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_socket(char *prefix);
73 static struct perf_evlist *evsel_list;
75 static struct perf_target target = {
79 static int run_count = 1;
80 static bool no_inherit = false;
81 static bool scale = true;
82 static bool no_aggr = false;
83 static bool aggr_socket = false;
84 static pid_t child_pid = -1;
85 static bool null_run = false;
86 static int detailed_run = 0;
87 static bool big_num = true;
88 static int big_num_opt = -1;
89 static const char *csv_sep = NULL;
90 static bool csv_output = false;
91 static bool group = false;
92 static FILE *output = NULL;
93 static const char *pre_cmd = NULL;
94 static const char *post_cmd = NULL;
95 static bool sync_run = false;
96 static unsigned int interval = 0;
97 static bool forever = false;
98 static struct timespec ref_time;
99 static struct cpu_map *sock_map;
101 static volatile int done = 0;
104 struct stats res_stats[3];
107 static inline void diff_timespec(struct timespec *r, struct timespec *a,
110 r->tv_sec = a->tv_sec - b->tv_sec;
111 if (a->tv_nsec < b->tv_nsec) {
112 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
115 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
119 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
121 return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
124 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
126 return perf_evsel__cpus(evsel)->nr;
129 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
131 memset(evsel->priv, 0, sizeof(struct perf_stat));
134 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
136 evsel->priv = zalloc(sizeof(struct perf_stat));
137 return evsel->priv == NULL ? -ENOMEM : 0;
140 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
146 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
151 sz = sizeof(*evsel->counts) +
152 (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
158 evsel->prev_raw_counts = addr;
163 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
165 free(evsel->prev_raw_counts);
166 evsel->prev_raw_counts = NULL;
169 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
170 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
171 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
172 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
173 static struct stats runtime_branches_stats[MAX_NR_CPUS];
174 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
175 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
176 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
177 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
178 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
179 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
180 static struct stats walltime_nsecs_stats;
182 static void reset_stats(void)
184 memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
185 memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
186 memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
187 memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
188 memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
189 memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
190 memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
191 memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
192 memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
193 memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
194 memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
195 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
198 static int create_perf_stat_counter(struct perf_evsel *evsel)
200 struct perf_event_attr *attr = &evsel->attr;
203 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
204 PERF_FORMAT_TOTAL_TIME_RUNNING;
206 attr->inherit = !no_inherit;
208 if (perf_target__has_cpu(&target))
209 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
211 if (!perf_target__has_task(&target) &&
212 perf_evsel__is_group_leader(evsel)) {
214 attr->enable_on_exec = 1;
217 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
221 * Does the counter have nsecs as a unit?
223 static inline int nsec_counter(struct perf_evsel *evsel)
225 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
226 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
233 * Update various tracking values we maintain to print
234 * more semantic information such as miss/hit ratios,
235 * instruction rates, etc:
237 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
239 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
240 update_stats(&runtime_nsecs_stats[0], count[0]);
241 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
242 update_stats(&runtime_cycles_stats[0], count[0]);
243 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
244 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
245 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
246 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
247 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
248 update_stats(&runtime_branches_stats[0], count[0]);
249 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
250 update_stats(&runtime_cacherefs_stats[0], count[0]);
251 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
252 update_stats(&runtime_l1_dcache_stats[0], count[0]);
253 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
254 update_stats(&runtime_l1_icache_stats[0], count[0]);
255 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
256 update_stats(&runtime_ll_cache_stats[0], count[0]);
257 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
258 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
259 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
260 update_stats(&runtime_itlb_cache_stats[0], count[0]);
264 * Read out the results of a single counter:
265 * aggregate counts across CPUs in system-wide mode
267 static int read_counter_aggr(struct perf_evsel *counter)
269 struct perf_stat *ps = counter->priv;
270 u64 *count = counter->counts->aggr.values;
273 if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
274 thread_map__nr(evsel_list->threads), scale) < 0)
277 for (i = 0; i < 3; i++)
278 update_stats(&ps->res_stats[i], count[i]);
281 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
282 perf_evsel__name(counter), count[0], count[1], count[2]);
286 * Save the full runtime - to allow normalization during printout:
288 update_shadow_stats(counter, count);
294 * Read out the results of a single counter:
295 * do not aggregate counts across CPUs in system-wide mode
297 static int read_counter(struct perf_evsel *counter)
302 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
303 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
306 count = counter->counts->cpu[cpu].values;
308 update_shadow_stats(counter, count);
314 static void print_interval(void)
316 static int num_print_interval;
317 struct perf_evsel *counter;
318 struct perf_stat *ps;
319 struct timespec ts, rs;
323 list_for_each_entry(counter, &evsel_list->entries, node) {
325 memset(ps->res_stats, 0, sizeof(ps->res_stats));
326 read_counter(counter);
329 list_for_each_entry(counter, &evsel_list->entries, node) {
331 memset(ps->res_stats, 0, sizeof(ps->res_stats));
332 read_counter_aggr(counter);
335 clock_gettime(CLOCK_MONOTONIC, &ts);
336 diff_timespec(&rs, &ts, &ref_time);
337 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
339 if (num_print_interval == 0 && !csv_output) {
341 fprintf(output, "# time socket cpus counts events\n");
343 fprintf(output, "# time CPU counts events\n");
345 fprintf(output, "# time counts events\n");
348 if (++num_print_interval == 25)
349 num_print_interval = 0;
352 print_aggr_socket(prefix);
354 list_for_each_entry(counter, &evsel_list->entries, node)
355 print_counter(counter, prefix);
357 list_for_each_entry(counter, &evsel_list->entries, node)
358 print_counter_aggr(counter, prefix);
362 static int __run_perf_stat(int argc, const char **argv)
365 unsigned long long t0, t1;
366 struct perf_evsel *counter;
369 const bool forks = (argc > 0);
372 ts.tv_sec = interval / 1000;
373 ts.tv_nsec = (interval % 1000) * 1000000;
380 && cpu_map__build_socket_map(evsel_list->cpus, &sock_map)) {
381 perror("cannot build socket map");
386 if (perf_evlist__prepare_workload(evsel_list, &target, argv,
388 perror("failed to prepare workload");
394 perf_evlist__set_leader(evsel_list);
396 list_for_each_entry(counter, &evsel_list->entries, node) {
397 if (create_perf_stat_counter(counter) < 0) {
399 * PPC returns ENXIO for HW counters until 2.6.37
400 * (behavior changed with commit b0a873e).
402 if (errno == EINVAL || errno == ENOSYS ||
403 errno == ENOENT || errno == EOPNOTSUPP ||
406 ui__warning("%s event is not supported by the kernel.\n",
407 perf_evsel__name(counter));
408 counter->supported = false;
412 perf_evsel__open_strerror(counter, &target,
413 errno, msg, sizeof(msg));
414 ui__error("%s\n", msg);
417 kill(child_pid, SIGTERM);
421 counter->supported = true;
424 if (perf_evlist__apply_filters(evsel_list)) {
425 error("failed to set filter with %d (%s)\n", errno,
431 * Enable counters and exec the command:
434 clock_gettime(CLOCK_MONOTONIC, &ref_time);
437 perf_evlist__start_workload(evsel_list);
440 while (!waitpid(child_pid, &status, WNOHANG)) {
441 nanosleep(&ts, NULL);
446 if (WIFSIGNALED(status))
447 psignal(WTERMSIG(status), argv[0]);
450 nanosleep(&ts, NULL);
458 update_stats(&walltime_nsecs_stats, t1 - t0);
461 list_for_each_entry(counter, &evsel_list->entries, node) {
462 read_counter(counter);
463 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
466 list_for_each_entry(counter, &evsel_list->entries, node) {
467 read_counter_aggr(counter);
468 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
469 thread_map__nr(evsel_list->threads));
473 return WEXITSTATUS(status);
476 static int run_perf_stat(int argc __maybe_unused, const char **argv)
481 ret = system(pre_cmd);
489 ret = __run_perf_stat(argc, argv);
494 ret = system(post_cmd);
502 static void print_noise_pct(double total, double avg)
504 double pct = rel_stddev_stats(total, avg);
507 fprintf(output, "%s%.2f%%", csv_sep, pct);
509 fprintf(output, " ( +-%6.2f%% )", pct);
512 static void print_noise(struct perf_evsel *evsel, double avg)
514 struct perf_stat *ps;
520 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
523 static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
525 double msecs = avg / 1e6;
526 char cpustr[16] = { '\0', };
527 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
530 sprintf(cpustr, "S%*d%s%*d%s",
538 sprintf(cpustr, "CPU%*d%s",
540 perf_evsel__cpus(evsel)->map[cpu], csv_sep);
542 fprintf(output, fmt, cpustr, msecs, csv_sep, perf_evsel__name(evsel));
545 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
547 if (csv_output || interval)
550 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
551 fprintf(output, " # %8.3f CPUs utilized ",
552 avg / avg_stats(&walltime_nsecs_stats));
554 fprintf(output, " ");
557 /* used for get_ratio_color() */
559 GRC_STALLED_CYCLES_FE,
560 GRC_STALLED_CYCLES_BE,
565 static const char *get_ratio_color(enum grc_type type, double ratio)
567 static const double grc_table[GRC_MAX_NR][3] = {
568 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
569 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
570 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
572 const char *color = PERF_COLOR_NORMAL;
574 if (ratio > grc_table[type][0])
575 color = PERF_COLOR_RED;
576 else if (ratio > grc_table[type][1])
577 color = PERF_COLOR_MAGENTA;
578 else if (ratio > grc_table[type][2])
579 color = PERF_COLOR_YELLOW;
584 static void print_stalled_cycles_frontend(int cpu,
585 struct perf_evsel *evsel
586 __maybe_unused, double avg)
588 double total, ratio = 0.0;
591 total = avg_stats(&runtime_cycles_stats[cpu]);
594 ratio = avg / total * 100.0;
596 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
598 fprintf(output, " # ");
599 color_fprintf(output, color, "%6.2f%%", ratio);
600 fprintf(output, " frontend cycles idle ");
603 static void print_stalled_cycles_backend(int cpu,
604 struct perf_evsel *evsel
605 __maybe_unused, double avg)
607 double total, ratio = 0.0;
610 total = avg_stats(&runtime_cycles_stats[cpu]);
613 ratio = avg / total * 100.0;
615 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
617 fprintf(output, " # ");
618 color_fprintf(output, color, "%6.2f%%", ratio);
619 fprintf(output, " backend cycles idle ");
622 static void print_branch_misses(int cpu,
623 struct perf_evsel *evsel __maybe_unused,
626 double total, ratio = 0.0;
629 total = avg_stats(&runtime_branches_stats[cpu]);
632 ratio = avg / total * 100.0;
634 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
636 fprintf(output, " # ");
637 color_fprintf(output, color, "%6.2f%%", ratio);
638 fprintf(output, " of all branches ");
641 static void print_l1_dcache_misses(int cpu,
642 struct perf_evsel *evsel __maybe_unused,
645 double total, ratio = 0.0;
648 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
651 ratio = avg / total * 100.0;
653 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
655 fprintf(output, " # ");
656 color_fprintf(output, color, "%6.2f%%", ratio);
657 fprintf(output, " of all L1-dcache hits ");
660 static void print_l1_icache_misses(int cpu,
661 struct perf_evsel *evsel __maybe_unused,
664 double total, ratio = 0.0;
667 total = avg_stats(&runtime_l1_icache_stats[cpu]);
670 ratio = avg / total * 100.0;
672 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
674 fprintf(output, " # ");
675 color_fprintf(output, color, "%6.2f%%", ratio);
676 fprintf(output, " of all L1-icache hits ");
679 static void print_dtlb_cache_misses(int cpu,
680 struct perf_evsel *evsel __maybe_unused,
683 double total, ratio = 0.0;
686 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
689 ratio = avg / total * 100.0;
691 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
693 fprintf(output, " # ");
694 color_fprintf(output, color, "%6.2f%%", ratio);
695 fprintf(output, " of all dTLB cache hits ");
698 static void print_itlb_cache_misses(int cpu,
699 struct perf_evsel *evsel __maybe_unused,
702 double total, ratio = 0.0;
705 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
708 ratio = avg / total * 100.0;
710 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
712 fprintf(output, " # ");
713 color_fprintf(output, color, "%6.2f%%", ratio);
714 fprintf(output, " of all iTLB cache hits ");
717 static void print_ll_cache_misses(int cpu,
718 struct perf_evsel *evsel __maybe_unused,
721 double total, ratio = 0.0;
724 total = avg_stats(&runtime_ll_cache_stats[cpu]);
727 ratio = avg / total * 100.0;
729 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
731 fprintf(output, " # ");
732 color_fprintf(output, color, "%6.2f%%", ratio);
733 fprintf(output, " of all LL-cache hits ");
736 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
738 double total, ratio = 0.0;
739 char cpustr[16] = { '\0', };
745 fmt = "%s%'18.0f%s%-25s";
747 fmt = "%s%18.0f%s%-25s";
750 sprintf(cpustr, "S%*d%s%*d%s",
758 sprintf(cpustr, "CPU%*d%s",
760 perf_evsel__cpus(evsel)->map[cpu], csv_sep);
764 fprintf(output, fmt, cpustr, avg, csv_sep, perf_evsel__name(evsel));
767 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
769 if (csv_output || interval)
772 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
773 total = avg_stats(&runtime_cycles_stats[cpu]);
777 fprintf(output, " # %5.2f insns per cycle ", ratio);
779 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
780 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
784 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
787 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
788 runtime_branches_stats[cpu].n != 0) {
789 print_branch_misses(cpu, evsel, avg);
791 evsel->attr.type == PERF_TYPE_HW_CACHE &&
792 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
793 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
794 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
795 runtime_l1_dcache_stats[cpu].n != 0) {
796 print_l1_dcache_misses(cpu, evsel, avg);
798 evsel->attr.type == PERF_TYPE_HW_CACHE &&
799 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
800 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
801 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
802 runtime_l1_icache_stats[cpu].n != 0) {
803 print_l1_icache_misses(cpu, evsel, avg);
805 evsel->attr.type == PERF_TYPE_HW_CACHE &&
806 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
807 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
808 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
809 runtime_dtlb_cache_stats[cpu].n != 0) {
810 print_dtlb_cache_misses(cpu, evsel, avg);
812 evsel->attr.type == PERF_TYPE_HW_CACHE &&
813 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
814 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
815 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
816 runtime_itlb_cache_stats[cpu].n != 0) {
817 print_itlb_cache_misses(cpu, evsel, avg);
819 evsel->attr.type == PERF_TYPE_HW_CACHE &&
820 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
821 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
822 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
823 runtime_ll_cache_stats[cpu].n != 0) {
824 print_ll_cache_misses(cpu, evsel, avg);
825 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
826 runtime_cacherefs_stats[cpu].n != 0) {
827 total = avg_stats(&runtime_cacherefs_stats[cpu]);
830 ratio = avg * 100 / total;
832 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
834 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
835 print_stalled_cycles_frontend(cpu, evsel, avg);
836 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
837 print_stalled_cycles_backend(cpu, evsel, avg);
838 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
839 total = avg_stats(&runtime_nsecs_stats[cpu]);
842 ratio = 1.0 * avg / total;
844 fprintf(output, " # %8.3f GHz ", ratio);
845 } else if (runtime_nsecs_stats[cpu].n != 0) {
848 total = avg_stats(&runtime_nsecs_stats[cpu]);
851 ratio = 1000.0 * avg / total;
857 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
859 fprintf(output, " ");
863 static void print_aggr_socket(char *prefix)
865 struct perf_evsel *counter;
867 int cpu, s, s2, sock, nr;
872 for (s = 0; s < sock_map->nr; s++) {
873 sock = cpu_map__socket(sock_map, s);
874 list_for_each_entry(counter, &evsel_list->entries, node) {
877 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
878 s2 = cpu_map__get_socket(evsel_list->cpus, cpu);
881 val += counter->counts->cpu[cpu].val;
882 ena += counter->counts->cpu[cpu].ena;
883 run += counter->counts->cpu[cpu].run;
887 fprintf(output, "%s", prefix);
889 if (run == 0 || ena == 0) {
890 fprintf(output, "S%*d%s%*d%s%*s%s%*s",
898 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
900 csv_output ? 0 : -24,
901 perf_evsel__name(counter));
903 fprintf(output, "%s%s",
904 csv_sep, counter->cgrp->name);
910 if (nsec_counter(counter))
911 nsec_printout(sock, nr, counter, val);
913 abs_printout(sock, nr, counter, val);
916 print_noise(counter, 1.0);
919 fprintf(output, " (%.2f%%)",
928 * Print out the results of a single counter:
929 * aggregated counts in system-wide mode
931 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
933 struct perf_stat *ps = counter->priv;
934 double avg = avg_stats(&ps->res_stats[0]);
935 int scaled = counter->counts->scaled;
938 fprintf(output, "%s", prefix);
941 fprintf(output, "%*s%s%*s",
943 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
945 csv_output ? 0 : -24,
946 perf_evsel__name(counter));
949 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
955 if (nsec_counter(counter))
956 nsec_printout(-1, 0, counter, avg);
958 abs_printout(-1, 0, counter, avg);
960 print_noise(counter, avg);
968 double avg_enabled, avg_running;
970 avg_enabled = avg_stats(&ps->res_stats[1]);
971 avg_running = avg_stats(&ps->res_stats[2]);
973 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
975 fprintf(output, "\n");
979 * Print out the results of a single counter:
980 * does not use aggregated count in system-wide
982 static void print_counter(struct perf_evsel *counter, char *prefix)
987 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
988 val = counter->counts->cpu[cpu].val;
989 ena = counter->counts->cpu[cpu].ena;
990 run = counter->counts->cpu[cpu].run;
993 fprintf(output, "%s", prefix);
995 if (run == 0 || ena == 0) {
996 fprintf(output, "CPU%*d%s%*s%s%*s",
998 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1000 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1002 csv_output ? 0 : -24,
1003 perf_evsel__name(counter));
1006 fprintf(output, "%s%s",
1007 csv_sep, counter->cgrp->name);
1009 fputc('\n', output);
1013 if (nsec_counter(counter))
1014 nsec_printout(cpu, 0, counter, val);
1016 abs_printout(cpu, 0, counter, val);
1019 print_noise(counter, 1.0);
1022 fprintf(output, " (%.2f%%)",
1025 fputc('\n', output);
1029 static void print_stat(int argc, const char **argv)
1031 struct perf_evsel *counter;
1037 fprintf(output, "\n");
1038 fprintf(output, " Performance counter stats for ");
1039 if (!perf_target__has_task(&target)) {
1040 fprintf(output, "\'%s", argv[0]);
1041 for (i = 1; i < argc; i++)
1042 fprintf(output, " %s", argv[i]);
1043 } else if (target.pid)
1044 fprintf(output, "process id \'%s", target.pid);
1046 fprintf(output, "thread id \'%s", target.tid);
1048 fprintf(output, "\'");
1050 fprintf(output, " (%d runs)", run_count);
1051 fprintf(output, ":\n\n");
1055 print_aggr_socket(NULL);
1057 list_for_each_entry(counter, &evsel_list->entries, node)
1058 print_counter(counter, NULL);
1060 list_for_each_entry(counter, &evsel_list->entries, node)
1061 print_counter_aggr(counter, NULL);
1066 fprintf(output, "\n");
1067 fprintf(output, " %17.9f seconds time elapsed",
1068 avg_stats(&walltime_nsecs_stats)/1e9);
1069 if (run_count > 1) {
1070 fprintf(output, " ");
1071 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1072 avg_stats(&walltime_nsecs_stats));
1074 fprintf(output, "\n\n");
1078 static volatile int signr = -1;
1080 static void skip_signal(int signo)
1082 if ((child_pid == -1) || interval)
1088 static void sig_atexit(void)
1090 if (child_pid != -1)
1091 kill(child_pid, SIGTERM);
1096 signal(signr, SIG_DFL);
1097 kill(getpid(), signr);
1100 static int stat__set_big_num(const struct option *opt __maybe_unused,
1101 const char *s __maybe_unused, int unset)
1103 big_num_opt = unset ? 0 : 1;
1108 * Add default attributes, if there were no attributes specified or
1109 * if -d/--detailed, -d -d or -d -d -d is used:
1111 static int add_default_attributes(void)
1113 struct perf_event_attr default_attrs[] = {
1115 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1116 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1117 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1118 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1120 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1121 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1122 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1123 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1124 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1125 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1130 * Detailed stats (-d), covering the L1 and last level data caches:
1132 struct perf_event_attr detailed_attrs[] = {
1134 { .type = PERF_TYPE_HW_CACHE,
1136 PERF_COUNT_HW_CACHE_L1D << 0 |
1137 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1138 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1140 { .type = PERF_TYPE_HW_CACHE,
1142 PERF_COUNT_HW_CACHE_L1D << 0 |
1143 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1144 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1146 { .type = PERF_TYPE_HW_CACHE,
1148 PERF_COUNT_HW_CACHE_LL << 0 |
1149 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1150 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1152 { .type = PERF_TYPE_HW_CACHE,
1154 PERF_COUNT_HW_CACHE_LL << 0 |
1155 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1156 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1160 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1162 struct perf_event_attr very_detailed_attrs[] = {
1164 { .type = PERF_TYPE_HW_CACHE,
1166 PERF_COUNT_HW_CACHE_L1I << 0 |
1167 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1168 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1170 { .type = PERF_TYPE_HW_CACHE,
1172 PERF_COUNT_HW_CACHE_L1I << 0 |
1173 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1174 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1176 { .type = PERF_TYPE_HW_CACHE,
1178 PERF_COUNT_HW_CACHE_DTLB << 0 |
1179 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1180 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1182 { .type = PERF_TYPE_HW_CACHE,
1184 PERF_COUNT_HW_CACHE_DTLB << 0 |
1185 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1186 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1188 { .type = PERF_TYPE_HW_CACHE,
1190 PERF_COUNT_HW_CACHE_ITLB << 0 |
1191 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1192 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1194 { .type = PERF_TYPE_HW_CACHE,
1196 PERF_COUNT_HW_CACHE_ITLB << 0 |
1197 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1198 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1203 * Very, very detailed stats (-d -d -d), adding prefetch events:
1205 struct perf_event_attr very_very_detailed_attrs[] = {
1207 { .type = PERF_TYPE_HW_CACHE,
1209 PERF_COUNT_HW_CACHE_L1D << 0 |
1210 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1211 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1213 { .type = PERF_TYPE_HW_CACHE,
1215 PERF_COUNT_HW_CACHE_L1D << 0 |
1216 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1217 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1220 /* Set attrs if no event is selected and !null_run: */
1224 if (!evsel_list->nr_entries) {
1225 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1229 /* Detailed events get appended to the event list: */
1231 if (detailed_run < 1)
1234 /* Append detailed run extra attributes: */
1235 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1238 if (detailed_run < 2)
1241 /* Append very detailed run extra attributes: */
1242 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1245 if (detailed_run < 3)
1248 /* Append very, very detailed run extra attributes: */
1249 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1252 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1254 bool append_file = false;
1256 const char *output_name = NULL;
1257 const struct option options[] = {
1258 OPT_CALLBACK('e', "event", &evsel_list, "event",
1259 "event selector. use 'perf list' to list available events",
1260 parse_events_option),
1261 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1262 "event filter", parse_filter),
1263 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1264 "child tasks do not inherit counters"),
1265 OPT_STRING('p', "pid", &target.pid, "pid",
1266 "stat events on existing process id"),
1267 OPT_STRING('t', "tid", &target.tid, "tid",
1268 "stat events on existing thread id"),
1269 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1270 "system-wide collection from all CPUs"),
1271 OPT_BOOLEAN('g', "group", &group,
1272 "put the counters into a counter group"),
1273 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1274 OPT_INCR('v', "verbose", &verbose,
1275 "be more verbose (show counter open errors, etc)"),
1276 OPT_INTEGER('r', "repeat", &run_count,
1277 "repeat command and print average + stddev (max: 100, forever: 0)"),
1278 OPT_BOOLEAN('n', "null", &null_run,
1279 "null run - dont start any counters"),
1280 OPT_INCR('d', "detailed", &detailed_run,
1281 "detailed run - start a lot of events"),
1282 OPT_BOOLEAN('S', "sync", &sync_run,
1283 "call sync() before starting a run"),
1284 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1285 "print large numbers with thousands\' separators",
1287 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1288 "list of cpus to monitor in system-wide"),
1289 OPT_BOOLEAN('A', "no-aggr", &no_aggr, "disable CPU count aggregation"),
1290 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1291 "print counts with custom separator"),
1292 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1293 "monitor event in cgroup name only", parse_cgroups),
1294 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1295 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1296 OPT_INTEGER(0, "log-fd", &output_fd,
1297 "log output to fd, instead of stderr"),
1298 OPT_STRING(0, "pre", &pre_cmd, "command",
1299 "command to run prior to the measured command"),
1300 OPT_STRING(0, "post", &post_cmd, "command",
1301 "command to run after to the measured command"),
1302 OPT_UINTEGER('I', "interval-print", &interval,
1303 "print counts at regular interval in ms (>= 100)"),
1304 OPT_BOOLEAN(0, "aggr-socket", &aggr_socket, "aggregate counts per processor socket"),
1307 const char * const stat_usage[] = {
1308 "perf stat [<options>] [<command>]",
1311 struct perf_evsel *pos;
1312 int status = -ENOMEM, run_idx;
1315 setlocale(LC_ALL, "");
1317 evsel_list = perf_evlist__new();
1318 if (evsel_list == NULL)
1321 argc = parse_options(argc, argv, options, stat_usage,
1322 PARSE_OPT_STOP_AT_NON_OPTION);
1325 if (output_name && strcmp(output_name, "-"))
1328 if (output_name && output_fd) {
1329 fprintf(stderr, "cannot use both --output and --log-fd\n");
1330 usage_with_options(stat_usage, options);
1333 if (output_fd < 0) {
1334 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1335 usage_with_options(stat_usage, options);
1340 mode = append_file ? "a" : "w";
1342 output = fopen(output_name, mode);
1344 perror("failed to create output file");
1347 clock_gettime(CLOCK_REALTIME, &tm);
1348 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1349 } else if (output_fd > 0) {
1350 mode = append_file ? "a" : "w";
1351 output = fdopen(output_fd, mode);
1353 perror("Failed opening logfd");
1360 if (!strcmp(csv_sep, "\\t"))
1363 csv_sep = DEFAULT_SEPARATOR;
1366 * let the spreadsheet do the pretty-printing
1369 /* User explicitly passed -B? */
1370 if (big_num_opt == 1) {
1371 fprintf(stderr, "-B option not supported with -x\n");
1372 usage_with_options(stat_usage, options);
1373 } else /* Nope, so disable big number formatting */
1375 } else if (big_num_opt == 0) /* User passed --no-big-num */
1378 if (!argc && !perf_target__has_task(&target))
1379 usage_with_options(stat_usage, options);
1380 if (run_count < 0) {
1381 usage_with_options(stat_usage, options);
1382 } else if (run_count == 0) {
1387 /* no_aggr, cgroup are for system-wide only */
1388 if ((no_aggr || nr_cgroups) && !perf_target__has_cpu(&target)) {
1389 fprintf(stderr, "both cgroup and no-aggregation "
1390 "modes only available in system-wide mode\n");
1392 usage_with_options(stat_usage, options);
1396 if (!perf_target__has_cpu(&target)) {
1397 fprintf(stderr, "--aggr-socket only available in system-wide mode (-a)\n");
1398 usage_with_options(stat_usage, options);
1403 if (add_default_attributes())
1406 perf_target__validate(&target);
1408 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1409 if (perf_target__has_task(&target))
1410 pr_err("Problems finding threads of monitor\n");
1411 if (perf_target__has_cpu(&target))
1412 perror("failed to parse CPUs map");
1414 usage_with_options(stat_usage, options);
1417 if (interval && interval < 100) {
1418 pr_err("print interval must be >= 100ms\n");
1419 usage_with_options(stat_usage, options);
1423 list_for_each_entry(pos, &evsel_list->entries, node) {
1424 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1425 perf_evsel__alloc_counts(pos, perf_evsel__nr_cpus(pos)) < 0)
1429 list_for_each_entry(pos, &evsel_list->entries, node) {
1430 if (perf_evsel__alloc_prev_raw_counts(pos) < 0)
1436 * We dont want to block the signals - that would cause
1437 * child tasks to inherit that and Ctrl-C would not work.
1438 * What we want is for Ctrl-C to work in the exec()-ed
1439 * task, but being ignored by perf stat itself:
1443 signal(SIGINT, skip_signal);
1444 signal(SIGCHLD, skip_signal);
1445 signal(SIGALRM, skip_signal);
1446 signal(SIGABRT, skip_signal);
1449 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1450 if (run_count != 1 && verbose)
1451 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1454 status = run_perf_stat(argc, argv);
1455 if (forever && status != -1) {
1456 print_stat(argc, argv);
1457 list_for_each_entry(pos, &evsel_list->entries, node) {
1458 perf_evsel__reset_stat_priv(pos);
1459 perf_evsel__reset_counts(pos, perf_evsel__nr_cpus(pos));
1465 if (!forever && status != -1 && !interval)
1466 print_stat(argc, argv);
1468 list_for_each_entry(pos, &evsel_list->entries, node) {
1469 perf_evsel__free_stat_priv(pos);
1470 perf_evsel__free_counts(pos);
1471 perf_evsel__free_prev_raw_counts(pos);
1473 perf_evlist__delete_maps(evsel_list);
1475 perf_evlist__delete(evsel_list);