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 perf_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 bool forever = false;
142 static struct timespec ref_time;
143 static struct cpu_map *aggr_map;
144 static int (*aggr_get_id)(struct cpu_map *m, int cpu);
146 static volatile int done = 0;
149 struct stats res_stats[3];
152 static inline void diff_timespec(struct timespec *r, struct timespec *a,
155 r->tv_sec = a->tv_sec - b->tv_sec;
156 if (a->tv_nsec < b->tv_nsec) {
157 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
160 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
164 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
166 return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
169 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
171 return perf_evsel__cpus(evsel)->nr;
174 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
176 memset(evsel->priv, 0, sizeof(struct perf_stat));
179 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
181 evsel->priv = zalloc(sizeof(struct perf_stat));
182 return evsel->priv == NULL ? -ENOMEM : 0;
185 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
191 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
196 sz = sizeof(*evsel->counts) +
197 (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
203 evsel->prev_raw_counts = addr;
208 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
210 free(evsel->prev_raw_counts);
211 evsel->prev_raw_counts = NULL;
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 (perf_target__has_cpu(&target))
298 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
300 if (!perf_target__has_task(&target) &&
301 perf_evsel__is_group_leader(evsel)) {
304 attr->enable_on_exec = 1;
307 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
311 * Does the counter have nsecs as a unit?
313 static inline int nsec_counter(struct perf_evsel *evsel)
315 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
316 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
322 static struct perf_evsel *nth_evsel(int n)
324 static struct perf_evsel **array;
325 static int array_len;
326 struct perf_evsel *ev;
329 /* Assumes this only called when evsel_list does not change anymore. */
331 list_for_each_entry(ev, &evsel_list->entries, node)
333 array = malloc(array_len * sizeof(void *));
337 list_for_each_entry(ev, &evsel_list->entries, node)
346 * Update various tracking values we maintain to print
347 * more semantic information such as miss/hit ratios,
348 * instruction rates, etc:
350 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
352 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
353 update_stats(&runtime_nsecs_stats[0], count[0]);
354 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
355 update_stats(&runtime_cycles_stats[0], count[0]);
356 else if (transaction_run &&
357 perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX)))
358 update_stats(&runtime_cycles_in_tx_stats[0], count[0]);
359 else if (transaction_run &&
360 perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START)))
361 update_stats(&runtime_transaction_stats[0], count[0]);
362 else if (transaction_run &&
363 perf_evsel__cmp(counter, nth_evsel(T_ELISION_START)))
364 update_stats(&runtime_elision_stats[0], count[0]);
365 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
366 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
367 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
368 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
369 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
370 update_stats(&runtime_branches_stats[0], count[0]);
371 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
372 update_stats(&runtime_cacherefs_stats[0], count[0]);
373 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
374 update_stats(&runtime_l1_dcache_stats[0], count[0]);
375 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
376 update_stats(&runtime_l1_icache_stats[0], count[0]);
377 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
378 update_stats(&runtime_ll_cache_stats[0], count[0]);
379 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
380 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
381 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
382 update_stats(&runtime_itlb_cache_stats[0], count[0]);
386 * Read out the results of a single counter:
387 * aggregate counts across CPUs in system-wide mode
389 static int read_counter_aggr(struct perf_evsel *counter)
391 struct perf_stat *ps = counter->priv;
392 u64 *count = counter->counts->aggr.values;
395 if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
396 thread_map__nr(evsel_list->threads), scale) < 0)
399 for (i = 0; i < 3; i++)
400 update_stats(&ps->res_stats[i], count[i]);
403 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
404 perf_evsel__name(counter), count[0], count[1], count[2]);
408 * Save the full runtime - to allow normalization during printout:
410 update_shadow_stats(counter, count);
416 * Read out the results of a single counter:
417 * do not aggregate counts across CPUs in system-wide mode
419 static int read_counter(struct perf_evsel *counter)
424 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
425 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
428 count = counter->counts->cpu[cpu].values;
430 update_shadow_stats(counter, count);
436 static void print_interval(void)
438 static int num_print_interval;
439 struct perf_evsel *counter;
440 struct perf_stat *ps;
441 struct timespec ts, rs;
444 if (aggr_mode == AGGR_GLOBAL) {
445 list_for_each_entry(counter, &evsel_list->entries, node) {
447 memset(ps->res_stats, 0, sizeof(ps->res_stats));
448 read_counter_aggr(counter);
451 list_for_each_entry(counter, &evsel_list->entries, node) {
453 memset(ps->res_stats, 0, sizeof(ps->res_stats));
454 read_counter(counter);
458 clock_gettime(CLOCK_MONOTONIC, &ts);
459 diff_timespec(&rs, &ts, &ref_time);
460 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
462 if (num_print_interval == 0 && !csv_output) {
465 fprintf(output, "# time socket cpus counts events\n");
468 fprintf(output, "# time core cpus counts events\n");
471 fprintf(output, "# time CPU counts events\n");
475 fprintf(output, "# time counts events\n");
479 if (++num_print_interval == 25)
480 num_print_interval = 0;
488 list_for_each_entry(counter, &evsel_list->entries, node)
489 print_counter(counter, prefix);
493 list_for_each_entry(counter, &evsel_list->entries, node)
494 print_counter_aggr(counter, prefix);
500 static void handle_initial_delay(void)
502 struct perf_evsel *counter;
505 const int ncpus = cpu_map__nr(evsel_list->cpus),
506 nthreads = thread_map__nr(evsel_list->threads);
508 usleep(initial_delay * 1000);
509 list_for_each_entry(counter, &evsel_list->entries, node)
510 perf_evsel__enable(counter, ncpus, nthreads);
514 static int __run_perf_stat(int argc, const char **argv)
517 unsigned long long t0, t1;
518 struct perf_evsel *counter;
521 const bool forks = (argc > 0);
524 ts.tv_sec = interval / 1000;
525 ts.tv_nsec = (interval % 1000) * 1000000;
532 if (perf_evlist__prepare_workload(evsel_list, &target, argv,
534 perror("failed to prepare workload");
537 child_pid = evsel_list->workload.pid;
541 perf_evlist__set_leader(evsel_list);
543 list_for_each_entry(counter, &evsel_list->entries, node) {
544 if (create_perf_stat_counter(counter) < 0) {
546 * PPC returns ENXIO for HW counters until 2.6.37
547 * (behavior changed with commit b0a873e).
549 if (errno == EINVAL || errno == ENOSYS ||
550 errno == ENOENT || errno == EOPNOTSUPP ||
553 ui__warning("%s event is not supported by the kernel.\n",
554 perf_evsel__name(counter));
555 counter->supported = false;
559 perf_evsel__open_strerror(counter, &target,
560 errno, msg, sizeof(msg));
561 ui__error("%s\n", msg);
564 kill(child_pid, SIGTERM);
568 counter->supported = true;
571 if (perf_evlist__apply_filters(evsel_list)) {
572 error("failed to set filter with %d (%s)\n", errno,
578 * Enable counters and exec the command:
581 clock_gettime(CLOCK_MONOTONIC, &ref_time);
584 perf_evlist__start_workload(evsel_list);
585 handle_initial_delay();
588 while (!waitpid(child_pid, &status, WNOHANG)) {
589 nanosleep(&ts, NULL);
594 if (WIFSIGNALED(status))
595 psignal(WTERMSIG(status), argv[0]);
597 handle_initial_delay();
599 nanosleep(&ts, NULL);
607 update_stats(&walltime_nsecs_stats, t1 - t0);
609 if (aggr_mode == AGGR_GLOBAL) {
610 list_for_each_entry(counter, &evsel_list->entries, node) {
611 read_counter_aggr(counter);
612 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
613 thread_map__nr(evsel_list->threads));
616 list_for_each_entry(counter, &evsel_list->entries, node) {
617 read_counter(counter);
618 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
622 return WEXITSTATUS(status);
625 static int run_perf_stat(int argc __maybe_unused, const char **argv)
630 ret = system(pre_cmd);
638 ret = __run_perf_stat(argc, argv);
643 ret = system(post_cmd);
651 static void print_noise_pct(double total, double avg)
653 double pct = rel_stddev_stats(total, avg);
656 fprintf(output, "%s%.2f%%", csv_sep, pct);
658 fprintf(output, " ( +-%6.2f%% )", pct);
661 static void print_noise(struct perf_evsel *evsel, double avg)
663 struct perf_stat *ps;
669 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
672 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
676 fprintf(output, "S%d-C%*d%s%*d%s",
677 cpu_map__id_to_socket(id),
679 cpu_map__id_to_cpu(id),
686 fprintf(output, "S%*d%s%*d%s",
695 fprintf(output, "CPU%*d%s",
697 perf_evsel__cpus(evsel)->map[id], csv_sep);
705 static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
707 double msecs = avg / 1e6;
708 const char *fmt = csv_output ? "%.6f%s%s" : "%18.6f%s%-25s";
710 aggr_printout(evsel, cpu, nr);
712 fprintf(output, fmt, msecs, csv_sep, perf_evsel__name(evsel));
715 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
717 if (csv_output || interval)
720 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
721 fprintf(output, " # %8.3f CPUs utilized ",
722 avg / avg_stats(&walltime_nsecs_stats));
724 fprintf(output, " ");
727 /* used for get_ratio_color() */
729 GRC_STALLED_CYCLES_FE,
730 GRC_STALLED_CYCLES_BE,
735 static const char *get_ratio_color(enum grc_type type, double ratio)
737 static const double grc_table[GRC_MAX_NR][3] = {
738 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
739 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
740 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
742 const char *color = PERF_COLOR_NORMAL;
744 if (ratio > grc_table[type][0])
745 color = PERF_COLOR_RED;
746 else if (ratio > grc_table[type][1])
747 color = PERF_COLOR_MAGENTA;
748 else if (ratio > grc_table[type][2])
749 color = PERF_COLOR_YELLOW;
754 static void print_stalled_cycles_frontend(int cpu,
755 struct perf_evsel *evsel
756 __maybe_unused, double avg)
758 double total, ratio = 0.0;
761 total = avg_stats(&runtime_cycles_stats[cpu]);
764 ratio = avg / total * 100.0;
766 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
768 fprintf(output, " # ");
769 color_fprintf(output, color, "%6.2f%%", ratio);
770 fprintf(output, " frontend cycles idle ");
773 static void print_stalled_cycles_backend(int cpu,
774 struct perf_evsel *evsel
775 __maybe_unused, double avg)
777 double total, ratio = 0.0;
780 total = avg_stats(&runtime_cycles_stats[cpu]);
783 ratio = avg / total * 100.0;
785 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
787 fprintf(output, " # ");
788 color_fprintf(output, color, "%6.2f%%", ratio);
789 fprintf(output, " backend cycles idle ");
792 static void print_branch_misses(int cpu,
793 struct perf_evsel *evsel __maybe_unused,
796 double total, ratio = 0.0;
799 total = avg_stats(&runtime_branches_stats[cpu]);
802 ratio = avg / total * 100.0;
804 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
806 fprintf(output, " # ");
807 color_fprintf(output, color, "%6.2f%%", ratio);
808 fprintf(output, " of all branches ");
811 static void print_l1_dcache_misses(int cpu,
812 struct perf_evsel *evsel __maybe_unused,
815 double total, ratio = 0.0;
818 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
821 ratio = avg / total * 100.0;
823 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
825 fprintf(output, " # ");
826 color_fprintf(output, color, "%6.2f%%", ratio);
827 fprintf(output, " of all L1-dcache hits ");
830 static void print_l1_icache_misses(int cpu,
831 struct perf_evsel *evsel __maybe_unused,
834 double total, ratio = 0.0;
837 total = avg_stats(&runtime_l1_icache_stats[cpu]);
840 ratio = avg / total * 100.0;
842 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
844 fprintf(output, " # ");
845 color_fprintf(output, color, "%6.2f%%", ratio);
846 fprintf(output, " of all L1-icache hits ");
849 static void print_dtlb_cache_misses(int cpu,
850 struct perf_evsel *evsel __maybe_unused,
853 double total, ratio = 0.0;
856 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
859 ratio = avg / total * 100.0;
861 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
863 fprintf(output, " # ");
864 color_fprintf(output, color, "%6.2f%%", ratio);
865 fprintf(output, " of all dTLB cache hits ");
868 static void print_itlb_cache_misses(int cpu,
869 struct perf_evsel *evsel __maybe_unused,
872 double total, ratio = 0.0;
875 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
878 ratio = avg / total * 100.0;
880 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
882 fprintf(output, " # ");
883 color_fprintf(output, color, "%6.2f%%", ratio);
884 fprintf(output, " of all iTLB cache hits ");
887 static void print_ll_cache_misses(int cpu,
888 struct perf_evsel *evsel __maybe_unused,
891 double total, ratio = 0.0;
894 total = avg_stats(&runtime_ll_cache_stats[cpu]);
897 ratio = avg / total * 100.0;
899 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
901 fprintf(output, " # ");
902 color_fprintf(output, color, "%6.2f%%", ratio);
903 fprintf(output, " of all LL-cache hits ");
906 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
908 double total, ratio = 0.0, total2;
914 fmt = "%'18.0f%s%-25s";
916 fmt = "%18.0f%s%-25s";
918 aggr_printout(evsel, cpu, nr);
920 if (aggr_mode == AGGR_GLOBAL)
923 fprintf(output, fmt, avg, csv_sep, perf_evsel__name(evsel));
926 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
928 if (csv_output || interval)
931 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
932 total = avg_stats(&runtime_cycles_stats[cpu]);
935 fprintf(output, " # %5.2f insns per cycle ", ratio);
937 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
938 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
942 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
945 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
946 runtime_branches_stats[cpu].n != 0) {
947 print_branch_misses(cpu, evsel, avg);
949 evsel->attr.type == PERF_TYPE_HW_CACHE &&
950 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
951 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
952 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
953 runtime_l1_dcache_stats[cpu].n != 0) {
954 print_l1_dcache_misses(cpu, evsel, avg);
956 evsel->attr.type == PERF_TYPE_HW_CACHE &&
957 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
958 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
959 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
960 runtime_l1_icache_stats[cpu].n != 0) {
961 print_l1_icache_misses(cpu, evsel, avg);
963 evsel->attr.type == PERF_TYPE_HW_CACHE &&
964 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
965 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
966 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
967 runtime_dtlb_cache_stats[cpu].n != 0) {
968 print_dtlb_cache_misses(cpu, evsel, avg);
970 evsel->attr.type == PERF_TYPE_HW_CACHE &&
971 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
972 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
973 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
974 runtime_itlb_cache_stats[cpu].n != 0) {
975 print_itlb_cache_misses(cpu, evsel, avg);
977 evsel->attr.type == PERF_TYPE_HW_CACHE &&
978 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
979 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
980 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
981 runtime_ll_cache_stats[cpu].n != 0) {
982 print_ll_cache_misses(cpu, evsel, avg);
983 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
984 runtime_cacherefs_stats[cpu].n != 0) {
985 total = avg_stats(&runtime_cacherefs_stats[cpu]);
988 ratio = avg * 100 / total;
990 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
992 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
993 print_stalled_cycles_frontend(cpu, evsel, avg);
994 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
995 print_stalled_cycles_backend(cpu, evsel, avg);
996 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
997 total = avg_stats(&runtime_nsecs_stats[cpu]);
1000 ratio = avg / total;
1001 fprintf(output, " # %8.3f GHz ", ratio);
1003 } else if (transaction_run &&
1004 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1005 total = avg_stats(&runtime_cycles_stats[cpu]);
1008 " # %5.2f%% transactional cycles ",
1009 100.0 * (avg / total));
1010 } else if (transaction_run &&
1011 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1012 total = avg_stats(&runtime_cycles_stats[cpu]);
1013 total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1018 " # %5.2f%% aborted cycles ",
1019 100.0 * ((total2-avg) / total));
1020 } else if (transaction_run &&
1021 perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1023 runtime_cycles_in_tx_stats[cpu].n != 0) {
1024 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1027 ratio = total / avg;
1029 fprintf(output, " # %8.0f cycles / transaction ", ratio);
1030 } else if (transaction_run &&
1031 perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1033 runtime_cycles_in_tx_stats[cpu].n != 0) {
1034 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1037 ratio = total / avg;
1039 fprintf(output, " # %8.0f cycles / elision ", ratio);
1040 } else if (runtime_nsecs_stats[cpu].n != 0) {
1043 total = avg_stats(&runtime_nsecs_stats[cpu]);
1046 ratio = 1000.0 * avg / total;
1047 if (ratio < 0.001) {
1052 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
1054 fprintf(output, " ");
1058 static void print_aggr(char *prefix)
1060 struct perf_evsel *counter;
1061 int cpu, cpu2, s, s2, id, nr;
1064 if (!(aggr_map || aggr_get_id))
1067 for (s = 0; s < aggr_map->nr; s++) {
1068 id = aggr_map->map[s];
1069 list_for_each_entry(counter, &evsel_list->entries, node) {
1070 val = ena = run = 0;
1072 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1073 cpu2 = perf_evsel__cpus(counter)->map[cpu];
1074 s2 = aggr_get_id(evsel_list->cpus, cpu2);
1077 val += counter->counts->cpu[cpu].val;
1078 ena += counter->counts->cpu[cpu].ena;
1079 run += counter->counts->cpu[cpu].run;
1083 fprintf(output, "%s", prefix);
1085 if (run == 0 || ena == 0) {
1086 aggr_printout(counter, id, nr);
1088 fprintf(output, "%*s%s%*s",
1089 csv_output ? 0 : 18,
1090 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1092 csv_output ? 0 : -24,
1093 perf_evsel__name(counter));
1096 fprintf(output, "%s%s",
1097 csv_sep, counter->cgrp->name);
1099 fputc('\n', output);
1103 if (nsec_counter(counter))
1104 nsec_printout(id, nr, counter, val);
1106 abs_printout(id, nr, counter, val);
1109 print_noise(counter, 1.0);
1112 fprintf(output, " (%.2f%%)",
1115 fputc('\n', output);
1121 * Print out the results of a single counter:
1122 * aggregated counts in system-wide mode
1124 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1126 struct perf_stat *ps = counter->priv;
1127 double avg = avg_stats(&ps->res_stats[0]);
1128 int scaled = counter->counts->scaled;
1131 fprintf(output, "%s", prefix);
1134 fprintf(output, "%*s%s%*s",
1135 csv_output ? 0 : 18,
1136 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1138 csv_output ? 0 : -24,
1139 perf_evsel__name(counter));
1142 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1144 fputc('\n', output);
1148 if (nsec_counter(counter))
1149 nsec_printout(-1, 0, counter, avg);
1151 abs_printout(-1, 0, counter, avg);
1153 print_noise(counter, avg);
1156 fputc('\n', output);
1161 double avg_enabled, avg_running;
1163 avg_enabled = avg_stats(&ps->res_stats[1]);
1164 avg_running = avg_stats(&ps->res_stats[2]);
1166 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1168 fprintf(output, "\n");
1172 * Print out the results of a single counter:
1173 * does not use aggregated count in system-wide
1175 static void print_counter(struct perf_evsel *counter, char *prefix)
1180 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1181 val = counter->counts->cpu[cpu].val;
1182 ena = counter->counts->cpu[cpu].ena;
1183 run = counter->counts->cpu[cpu].run;
1186 fprintf(output, "%s", prefix);
1188 if (run == 0 || ena == 0) {
1189 fprintf(output, "CPU%*d%s%*s%s%*s",
1190 csv_output ? 0 : -4,
1191 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1192 csv_output ? 0 : 18,
1193 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1195 csv_output ? 0 : -24,
1196 perf_evsel__name(counter));
1199 fprintf(output, "%s%s",
1200 csv_sep, counter->cgrp->name);
1202 fputc('\n', output);
1206 if (nsec_counter(counter))
1207 nsec_printout(cpu, 0, counter, val);
1209 abs_printout(cpu, 0, counter, val);
1212 print_noise(counter, 1.0);
1215 fprintf(output, " (%.2f%%)",
1218 fputc('\n', output);
1222 static void print_stat(int argc, const char **argv)
1224 struct perf_evsel *counter;
1230 fprintf(output, "\n");
1231 fprintf(output, " Performance counter stats for ");
1232 if (target.system_wide)
1233 fprintf(output, "\'system wide");
1234 else if (target.cpu_list)
1235 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1236 else if (!perf_target__has_task(&target)) {
1237 fprintf(output, "\'%s", argv[0]);
1238 for (i = 1; i < argc; i++)
1239 fprintf(output, " %s", argv[i]);
1240 } else if (target.pid)
1241 fprintf(output, "process id \'%s", target.pid);
1243 fprintf(output, "thread id \'%s", target.tid);
1245 fprintf(output, "\'");
1247 fprintf(output, " (%d runs)", run_count);
1248 fprintf(output, ":\n\n");
1251 switch (aggr_mode) {
1257 list_for_each_entry(counter, &evsel_list->entries, node)
1258 print_counter_aggr(counter, NULL);
1261 list_for_each_entry(counter, &evsel_list->entries, node)
1262 print_counter(counter, NULL);
1270 fprintf(output, "\n");
1271 fprintf(output, " %17.9f seconds time elapsed",
1272 avg_stats(&walltime_nsecs_stats)/1e9);
1273 if (run_count > 1) {
1274 fprintf(output, " ");
1275 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1276 avg_stats(&walltime_nsecs_stats));
1278 fprintf(output, "\n\n");
1282 static volatile int signr = -1;
1284 static void skip_signal(int signo)
1286 if ((child_pid == -1) || interval)
1291 * render child_pid harmless
1292 * won't send SIGTERM to a random
1293 * process in case of race condition
1294 * and fast PID recycling
1299 static void sig_atexit(void)
1304 * avoid race condition with SIGCHLD handler
1305 * in skip_signal() which is modifying child_pid
1306 * goal is to avoid send SIGTERM to a random
1310 sigaddset(&set, SIGCHLD);
1311 sigprocmask(SIG_BLOCK, &set, &oset);
1313 if (child_pid != -1)
1314 kill(child_pid, SIGTERM);
1316 sigprocmask(SIG_SETMASK, &oset, NULL);
1321 signal(signr, SIG_DFL);
1322 kill(getpid(), signr);
1325 static int stat__set_big_num(const struct option *opt __maybe_unused,
1326 const char *s __maybe_unused, int unset)
1328 big_num_opt = unset ? 0 : 1;
1332 static int perf_stat_init_aggr_mode(void)
1334 switch (aggr_mode) {
1336 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1337 perror("cannot build socket map");
1340 aggr_get_id = cpu_map__get_socket;
1343 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1344 perror("cannot build core map");
1347 aggr_get_id = cpu_map__get_core;
1357 static int setup_events(const char * const *attrs, unsigned len)
1361 for (i = 0; i < len; i++) {
1362 if (parse_events(evsel_list, attrs[i]))
1369 * Add default attributes, if there were no attributes specified or
1370 * if -d/--detailed, -d -d or -d -d -d is used:
1372 static int add_default_attributes(void)
1374 struct perf_event_attr default_attrs[] = {
1376 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1377 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1378 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1379 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1381 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1382 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1383 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1384 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1385 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1386 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1391 * Detailed stats (-d), covering the L1 and last level data caches:
1393 struct perf_event_attr detailed_attrs[] = {
1395 { .type = PERF_TYPE_HW_CACHE,
1397 PERF_COUNT_HW_CACHE_L1D << 0 |
1398 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1399 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1401 { .type = PERF_TYPE_HW_CACHE,
1403 PERF_COUNT_HW_CACHE_L1D << 0 |
1404 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1405 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1407 { .type = PERF_TYPE_HW_CACHE,
1409 PERF_COUNT_HW_CACHE_LL << 0 |
1410 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1411 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1413 { .type = PERF_TYPE_HW_CACHE,
1415 PERF_COUNT_HW_CACHE_LL << 0 |
1416 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1417 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1421 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1423 struct perf_event_attr very_detailed_attrs[] = {
1425 { .type = PERF_TYPE_HW_CACHE,
1427 PERF_COUNT_HW_CACHE_L1I << 0 |
1428 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1429 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1431 { .type = PERF_TYPE_HW_CACHE,
1433 PERF_COUNT_HW_CACHE_L1I << 0 |
1434 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1435 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1437 { .type = PERF_TYPE_HW_CACHE,
1439 PERF_COUNT_HW_CACHE_DTLB << 0 |
1440 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1441 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1443 { .type = PERF_TYPE_HW_CACHE,
1445 PERF_COUNT_HW_CACHE_DTLB << 0 |
1446 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1447 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1449 { .type = PERF_TYPE_HW_CACHE,
1451 PERF_COUNT_HW_CACHE_ITLB << 0 |
1452 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1453 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1455 { .type = PERF_TYPE_HW_CACHE,
1457 PERF_COUNT_HW_CACHE_ITLB << 0 |
1458 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1459 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1464 * Very, very detailed stats (-d -d -d), adding prefetch events:
1466 struct perf_event_attr very_very_detailed_attrs[] = {
1468 { .type = PERF_TYPE_HW_CACHE,
1470 PERF_COUNT_HW_CACHE_L1D << 0 |
1471 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1472 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1474 { .type = PERF_TYPE_HW_CACHE,
1476 PERF_COUNT_HW_CACHE_L1D << 0 |
1477 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1478 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1481 /* Set attrs if no event is selected and !null_run: */
1485 if (transaction_run) {
1487 if (pmu_have_event("cpu", "cycles-ct") &&
1488 pmu_have_event("cpu", "el-start"))
1489 err = setup_events(transaction_attrs,
1490 ARRAY_SIZE(transaction_attrs));
1492 err = setup_events(transaction_limited_attrs,
1493 ARRAY_SIZE(transaction_limited_attrs));
1495 fprintf(stderr, "Cannot set up transaction events\n");
1501 if (!evsel_list->nr_entries) {
1502 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1506 /* Detailed events get appended to the event list: */
1508 if (detailed_run < 1)
1511 /* Append detailed run extra attributes: */
1512 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1515 if (detailed_run < 2)
1518 /* Append very detailed run extra attributes: */
1519 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1522 if (detailed_run < 3)
1525 /* Append very, very detailed run extra attributes: */
1526 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1529 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1531 bool append_file = false;
1533 const char *output_name = NULL;
1534 const struct option options[] = {
1535 OPT_BOOLEAN('T', "transaction", &transaction_run,
1536 "hardware transaction statistics"),
1537 OPT_CALLBACK('e', "event", &evsel_list, "event",
1538 "event selector. use 'perf list' to list available events",
1539 parse_events_option),
1540 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1541 "event filter", parse_filter),
1542 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1543 "child tasks do not inherit counters"),
1544 OPT_STRING('p', "pid", &target.pid, "pid",
1545 "stat events on existing process id"),
1546 OPT_STRING('t', "tid", &target.tid, "tid",
1547 "stat events on existing thread id"),
1548 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1549 "system-wide collection from all CPUs"),
1550 OPT_BOOLEAN('g', "group", &group,
1551 "put the counters into a counter group"),
1552 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1553 OPT_INCR('v', "verbose", &verbose,
1554 "be more verbose (show counter open errors, etc)"),
1555 OPT_INTEGER('r', "repeat", &run_count,
1556 "repeat command and print average + stddev (max: 100, forever: 0)"),
1557 OPT_BOOLEAN('n', "null", &null_run,
1558 "null run - dont start any counters"),
1559 OPT_INCR('d', "detailed", &detailed_run,
1560 "detailed run - start a lot of events"),
1561 OPT_BOOLEAN('S', "sync", &sync_run,
1562 "call sync() before starting a run"),
1563 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1564 "print large numbers with thousands\' separators",
1566 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1567 "list of cpus to monitor in system-wide"),
1568 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1569 "disable CPU count aggregation", AGGR_NONE),
1570 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1571 "print counts with custom separator"),
1572 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1573 "monitor event in cgroup name only", parse_cgroups),
1574 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1575 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1576 OPT_INTEGER(0, "log-fd", &output_fd,
1577 "log output to fd, instead of stderr"),
1578 OPT_STRING(0, "pre", &pre_cmd, "command",
1579 "command to run prior to the measured command"),
1580 OPT_STRING(0, "post", &post_cmd, "command",
1581 "command to run after to the measured command"),
1582 OPT_UINTEGER('I', "interval-print", &interval,
1583 "print counts at regular interval in ms (>= 100)"),
1584 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1585 "aggregate counts per processor socket", AGGR_SOCKET),
1586 OPT_SET_UINT(0, "per-core", &aggr_mode,
1587 "aggregate counts per physical processor core", AGGR_CORE),
1588 OPT_UINTEGER('D', "delay", &initial_delay,
1589 "ms to wait before starting measurement after program start"),
1592 const char * const stat_usage[] = {
1593 "perf stat [<options>] [<command>]",
1596 int status = -ENOMEM, run_idx;
1599 setlocale(LC_ALL, "");
1601 evsel_list = perf_evlist__new();
1602 if (evsel_list == NULL)
1605 argc = parse_options(argc, argv, options, stat_usage,
1606 PARSE_OPT_STOP_AT_NON_OPTION);
1609 if (output_name && strcmp(output_name, "-"))
1612 if (output_name && output_fd) {
1613 fprintf(stderr, "cannot use both --output and --log-fd\n");
1614 usage_with_options(stat_usage, options);
1617 if (output_fd < 0) {
1618 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1619 usage_with_options(stat_usage, options);
1624 mode = append_file ? "a" : "w";
1626 output = fopen(output_name, mode);
1628 perror("failed to create output file");
1631 clock_gettime(CLOCK_REALTIME, &tm);
1632 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1633 } else if (output_fd > 0) {
1634 mode = append_file ? "a" : "w";
1635 output = fdopen(output_fd, mode);
1637 perror("Failed opening logfd");
1644 if (!strcmp(csv_sep, "\\t"))
1647 csv_sep = DEFAULT_SEPARATOR;
1650 * let the spreadsheet do the pretty-printing
1653 /* User explicitly passed -B? */
1654 if (big_num_opt == 1) {
1655 fprintf(stderr, "-B option not supported with -x\n");
1656 usage_with_options(stat_usage, options);
1657 } else /* Nope, so disable big number formatting */
1659 } else if (big_num_opt == 0) /* User passed --no-big-num */
1662 if (!argc && perf_target__none(&target))
1663 usage_with_options(stat_usage, options);
1665 if (run_count < 0) {
1666 usage_with_options(stat_usage, options);
1667 } else if (run_count == 0) {
1672 /* no_aggr, cgroup are for system-wide only */
1673 if ((aggr_mode != AGGR_GLOBAL || nr_cgroups)
1674 && !perf_target__has_cpu(&target)) {
1675 fprintf(stderr, "both cgroup and no-aggregation "
1676 "modes only available in system-wide mode\n");
1678 usage_with_options(stat_usage, options);
1682 if (add_default_attributes())
1685 perf_target__validate(&target);
1687 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1688 if (perf_target__has_task(&target))
1689 pr_err("Problems finding threads of monitor\n");
1690 if (perf_target__has_cpu(&target))
1691 perror("failed to parse CPUs map");
1693 usage_with_options(stat_usage, options);
1696 if (interval && interval < 100) {
1697 pr_err("print interval must be >= 100ms\n");
1698 usage_with_options(stat_usage, options);
1702 if (perf_evlist__alloc_stats(evsel_list, interval))
1705 if (perf_stat_init_aggr_mode())
1709 * We dont want to block the signals - that would cause
1710 * child tasks to inherit that and Ctrl-C would not work.
1711 * What we want is for Ctrl-C to work in the exec()-ed
1712 * task, but being ignored by perf stat itself:
1716 signal(SIGINT, skip_signal);
1717 signal(SIGCHLD, skip_signal);
1718 signal(SIGALRM, skip_signal);
1719 signal(SIGABRT, skip_signal);
1722 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1723 if (run_count != 1 && verbose)
1724 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1727 status = run_perf_stat(argc, argv);
1728 if (forever && status != -1) {
1729 print_stat(argc, argv);
1730 perf_stat__reset_stats(evsel_list);
1734 if (!forever && status != -1 && !interval)
1735 print_stat(argc, argv);
1737 perf_evlist__free_stats(evsel_list);
1739 perf_evlist__delete_maps(evsel_list);
1741 perf_evlist__delete(evsel_list);