4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ./hackbench 10
13 Performance counter stats for './hackbench 10':
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
27 0.154822978 seconds time elapsed
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 * Improvements and fixes by:
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
41 * Released under the GPL v2. (and only v2, not any later version)
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include "util/parse-options.h"
49 #include "util/parse-events.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/color.h"
56 #include "util/stat.h"
57 #include "util/header.h"
58 #include "util/cpumap.h"
59 #include "util/thread.h"
60 #include "util/thread_map.h"
63 #include <sys/prctl.h>
66 #define DEFAULT_SEPARATOR " "
67 #define CNTR_NOT_SUPPORTED "<not supported>"
68 #define CNTR_NOT_COUNTED "<not counted>"
70 static void print_stat(int argc, const char **argv);
71 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
72 static void print_counter(struct perf_evsel *counter, char *prefix);
73 static void print_aggr(char *prefix);
75 /* Default events used for perf stat -T */
76 static const char * const transaction_attrs[] = {
88 /* More limited version when the CPU does not have all events. */
89 static const char * const transaction_limited_attrs[] = {
99 /* must match transaction_attrs and the beginning limited_attrs */
110 static struct perf_evlist *evsel_list;
112 static struct target target = {
123 static int run_count = 1;
124 static bool no_inherit = false;
125 static bool scale = true;
126 static enum aggr_mode aggr_mode = AGGR_GLOBAL;
127 static volatile pid_t child_pid = -1;
128 static bool null_run = false;
129 static int detailed_run = 0;
130 static bool transaction_run;
131 static bool big_num = true;
132 static int big_num_opt = -1;
133 static const char *csv_sep = NULL;
134 static bool csv_output = false;
135 static bool group = false;
136 static FILE *output = NULL;
137 static const char *pre_cmd = NULL;
138 static const char *post_cmd = NULL;
139 static bool sync_run = false;
140 static unsigned int interval = 0;
141 static unsigned int initial_delay = 0;
142 static unsigned int unit_width = 4; /* strlen("unit") */
143 static bool forever = false;
144 static struct timespec ref_time;
145 static struct cpu_map *aggr_map;
146 static int (*aggr_get_id)(struct cpu_map *m, int cpu);
148 static volatile int done = 0;
151 struct stats res_stats[3];
154 static inline void diff_timespec(struct timespec *r, struct timespec *a,
157 r->tv_sec = a->tv_sec - b->tv_sec;
158 if (a->tv_nsec < b->tv_nsec) {
159 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
162 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
166 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
168 return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
171 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
173 return perf_evsel__cpus(evsel)->nr;
176 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
179 struct perf_stat *ps = evsel->priv;
181 for (i = 0; i < 3; i++)
182 init_stats(&ps->res_stats[i]);
185 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
187 evsel->priv = zalloc(sizeof(struct perf_stat));
188 if (evsel->priv == NULL)
190 perf_evsel__reset_stat_priv(evsel);
194 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
199 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
204 sz = sizeof(*evsel->counts) +
205 (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
211 evsel->prev_raw_counts = addr;
216 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
218 zfree(&evsel->prev_raw_counts);
221 static void perf_evlist__free_stats(struct perf_evlist *evlist)
223 struct perf_evsel *evsel;
225 evlist__for_each(evlist, evsel) {
226 perf_evsel__free_stat_priv(evsel);
227 perf_evsel__free_counts(evsel);
228 perf_evsel__free_prev_raw_counts(evsel);
232 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
234 struct perf_evsel *evsel;
236 evlist__for_each(evlist, evsel) {
237 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
238 perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
239 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
246 perf_evlist__free_stats(evlist);
250 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
251 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
252 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
253 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
254 static struct stats runtime_branches_stats[MAX_NR_CPUS];
255 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
256 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
257 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
258 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
259 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
260 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
261 static struct stats runtime_cycles_in_tx_stats[MAX_NR_CPUS];
262 static struct stats walltime_nsecs_stats;
263 static struct stats runtime_transaction_stats[MAX_NR_CPUS];
264 static struct stats runtime_elision_stats[MAX_NR_CPUS];
266 static void perf_stat__reset_stats(struct perf_evlist *evlist)
268 struct perf_evsel *evsel;
270 evlist__for_each(evlist, evsel) {
271 perf_evsel__reset_stat_priv(evsel);
272 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
275 memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
276 memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
277 memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
278 memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
279 memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
280 memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
281 memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
282 memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
283 memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
284 memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
285 memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
286 memset(runtime_cycles_in_tx_stats, 0,
287 sizeof(runtime_cycles_in_tx_stats));
288 memset(runtime_transaction_stats, 0,
289 sizeof(runtime_transaction_stats));
290 memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
291 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
294 static int create_perf_stat_counter(struct perf_evsel *evsel)
296 struct perf_event_attr *attr = &evsel->attr;
299 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
300 PERF_FORMAT_TOTAL_TIME_RUNNING;
302 attr->inherit = !no_inherit;
304 if (target__has_cpu(&target))
305 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
307 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
310 attr->enable_on_exec = 1;
313 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
317 * Does the counter have nsecs as a unit?
319 static inline int nsec_counter(struct perf_evsel *evsel)
321 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
322 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
328 static struct perf_evsel *nth_evsel(int n)
330 static struct perf_evsel **array;
331 static int array_len;
332 struct perf_evsel *ev;
335 /* Assumes this only called when evsel_list does not change anymore. */
337 evlist__for_each(evsel_list, ev)
339 array = malloc(array_len * sizeof(void *));
343 evlist__for_each(evsel_list, ev)
352 * Update various tracking values we maintain to print
353 * more semantic information such as miss/hit ratios,
354 * instruction rates, etc:
356 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
358 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
359 update_stats(&runtime_nsecs_stats[0], count[0]);
360 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
361 update_stats(&runtime_cycles_stats[0], count[0]);
362 else if (transaction_run &&
363 perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX)))
364 update_stats(&runtime_cycles_in_tx_stats[0], count[0]);
365 else if (transaction_run &&
366 perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START)))
367 update_stats(&runtime_transaction_stats[0], count[0]);
368 else if (transaction_run &&
369 perf_evsel__cmp(counter, nth_evsel(T_ELISION_START)))
370 update_stats(&runtime_elision_stats[0], count[0]);
371 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
372 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
373 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
374 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
375 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
376 update_stats(&runtime_branches_stats[0], count[0]);
377 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
378 update_stats(&runtime_cacherefs_stats[0], count[0]);
379 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
380 update_stats(&runtime_l1_dcache_stats[0], count[0]);
381 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
382 update_stats(&runtime_l1_icache_stats[0], count[0]);
383 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
384 update_stats(&runtime_ll_cache_stats[0], count[0]);
385 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
386 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
387 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
388 update_stats(&runtime_itlb_cache_stats[0], count[0]);
392 * Read out the results of a single counter:
393 * aggregate counts across CPUs in system-wide mode
395 static int read_counter_aggr(struct perf_evsel *counter)
397 struct perf_stat *ps = counter->priv;
398 u64 *count = counter->counts->aggr.values;
401 if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
402 thread_map__nr(evsel_list->threads), scale) < 0)
405 for (i = 0; i < 3; i++)
406 update_stats(&ps->res_stats[i], count[i]);
409 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
410 perf_evsel__name(counter), count[0], count[1], count[2]);
414 * Save the full runtime - to allow normalization during printout:
416 update_shadow_stats(counter, count);
422 * Read out the results of a single counter:
423 * do not aggregate counts across CPUs in system-wide mode
425 static int read_counter(struct perf_evsel *counter)
430 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
431 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
434 count = counter->counts->cpu[cpu].values;
436 update_shadow_stats(counter, count);
442 static void print_interval(void)
444 static int num_print_interval;
445 struct perf_evsel *counter;
446 struct perf_stat *ps;
447 struct timespec ts, rs;
450 if (aggr_mode == AGGR_GLOBAL) {
451 evlist__for_each(evsel_list, counter) {
453 memset(ps->res_stats, 0, sizeof(ps->res_stats));
454 read_counter_aggr(counter);
457 evlist__for_each(evsel_list, counter) {
459 memset(ps->res_stats, 0, sizeof(ps->res_stats));
460 read_counter(counter);
464 clock_gettime(CLOCK_MONOTONIC, &ts);
465 diff_timespec(&rs, &ts, &ref_time);
466 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
468 if (num_print_interval == 0 && !csv_output) {
471 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
474 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
477 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
481 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
485 if (++num_print_interval == 25)
486 num_print_interval = 0;
494 evlist__for_each(evsel_list, counter)
495 print_counter(counter, prefix);
499 evlist__for_each(evsel_list, counter)
500 print_counter_aggr(counter, prefix);
506 static void handle_initial_delay(void)
508 struct perf_evsel *counter;
511 const int ncpus = cpu_map__nr(evsel_list->cpus),
512 nthreads = thread_map__nr(evsel_list->threads);
514 usleep(initial_delay * 1000);
515 evlist__for_each(evsel_list, counter)
516 perf_evsel__enable(counter, ncpus, nthreads);
520 static volatile int workload_exec_errno;
523 * perf_evlist__prepare_workload will send a SIGUSR1
524 * if the fork fails, since we asked by setting its
525 * want_signal to true.
527 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
528 void *ucontext __maybe_unused)
530 workload_exec_errno = info->si_value.sival_int;
533 static int __run_perf_stat(int argc, const char **argv)
536 unsigned long long t0, t1;
537 struct perf_evsel *counter;
541 const bool forks = (argc > 0);
544 ts.tv_sec = interval / 1000;
545 ts.tv_nsec = (interval % 1000) * 1000000;
552 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
553 workload_exec_failed_signal) < 0) {
554 perror("failed to prepare workload");
557 child_pid = evsel_list->workload.pid;
561 perf_evlist__set_leader(evsel_list);
563 evlist__for_each(evsel_list, counter) {
564 if (create_perf_stat_counter(counter) < 0) {
566 * PPC returns ENXIO for HW counters until 2.6.37
567 * (behavior changed with commit b0a873e).
569 if (errno == EINVAL || errno == ENOSYS ||
570 errno == ENOENT || errno == EOPNOTSUPP ||
573 ui__warning("%s event is not supported by the kernel.\n",
574 perf_evsel__name(counter));
575 counter->supported = false;
579 perf_evsel__open_strerror(counter, &target,
580 errno, msg, sizeof(msg));
581 ui__error("%s\n", msg);
584 kill(child_pid, SIGTERM);
588 counter->supported = true;
590 l = strlen(counter->unit);
595 if (perf_evlist__apply_filters(evsel_list)) {
596 error("failed to set filter with %d (%s)\n", errno,
597 strerror_r(errno, msg, sizeof(msg)));
602 * Enable counters and exec the command:
605 clock_gettime(CLOCK_MONOTONIC, &ref_time);
608 perf_evlist__start_workload(evsel_list);
609 handle_initial_delay();
612 while (!waitpid(child_pid, &status, WNOHANG)) {
613 nanosleep(&ts, NULL);
619 if (workload_exec_errno) {
620 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
621 pr_err("Workload failed: %s\n", emsg);
625 if (WIFSIGNALED(status))
626 psignal(WTERMSIG(status), argv[0]);
628 handle_initial_delay();
630 nanosleep(&ts, NULL);
638 update_stats(&walltime_nsecs_stats, t1 - t0);
640 if (aggr_mode == AGGR_GLOBAL) {
641 evlist__for_each(evsel_list, counter) {
642 read_counter_aggr(counter);
643 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
644 thread_map__nr(evsel_list->threads));
647 evlist__for_each(evsel_list, counter) {
648 read_counter(counter);
649 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
653 return WEXITSTATUS(status);
656 static int run_perf_stat(int argc, const char **argv)
661 ret = system(pre_cmd);
669 ret = __run_perf_stat(argc, argv);
674 ret = system(post_cmd);
682 static void print_noise_pct(double total, double avg)
684 double pct = rel_stddev_stats(total, avg);
687 fprintf(output, "%s%.2f%%", csv_sep, pct);
689 fprintf(output, " ( +-%6.2f%% )", pct);
692 static void print_noise(struct perf_evsel *evsel, double avg)
694 struct perf_stat *ps;
700 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
703 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
707 fprintf(output, "S%d-C%*d%s%*d%s",
708 cpu_map__id_to_socket(id),
710 cpu_map__id_to_cpu(id),
717 fprintf(output, "S%*d%s%*d%s",
726 fprintf(output, "CPU%*d%s",
728 perf_evsel__cpus(evsel)->map[id], csv_sep);
736 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
738 double msecs = avg / 1e6;
739 const char *fmt_v, *fmt_n;
742 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
743 fmt_n = csv_output ? "%s" : "%-25s";
745 aggr_printout(evsel, id, nr);
747 scnprintf(name, sizeof(name), "%s%s",
748 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
750 fprintf(output, fmt_v, msecs, csv_sep);
753 fprintf(output, "%s%s", evsel->unit, csv_sep);
755 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
757 fprintf(output, fmt_n, name);
760 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
762 if (csv_output || interval)
765 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
766 fprintf(output, " # %8.3f CPUs utilized ",
767 avg / avg_stats(&walltime_nsecs_stats));
769 fprintf(output, " ");
772 /* used for get_ratio_color() */
774 GRC_STALLED_CYCLES_FE,
775 GRC_STALLED_CYCLES_BE,
780 static const char *get_ratio_color(enum grc_type type, double ratio)
782 static const double grc_table[GRC_MAX_NR][3] = {
783 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
784 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
785 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
787 const char *color = PERF_COLOR_NORMAL;
789 if (ratio > grc_table[type][0])
790 color = PERF_COLOR_RED;
791 else if (ratio > grc_table[type][1])
792 color = PERF_COLOR_MAGENTA;
793 else if (ratio > grc_table[type][2])
794 color = PERF_COLOR_YELLOW;
799 static void print_stalled_cycles_frontend(int cpu,
800 struct perf_evsel *evsel
801 __maybe_unused, double avg)
803 double total, ratio = 0.0;
806 total = avg_stats(&runtime_cycles_stats[cpu]);
809 ratio = avg / total * 100.0;
811 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
813 fprintf(output, " # ");
814 color_fprintf(output, color, "%6.2f%%", ratio);
815 fprintf(output, " frontend cycles idle ");
818 static void print_stalled_cycles_backend(int cpu,
819 struct perf_evsel *evsel
820 __maybe_unused, double avg)
822 double total, ratio = 0.0;
825 total = avg_stats(&runtime_cycles_stats[cpu]);
828 ratio = avg / total * 100.0;
830 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
832 fprintf(output, " # ");
833 color_fprintf(output, color, "%6.2f%%", ratio);
834 fprintf(output, " backend cycles idle ");
837 static void print_branch_misses(int cpu,
838 struct perf_evsel *evsel __maybe_unused,
841 double total, ratio = 0.0;
844 total = avg_stats(&runtime_branches_stats[cpu]);
847 ratio = avg / total * 100.0;
849 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
851 fprintf(output, " # ");
852 color_fprintf(output, color, "%6.2f%%", ratio);
853 fprintf(output, " of all branches ");
856 static void print_l1_dcache_misses(int cpu,
857 struct perf_evsel *evsel __maybe_unused,
860 double total, ratio = 0.0;
863 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
866 ratio = avg / total * 100.0;
868 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
870 fprintf(output, " # ");
871 color_fprintf(output, color, "%6.2f%%", ratio);
872 fprintf(output, " of all L1-dcache hits ");
875 static void print_l1_icache_misses(int cpu,
876 struct perf_evsel *evsel __maybe_unused,
879 double total, ratio = 0.0;
882 total = avg_stats(&runtime_l1_icache_stats[cpu]);
885 ratio = avg / total * 100.0;
887 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
889 fprintf(output, " # ");
890 color_fprintf(output, color, "%6.2f%%", ratio);
891 fprintf(output, " of all L1-icache hits ");
894 static void print_dtlb_cache_misses(int cpu,
895 struct perf_evsel *evsel __maybe_unused,
898 double total, ratio = 0.0;
901 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
904 ratio = avg / total * 100.0;
906 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
908 fprintf(output, " # ");
909 color_fprintf(output, color, "%6.2f%%", ratio);
910 fprintf(output, " of all dTLB cache hits ");
913 static void print_itlb_cache_misses(int cpu,
914 struct perf_evsel *evsel __maybe_unused,
917 double total, ratio = 0.0;
920 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
923 ratio = avg / total * 100.0;
925 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
927 fprintf(output, " # ");
928 color_fprintf(output, color, "%6.2f%%", ratio);
929 fprintf(output, " of all iTLB cache hits ");
932 static void print_ll_cache_misses(int cpu,
933 struct perf_evsel *evsel __maybe_unused,
936 double total, ratio = 0.0;
939 total = avg_stats(&runtime_ll_cache_stats[cpu]);
942 ratio = avg / total * 100.0;
944 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
946 fprintf(output, " # ");
947 color_fprintf(output, color, "%6.2f%%", ratio);
948 fprintf(output, " of all LL-cache hits ");
951 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
953 double total, ratio = 0.0, total2;
954 double sc = evsel->scale;
956 int cpu = cpu_map__id_to_cpu(id);
959 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
962 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
964 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
967 aggr_printout(evsel, id, nr);
969 if (aggr_mode == AGGR_GLOBAL)
972 fprintf(output, fmt, avg, csv_sep);
975 fprintf(output, "%-*s%s",
976 csv_output ? 0 : unit_width,
977 evsel->unit, csv_sep);
979 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
982 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
984 if (csv_output || interval)
987 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
988 total = avg_stats(&runtime_cycles_stats[cpu]);
991 fprintf(output, " # %5.2f insns per cycle ", ratio);
993 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
994 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
998 fprintf(output, "\n");
999 if (aggr_mode == AGGR_NONE)
1000 fprintf(output, " ");
1001 fprintf(output, " # %5.2f stalled cycles per insn", ratio);
1004 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
1005 runtime_branches_stats[cpu].n != 0) {
1006 print_branch_misses(cpu, evsel, avg);
1008 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1009 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
1010 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1011 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1012 runtime_l1_dcache_stats[cpu].n != 0) {
1013 print_l1_dcache_misses(cpu, evsel, avg);
1015 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1016 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
1017 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1018 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1019 runtime_l1_icache_stats[cpu].n != 0) {
1020 print_l1_icache_misses(cpu, evsel, avg);
1022 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1023 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
1024 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1025 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1026 runtime_dtlb_cache_stats[cpu].n != 0) {
1027 print_dtlb_cache_misses(cpu, evsel, avg);
1029 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1030 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
1031 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1032 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1033 runtime_itlb_cache_stats[cpu].n != 0) {
1034 print_itlb_cache_misses(cpu, evsel, avg);
1036 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1037 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
1038 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1039 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1040 runtime_ll_cache_stats[cpu].n != 0) {
1041 print_ll_cache_misses(cpu, evsel, avg);
1042 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
1043 runtime_cacherefs_stats[cpu].n != 0) {
1044 total = avg_stats(&runtime_cacherefs_stats[cpu]);
1047 ratio = avg * 100 / total;
1049 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
1051 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1052 print_stalled_cycles_frontend(cpu, evsel, avg);
1053 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1054 print_stalled_cycles_backend(cpu, evsel, avg);
1055 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1056 total = avg_stats(&runtime_nsecs_stats[cpu]);
1059 ratio = avg / total;
1060 fprintf(output, " # %8.3f GHz ", ratio);
1062 } else if (transaction_run &&
1063 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1064 total = avg_stats(&runtime_cycles_stats[cpu]);
1067 " # %5.2f%% transactional cycles ",
1068 100.0 * (avg / total));
1069 } else if (transaction_run &&
1070 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1071 total = avg_stats(&runtime_cycles_stats[cpu]);
1072 total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1077 " # %5.2f%% aborted cycles ",
1078 100.0 * ((total2-avg) / total));
1079 } else if (transaction_run &&
1080 perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1082 runtime_cycles_in_tx_stats[cpu].n != 0) {
1083 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1086 ratio = total / avg;
1088 fprintf(output, " # %8.0f cycles / transaction ", ratio);
1089 } else if (transaction_run &&
1090 perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1092 runtime_cycles_in_tx_stats[cpu].n != 0) {
1093 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1096 ratio = total / avg;
1098 fprintf(output, " # %8.0f cycles / elision ", ratio);
1099 } else if (runtime_nsecs_stats[cpu].n != 0) {
1102 total = avg_stats(&runtime_nsecs_stats[cpu]);
1105 ratio = 1000.0 * avg / total;
1106 if (ratio < 0.001) {
1111 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
1113 fprintf(output, " ");
1117 static void print_aggr(char *prefix)
1119 struct perf_evsel *counter;
1120 int cpu, cpu2, s, s2, id, nr;
1124 if (!(aggr_map || aggr_get_id))
1127 for (s = 0; s < aggr_map->nr; s++) {
1128 id = aggr_map->map[s];
1129 evlist__for_each(evsel_list, counter) {
1130 val = ena = run = 0;
1132 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1133 cpu2 = perf_evsel__cpus(counter)->map[cpu];
1134 s2 = aggr_get_id(evsel_list->cpus, cpu2);
1137 val += counter->counts->cpu[cpu].val;
1138 ena += counter->counts->cpu[cpu].ena;
1139 run += counter->counts->cpu[cpu].run;
1143 fprintf(output, "%s", prefix);
1145 if (run == 0 || ena == 0) {
1146 aggr_printout(counter, id, nr);
1148 fprintf(output, "%*s%s",
1149 csv_output ? 0 : 18,
1150 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1153 fprintf(output, "%-*s%s",
1154 csv_output ? 0 : unit_width,
1155 counter->unit, csv_sep);
1157 fprintf(output, "%*s",
1158 csv_output ? 0 : -25,
1159 perf_evsel__name(counter));
1162 fprintf(output, "%s%s",
1163 csv_sep, counter->cgrp->name);
1165 fputc('\n', output);
1168 uval = val * counter->scale;
1170 if (nsec_counter(counter))
1171 nsec_printout(id, nr, counter, uval);
1173 abs_printout(id, nr, counter, uval);
1176 print_noise(counter, 1.0);
1179 fprintf(output, " (%.2f%%)",
1182 fputc('\n', output);
1188 * Print out the results of a single counter:
1189 * aggregated counts in system-wide mode
1191 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1193 struct perf_stat *ps = counter->priv;
1194 double avg = avg_stats(&ps->res_stats[0]);
1195 int scaled = counter->counts->scaled;
1199 fprintf(output, "%s", prefix);
1202 fprintf(output, "%*s%s",
1203 csv_output ? 0 : 18,
1204 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1206 fprintf(output, "%-*s%s",
1207 csv_output ? 0 : unit_width,
1208 counter->unit, csv_sep);
1209 fprintf(output, "%*s",
1210 csv_output ? 0 : -25,
1211 perf_evsel__name(counter));
1214 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1216 fputc('\n', output);
1220 uval = avg * counter->scale;
1222 if (nsec_counter(counter))
1223 nsec_printout(-1, 0, counter, uval);
1225 abs_printout(-1, 0, counter, uval);
1227 print_noise(counter, avg);
1230 fputc('\n', output);
1235 double avg_enabled, avg_running;
1237 avg_enabled = avg_stats(&ps->res_stats[1]);
1238 avg_running = avg_stats(&ps->res_stats[2]);
1240 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1242 fprintf(output, "\n");
1246 * Print out the results of a single counter:
1247 * does not use aggregated count in system-wide
1249 static void print_counter(struct perf_evsel *counter, char *prefix)
1255 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1256 val = counter->counts->cpu[cpu].val;
1257 ena = counter->counts->cpu[cpu].ena;
1258 run = counter->counts->cpu[cpu].run;
1261 fprintf(output, "%s", prefix);
1263 if (run == 0 || ena == 0) {
1264 fprintf(output, "CPU%*d%s%*s%s",
1265 csv_output ? 0 : -4,
1266 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1267 csv_output ? 0 : 18,
1268 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1271 fprintf(output, "%-*s%s",
1272 csv_output ? 0 : unit_width,
1273 counter->unit, csv_sep);
1275 fprintf(output, "%*s",
1276 csv_output ? 0 : -25,
1277 perf_evsel__name(counter));
1280 fprintf(output, "%s%s",
1281 csv_sep, counter->cgrp->name);
1283 fputc('\n', output);
1287 uval = val * counter->scale;
1289 if (nsec_counter(counter))
1290 nsec_printout(cpu, 0, counter, uval);
1292 abs_printout(cpu, 0, counter, uval);
1295 print_noise(counter, 1.0);
1298 fprintf(output, " (%.2f%%)",
1301 fputc('\n', output);
1305 static void print_stat(int argc, const char **argv)
1307 struct perf_evsel *counter;
1313 fprintf(output, "\n");
1314 fprintf(output, " Performance counter stats for ");
1315 if (target.system_wide)
1316 fprintf(output, "\'system wide");
1317 else if (target.cpu_list)
1318 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1319 else if (!target__has_task(&target)) {
1320 fprintf(output, "\'%s", argv[0]);
1321 for (i = 1; i < argc; i++)
1322 fprintf(output, " %s", argv[i]);
1323 } else if (target.pid)
1324 fprintf(output, "process id \'%s", target.pid);
1326 fprintf(output, "thread id \'%s", target.tid);
1328 fprintf(output, "\'");
1330 fprintf(output, " (%d runs)", run_count);
1331 fprintf(output, ":\n\n");
1334 switch (aggr_mode) {
1340 evlist__for_each(evsel_list, counter)
1341 print_counter_aggr(counter, NULL);
1344 evlist__for_each(evsel_list, counter)
1345 print_counter(counter, NULL);
1353 fprintf(output, "\n");
1354 fprintf(output, " %17.9f seconds time elapsed",
1355 avg_stats(&walltime_nsecs_stats)/1e9);
1356 if (run_count > 1) {
1357 fprintf(output, " ");
1358 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1359 avg_stats(&walltime_nsecs_stats));
1361 fprintf(output, "\n\n");
1365 static volatile int signr = -1;
1367 static void skip_signal(int signo)
1369 if ((child_pid == -1) || interval)
1374 * render child_pid harmless
1375 * won't send SIGTERM to a random
1376 * process in case of race condition
1377 * and fast PID recycling
1382 static void sig_atexit(void)
1387 * avoid race condition with SIGCHLD handler
1388 * in skip_signal() which is modifying child_pid
1389 * goal is to avoid send SIGTERM to a random
1393 sigaddset(&set, SIGCHLD);
1394 sigprocmask(SIG_BLOCK, &set, &oset);
1396 if (child_pid != -1)
1397 kill(child_pid, SIGTERM);
1399 sigprocmask(SIG_SETMASK, &oset, NULL);
1404 signal(signr, SIG_DFL);
1405 kill(getpid(), signr);
1408 static int stat__set_big_num(const struct option *opt __maybe_unused,
1409 const char *s __maybe_unused, int unset)
1411 big_num_opt = unset ? 0 : 1;
1415 static int perf_stat_init_aggr_mode(void)
1417 switch (aggr_mode) {
1419 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1420 perror("cannot build socket map");
1423 aggr_get_id = cpu_map__get_socket;
1426 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1427 perror("cannot build core map");
1430 aggr_get_id = cpu_map__get_core;
1440 static int setup_events(const char * const *attrs, unsigned len)
1444 for (i = 0; i < len; i++) {
1445 if (parse_events(evsel_list, attrs[i]))
1452 * Add default attributes, if there were no attributes specified or
1453 * if -d/--detailed, -d -d or -d -d -d is used:
1455 static int add_default_attributes(void)
1457 struct perf_event_attr default_attrs[] = {
1459 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1460 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1461 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1462 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1464 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1465 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1466 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1467 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1468 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1469 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1474 * Detailed stats (-d), covering the L1 and last level data caches:
1476 struct perf_event_attr detailed_attrs[] = {
1478 { .type = PERF_TYPE_HW_CACHE,
1480 PERF_COUNT_HW_CACHE_L1D << 0 |
1481 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1482 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1484 { .type = PERF_TYPE_HW_CACHE,
1486 PERF_COUNT_HW_CACHE_L1D << 0 |
1487 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1488 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1490 { .type = PERF_TYPE_HW_CACHE,
1492 PERF_COUNT_HW_CACHE_LL << 0 |
1493 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1494 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1496 { .type = PERF_TYPE_HW_CACHE,
1498 PERF_COUNT_HW_CACHE_LL << 0 |
1499 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1500 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1504 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1506 struct perf_event_attr very_detailed_attrs[] = {
1508 { .type = PERF_TYPE_HW_CACHE,
1510 PERF_COUNT_HW_CACHE_L1I << 0 |
1511 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1512 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1514 { .type = PERF_TYPE_HW_CACHE,
1516 PERF_COUNT_HW_CACHE_L1I << 0 |
1517 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1518 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1520 { .type = PERF_TYPE_HW_CACHE,
1522 PERF_COUNT_HW_CACHE_DTLB << 0 |
1523 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1524 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1526 { .type = PERF_TYPE_HW_CACHE,
1528 PERF_COUNT_HW_CACHE_DTLB << 0 |
1529 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1530 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1532 { .type = PERF_TYPE_HW_CACHE,
1534 PERF_COUNT_HW_CACHE_ITLB << 0 |
1535 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1536 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1538 { .type = PERF_TYPE_HW_CACHE,
1540 PERF_COUNT_HW_CACHE_ITLB << 0 |
1541 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1542 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1547 * Very, very detailed stats (-d -d -d), adding prefetch events:
1549 struct perf_event_attr very_very_detailed_attrs[] = {
1551 { .type = PERF_TYPE_HW_CACHE,
1553 PERF_COUNT_HW_CACHE_L1D << 0 |
1554 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1555 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1557 { .type = PERF_TYPE_HW_CACHE,
1559 PERF_COUNT_HW_CACHE_L1D << 0 |
1560 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1561 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1564 /* Set attrs if no event is selected and !null_run: */
1568 if (transaction_run) {
1570 if (pmu_have_event("cpu", "cycles-ct") &&
1571 pmu_have_event("cpu", "el-start"))
1572 err = setup_events(transaction_attrs,
1573 ARRAY_SIZE(transaction_attrs));
1575 err = setup_events(transaction_limited_attrs,
1576 ARRAY_SIZE(transaction_limited_attrs));
1578 fprintf(stderr, "Cannot set up transaction events\n");
1584 if (!evsel_list->nr_entries) {
1585 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1589 /* Detailed events get appended to the event list: */
1591 if (detailed_run < 1)
1594 /* Append detailed run extra attributes: */
1595 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1598 if (detailed_run < 2)
1601 /* Append very detailed run extra attributes: */
1602 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1605 if (detailed_run < 3)
1608 /* Append very, very detailed run extra attributes: */
1609 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1612 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1614 bool append_file = false;
1616 const char *output_name = NULL;
1617 const struct option options[] = {
1618 OPT_BOOLEAN('T', "transaction", &transaction_run,
1619 "hardware transaction statistics"),
1620 OPT_CALLBACK('e', "event", &evsel_list, "event",
1621 "event selector. use 'perf list' to list available events",
1622 parse_events_option),
1623 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1624 "event filter", parse_filter),
1625 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1626 "child tasks do not inherit counters"),
1627 OPT_STRING('p', "pid", &target.pid, "pid",
1628 "stat events on existing process id"),
1629 OPT_STRING('t', "tid", &target.tid, "tid",
1630 "stat events on existing thread id"),
1631 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1632 "system-wide collection from all CPUs"),
1633 OPT_BOOLEAN('g', "group", &group,
1634 "put the counters into a counter group"),
1635 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1636 OPT_INCR('v', "verbose", &verbose,
1637 "be more verbose (show counter open errors, etc)"),
1638 OPT_INTEGER('r', "repeat", &run_count,
1639 "repeat command and print average + stddev (max: 100, forever: 0)"),
1640 OPT_BOOLEAN('n', "null", &null_run,
1641 "null run - dont start any counters"),
1642 OPT_INCR('d', "detailed", &detailed_run,
1643 "detailed run - start a lot of events"),
1644 OPT_BOOLEAN('S', "sync", &sync_run,
1645 "call sync() before starting a run"),
1646 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1647 "print large numbers with thousands\' separators",
1649 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1650 "list of cpus to monitor in system-wide"),
1651 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1652 "disable CPU count aggregation", AGGR_NONE),
1653 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1654 "print counts with custom separator"),
1655 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1656 "monitor event in cgroup name only", parse_cgroups),
1657 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1658 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1659 OPT_INTEGER(0, "log-fd", &output_fd,
1660 "log output to fd, instead of stderr"),
1661 OPT_STRING(0, "pre", &pre_cmd, "command",
1662 "command to run prior to the measured command"),
1663 OPT_STRING(0, "post", &post_cmd, "command",
1664 "command to run after to the measured command"),
1665 OPT_UINTEGER('I', "interval-print", &interval,
1666 "print counts at regular interval in ms (>= 100)"),
1667 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1668 "aggregate counts per processor socket", AGGR_SOCKET),
1669 OPT_SET_UINT(0, "per-core", &aggr_mode,
1670 "aggregate counts per physical processor core", AGGR_CORE),
1671 OPT_UINTEGER('D', "delay", &initial_delay,
1672 "ms to wait before starting measurement after program start"),
1675 const char * const stat_usage[] = {
1676 "perf stat [<options>] [<command>]",
1679 int status = -EINVAL, run_idx;
1682 setlocale(LC_ALL, "");
1684 evsel_list = perf_evlist__new();
1685 if (evsel_list == NULL)
1688 argc = parse_options(argc, argv, options, stat_usage,
1689 PARSE_OPT_STOP_AT_NON_OPTION);
1692 if (output_name && strcmp(output_name, "-"))
1695 if (output_name && output_fd) {
1696 fprintf(stderr, "cannot use both --output and --log-fd\n");
1697 parse_options_usage(stat_usage, options, "o", 1);
1698 parse_options_usage(NULL, options, "log-fd", 0);
1702 if (output_fd < 0) {
1703 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1704 parse_options_usage(stat_usage, options, "log-fd", 0);
1710 mode = append_file ? "a" : "w";
1712 output = fopen(output_name, mode);
1714 perror("failed to create output file");
1717 clock_gettime(CLOCK_REALTIME, &tm);
1718 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1719 } else if (output_fd > 0) {
1720 mode = append_file ? "a" : "w";
1721 output = fdopen(output_fd, mode);
1723 perror("Failed opening logfd");
1730 if (!strcmp(csv_sep, "\\t"))
1733 csv_sep = DEFAULT_SEPARATOR;
1736 * let the spreadsheet do the pretty-printing
1739 /* User explicitly passed -B? */
1740 if (big_num_opt == 1) {
1741 fprintf(stderr, "-B option not supported with -x\n");
1742 parse_options_usage(stat_usage, options, "B", 1);
1743 parse_options_usage(NULL, options, "x", 1);
1745 } else /* Nope, so disable big number formatting */
1747 } else if (big_num_opt == 0) /* User passed --no-big-num */
1750 if (!argc && target__none(&target))
1751 usage_with_options(stat_usage, options);
1753 if (run_count < 0) {
1754 pr_err("Run count must be a positive number\n");
1755 parse_options_usage(stat_usage, options, "r", 1);
1757 } else if (run_count == 0) {
1762 /* no_aggr, cgroup are for system-wide only */
1763 if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1764 !target__has_cpu(&target)) {
1765 fprintf(stderr, "both cgroup and no-aggregation "
1766 "modes only available in system-wide mode\n");
1768 parse_options_usage(stat_usage, options, "G", 1);
1769 parse_options_usage(NULL, options, "A", 1);
1770 parse_options_usage(NULL, options, "a", 1);
1774 if (add_default_attributes())
1777 target__validate(&target);
1779 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1780 if (target__has_task(&target)) {
1781 pr_err("Problems finding threads of monitor\n");
1782 parse_options_usage(stat_usage, options, "p", 1);
1783 parse_options_usage(NULL, options, "t", 1);
1784 } else if (target__has_cpu(&target)) {
1785 perror("failed to parse CPUs map");
1786 parse_options_usage(stat_usage, options, "C", 1);
1787 parse_options_usage(NULL, options, "a", 1);
1791 if (interval && interval < 100) {
1792 pr_err("print interval must be >= 100ms\n");
1793 parse_options_usage(stat_usage, options, "I", 1);
1797 if (perf_evlist__alloc_stats(evsel_list, interval))
1800 if (perf_stat_init_aggr_mode())
1804 * We dont want to block the signals - that would cause
1805 * child tasks to inherit that and Ctrl-C would not work.
1806 * What we want is for Ctrl-C to work in the exec()-ed
1807 * task, but being ignored by perf stat itself:
1811 signal(SIGINT, skip_signal);
1812 signal(SIGCHLD, skip_signal);
1813 signal(SIGALRM, skip_signal);
1814 signal(SIGABRT, skip_signal);
1817 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1818 if (run_count != 1 && verbose)
1819 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1822 status = run_perf_stat(argc, argv);
1823 if (forever && status != -1) {
1824 print_stat(argc, argv);
1825 perf_stat__reset_stats(evsel_list);
1829 if (!forever && status != -1 && !interval)
1830 print_stat(argc, argv);
1832 perf_evlist__free_stats(evsel_list);
1834 perf_evlist__delete(evsel_list);