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 void perf_evlist__free_stats(struct perf_evlist *evlist)
171 struct perf_evsel *evsel;
173 list_for_each_entry(evsel, &evlist->entries, node) {
174 perf_evsel__free_stat_priv(evsel);
175 perf_evsel__free_counts(evsel);
176 perf_evsel__free_prev_raw_counts(evsel);
180 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
182 struct perf_evsel *evsel;
184 list_for_each_entry(evsel, &evlist->entries, node) {
185 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
186 perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
187 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
194 perf_evlist__free_stats(evlist);
198 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
199 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
200 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
201 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
202 static struct stats runtime_branches_stats[MAX_NR_CPUS];
203 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
204 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
205 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
206 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
207 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
208 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
209 static struct stats walltime_nsecs_stats;
211 static void perf_stat__reset_stats(struct perf_evlist *evlist)
213 struct perf_evsel *evsel;
215 list_for_each_entry(evsel, &evlist->entries, node) {
216 perf_evsel__reset_stat_priv(evsel);
217 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
220 memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
221 memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
222 memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
223 memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
224 memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
225 memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
226 memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
227 memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
228 memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
229 memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
230 memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
231 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
234 static int create_perf_stat_counter(struct perf_evsel *evsel)
236 struct perf_event_attr *attr = &evsel->attr;
239 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
240 PERF_FORMAT_TOTAL_TIME_RUNNING;
242 attr->inherit = !no_inherit;
244 if (perf_target__has_cpu(&target))
245 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
247 if (!perf_target__has_task(&target) &&
248 perf_evsel__is_group_leader(evsel)) {
250 attr->enable_on_exec = 1;
253 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
257 * Does the counter have nsecs as a unit?
259 static inline int nsec_counter(struct perf_evsel *evsel)
261 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
262 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
269 * Update various tracking values we maintain to print
270 * more semantic information such as miss/hit ratios,
271 * instruction rates, etc:
273 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
275 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
276 update_stats(&runtime_nsecs_stats[0], count[0]);
277 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
278 update_stats(&runtime_cycles_stats[0], count[0]);
279 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
280 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
281 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
282 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
283 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
284 update_stats(&runtime_branches_stats[0], count[0]);
285 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
286 update_stats(&runtime_cacherefs_stats[0], count[0]);
287 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
288 update_stats(&runtime_l1_dcache_stats[0], count[0]);
289 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
290 update_stats(&runtime_l1_icache_stats[0], count[0]);
291 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
292 update_stats(&runtime_ll_cache_stats[0], count[0]);
293 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
294 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
295 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
296 update_stats(&runtime_itlb_cache_stats[0], count[0]);
300 * Read out the results of a single counter:
301 * aggregate counts across CPUs in system-wide mode
303 static int read_counter_aggr(struct perf_evsel *counter)
305 struct perf_stat *ps = counter->priv;
306 u64 *count = counter->counts->aggr.values;
309 if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
310 thread_map__nr(evsel_list->threads), scale) < 0)
313 for (i = 0; i < 3; i++)
314 update_stats(&ps->res_stats[i], count[i]);
317 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
318 perf_evsel__name(counter), count[0], count[1], count[2]);
322 * Save the full runtime - to allow normalization during printout:
324 update_shadow_stats(counter, count);
330 * Read out the results of a single counter:
331 * do not aggregate counts across CPUs in system-wide mode
333 static int read_counter(struct perf_evsel *counter)
338 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
339 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
342 count = counter->counts->cpu[cpu].values;
344 update_shadow_stats(counter, count);
350 static void print_interval(void)
352 static int num_print_interval;
353 struct perf_evsel *counter;
354 struct perf_stat *ps;
355 struct timespec ts, rs;
359 list_for_each_entry(counter, &evsel_list->entries, node) {
361 memset(ps->res_stats, 0, sizeof(ps->res_stats));
362 read_counter(counter);
365 list_for_each_entry(counter, &evsel_list->entries, node) {
367 memset(ps->res_stats, 0, sizeof(ps->res_stats));
368 read_counter_aggr(counter);
371 clock_gettime(CLOCK_MONOTONIC, &ts);
372 diff_timespec(&rs, &ts, &ref_time);
373 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
375 if (num_print_interval == 0 && !csv_output) {
377 fprintf(output, "# time socket cpus counts events\n");
379 fprintf(output, "# time CPU counts events\n");
381 fprintf(output, "# time counts events\n");
384 if (++num_print_interval == 25)
385 num_print_interval = 0;
388 print_aggr_socket(prefix);
390 list_for_each_entry(counter, &evsel_list->entries, node)
391 print_counter(counter, prefix);
393 list_for_each_entry(counter, &evsel_list->entries, node)
394 print_counter_aggr(counter, prefix);
398 static int __run_perf_stat(int argc, const char **argv)
401 unsigned long long t0, t1;
402 struct perf_evsel *counter;
405 const bool forks = (argc > 0);
408 ts.tv_sec = interval / 1000;
409 ts.tv_nsec = (interval % 1000) * 1000000;
416 && cpu_map__build_socket_map(evsel_list->cpus, &sock_map)) {
417 perror("cannot build socket map");
422 if (perf_evlist__prepare_workload(evsel_list, &target, argv,
424 perror("failed to prepare workload");
430 perf_evlist__set_leader(evsel_list);
432 list_for_each_entry(counter, &evsel_list->entries, node) {
433 if (create_perf_stat_counter(counter) < 0) {
435 * PPC returns ENXIO for HW counters until 2.6.37
436 * (behavior changed with commit b0a873e).
438 if (errno == EINVAL || errno == ENOSYS ||
439 errno == ENOENT || errno == EOPNOTSUPP ||
442 ui__warning("%s event is not supported by the kernel.\n",
443 perf_evsel__name(counter));
444 counter->supported = false;
448 perf_evsel__open_strerror(counter, &target,
449 errno, msg, sizeof(msg));
450 ui__error("%s\n", msg);
453 kill(child_pid, SIGTERM);
457 counter->supported = true;
460 if (perf_evlist__apply_filters(evsel_list)) {
461 error("failed to set filter with %d (%s)\n", errno,
467 * Enable counters and exec the command:
470 clock_gettime(CLOCK_MONOTONIC, &ref_time);
473 perf_evlist__start_workload(evsel_list);
476 while (!waitpid(child_pid, &status, WNOHANG)) {
477 nanosleep(&ts, NULL);
482 if (WIFSIGNALED(status))
483 psignal(WTERMSIG(status), argv[0]);
486 nanosleep(&ts, NULL);
494 update_stats(&walltime_nsecs_stats, t1 - t0);
497 list_for_each_entry(counter, &evsel_list->entries, node) {
498 read_counter(counter);
499 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
502 list_for_each_entry(counter, &evsel_list->entries, node) {
503 read_counter_aggr(counter);
504 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
505 thread_map__nr(evsel_list->threads));
509 return WEXITSTATUS(status);
512 static int run_perf_stat(int argc __maybe_unused, const char **argv)
517 ret = system(pre_cmd);
525 ret = __run_perf_stat(argc, argv);
530 ret = system(post_cmd);
538 static void print_noise_pct(double total, double avg)
540 double pct = rel_stddev_stats(total, avg);
543 fprintf(output, "%s%.2f%%", csv_sep, pct);
545 fprintf(output, " ( +-%6.2f%% )", pct);
548 static void print_noise(struct perf_evsel *evsel, double avg)
550 struct perf_stat *ps;
556 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
559 static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
561 double msecs = avg / 1e6;
562 char cpustr[16] = { '\0', };
563 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
566 sprintf(cpustr, "S%*d%s%*d%s",
574 sprintf(cpustr, "CPU%*d%s",
576 perf_evsel__cpus(evsel)->map[cpu], csv_sep);
578 fprintf(output, fmt, cpustr, msecs, csv_sep, perf_evsel__name(evsel));
581 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
583 if (csv_output || interval)
586 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
587 fprintf(output, " # %8.3f CPUs utilized ",
588 avg / avg_stats(&walltime_nsecs_stats));
590 fprintf(output, " ");
593 /* used for get_ratio_color() */
595 GRC_STALLED_CYCLES_FE,
596 GRC_STALLED_CYCLES_BE,
601 static const char *get_ratio_color(enum grc_type type, double ratio)
603 static const double grc_table[GRC_MAX_NR][3] = {
604 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
605 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
606 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
608 const char *color = PERF_COLOR_NORMAL;
610 if (ratio > grc_table[type][0])
611 color = PERF_COLOR_RED;
612 else if (ratio > grc_table[type][1])
613 color = PERF_COLOR_MAGENTA;
614 else if (ratio > grc_table[type][2])
615 color = PERF_COLOR_YELLOW;
620 static void print_stalled_cycles_frontend(int cpu,
621 struct perf_evsel *evsel
622 __maybe_unused, double avg)
624 double total, ratio = 0.0;
627 total = avg_stats(&runtime_cycles_stats[cpu]);
630 ratio = avg / total * 100.0;
632 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
634 fprintf(output, " # ");
635 color_fprintf(output, color, "%6.2f%%", ratio);
636 fprintf(output, " frontend cycles idle ");
639 static void print_stalled_cycles_backend(int cpu,
640 struct perf_evsel *evsel
641 __maybe_unused, double avg)
643 double total, ratio = 0.0;
646 total = avg_stats(&runtime_cycles_stats[cpu]);
649 ratio = avg / total * 100.0;
651 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
653 fprintf(output, " # ");
654 color_fprintf(output, color, "%6.2f%%", ratio);
655 fprintf(output, " backend cycles idle ");
658 static void print_branch_misses(int cpu,
659 struct perf_evsel *evsel __maybe_unused,
662 double total, ratio = 0.0;
665 total = avg_stats(&runtime_branches_stats[cpu]);
668 ratio = avg / total * 100.0;
670 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
672 fprintf(output, " # ");
673 color_fprintf(output, color, "%6.2f%%", ratio);
674 fprintf(output, " of all branches ");
677 static void print_l1_dcache_misses(int cpu,
678 struct perf_evsel *evsel __maybe_unused,
681 double total, ratio = 0.0;
684 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
687 ratio = avg / total * 100.0;
689 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
691 fprintf(output, " # ");
692 color_fprintf(output, color, "%6.2f%%", ratio);
693 fprintf(output, " of all L1-dcache hits ");
696 static void print_l1_icache_misses(int cpu,
697 struct perf_evsel *evsel __maybe_unused,
700 double total, ratio = 0.0;
703 total = avg_stats(&runtime_l1_icache_stats[cpu]);
706 ratio = avg / total * 100.0;
708 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
710 fprintf(output, " # ");
711 color_fprintf(output, color, "%6.2f%%", ratio);
712 fprintf(output, " of all L1-icache hits ");
715 static void print_dtlb_cache_misses(int cpu,
716 struct perf_evsel *evsel __maybe_unused,
719 double total, ratio = 0.0;
722 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
725 ratio = avg / total * 100.0;
727 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
729 fprintf(output, " # ");
730 color_fprintf(output, color, "%6.2f%%", ratio);
731 fprintf(output, " of all dTLB cache hits ");
734 static void print_itlb_cache_misses(int cpu,
735 struct perf_evsel *evsel __maybe_unused,
738 double total, ratio = 0.0;
741 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
744 ratio = avg / total * 100.0;
746 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
748 fprintf(output, " # ");
749 color_fprintf(output, color, "%6.2f%%", ratio);
750 fprintf(output, " of all iTLB cache hits ");
753 static void print_ll_cache_misses(int cpu,
754 struct perf_evsel *evsel __maybe_unused,
757 double total, ratio = 0.0;
760 total = avg_stats(&runtime_ll_cache_stats[cpu]);
763 ratio = avg / total * 100.0;
765 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
767 fprintf(output, " # ");
768 color_fprintf(output, color, "%6.2f%%", ratio);
769 fprintf(output, " of all LL-cache hits ");
772 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
774 double total, ratio = 0.0;
775 char cpustr[16] = { '\0', };
781 fmt = "%s%'18.0f%s%-25s";
783 fmt = "%s%18.0f%s%-25s";
786 sprintf(cpustr, "S%*d%s%*d%s",
794 sprintf(cpustr, "CPU%*d%s",
796 perf_evsel__cpus(evsel)->map[cpu], csv_sep);
800 fprintf(output, fmt, cpustr, avg, csv_sep, perf_evsel__name(evsel));
803 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
805 if (csv_output || interval)
808 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
809 total = avg_stats(&runtime_cycles_stats[cpu]);
813 fprintf(output, " # %5.2f insns per cycle ", ratio);
815 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
816 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
820 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
823 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
824 runtime_branches_stats[cpu].n != 0) {
825 print_branch_misses(cpu, evsel, avg);
827 evsel->attr.type == PERF_TYPE_HW_CACHE &&
828 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
829 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
830 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
831 runtime_l1_dcache_stats[cpu].n != 0) {
832 print_l1_dcache_misses(cpu, evsel, avg);
834 evsel->attr.type == PERF_TYPE_HW_CACHE &&
835 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
836 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
837 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
838 runtime_l1_icache_stats[cpu].n != 0) {
839 print_l1_icache_misses(cpu, evsel, avg);
841 evsel->attr.type == PERF_TYPE_HW_CACHE &&
842 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
843 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
844 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
845 runtime_dtlb_cache_stats[cpu].n != 0) {
846 print_dtlb_cache_misses(cpu, evsel, avg);
848 evsel->attr.type == PERF_TYPE_HW_CACHE &&
849 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
850 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
851 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
852 runtime_itlb_cache_stats[cpu].n != 0) {
853 print_itlb_cache_misses(cpu, evsel, avg);
855 evsel->attr.type == PERF_TYPE_HW_CACHE &&
856 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
857 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
858 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
859 runtime_ll_cache_stats[cpu].n != 0) {
860 print_ll_cache_misses(cpu, evsel, avg);
861 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
862 runtime_cacherefs_stats[cpu].n != 0) {
863 total = avg_stats(&runtime_cacherefs_stats[cpu]);
866 ratio = avg * 100 / total;
868 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
870 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
871 print_stalled_cycles_frontend(cpu, evsel, avg);
872 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
873 print_stalled_cycles_backend(cpu, evsel, avg);
874 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
875 total = avg_stats(&runtime_nsecs_stats[cpu]);
878 ratio = 1.0 * avg / total;
880 fprintf(output, " # %8.3f GHz ", ratio);
881 } else if (runtime_nsecs_stats[cpu].n != 0) {
884 total = avg_stats(&runtime_nsecs_stats[cpu]);
887 ratio = 1000.0 * avg / total;
893 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
895 fprintf(output, " ");
899 static void print_aggr_socket(char *prefix)
901 struct perf_evsel *counter;
903 int cpu, s, s2, sock, nr;
908 for (s = 0; s < sock_map->nr; s++) {
909 sock = cpu_map__socket(sock_map, s);
910 list_for_each_entry(counter, &evsel_list->entries, node) {
913 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
914 s2 = cpu_map__get_socket(evsel_list->cpus, cpu);
917 val += counter->counts->cpu[cpu].val;
918 ena += counter->counts->cpu[cpu].ena;
919 run += counter->counts->cpu[cpu].run;
923 fprintf(output, "%s", prefix);
925 if (run == 0 || ena == 0) {
926 fprintf(output, "S%*d%s%*d%s%*s%s%*s",
934 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
936 csv_output ? 0 : -24,
937 perf_evsel__name(counter));
939 fprintf(output, "%s%s",
940 csv_sep, counter->cgrp->name);
946 if (nsec_counter(counter))
947 nsec_printout(sock, nr, counter, val);
949 abs_printout(sock, nr, counter, val);
952 print_noise(counter, 1.0);
955 fprintf(output, " (%.2f%%)",
964 * Print out the results of a single counter:
965 * aggregated counts in system-wide mode
967 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
969 struct perf_stat *ps = counter->priv;
970 double avg = avg_stats(&ps->res_stats[0]);
971 int scaled = counter->counts->scaled;
974 fprintf(output, "%s", prefix);
977 fprintf(output, "%*s%s%*s",
979 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
981 csv_output ? 0 : -24,
982 perf_evsel__name(counter));
985 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
991 if (nsec_counter(counter))
992 nsec_printout(-1, 0, counter, avg);
994 abs_printout(-1, 0, counter, avg);
996 print_noise(counter, avg);
1004 double avg_enabled, avg_running;
1006 avg_enabled = avg_stats(&ps->res_stats[1]);
1007 avg_running = avg_stats(&ps->res_stats[2]);
1009 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1011 fprintf(output, "\n");
1015 * Print out the results of a single counter:
1016 * does not use aggregated count in system-wide
1018 static void print_counter(struct perf_evsel *counter, char *prefix)
1023 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1024 val = counter->counts->cpu[cpu].val;
1025 ena = counter->counts->cpu[cpu].ena;
1026 run = counter->counts->cpu[cpu].run;
1029 fprintf(output, "%s", prefix);
1031 if (run == 0 || ena == 0) {
1032 fprintf(output, "CPU%*d%s%*s%s%*s",
1033 csv_output ? 0 : -4,
1034 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1035 csv_output ? 0 : 18,
1036 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1038 csv_output ? 0 : -24,
1039 perf_evsel__name(counter));
1042 fprintf(output, "%s%s",
1043 csv_sep, counter->cgrp->name);
1045 fputc('\n', output);
1049 if (nsec_counter(counter))
1050 nsec_printout(cpu, 0, counter, val);
1052 abs_printout(cpu, 0, counter, val);
1055 print_noise(counter, 1.0);
1058 fprintf(output, " (%.2f%%)",
1061 fputc('\n', output);
1065 static void print_stat(int argc, const char **argv)
1067 struct perf_evsel *counter;
1073 fprintf(output, "\n");
1074 fprintf(output, " Performance counter stats for ");
1075 if (!perf_target__has_task(&target)) {
1076 fprintf(output, "\'%s", argv[0]);
1077 for (i = 1; i < argc; i++)
1078 fprintf(output, " %s", argv[i]);
1079 } else if (target.pid)
1080 fprintf(output, "process id \'%s", target.pid);
1082 fprintf(output, "thread id \'%s", target.tid);
1084 fprintf(output, "\'");
1086 fprintf(output, " (%d runs)", run_count);
1087 fprintf(output, ":\n\n");
1091 print_aggr_socket(NULL);
1093 list_for_each_entry(counter, &evsel_list->entries, node)
1094 print_counter(counter, NULL);
1096 list_for_each_entry(counter, &evsel_list->entries, node)
1097 print_counter_aggr(counter, NULL);
1102 fprintf(output, "\n");
1103 fprintf(output, " %17.9f seconds time elapsed",
1104 avg_stats(&walltime_nsecs_stats)/1e9);
1105 if (run_count > 1) {
1106 fprintf(output, " ");
1107 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1108 avg_stats(&walltime_nsecs_stats));
1110 fprintf(output, "\n\n");
1114 static volatile int signr = -1;
1116 static void skip_signal(int signo)
1118 if ((child_pid == -1) || interval)
1124 static void sig_atexit(void)
1126 if (child_pid != -1)
1127 kill(child_pid, SIGTERM);
1132 signal(signr, SIG_DFL);
1133 kill(getpid(), signr);
1136 static int stat__set_big_num(const struct option *opt __maybe_unused,
1137 const char *s __maybe_unused, int unset)
1139 big_num_opt = unset ? 0 : 1;
1144 * Add default attributes, if there were no attributes specified or
1145 * if -d/--detailed, -d -d or -d -d -d is used:
1147 static int add_default_attributes(void)
1149 struct perf_event_attr default_attrs[] = {
1151 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1152 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1153 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1154 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1156 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1157 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1158 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1159 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1160 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1161 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1166 * Detailed stats (-d), covering the L1 and last level data caches:
1168 struct perf_event_attr detailed_attrs[] = {
1170 { .type = PERF_TYPE_HW_CACHE,
1172 PERF_COUNT_HW_CACHE_L1D << 0 |
1173 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1174 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1176 { .type = PERF_TYPE_HW_CACHE,
1178 PERF_COUNT_HW_CACHE_L1D << 0 |
1179 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1180 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1182 { .type = PERF_TYPE_HW_CACHE,
1184 PERF_COUNT_HW_CACHE_LL << 0 |
1185 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1186 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1188 { .type = PERF_TYPE_HW_CACHE,
1190 PERF_COUNT_HW_CACHE_LL << 0 |
1191 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1192 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1196 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1198 struct perf_event_attr very_detailed_attrs[] = {
1200 { .type = PERF_TYPE_HW_CACHE,
1202 PERF_COUNT_HW_CACHE_L1I << 0 |
1203 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1204 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1206 { .type = PERF_TYPE_HW_CACHE,
1208 PERF_COUNT_HW_CACHE_L1I << 0 |
1209 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1210 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1212 { .type = PERF_TYPE_HW_CACHE,
1214 PERF_COUNT_HW_CACHE_DTLB << 0 |
1215 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1216 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1218 { .type = PERF_TYPE_HW_CACHE,
1220 PERF_COUNT_HW_CACHE_DTLB << 0 |
1221 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1222 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1224 { .type = PERF_TYPE_HW_CACHE,
1226 PERF_COUNT_HW_CACHE_ITLB << 0 |
1227 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1228 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1230 { .type = PERF_TYPE_HW_CACHE,
1232 PERF_COUNT_HW_CACHE_ITLB << 0 |
1233 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1234 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1239 * Very, very detailed stats (-d -d -d), adding prefetch events:
1241 struct perf_event_attr very_very_detailed_attrs[] = {
1243 { .type = PERF_TYPE_HW_CACHE,
1245 PERF_COUNT_HW_CACHE_L1D << 0 |
1246 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1247 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1249 { .type = PERF_TYPE_HW_CACHE,
1251 PERF_COUNT_HW_CACHE_L1D << 0 |
1252 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1253 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1256 /* Set attrs if no event is selected and !null_run: */
1260 if (!evsel_list->nr_entries) {
1261 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1265 /* Detailed events get appended to the event list: */
1267 if (detailed_run < 1)
1270 /* Append detailed run extra attributes: */
1271 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1274 if (detailed_run < 2)
1277 /* Append very detailed run extra attributes: */
1278 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1281 if (detailed_run < 3)
1284 /* Append very, very detailed run extra attributes: */
1285 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1288 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1290 bool append_file = false;
1292 const char *output_name = NULL;
1293 const struct option options[] = {
1294 OPT_CALLBACK('e', "event", &evsel_list, "event",
1295 "event selector. use 'perf list' to list available events",
1296 parse_events_option),
1297 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1298 "event filter", parse_filter),
1299 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1300 "child tasks do not inherit counters"),
1301 OPT_STRING('p', "pid", &target.pid, "pid",
1302 "stat events on existing process id"),
1303 OPT_STRING('t', "tid", &target.tid, "tid",
1304 "stat events on existing thread id"),
1305 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1306 "system-wide collection from all CPUs"),
1307 OPT_BOOLEAN('g', "group", &group,
1308 "put the counters into a counter group"),
1309 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1310 OPT_INCR('v', "verbose", &verbose,
1311 "be more verbose (show counter open errors, etc)"),
1312 OPT_INTEGER('r', "repeat", &run_count,
1313 "repeat command and print average + stddev (max: 100, forever: 0)"),
1314 OPT_BOOLEAN('n', "null", &null_run,
1315 "null run - dont start any counters"),
1316 OPT_INCR('d', "detailed", &detailed_run,
1317 "detailed run - start a lot of events"),
1318 OPT_BOOLEAN('S', "sync", &sync_run,
1319 "call sync() before starting a run"),
1320 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1321 "print large numbers with thousands\' separators",
1323 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1324 "list of cpus to monitor in system-wide"),
1325 OPT_BOOLEAN('A', "no-aggr", &no_aggr, "disable CPU count aggregation"),
1326 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1327 "print counts with custom separator"),
1328 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1329 "monitor event in cgroup name only", parse_cgroups),
1330 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1331 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1332 OPT_INTEGER(0, "log-fd", &output_fd,
1333 "log output to fd, instead of stderr"),
1334 OPT_STRING(0, "pre", &pre_cmd, "command",
1335 "command to run prior to the measured command"),
1336 OPT_STRING(0, "post", &post_cmd, "command",
1337 "command to run after to the measured command"),
1338 OPT_UINTEGER('I', "interval-print", &interval,
1339 "print counts at regular interval in ms (>= 100)"),
1340 OPT_BOOLEAN(0, "aggr-socket", &aggr_socket, "aggregate counts per processor socket"),
1343 const char * const stat_usage[] = {
1344 "perf stat [<options>] [<command>]",
1347 int status = -ENOMEM, run_idx;
1350 setlocale(LC_ALL, "");
1352 evsel_list = perf_evlist__new();
1353 if (evsel_list == NULL)
1356 argc = parse_options(argc, argv, options, stat_usage,
1357 PARSE_OPT_STOP_AT_NON_OPTION);
1360 if (output_name && strcmp(output_name, "-"))
1363 if (output_name && output_fd) {
1364 fprintf(stderr, "cannot use both --output and --log-fd\n");
1365 usage_with_options(stat_usage, options);
1368 if (output_fd < 0) {
1369 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1370 usage_with_options(stat_usage, options);
1375 mode = append_file ? "a" : "w";
1377 output = fopen(output_name, mode);
1379 perror("failed to create output file");
1382 clock_gettime(CLOCK_REALTIME, &tm);
1383 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1384 } else if (output_fd > 0) {
1385 mode = append_file ? "a" : "w";
1386 output = fdopen(output_fd, mode);
1388 perror("Failed opening logfd");
1395 if (!strcmp(csv_sep, "\\t"))
1398 csv_sep = DEFAULT_SEPARATOR;
1401 * let the spreadsheet do the pretty-printing
1404 /* User explicitly passed -B? */
1405 if (big_num_opt == 1) {
1406 fprintf(stderr, "-B option not supported with -x\n");
1407 usage_with_options(stat_usage, options);
1408 } else /* Nope, so disable big number formatting */
1410 } else if (big_num_opt == 0) /* User passed --no-big-num */
1413 if (!argc && !perf_target__has_task(&target))
1414 usage_with_options(stat_usage, options);
1415 if (run_count < 0) {
1416 usage_with_options(stat_usage, options);
1417 } else if (run_count == 0) {
1422 /* no_aggr, cgroup are for system-wide only */
1423 if ((no_aggr || nr_cgroups) && !perf_target__has_cpu(&target)) {
1424 fprintf(stderr, "both cgroup and no-aggregation "
1425 "modes only available in system-wide mode\n");
1427 usage_with_options(stat_usage, options);
1431 if (!perf_target__has_cpu(&target)) {
1432 fprintf(stderr, "--aggr-socket only available in system-wide mode (-a)\n");
1433 usage_with_options(stat_usage, options);
1438 if (add_default_attributes())
1441 perf_target__validate(&target);
1443 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1444 if (perf_target__has_task(&target))
1445 pr_err("Problems finding threads of monitor\n");
1446 if (perf_target__has_cpu(&target))
1447 perror("failed to parse CPUs map");
1449 usage_with_options(stat_usage, options);
1452 if (interval && interval < 100) {
1453 pr_err("print interval must be >= 100ms\n");
1454 usage_with_options(stat_usage, options);
1458 if (perf_evlist__alloc_stats(evsel_list, interval))
1462 * We dont want to block the signals - that would cause
1463 * child tasks to inherit that and Ctrl-C would not work.
1464 * What we want is for Ctrl-C to work in the exec()-ed
1465 * task, but being ignored by perf stat itself:
1469 signal(SIGINT, skip_signal);
1470 signal(SIGCHLD, skip_signal);
1471 signal(SIGALRM, skip_signal);
1472 signal(SIGABRT, skip_signal);
1475 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1476 if (run_count != 1 && verbose)
1477 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1480 status = run_perf_stat(argc, argv);
1481 if (forever && status != -1) {
1482 print_stat(argc, argv);
1483 perf_stat__reset_stats(evsel_list);
1487 if (!forever && status != -1 && !interval)
1488 print_stat(argc, argv);
1490 perf_evlist__free_stats(evsel_list);
1492 perf_evlist__delete_maps(evsel_list);
1494 perf_evlist__delete(evsel_list);