_CACHE,
1532 PERF_COUNT_HW_CACHE_ITLB << 0 |
1533 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1534 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 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_MISS << 16) },
1545 * Very, very detailed stats (-d -d -d), adding prefetch events:
1547 struct perf_event_attr very_very_detailed_attrs[] = {
1549 { .type = PERF_TYPE_HW_CACHE,
1551 PERF_COUNT_HW_CACHE_L1D << 0 |
1552 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1553 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
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_MISS << 16) },
1562 /* Set attrs if no event is selected and !null_run: */
1566 if (transaction_run) {
1568 if (pmu_have_event("cpu", "cycles-ct") &&
1569 pmu_have_event("cpu", "el-start"))
1570 err = setup_events(transaction_attrs,
1571 ARRAY_SIZE(transaction_attrs));
1573 err = setup_events(transaction_limited_attrs,
1574 ARRAY_SIZE(transaction_limited_attrs));
1576 fprintf(stderr, "Cannot set up transaction events\n");
1582 if (!evsel_list->nr_entries) {
1583 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1587 /* Detailed events get appended to the event list: */
1589 if (detailed_run < 1)
1592 /* Append detailed run extra attributes: */
1593 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1596 if (detailed_run < 2)
1599 /* Append very detailed run extra attributes: */
1600 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1603 if (detailed_run < 3)
1606 /* Append very, very detailed run extra attributes: */
1607 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1610 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1612 bool append_file = false;
1614 const char *output_name = NULL;
1615 const struct option options[] = {
1616 OPT_BOOLEAN('T', "transaction", &transaction_run,
1617 "hardware transaction statistics"),
1618 OPT_CALLBACK('e', "event", &evsel_list, "event",
1619 "event selector. use 'perf list' to list available events",
1620 parse_events_option),
1621 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1622 "event filter", parse_filter),
1623 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1624 "child tasks do not inherit counters"),
1625 OPT_STRING('p', "pid", &target.pid, "pid",
1626 "stat events on existing process id"),
1627 OPT_STRING('t', "tid", &target.tid, "tid",
1628 "stat events on existing thread id"),
1629 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1630 "system-wide collection from all CPUs"),
1631 OPT_BOOLEAN('g', "group", &group,
1632 "put the counters into a counter group"),
1633 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1634 OPT_INCR('v', "verbose", &verbose,
1635 "be more verbose (show counter open errors, etc)"),
1636 OPT_INTEGER('r', "repeat", &run_count,
1637 "repeat command and print average + stddev (max: 100, forever: 0)"),
1638 OPT_BOOLEAN('n', "null", &null_run,
1639 "null run - dont start any counters"),
1640 OPT_INCR('d', "detailed", &detailed_run,
1641 "detailed run - start a lot of events"),
1642 OPT_BOOLEAN('S', "sync", &sync_run,
1643 "call sync() before starting a run"),
1644 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1645 "print large numbers with thousands\' separators",
1647 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1648 "list of cpus to monitor in system-wide"),
1649 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1650 "disable CPU count aggregation", AGGR_NONE),
1651 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1652 "print counts with custom separator"),
1653 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1654 "monitor event in cgroup name only", parse_cgroups),
1655 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1656 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1657 OPT_INTEGER(0, "log-fd", &output_fd,
1658 "log output to fd, instead of stderr"),
1659 OPT_STRING(0, "pre", &pre_cmd, "command",
1660 "command to run prior to the measured command"),
1661 OPT_STRING(0, "post", &post_cmd, "command",
1662 "command to run after to the measured command"),
1663 OPT_UINTEGER('I', "interval-print", &interval,
1664 "print counts at regular interval in ms (>= 100)"),
1665 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1666 "aggregate counts per processor socket", AGGR_SOCKET),
1667 OPT_SET_UINT(0, "per-core", &aggr_mode,
1668 "aggregate counts per physical processor core", AGGR_CORE),
1669 OPT_UINTEGER('D', "delay", &initial_delay,
1670 "ms to wait before starting measurement after program start"),
1673 const char * const stat_usage[] = {
1674 "perf stat [<options>] [<command>]",
1677 int status = -EINVAL, run_idx;
1680 setlocale(LC_ALL, "");
1682 evsel_list = perf_evlist__new();
1683 if (evsel_list == NULL)
1686 argc = parse_options(argc, argv, options, stat_usage,
1687 PARSE_OPT_STOP_AT_NON_OPTION);
1690 if (output_name && strcmp(output_name, "-"))
1693 if (output_name && output_fd) {
1694 fprintf(stderr, "cannot use both --output and --log-fd\n");
1695 parse_options_usage(stat_usage, options, "o", 1);
1696 parse_options_usage(NULL, options, "log-fd", 0);
1700 if (output_fd < 0) {
1701 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1702 parse_options_usage(stat_usage, options, "log-fd", 0);
1708 mode = append_file ? "a" : "w";
1710 output = fopen(output_name, mode);
1712 perror("failed to create output file");
1715 clock_gettime(CLOCK_REALTIME, &tm);
1716 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1717 } else if (output_fd > 0) {
1718 mode = append_file ? "a" : "w";
1719 output = fdopen(output_fd, mode);
1721 perror("Failed opening logfd");
1728 if (!strcmp(csv_sep, "\\t"))
1731 csv_sep = DEFAULT_SEPARATOR;
1734 * let the spreadsheet do the pretty-printing
1737 /* User explicitly passed -B? */
1738 if (big_num_opt == 1) {
1739 fprintf(stderr, "-B option not supported with -x\n");
1740 parse_options_usage(stat_usage, options, "B", 1);
1741 parse_options_usage(NULL, options, "x", 1);
1743 } else /* Nope, so disable big number formatting */
1745 } else if (big_num_opt == 0) /* User passed --no-big-num */
1748 if (!argc && target__none(&target))
1749 usage_with_options(stat_usage, options);
1751 if (run_count < 0) {
1752 pr_err("Run count must be a positive number\n");
1753 parse_options_usage(stat_usage, options, "r", 1);
1755 } else if (run_count == 0) {
1760 /* no_aggr, cgroup are for system-wide only */
1761 if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1762 !target__has_cpu(&target)) {
1763 fprintf(stderr, "both cgroup and no-aggregation "
1764 "modes only available in system-wide mode\n");
1766 parse_options_usage(stat_usage, options, "G", 1);
1767 parse_options_usage(NULL, options, "A", 1);
1768 parse_options_usage(NULL, options, "a", 1);
1772 if (add_default_attributes())
1775 target__validate(&target);
1777 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1778 if (target__has_task(&target)) {
1779 pr_err("Problems finding threads of monitor\n");
1780 parse_options_usage(stat_usage, options, "p", 1);
1781 parse_options_usage(NULL, options, "t", 1);
1782 } else if (target__has_cpu(&target)) {
1783 perror("failed to parse CPUs map");
1784 parse_options_usage(stat_usage, options, "C", 1);
1785 parse_options_usage(NULL, options, "a", 1);
1789 if (interval && interval < 100) {
1790 pr_err("print interval must be >= 100ms\n");
1791 parse_options_usage(stat_usage, options, "I", 1);
1795 if (perf_evlist__alloc_stats(evsel_list, interval))
1798 if (perf_stat_init_aggr_mode())
1802 * We dont want to block the signals - that would cause
1803 * child tasks to inherit that and Ctrl-C would not work.
1804 * What we want is for Ctrl-C to work in the exec()-ed
1805 * task, but being ignored by perf stat itself:
1809 signal(SIGINT, skip_signal);
1810 signal(SIGCHLD, skip_signal);
1811 signal(SIGALRM, skip_signal);
1812 signal(SIGABRT, skip_signal);
1815 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1816 if (run_count != 1 && verbose)
1817 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1820 status = run_perf_stat(argc, argv);
1821 if (forever && status != -1) {
1822 print_stat(argc, argv);
1823 perf_stat__reset_stats(evsel_list);
1827 if (!forever && status != -1 && !interval)
1828 print_stat(argc, argv);
1830 perf_evlist__free_stats(evsel_list);
1832 perf_evlist__delete(evsel_list);