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"
49 #include "util/event.h"
50 #include "util/evlist.h"
51 #include "util/evsel.h"
52 #include "util/debug.h"
53 #include "util/color.h"
54 #include "util/stat.h"
55 #include "util/header.h"
56 #include "util/cpumap.h"
57 #include "util/thread.h"
58 #include "util/thread_map.h"
61 #include <sys/prctl.h>
64 #define DEFAULT_SEPARATOR " "
65 #define CNTR_NOT_SUPPORTED "<not supported>"
66 #define CNTR_NOT_COUNTED "<not counted>"
68 static void print_stat(int argc, const char **argv);
69 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
70 static void print_counter(struct perf_evsel *counter, char *prefix);
71 static void print_aggr(char *prefix);
73 static struct perf_evlist *evsel_list;
75 static struct perf_target target = {
86 static int run_count = 1;
87 static bool no_inherit = false;
88 static bool scale = true;
89 static enum aggr_mode aggr_mode = AGGR_GLOBAL;
90 static volatile pid_t child_pid = -1;
91 static bool null_run = false;
92 static int detailed_run = 0;
93 static bool big_num = true;
94 static int big_num_opt = -1;
95 static const char *csv_sep = NULL;
96 static bool csv_output = false;
97 static bool group = false;
98 static FILE *output = NULL;
99 static const char *pre_cmd = NULL;
100 static const char *post_cmd = NULL;
101 static bool sync_run = false;
102 static unsigned int interval = 0;
103 static unsigned int initial_delay = 0;
104 static bool forever = false;
105 static struct timespec ref_time;
106 static struct cpu_map *aggr_map;
107 static int (*aggr_get_id)(struct cpu_map *m, int cpu);
109 static volatile int done = 0;
112 struct stats res_stats[3];
115 static inline void diff_timespec(struct timespec *r, struct timespec *a,
118 r->tv_sec = a->tv_sec - b->tv_sec;
119 if (a->tv_nsec < b->tv_nsec) {
120 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
123 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
127 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
129 return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
132 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
134 return perf_evsel__cpus(evsel)->nr;
137 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
139 memset(evsel->priv, 0, sizeof(struct perf_stat));
142 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
144 evsel->priv = zalloc(sizeof(struct perf_stat));
145 return evsel->priv == NULL ? -ENOMEM : 0;
148 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
154 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
159 sz = sizeof(*evsel->counts) +
160 (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
166 evsel->prev_raw_counts = addr;
171 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
173 free(evsel->prev_raw_counts);
174 evsel->prev_raw_counts = NULL;
177 static void perf_evlist__free_stats(struct perf_evlist *evlist)
179 struct perf_evsel *evsel;
181 list_for_each_entry(evsel, &evlist->entries, node) {
182 perf_evsel__free_stat_priv(evsel);
183 perf_evsel__free_counts(evsel);
184 perf_evsel__free_prev_raw_counts(evsel);
188 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
190 struct perf_evsel *evsel;
192 list_for_each_entry(evsel, &evlist->entries, node) {
193 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
194 perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
195 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
202 perf_evlist__free_stats(evlist);
206 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
207 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
208 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
209 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
210 static struct stats runtime_branches_stats[MAX_NR_CPUS];
211 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
212 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
213 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
214 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
215 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
216 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
217 static struct stats walltime_nsecs_stats;
219 static void perf_stat__reset_stats(struct perf_evlist *evlist)
221 struct perf_evsel *evsel;
223 list_for_each_entry(evsel, &evlist->entries, node) {
224 perf_evsel__reset_stat_priv(evsel);
225 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
228 memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
229 memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
230 memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
231 memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
232 memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
233 memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
234 memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
235 memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
236 memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
237 memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
238 memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
239 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
242 static int create_perf_stat_counter(struct perf_evsel *evsel)
244 struct perf_event_attr *attr = &evsel->attr;
247 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
248 PERF_FORMAT_TOTAL_TIME_RUNNING;
250 attr->inherit = !no_inherit;
252 if (perf_target__has_cpu(&target))
253 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
255 if (!perf_target__has_task(&target) &&
256 perf_evsel__is_group_leader(evsel)) {
259 attr->enable_on_exec = 1;
262 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
266 * Does the counter have nsecs as a unit?
268 static inline int nsec_counter(struct perf_evsel *evsel)
270 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
271 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
278 * Update various tracking values we maintain to print
279 * more semantic information such as miss/hit ratios,
280 * instruction rates, etc:
282 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
284 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
285 update_stats(&runtime_nsecs_stats[0], count[0]);
286 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
287 update_stats(&runtime_cycles_stats[0], count[0]);
288 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
289 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
290 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
291 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
292 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
293 update_stats(&runtime_branches_stats[0], count[0]);
294 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
295 update_stats(&runtime_cacherefs_stats[0], count[0]);
296 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
297 update_stats(&runtime_l1_dcache_stats[0], count[0]);
298 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
299 update_stats(&runtime_l1_icache_stats[0], count[0]);
300 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
301 update_stats(&runtime_ll_cache_stats[0], count[0]);
302 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
303 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
304 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
305 update_stats(&runtime_itlb_cache_stats[0], count[0]);
309 * Read out the results of a single counter:
310 * aggregate counts across CPUs in system-wide mode
312 static int read_counter_aggr(struct perf_evsel *counter)
314 struct perf_stat *ps = counter->priv;
315 u64 *count = counter->counts->aggr.values;
318 if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
319 thread_map__nr(evsel_list->threads), scale) < 0)
322 for (i = 0; i < 3; i++)
323 update_stats(&ps->res_stats[i], count[i]);
326 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
327 perf_evsel__name(counter), count[0], count[1], count[2]);
331 * Save the full runtime - to allow normalization during printout:
333 update_shadow_stats(counter, count);
339 * Read out the results of a single counter:
340 * do not aggregate counts across CPUs in system-wide mode
342 static int read_counter(struct perf_evsel *counter)
347 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
348 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
351 count = counter->counts->cpu[cpu].values;
353 update_shadow_stats(counter, count);
359 static void print_interval(void)
361 static int num_print_interval;
362 struct perf_evsel *counter;
363 struct perf_stat *ps;
364 struct timespec ts, rs;
367 if (aggr_mode == AGGR_GLOBAL) {
368 list_for_each_entry(counter, &evsel_list->entries, node) {
370 memset(ps->res_stats, 0, sizeof(ps->res_stats));
371 read_counter_aggr(counter);
374 list_for_each_entry(counter, &evsel_list->entries, node) {
376 memset(ps->res_stats, 0, sizeof(ps->res_stats));
377 read_counter(counter);
381 clock_gettime(CLOCK_MONOTONIC, &ts);
382 diff_timespec(&rs, &ts, &ref_time);
383 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
385 if (num_print_interval == 0 && !csv_output) {
388 fprintf(output, "# time socket cpus counts events\n");
391 fprintf(output, "# time core cpus counts events\n");
394 fprintf(output, "# time CPU counts events\n");
398 fprintf(output, "# time counts events\n");
402 if (++num_print_interval == 25)
403 num_print_interval = 0;
411 list_for_each_entry(counter, &evsel_list->entries, node)
412 print_counter(counter, prefix);
416 list_for_each_entry(counter, &evsel_list->entries, node)
417 print_counter_aggr(counter, prefix);
423 static void handle_initial_delay(void)
425 struct perf_evsel *counter;
428 const int ncpus = cpu_map__nr(evsel_list->cpus),
429 nthreads = thread_map__nr(evsel_list->threads);
431 usleep(initial_delay * 1000);
432 list_for_each_entry(counter, &evsel_list->entries, node)
433 perf_evsel__enable(counter, ncpus, nthreads);
437 static int __run_perf_stat(int argc, const char **argv)
440 unsigned long long t0, t1;
441 struct perf_evsel *counter;
444 const bool forks = (argc > 0);
447 ts.tv_sec = interval / 1000;
448 ts.tv_nsec = (interval % 1000) * 1000000;
455 if (perf_evlist__prepare_workload(evsel_list, &target, argv,
457 perror("failed to prepare workload");
463 perf_evlist__set_leader(evsel_list);
465 list_for_each_entry(counter, &evsel_list->entries, node) {
466 if (create_perf_stat_counter(counter) < 0) {
468 * PPC returns ENXIO for HW counters until 2.6.37
469 * (behavior changed with commit b0a873e).
471 if (errno == EINVAL || errno == ENOSYS ||
472 errno == ENOENT || errno == EOPNOTSUPP ||
475 ui__warning("%s event is not supported by the kernel.\n",
476 perf_evsel__name(counter));
477 counter->supported = false;
481 perf_evsel__open_strerror(counter, &target,
482 errno, msg, sizeof(msg));
483 ui__error("%s\n", msg);
486 kill(child_pid, SIGTERM);
490 counter->supported = true;
493 if (perf_evlist__apply_filters(evsel_list)) {
494 error("failed to set filter with %d (%s)\n", errno,
500 * Enable counters and exec the command:
503 clock_gettime(CLOCK_MONOTONIC, &ref_time);
506 perf_evlist__start_workload(evsel_list);
507 handle_initial_delay();
510 while (!waitpid(child_pid, &status, WNOHANG)) {
511 nanosleep(&ts, NULL);
516 if (WIFSIGNALED(status))
517 psignal(WTERMSIG(status), argv[0]);
519 handle_initial_delay();
521 nanosleep(&ts, NULL);
529 update_stats(&walltime_nsecs_stats, t1 - t0);
531 if (aggr_mode == AGGR_GLOBAL) {
532 list_for_each_entry(counter, &evsel_list->entries, node) {
533 read_counter_aggr(counter);
534 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
535 thread_map__nr(evsel_list->threads));
538 list_for_each_entry(counter, &evsel_list->entries, node) {
539 read_counter(counter);
540 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
544 return WEXITSTATUS(status);
547 static int run_perf_stat(int argc __maybe_unused, const char **argv)
552 ret = system(pre_cmd);
560 ret = __run_perf_stat(argc, argv);
565 ret = system(post_cmd);
573 static void print_noise_pct(double total, double avg)
575 double pct = rel_stddev_stats(total, avg);
578 fprintf(output, "%s%.2f%%", csv_sep, pct);
580 fprintf(output, " ( +-%6.2f%% )", pct);
583 static void print_noise(struct perf_evsel *evsel, double avg)
585 struct perf_stat *ps;
591 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
594 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
598 fprintf(output, "S%d-C%*d%s%*d%s",
599 cpu_map__id_to_socket(id),
601 cpu_map__id_to_cpu(id),
608 fprintf(output, "S%*d%s%*d%s",
617 fprintf(output, "CPU%*d%s",
619 perf_evsel__cpus(evsel)->map[id], csv_sep);
627 static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
629 double msecs = avg / 1e6;
630 const char *fmt = csv_output ? "%.6f%s%s" : "%18.6f%s%-25s";
632 aggr_printout(evsel, cpu, nr);
634 fprintf(output, fmt, msecs, csv_sep, perf_evsel__name(evsel));
637 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
639 if (csv_output || interval)
642 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
643 fprintf(output, " # %8.3f CPUs utilized ",
644 avg / avg_stats(&walltime_nsecs_stats));
646 fprintf(output, " ");
649 /* used for get_ratio_color() */
651 GRC_STALLED_CYCLES_FE,
652 GRC_STALLED_CYCLES_BE,
657 static const char *get_ratio_color(enum grc_type type, double ratio)
659 static const double grc_table[GRC_MAX_NR][3] = {
660 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
661 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
662 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
664 const char *color = PERF_COLOR_NORMAL;
666 if (ratio > grc_table[type][0])
667 color = PERF_COLOR_RED;
668 else if (ratio > grc_table[type][1])
669 color = PERF_COLOR_MAGENTA;
670 else if (ratio > grc_table[type][2])
671 color = PERF_COLOR_YELLOW;
676 static void print_stalled_cycles_frontend(int cpu,
677 struct perf_evsel *evsel
678 __maybe_unused, double avg)
680 double total, ratio = 0.0;
683 total = avg_stats(&runtime_cycles_stats[cpu]);
686 ratio = avg / total * 100.0;
688 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
690 fprintf(output, " # ");
691 color_fprintf(output, color, "%6.2f%%", ratio);
692 fprintf(output, " frontend cycles idle ");
695 static void print_stalled_cycles_backend(int cpu,
696 struct perf_evsel *evsel
697 __maybe_unused, double avg)
699 double total, ratio = 0.0;
702 total = avg_stats(&runtime_cycles_stats[cpu]);
705 ratio = avg / total * 100.0;
707 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
709 fprintf(output, " # ");
710 color_fprintf(output, color, "%6.2f%%", ratio);
711 fprintf(output, " backend cycles idle ");
714 static void print_branch_misses(int cpu,
715 struct perf_evsel *evsel __maybe_unused,
718 double total, ratio = 0.0;
721 total = avg_stats(&runtime_branches_stats[cpu]);
724 ratio = avg / total * 100.0;
726 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
728 fprintf(output, " # ");
729 color_fprintf(output, color, "%6.2f%%", ratio);
730 fprintf(output, " of all branches ");
733 static void print_l1_dcache_misses(int cpu,
734 struct perf_evsel *evsel __maybe_unused,
737 double total, ratio = 0.0;
740 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
743 ratio = avg / total * 100.0;
745 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
747 fprintf(output, " # ");
748 color_fprintf(output, color, "%6.2f%%", ratio);
749 fprintf(output, " of all L1-dcache hits ");
752 static void print_l1_icache_misses(int cpu,
753 struct perf_evsel *evsel __maybe_unused,
756 double total, ratio = 0.0;
759 total = avg_stats(&runtime_l1_icache_stats[cpu]);
762 ratio = avg / total * 100.0;
764 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
766 fprintf(output, " # ");
767 color_fprintf(output, color, "%6.2f%%", ratio);
768 fprintf(output, " of all L1-icache hits ");
771 static void print_dtlb_cache_misses(int cpu,
772 struct perf_evsel *evsel __maybe_unused,
775 double total, ratio = 0.0;
778 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
781 ratio = avg / total * 100.0;
783 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
785 fprintf(output, " # ");
786 color_fprintf(output, color, "%6.2f%%", ratio);
787 fprintf(output, " of all dTLB cache hits ");
790 static void print_itlb_cache_misses(int cpu,
791 struct perf_evsel *evsel __maybe_unused,
794 double total, ratio = 0.0;
797 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
800 ratio = avg / total * 100.0;
802 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
804 fprintf(output, " # ");
805 color_fprintf(output, color, "%6.2f%%", ratio);
806 fprintf(output, " of all iTLB cache hits ");
809 static void print_ll_cache_misses(int cpu,
810 struct perf_evsel *evsel __maybe_unused,
813 double total, ratio = 0.0;
816 total = avg_stats(&runtime_ll_cache_stats[cpu]);
819 ratio = avg / total * 100.0;
821 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
823 fprintf(output, " # ");
824 color_fprintf(output, color, "%6.2f%%", ratio);
825 fprintf(output, " of all LL-cache hits ");
828 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
830 double total, ratio = 0.0;
836 fmt = "%'18.0f%s%-25s";
838 fmt = "%18.0f%s%-25s";
840 aggr_printout(evsel, cpu, nr);
842 if (aggr_mode == AGGR_GLOBAL)
845 fprintf(output, fmt, avg, csv_sep, perf_evsel__name(evsel));
848 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
850 if (csv_output || interval)
853 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
854 total = avg_stats(&runtime_cycles_stats[cpu]);
858 fprintf(output, " # %5.2f insns per cycle ", ratio);
860 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
861 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
865 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
868 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
869 runtime_branches_stats[cpu].n != 0) {
870 print_branch_misses(cpu, evsel, avg);
872 evsel->attr.type == PERF_TYPE_HW_CACHE &&
873 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
874 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
875 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
876 runtime_l1_dcache_stats[cpu].n != 0) {
877 print_l1_dcache_misses(cpu, evsel, avg);
879 evsel->attr.type == PERF_TYPE_HW_CACHE &&
880 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
881 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
882 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
883 runtime_l1_icache_stats[cpu].n != 0) {
884 print_l1_icache_misses(cpu, evsel, avg);
886 evsel->attr.type == PERF_TYPE_HW_CACHE &&
887 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
888 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
889 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
890 runtime_dtlb_cache_stats[cpu].n != 0) {
891 print_dtlb_cache_misses(cpu, evsel, avg);
893 evsel->attr.type == PERF_TYPE_HW_CACHE &&
894 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
895 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
896 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
897 runtime_itlb_cache_stats[cpu].n != 0) {
898 print_itlb_cache_misses(cpu, evsel, avg);
900 evsel->attr.type == PERF_TYPE_HW_CACHE &&
901 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
902 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
903 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
904 runtime_ll_cache_stats[cpu].n != 0) {
905 print_ll_cache_misses(cpu, evsel, avg);
906 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
907 runtime_cacherefs_stats[cpu].n != 0) {
908 total = avg_stats(&runtime_cacherefs_stats[cpu]);
911 ratio = avg * 100 / total;
913 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
915 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
916 print_stalled_cycles_frontend(cpu, evsel, avg);
917 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
918 print_stalled_cycles_backend(cpu, evsel, avg);
919 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
920 total = avg_stats(&runtime_nsecs_stats[cpu]);
923 ratio = 1.0 * avg / total;
925 fprintf(output, " # %8.3f GHz ", ratio);
926 } else if (runtime_nsecs_stats[cpu].n != 0) {
929 total = avg_stats(&runtime_nsecs_stats[cpu]);
932 ratio = 1000.0 * avg / total;
938 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
940 fprintf(output, " ");
944 static void print_aggr(char *prefix)
946 struct perf_evsel *counter;
947 int cpu, cpu2, s, s2, id, nr;
950 if (!(aggr_map || aggr_get_id))
953 for (s = 0; s < aggr_map->nr; s++) {
954 id = aggr_map->map[s];
955 list_for_each_entry(counter, &evsel_list->entries, node) {
958 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
959 cpu2 = perf_evsel__cpus(counter)->map[cpu];
960 s2 = aggr_get_id(evsel_list->cpus, cpu2);
963 val += counter->counts->cpu[cpu].val;
964 ena += counter->counts->cpu[cpu].ena;
965 run += counter->counts->cpu[cpu].run;
969 fprintf(output, "%s", prefix);
971 if (run == 0 || ena == 0) {
972 aggr_printout(counter, id, nr);
974 fprintf(output, "%*s%s%*s",
976 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
978 csv_output ? 0 : -24,
979 perf_evsel__name(counter));
982 fprintf(output, "%s%s",
983 csv_sep, counter->cgrp->name);
989 if (nsec_counter(counter))
990 nsec_printout(id, nr, counter, val);
992 abs_printout(id, nr, counter, val);
995 print_noise(counter, 1.0);
998 fprintf(output, " (%.2f%%)",
1001 fputc('\n', output);
1007 * Print out the results of a single counter:
1008 * aggregated counts in system-wide mode
1010 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1012 struct perf_stat *ps = counter->priv;
1013 double avg = avg_stats(&ps->res_stats[0]);
1014 int scaled = counter->counts->scaled;
1017 fprintf(output, "%s", prefix);
1020 fprintf(output, "%*s%s%*s",
1021 csv_output ? 0 : 18,
1022 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1024 csv_output ? 0 : -24,
1025 perf_evsel__name(counter));
1028 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1030 fputc('\n', output);
1034 if (nsec_counter(counter))
1035 nsec_printout(-1, 0, counter, avg);
1037 abs_printout(-1, 0, counter, avg);
1039 print_noise(counter, avg);
1042 fputc('\n', output);
1047 double avg_enabled, avg_running;
1049 avg_enabled = avg_stats(&ps->res_stats[1]);
1050 avg_running = avg_stats(&ps->res_stats[2]);
1052 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1054 fprintf(output, "\n");
1058 * Print out the results of a single counter:
1059 * does not use aggregated count in system-wide
1061 static void print_counter(struct perf_evsel *counter, char *prefix)
1066 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1067 val = counter->counts->cpu[cpu].val;
1068 ena = counter->counts->cpu[cpu].ena;
1069 run = counter->counts->cpu[cpu].run;
1072 fprintf(output, "%s", prefix);
1074 if (run == 0 || ena == 0) {
1075 fprintf(output, "CPU%*d%s%*s%s%*s",
1076 csv_output ? 0 : -4,
1077 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1078 csv_output ? 0 : 18,
1079 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1081 csv_output ? 0 : -24,
1082 perf_evsel__name(counter));
1085 fprintf(output, "%s%s",
1086 csv_sep, counter->cgrp->name);
1088 fputc('\n', output);
1092 if (nsec_counter(counter))
1093 nsec_printout(cpu, 0, counter, val);
1095 abs_printout(cpu, 0, counter, val);
1098 print_noise(counter, 1.0);
1101 fprintf(output, " (%.2f%%)",
1104 fputc('\n', output);
1108 static void print_stat(int argc, const char **argv)
1110 struct perf_evsel *counter;
1116 fprintf(output, "\n");
1117 fprintf(output, " Performance counter stats for ");
1118 if (!perf_target__has_task(&target)) {
1119 fprintf(output, "\'%s", argv[0]);
1120 for (i = 1; i < argc; i++)
1121 fprintf(output, " %s", argv[i]);
1122 } else if (target.pid)
1123 fprintf(output, "process id \'%s", target.pid);
1125 fprintf(output, "thread id \'%s", target.tid);
1127 fprintf(output, "\'");
1129 fprintf(output, " (%d runs)", run_count);
1130 fprintf(output, ":\n\n");
1133 switch (aggr_mode) {
1139 list_for_each_entry(counter, &evsel_list->entries, node)
1140 print_counter_aggr(counter, NULL);
1143 list_for_each_entry(counter, &evsel_list->entries, node)
1144 print_counter(counter, NULL);
1152 fprintf(output, "\n");
1153 fprintf(output, " %17.9f seconds time elapsed",
1154 avg_stats(&walltime_nsecs_stats)/1e9);
1155 if (run_count > 1) {
1156 fprintf(output, " ");
1157 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1158 avg_stats(&walltime_nsecs_stats));
1160 fprintf(output, "\n\n");
1164 static volatile int signr = -1;
1166 static void skip_signal(int signo)
1168 if ((child_pid == -1) || interval)
1173 * render child_pid harmless
1174 * won't send SIGTERM to a random
1175 * process in case of race condition
1176 * and fast PID recycling
1181 static void sig_atexit(void)
1186 * avoid race condition with SIGCHLD handler
1187 * in skip_signal() which is modifying child_pid
1188 * goal is to avoid send SIGTERM to a random
1192 sigaddset(&set, SIGCHLD);
1193 sigprocmask(SIG_BLOCK, &set, &oset);
1195 if (child_pid != -1)
1196 kill(child_pid, SIGTERM);
1198 sigprocmask(SIG_SETMASK, &oset, NULL);
1203 signal(signr, SIG_DFL);
1204 kill(getpid(), signr);
1207 static int stat__set_big_num(const struct option *opt __maybe_unused,
1208 const char *s __maybe_unused, int unset)
1210 big_num_opt = unset ? 0 : 1;
1214 static int perf_stat_init_aggr_mode(void)
1216 switch (aggr_mode) {
1218 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1219 perror("cannot build socket map");
1222 aggr_get_id = cpu_map__get_socket;
1225 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1226 perror("cannot build core map");
1229 aggr_get_id = cpu_map__get_core;
1241 * Add default attributes, if there were no attributes specified or
1242 * if -d/--detailed, -d -d or -d -d -d is used:
1244 static int add_default_attributes(void)
1246 struct perf_event_attr default_attrs[] = {
1248 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1249 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1250 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1251 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1253 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1254 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1255 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1256 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1257 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1258 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1263 * Detailed stats (-d), covering the L1 and last level data caches:
1265 struct perf_event_attr detailed_attrs[] = {
1267 { .type = PERF_TYPE_HW_CACHE,
1269 PERF_COUNT_HW_CACHE_L1D << 0 |
1270 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1271 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1273 { .type = PERF_TYPE_HW_CACHE,
1275 PERF_COUNT_HW_CACHE_L1D << 0 |
1276 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1277 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1279 { .type = PERF_TYPE_HW_CACHE,
1281 PERF_COUNT_HW_CACHE_LL << 0 |
1282 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1283 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1285 { .type = PERF_TYPE_HW_CACHE,
1287 PERF_COUNT_HW_CACHE_LL << 0 |
1288 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1289 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1293 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1295 struct perf_event_attr very_detailed_attrs[] = {
1297 { .type = PERF_TYPE_HW_CACHE,
1299 PERF_COUNT_HW_CACHE_L1I << 0 |
1300 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1301 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1303 { .type = PERF_TYPE_HW_CACHE,
1305 PERF_COUNT_HW_CACHE_L1I << 0 |
1306 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1307 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1309 { .type = PERF_TYPE_HW_CACHE,
1311 PERF_COUNT_HW_CACHE_DTLB << 0 |
1312 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1313 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1315 { .type = PERF_TYPE_HW_CACHE,
1317 PERF_COUNT_HW_CACHE_DTLB << 0 |
1318 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1319 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1321 { .type = PERF_TYPE_HW_CACHE,
1323 PERF_COUNT_HW_CACHE_ITLB << 0 |
1324 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1325 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1327 { .type = PERF_TYPE_HW_CACHE,
1329 PERF_COUNT_HW_CACHE_ITLB << 0 |
1330 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1331 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1336 * Very, very detailed stats (-d -d -d), adding prefetch events:
1338 struct perf_event_attr very_very_detailed_attrs[] = {
1340 { .type = PERF_TYPE_HW_CACHE,
1342 PERF_COUNT_HW_CACHE_L1D << 0 |
1343 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1344 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1346 { .type = PERF_TYPE_HW_CACHE,
1348 PERF_COUNT_HW_CACHE_L1D << 0 |
1349 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1350 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1353 /* Set attrs if no event is selected and !null_run: */
1357 if (!evsel_list->nr_entries) {
1358 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1362 /* Detailed events get appended to the event list: */
1364 if (detailed_run < 1)
1367 /* Append detailed run extra attributes: */
1368 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1371 if (detailed_run < 2)
1374 /* Append very detailed run extra attributes: */
1375 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1378 if (detailed_run < 3)
1381 /* Append very, very detailed run extra attributes: */
1382 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1385 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1387 bool append_file = false;
1389 const char *output_name = NULL;
1390 const struct option options[] = {
1391 OPT_CALLBACK('e', "event", &evsel_list, "event",
1392 "event selector. use 'perf list' to list available events",
1393 parse_events_option),
1394 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1395 "event filter", parse_filter),
1396 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1397 "child tasks do not inherit counters"),
1398 OPT_STRING('p', "pid", &target.pid, "pid",
1399 "stat events on existing process id"),
1400 OPT_STRING('t', "tid", &target.tid, "tid",
1401 "stat events on existing thread id"),
1402 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1403 "system-wide collection from all CPUs"),
1404 OPT_BOOLEAN('g', "group", &group,
1405 "put the counters into a counter group"),
1406 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1407 OPT_INCR('v', "verbose", &verbose,
1408 "be more verbose (show counter open errors, etc)"),
1409 OPT_INTEGER('r', "repeat", &run_count,
1410 "repeat command and print average + stddev (max: 100, forever: 0)"),
1411 OPT_BOOLEAN('n', "null", &null_run,
1412 "null run - dont start any counters"),
1413 OPT_INCR('d', "detailed", &detailed_run,
1414 "detailed run - start a lot of events"),
1415 OPT_BOOLEAN('S', "sync", &sync_run,
1416 "call sync() before starting a run"),
1417 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1418 "print large numbers with thousands\' separators",
1420 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1421 "list of cpus to monitor in system-wide"),
1422 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1423 "disable CPU count aggregation", AGGR_NONE),
1424 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1425 "print counts with custom separator"),
1426 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1427 "monitor event in cgroup name only", parse_cgroups),
1428 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1429 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1430 OPT_INTEGER(0, "log-fd", &output_fd,
1431 "log output to fd, instead of stderr"),
1432 OPT_STRING(0, "pre", &pre_cmd, "command",
1433 "command to run prior to the measured command"),
1434 OPT_STRING(0, "post", &post_cmd, "command",
1435 "command to run after to the measured command"),
1436 OPT_UINTEGER('I', "interval-print", &interval,
1437 "print counts at regular interval in ms (>= 100)"),
1438 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1439 "aggregate counts per processor socket", AGGR_SOCKET),
1440 OPT_SET_UINT(0, "per-core", &aggr_mode,
1441 "aggregate counts per physical processor core", AGGR_CORE),
1442 OPT_UINTEGER('D', "delay", &initial_delay,
1443 "ms to wait before starting measurement after program start"),
1446 const char * const stat_usage[] = {
1447 "perf stat [<options>] [<command>]",
1450 int status = -ENOMEM, run_idx;
1453 setlocale(LC_ALL, "");
1455 evsel_list = perf_evlist__new();
1456 if (evsel_list == NULL)
1459 argc = parse_options(argc, argv, options, stat_usage,
1460 PARSE_OPT_STOP_AT_NON_OPTION);
1463 if (output_name && strcmp(output_name, "-"))
1466 if (output_name && output_fd) {
1467 fprintf(stderr, "cannot use both --output and --log-fd\n");
1468 usage_with_options(stat_usage, options);
1471 if (output_fd < 0) {
1472 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1473 usage_with_options(stat_usage, options);
1478 mode = append_file ? "a" : "w";
1480 output = fopen(output_name, mode);
1482 perror("failed to create output file");
1485 clock_gettime(CLOCK_REALTIME, &tm);
1486 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1487 } else if (output_fd > 0) {
1488 mode = append_file ? "a" : "w";
1489 output = fdopen(output_fd, mode);
1491 perror("Failed opening logfd");
1498 if (!strcmp(csv_sep, "\\t"))
1501 csv_sep = DEFAULT_SEPARATOR;
1504 * let the spreadsheet do the pretty-printing
1507 /* User explicitly passed -B? */
1508 if (big_num_opt == 1) {
1509 fprintf(stderr, "-B option not supported with -x\n");
1510 usage_with_options(stat_usage, options);
1511 } else /* Nope, so disable big number formatting */
1513 } else if (big_num_opt == 0) /* User passed --no-big-num */
1516 if (!argc && !perf_target__has_task(&target))
1517 usage_with_options(stat_usage, options);
1518 if (run_count < 0) {
1519 usage_with_options(stat_usage, options);
1520 } else if (run_count == 0) {
1525 /* no_aggr, cgroup are for system-wide only */
1526 if ((aggr_mode != AGGR_GLOBAL || nr_cgroups)
1527 && !perf_target__has_cpu(&target)) {
1528 fprintf(stderr, "both cgroup and no-aggregation "
1529 "modes only available in system-wide mode\n");
1531 usage_with_options(stat_usage, options);
1535 if (add_default_attributes())
1538 perf_target__validate(&target);
1540 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1541 if (perf_target__has_task(&target))
1542 pr_err("Problems finding threads of monitor\n");
1543 if (perf_target__has_cpu(&target))
1544 perror("failed to parse CPUs map");
1546 usage_with_options(stat_usage, options);
1549 if (interval && interval < 100) {
1550 pr_err("print interval must be >= 100ms\n");
1551 usage_with_options(stat_usage, options);
1555 if (perf_evlist__alloc_stats(evsel_list, interval))
1558 if (perf_stat_init_aggr_mode())
1562 * We dont want to block the signals - that would cause
1563 * child tasks to inherit that and Ctrl-C would not work.
1564 * What we want is for Ctrl-C to work in the exec()-ed
1565 * task, but being ignored by perf stat itself:
1569 signal(SIGINT, skip_signal);
1570 signal(SIGCHLD, skip_signal);
1571 signal(SIGALRM, skip_signal);
1572 signal(SIGABRT, skip_signal);
1575 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1576 if (run_count != 1 && verbose)
1577 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1580 status = run_perf_stat(argc, argv);
1581 if (forever && status != -1) {
1582 print_stat(argc, argv);
1583 perf_stat__reset_stats(evsel_list);
1587 if (!forever && status != -1 && !interval)
1588 print_stat(argc, argv);
1590 perf_evlist__free_stats(evsel_list);
1592 perf_evlist__delete_maps(evsel_list);
1594 perf_evlist__delete(evsel_list);