4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ./hackbench 10
13 Performance counter stats for './hackbench 10':
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
27 0.154822978 seconds time elapsed
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 * Improvements and fixes by:
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
41 * Released under the GPL v2. (and only v2, not any later version)
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include "util/parse-options.h"
49 #include "util/parse-events.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/color.h"
56 #include "util/stat.h"
57 #include "util/header.h"
58 #include "util/cpumap.h"
59 #include "util/thread.h"
60 #include "util/thread_map.h"
63 #include <sys/prctl.h>
66 #define DEFAULT_SEPARATOR " "
67 #define CNTR_NOT_SUPPORTED "<not supported>"
68 #define CNTR_NOT_COUNTED "<not counted>"
70 static void print_stat(int argc, const char **argv);
71 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
72 static void print_counter(struct perf_evsel *counter, char *prefix);
73 static void print_aggr(char *prefix);
75 /* Default events used for perf stat -T */
76 static const char *transaction_attrs = {
88 /* More limited version when the CPU does not have all events. */
89 static const char * transaction_limited_attrs = {
99 static struct perf_evlist *evsel_list;
101 static struct target target = {
112 static int run_count = 1;
113 static bool no_inherit = false;
114 static bool scale = true;
115 static enum aggr_mode aggr_mode = AGGR_GLOBAL;
116 static volatile pid_t child_pid = -1;
117 static bool null_run = false;
118 static int detailed_run = 0;
119 static bool transaction_run;
120 static bool big_num = true;
121 static int big_num_opt = -1;
122 static const char *csv_sep = NULL;
123 static bool csv_output = false;
124 static bool group = false;
125 static FILE *output = NULL;
126 static const char *pre_cmd = NULL;
127 static const char *post_cmd = NULL;
128 static bool sync_run = false;
129 static unsigned int interval = 0;
130 static unsigned int initial_delay = 0;
131 static unsigned int unit_width = 4; /* strlen("unit") */
132 static bool forever = false;
133 static struct timespec ref_time;
134 static struct cpu_map *aggr_map;
135 static int (*aggr_get_id)(struct cpu_map *m, int cpu);
137 static volatile int done = 0;
139 static inline void diff_timespec(struct timespec *r, struct timespec *a,
142 r->tv_sec = a->tv_sec - b->tv_sec;
143 if (a->tv_nsec < b->tv_nsec) {
144 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
147 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
151 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
153 return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
156 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
158 return perf_evsel__cpus(evsel)->nr;
161 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
164 struct perf_stat *ps = evsel->priv;
166 for (i = 0; i < 3; i++)
167 init_stats(&ps->res_stats[i]);
169 perf_stat_evsel_id_init(evsel);
172 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
174 evsel->priv = zalloc(sizeof(struct perf_stat));
175 if (evsel->priv == NULL)
177 perf_evsel__reset_stat_priv(evsel);
181 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
186 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
191 sz = sizeof(*evsel->counts) +
192 (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
198 evsel->prev_raw_counts = addr;
203 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
205 zfree(&evsel->prev_raw_counts);
208 static void perf_evlist__free_stats(struct perf_evlist *evlist)
210 struct perf_evsel *evsel;
212 evlist__for_each(evlist, evsel) {
213 perf_evsel__free_stat_priv(evsel);
214 perf_evsel__free_counts(evsel);
215 perf_evsel__free_prev_raw_counts(evsel);
219 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
221 struct perf_evsel *evsel;
223 evlist__for_each(evlist, evsel) {
224 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
225 perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
226 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
233 perf_evlist__free_stats(evlist);
238 CTX_BIT_USER = 1 << 0,
239 CTX_BIT_KERNEL = 1 << 1,
241 CTX_BIT_HOST = 1 << 3,
242 CTX_BIT_IDLE = 1 << 4,
243 CTX_BIT_MAX = 1 << 5,
246 #define NUM_CTX CTX_BIT_MAX
248 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
249 static struct stats runtime_cycles_stats[NUM_CTX][MAX_NR_CPUS];
250 static struct stats runtime_stalled_cycles_front_stats[NUM_CTX][MAX_NR_CPUS];
251 static struct stats runtime_stalled_cycles_back_stats[NUM_CTX][MAX_NR_CPUS];
252 static struct stats runtime_branches_stats[NUM_CTX][MAX_NR_CPUS];
253 static struct stats runtime_cacherefs_stats[NUM_CTX][MAX_NR_CPUS];
254 static struct stats runtime_l1_dcache_stats[NUM_CTX][MAX_NR_CPUS];
255 static struct stats runtime_l1_icache_stats[NUM_CTX][MAX_NR_CPUS];
256 static struct stats runtime_ll_cache_stats[NUM_CTX][MAX_NR_CPUS];
257 static struct stats runtime_itlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
258 static struct stats runtime_dtlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
259 static struct stats runtime_cycles_in_tx_stats[NUM_CTX][MAX_NR_CPUS];
260 static struct stats walltime_nsecs_stats;
261 static struct stats runtime_transaction_stats[NUM_CTX][MAX_NR_CPUS];
262 static struct stats runtime_elision_stats[NUM_CTX][MAX_NR_CPUS];
264 static int evsel_context(struct perf_evsel *evsel)
268 if (evsel->attr.exclude_kernel)
269 ctx |= CTX_BIT_KERNEL;
270 if (evsel->attr.exclude_user)
272 if (evsel->attr.exclude_hv)
274 if (evsel->attr.exclude_host)
276 if (evsel->attr.exclude_idle)
282 static void reset_shadow_stats(void)
284 memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
285 memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
286 memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
287 memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
288 memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
289 memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
290 memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
291 memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
292 memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
293 memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
294 memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
295 memset(runtime_cycles_in_tx_stats, 0,
296 sizeof(runtime_cycles_in_tx_stats));
297 memset(runtime_transaction_stats, 0,
298 sizeof(runtime_transaction_stats));
299 memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
300 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
303 static void perf_stat__reset_stats(struct perf_evlist *evlist)
305 struct perf_evsel *evsel;
307 evlist__for_each(evlist, evsel) {
308 perf_evsel__reset_stat_priv(evsel);
309 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
312 reset_shadow_stats();
315 static int create_perf_stat_counter(struct perf_evsel *evsel)
317 struct perf_event_attr *attr = &evsel->attr;
320 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
321 PERF_FORMAT_TOTAL_TIME_RUNNING;
323 attr->inherit = !no_inherit;
325 if (target__has_cpu(&target))
326 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
328 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
331 attr->enable_on_exec = 1;
334 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
338 * Does the counter have nsecs as a unit?
340 static inline int nsec_counter(struct perf_evsel *evsel)
342 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
343 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
350 * Update various tracking values we maintain to print
351 * more semantic information such as miss/hit ratios,
352 * instruction rates, etc:
354 static void update_shadow_stats(struct perf_evsel *counter, u64 *count,
357 int ctx = evsel_context(counter);
359 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
360 update_stats(&runtime_nsecs_stats[cpu], count[0]);
361 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
362 update_stats(&runtime_cycles_stats[ctx][cpu], count[0]);
363 else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
364 update_stats(&runtime_transaction_stats[ctx][cpu], count[0]);
365 else if (perf_stat_evsel__is(counter, TRANSACTION_START))
366 update_stats(&runtime_transaction_stats[ctx][cpu], count[0]);
367 else if (perf_stat_evsel__is(counter, ELISION_START))
368 update_stats(&runtime_elision_stats[ctx][cpu], count[0]);
369 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
370 update_stats(&runtime_stalled_cycles_front_stats[ctx][cpu], count[0]);
371 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
372 update_stats(&runtime_stalled_cycles_back_stats[ctx][cpu], count[0]);
373 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
374 update_stats(&runtime_branches_stats[ctx][cpu], count[0]);
375 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
376 update_stats(&runtime_cacherefs_stats[ctx][cpu], count[0]);
377 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
378 update_stats(&runtime_l1_dcache_stats[ctx][cpu], count[0]);
379 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
380 update_stats(&runtime_ll_cache_stats[ctx][cpu], count[0]);
381 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
382 update_stats(&runtime_ll_cache_stats[ctx][cpu], count[0]);
383 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
384 update_stats(&runtime_dtlb_cache_stats[ctx][cpu], count[0]);
385 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
386 update_stats(&runtime_itlb_cache_stats[ctx][cpu], count[0]);
389 static void zero_per_pkg(struct perf_evsel *counter)
391 if (counter->per_pkg_mask)
392 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
395 static int check_per_pkg(struct perf_evsel *counter, int cpu, bool *skip)
397 unsigned long *mask = counter->per_pkg_mask;
398 struct cpu_map *cpus = perf_evsel__cpus(counter);
403 if (!counter->per_pkg)
406 if (cpu_map__empty(cpus))
410 mask = zalloc(MAX_NR_CPUS);
414 counter->per_pkg_mask = mask;
417 s = cpu_map__get_socket(cpus, cpu);
421 *skip = test_and_set_bit(s, mask) == 1;
425 static int read_cb(struct perf_evsel *evsel, int cpu, int thread __maybe_unused,
426 struct perf_counts_values *count)
428 struct perf_counts_values *aggr = &evsel->counts->aggr;
429 static struct perf_counts_values zero;
432 if (check_per_pkg(evsel, cpu, &skip)) {
433 pr_err("failed to read per-pkg counter\n");
444 if (!evsel->snapshot)
445 perf_evsel__compute_deltas(evsel, cpu, count);
446 perf_counts_values__scale(count, scale, NULL);
447 evsel->counts->cpu[cpu] = *count;
448 if (aggr_mode == AGGR_NONE)
449 update_shadow_stats(evsel, count->values, cpu);
452 aggr->val += count->val;
454 aggr->ena += count->ena;
455 aggr->run += count->run;
464 static int read_counter(struct perf_evsel *counter);
467 * Read out the results of a single counter:
468 * aggregate counts across CPUs in system-wide mode
470 static int read_counter_aggr(struct perf_evsel *counter)
472 struct perf_counts_values *aggr = &counter->counts->aggr;
473 struct perf_stat *ps = counter->priv;
474 u64 *count = counter->counts->aggr.values;
477 aggr->val = aggr->ena = aggr->run = 0;
479 if (read_counter(counter))
482 if (!counter->snapshot)
483 perf_evsel__compute_deltas(counter, -1, aggr);
484 perf_counts_values__scale(aggr, scale, &counter->counts->scaled);
486 for (i = 0; i < 3; i++)
487 update_stats(&ps->res_stats[i], count[i]);
490 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
491 perf_evsel__name(counter), count[0], count[1], count[2]);
495 * Save the full runtime - to allow normalization during printout:
497 update_shadow_stats(counter, count, 0);
503 * Read out the results of a single counter:
504 * do not aggregate counts across CPUs in system-wide mode
506 static int read_counter(struct perf_evsel *counter)
508 int nthreads = thread_map__nr(evsel_list->threads);
509 int ncpus = perf_evsel__nr_cpus(counter);
512 if (!counter->supported)
515 if (counter->system_wide)
518 if (counter->per_pkg)
519 zero_per_pkg(counter);
521 for (thread = 0; thread < nthreads; thread++) {
522 for (cpu = 0; cpu < ncpus; cpu++) {
523 if (perf_evsel__read_cb(counter, cpu, thread, read_cb))
531 static void print_interval(void)
533 static int num_print_interval;
534 struct perf_evsel *counter;
535 struct perf_stat *ps;
536 struct timespec ts, rs;
539 if (aggr_mode == AGGR_GLOBAL) {
540 evlist__for_each(evsel_list, counter) {
542 memset(ps->res_stats, 0, sizeof(ps->res_stats));
543 read_counter_aggr(counter);
546 evlist__for_each(evsel_list, counter) {
548 memset(ps->res_stats, 0, sizeof(ps->res_stats));
549 read_counter(counter);
553 clock_gettime(CLOCK_MONOTONIC, &ts);
554 diff_timespec(&rs, &ts, &ref_time);
555 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
557 if (num_print_interval == 0 && !csv_output) {
560 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
563 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
566 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
570 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
574 if (++num_print_interval == 25)
575 num_print_interval = 0;
583 evlist__for_each(evsel_list, counter)
584 print_counter(counter, prefix);
588 evlist__for_each(evsel_list, counter)
589 print_counter_aggr(counter, prefix);
595 static void handle_initial_delay(void)
597 struct perf_evsel *counter;
600 const int ncpus = cpu_map__nr(evsel_list->cpus),
601 nthreads = thread_map__nr(evsel_list->threads);
603 usleep(initial_delay * 1000);
604 evlist__for_each(evsel_list, counter)
605 perf_evsel__enable(counter, ncpus, nthreads);
609 static volatile int workload_exec_errno;
612 * perf_evlist__prepare_workload will send a SIGUSR1
613 * if the fork fails, since we asked by setting its
614 * want_signal to true.
616 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
617 void *ucontext __maybe_unused)
619 workload_exec_errno = info->si_value.sival_int;
622 static int __run_perf_stat(int argc, const char **argv)
625 unsigned long long t0, t1;
626 struct perf_evsel *counter;
630 const bool forks = (argc > 0);
633 ts.tv_sec = interval / 1000;
634 ts.tv_nsec = (interval % 1000) * 1000000;
641 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
642 workload_exec_failed_signal) < 0) {
643 perror("failed to prepare workload");
646 child_pid = evsel_list->workload.pid;
650 perf_evlist__set_leader(evsel_list);
652 evlist__for_each(evsel_list, counter) {
653 if (create_perf_stat_counter(counter) < 0) {
655 * PPC returns ENXIO for HW counters until 2.6.37
656 * (behavior changed with commit b0a873e).
658 if (errno == EINVAL || errno == ENOSYS ||
659 errno == ENOENT || errno == EOPNOTSUPP ||
662 ui__warning("%s event is not supported by the kernel.\n",
663 perf_evsel__name(counter));
664 counter->supported = false;
668 perf_evsel__open_strerror(counter, &target,
669 errno, msg, sizeof(msg));
670 ui__error("%s\n", msg);
673 kill(child_pid, SIGTERM);
677 counter->supported = true;
679 l = strlen(counter->unit);
684 if (perf_evlist__apply_filters(evsel_list, &counter)) {
685 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
686 counter->filter, perf_evsel__name(counter), errno,
687 strerror_r(errno, msg, sizeof(msg)));
692 * Enable counters and exec the command:
695 clock_gettime(CLOCK_MONOTONIC, &ref_time);
698 perf_evlist__start_workload(evsel_list);
699 handle_initial_delay();
702 while (!waitpid(child_pid, &status, WNOHANG)) {
703 nanosleep(&ts, NULL);
709 if (workload_exec_errno) {
710 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
711 pr_err("Workload failed: %s\n", emsg);
715 if (WIFSIGNALED(status))
716 psignal(WTERMSIG(status), argv[0]);
718 handle_initial_delay();
720 nanosleep(&ts, NULL);
728 update_stats(&walltime_nsecs_stats, t1 - t0);
730 if (aggr_mode == AGGR_GLOBAL) {
731 evlist__for_each(evsel_list, counter) {
732 read_counter_aggr(counter);
733 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
734 thread_map__nr(evsel_list->threads));
737 evlist__for_each(evsel_list, counter) {
738 read_counter(counter);
739 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
743 return WEXITSTATUS(status);
746 static int run_perf_stat(int argc, const char **argv)
751 ret = system(pre_cmd);
759 ret = __run_perf_stat(argc, argv);
764 ret = system(post_cmd);
772 static void print_running(u64 run, u64 ena)
775 fprintf(output, "%s%" PRIu64 "%s%.2f",
779 ena ? 100.0 * run / ena : 100.0);
780 } else if (run != ena) {
781 fprintf(output, " (%.2f%%)", 100.0 * run / ena);
785 static void print_noise_pct(double total, double avg)
787 double pct = rel_stddev_stats(total, avg);
790 fprintf(output, "%s%.2f%%", csv_sep, pct);
792 fprintf(output, " ( +-%6.2f%% )", pct);
795 static void print_noise(struct perf_evsel *evsel, double avg)
797 struct perf_stat *ps;
803 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
806 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
810 fprintf(output, "S%d-C%*d%s%*d%s",
811 cpu_map__id_to_socket(id),
813 cpu_map__id_to_cpu(id),
820 fprintf(output, "S%*d%s%*d%s",
829 fprintf(output, "CPU%*d%s",
831 perf_evsel__cpus(evsel)->map[id], csv_sep);
839 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
841 double msecs = avg / 1e6;
842 const char *fmt_v, *fmt_n;
845 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
846 fmt_n = csv_output ? "%s" : "%-25s";
848 aggr_printout(evsel, id, nr);
850 scnprintf(name, sizeof(name), "%s%s",
851 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
853 fprintf(output, fmt_v, msecs, csv_sep);
856 fprintf(output, "%s%s", evsel->unit, csv_sep);
858 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
860 fprintf(output, fmt_n, name);
863 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
865 if (csv_output || interval)
868 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
869 fprintf(output, " # %8.3f CPUs utilized ",
870 avg / avg_stats(&walltime_nsecs_stats));
872 fprintf(output, " ");
875 /* used for get_ratio_color() */
877 GRC_STALLED_CYCLES_FE,
878 GRC_STALLED_CYCLES_BE,
883 static const char *get_ratio_color(enum grc_type type, double ratio)
885 static const double grc_table[GRC_MAX_NR][3] = {
886 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
887 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
888 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
890 const char *color = PERF_COLOR_NORMAL;
892 if (ratio > grc_table[type][0])
893 color = PERF_COLOR_RED;
894 else if (ratio > grc_table[type][1])
895 color = PERF_COLOR_MAGENTA;
896 else if (ratio > grc_table[type][2])
897 color = PERF_COLOR_YELLOW;
902 static void print_stalled_cycles_frontend(FILE *out, int cpu,
903 struct perf_evsel *evsel
904 __maybe_unused, double avg)
906 double total, ratio = 0.0;
908 int ctx = evsel_context(evsel);
910 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
913 ratio = avg / total * 100.0;
915 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
918 color_fprintf(out, color, "%6.2f%%", ratio);
919 fprintf(out, " frontend cycles idle ");
922 static void print_stalled_cycles_backend(FILE *out, int cpu,
923 struct perf_evsel *evsel
924 __maybe_unused, double avg)
926 double total, ratio = 0.0;
928 int ctx = evsel_context(evsel);
930 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
933 ratio = avg / total * 100.0;
935 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
938 color_fprintf(out, color, "%6.2f%%", ratio);
939 fprintf(out, " backend cycles idle ");
942 static void print_branch_misses(FILE *out, int cpu,
943 struct perf_evsel *evsel __maybe_unused,
946 double total, ratio = 0.0;
948 int ctx = evsel_context(evsel);
950 total = avg_stats(&runtime_branches_stats[ctx][cpu]);
953 ratio = avg / total * 100.0;
955 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
958 color_fprintf(out, color, "%6.2f%%", ratio);
959 fprintf(out, " of all branches ");
962 static void print_l1_dcache_misses(FILE *out, int cpu,
963 struct perf_evsel *evsel __maybe_unused,
966 double total, ratio = 0.0;
968 int ctx = evsel_context(evsel);
970 total = avg_stats(&runtime_l1_dcache_stats[ctx][cpu]);
973 ratio = avg / total * 100.0;
975 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
978 color_fprintf(out, color, "%6.2f%%", ratio);
979 fprintf(out, " of all L1-dcache hits ");
982 static void print_l1_icache_misses(FILE *out, int cpu,
983 struct perf_evsel *evsel __maybe_unused,
986 double total, ratio = 0.0;
988 int ctx = evsel_context(evsel);
990 total = avg_stats(&runtime_l1_icache_stats[ctx][cpu]);
993 ratio = avg / total * 100.0;
995 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
998 color_fprintf(out, color, "%6.2f%%", ratio);
999 fprintf(out, " of all L1-icache hits ");
1002 static void print_dtlb_cache_misses(FILE *out, int cpu,
1003 struct perf_evsel *evsel __maybe_unused,
1006 double total, ratio = 0.0;
1008 int ctx = evsel_context(evsel);
1010 total = avg_stats(&runtime_dtlb_cache_stats[ctx][cpu]);
1013 ratio = avg / total * 100.0;
1015 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
1017 fprintf(out, " # ");
1018 color_fprintf(out, color, "%6.2f%%", ratio);
1019 fprintf(out, " of all dTLB cache hits ");
1022 static void print_itlb_cache_misses(FILE *out, int cpu,
1023 struct perf_evsel *evsel __maybe_unused,
1026 double total, ratio = 0.0;
1028 int ctx = evsel_context(evsel);
1030 total = avg_stats(&runtime_itlb_cache_stats[ctx][cpu]);
1033 ratio = avg / total * 100.0;
1035 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
1037 fprintf(out, " # ");
1038 color_fprintf(out, color, "%6.2f%%", ratio);
1039 fprintf(out, " of all iTLB cache hits ");
1042 static void print_ll_cache_misses(FILE *out, int cpu,
1043 struct perf_evsel *evsel __maybe_unused,
1046 double total, ratio = 0.0;
1048 int ctx = evsel_context(evsel);
1050 total = avg_stats(&runtime_ll_cache_stats[ctx][cpu]);
1053 ratio = avg / total * 100.0;
1055 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
1057 fprintf(out, " # ");
1058 color_fprintf(out, color, "%6.2f%%", ratio);
1059 fprintf(out, " of all LL-cache hits ");
1062 static void print_shadow_stats(FILE *out, struct perf_evsel *evsel, double avg, int cpu)
1064 double total, ratio = 0.0, total2;
1065 int ctx = evsel_context(evsel);
1067 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
1068 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
1070 ratio = avg / total;
1071 fprintf(out, " # %5.2f insns per cycle ", ratio);
1075 total = avg_stats(&runtime_stalled_cycles_front_stats[ctx][cpu]);
1076 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[ctx][cpu]));
1079 ratio = total / avg;
1081 if (aggr_mode == AGGR_NONE)
1083 fprintf(out, " # %5.2f stalled cycles per insn", ratio);
1086 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
1087 runtime_branches_stats[ctx][cpu].n != 0) {
1088 print_branch_misses(out, cpu, evsel, avg);
1090 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1091 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
1092 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1093 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1094 runtime_l1_dcache_stats[ctx][cpu].n != 0) {
1095 print_l1_dcache_misses(out, cpu, evsel, avg);
1097 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1098 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
1099 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1100 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1101 runtime_l1_icache_stats[ctx][cpu].n != 0) {
1102 print_l1_icache_misses(out, cpu, evsel, avg);
1104 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1105 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
1106 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1107 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1108 runtime_dtlb_cache_stats[ctx][cpu].n != 0) {
1109 print_dtlb_cache_misses(out, cpu, evsel, avg);
1111 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1112 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
1113 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1114 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1115 runtime_itlb_cache_stats[ctx][cpu].n != 0) {
1116 print_itlb_cache_misses(out, cpu, evsel, avg);
1118 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1119 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
1120 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1121 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1122 runtime_ll_cache_stats[ctx][cpu].n != 0) {
1123 print_ll_cache_misses(out, cpu, evsel, avg);
1124 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
1125 runtime_cacherefs_stats[ctx][cpu].n != 0) {
1126 total = avg_stats(&runtime_cacherefs_stats[ctx][cpu]);
1129 ratio = avg * 100 / total;
1131 fprintf(out, " # %8.3f %% of all cache refs ", ratio);
1133 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1134 print_stalled_cycles_frontend(out, cpu, evsel, avg);
1135 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1136 print_stalled_cycles_backend(out, cpu, evsel, avg);
1137 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1138 total = avg_stats(&runtime_nsecs_stats[cpu]);
1141 ratio = avg / total;
1142 fprintf(out, " # %8.3f GHz ", ratio);
1146 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
1147 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
1150 " # %5.2f%% transactional cycles ",
1151 100.0 * (avg / total));
1152 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
1153 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
1154 total2 = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
1159 " # %5.2f%% aborted cycles ",
1160 100.0 * ((total2-avg) / total));
1161 } else if (perf_stat_evsel__is(evsel, TRANSACTION_START) &&
1163 runtime_cycles_in_tx_stats[ctx][cpu].n != 0) {
1164 total = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
1167 ratio = total / avg;
1169 fprintf(out, " # %8.0f cycles / transaction ", ratio);
1170 } else if (perf_stat_evsel__is(evsel, ELISION_START) &&
1172 runtime_cycles_in_tx_stats[ctx][cpu].n != 0) {
1173 total = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
1176 ratio = total / avg;
1178 fprintf(out, " # %8.0f cycles / elision ", ratio);
1179 } else if (runtime_nsecs_stats[cpu].n != 0) {
1182 total = avg_stats(&runtime_nsecs_stats[cpu]);
1185 ratio = 1000.0 * avg / total;
1186 if (ratio < 0.001) {
1191 fprintf(out, " # %8.3f %c/sec ", ratio, unit);
1197 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1199 double sc = evsel->scale;
1201 int cpu = cpu_map__id_to_cpu(id);
1204 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
1207 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
1209 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
1212 aggr_printout(evsel, id, nr);
1214 if (aggr_mode == AGGR_GLOBAL)
1217 fprintf(output, fmt, avg, csv_sep);
1220 fprintf(output, "%-*s%s",
1221 csv_output ? 0 : unit_width,
1222 evsel->unit, csv_sep);
1224 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1227 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1229 if (csv_output || interval)
1232 print_shadow_stats(output, evsel, avg, cpu);
1235 static void print_aggr(char *prefix)
1237 struct perf_evsel *counter;
1238 int cpu, cpu2, s, s2, id, nr;
1242 if (!(aggr_map || aggr_get_id))
1245 for (s = 0; s < aggr_map->nr; s++) {
1246 id = aggr_map->map[s];
1247 evlist__for_each(evsel_list, counter) {
1248 val = ena = run = 0;
1250 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1251 cpu2 = perf_evsel__cpus(counter)->map[cpu];
1252 s2 = aggr_get_id(evsel_list->cpus, cpu2);
1255 val += counter->counts->cpu[cpu].val;
1256 ena += counter->counts->cpu[cpu].ena;
1257 run += counter->counts->cpu[cpu].run;
1261 fprintf(output, "%s", prefix);
1263 if (run == 0 || ena == 0) {
1264 aggr_printout(counter, id, nr);
1266 fprintf(output, "%*s%s",
1267 csv_output ? 0 : 18,
1268 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1271 fprintf(output, "%-*s%s",
1272 csv_output ? 0 : unit_width,
1273 counter->unit, csv_sep);
1275 fprintf(output, "%*s",
1276 csv_output ? 0 : -25,
1277 perf_evsel__name(counter));
1280 fprintf(output, "%s%s",
1281 csv_sep, counter->cgrp->name);
1283 print_running(run, ena);
1284 fputc('\n', output);
1287 uval = val * counter->scale;
1289 if (nsec_counter(counter))
1290 nsec_printout(id, nr, counter, uval);
1292 abs_printout(id, nr, counter, uval);
1295 print_noise(counter, 1.0);
1297 print_running(run, ena);
1298 fputc('\n', output);
1304 * Print out the results of a single counter:
1305 * aggregated counts in system-wide mode
1307 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1309 struct perf_stat *ps = counter->priv;
1310 double avg = avg_stats(&ps->res_stats[0]);
1311 int scaled = counter->counts->scaled;
1313 double avg_enabled, avg_running;
1315 avg_enabled = avg_stats(&ps->res_stats[1]);
1316 avg_running = avg_stats(&ps->res_stats[2]);
1319 fprintf(output, "%s", prefix);
1321 if (scaled == -1 || !counter->supported) {
1322 fprintf(output, "%*s%s",
1323 csv_output ? 0 : 18,
1324 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1326 fprintf(output, "%-*s%s",
1327 csv_output ? 0 : unit_width,
1328 counter->unit, csv_sep);
1329 fprintf(output, "%*s",
1330 csv_output ? 0 : -25,
1331 perf_evsel__name(counter));
1334 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1336 print_running(avg_running, avg_enabled);
1337 fputc('\n', output);
1341 uval = avg * counter->scale;
1343 if (nsec_counter(counter))
1344 nsec_printout(-1, 0, counter, uval);
1346 abs_printout(-1, 0, counter, uval);
1348 print_noise(counter, avg);
1350 print_running(avg_running, avg_enabled);
1351 fprintf(output, "\n");
1355 * Print out the results of a single counter:
1356 * does not use aggregated count in system-wide
1358 static void print_counter(struct perf_evsel *counter, char *prefix)
1364 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1365 val = counter->counts->cpu[cpu].val;
1366 ena = counter->counts->cpu[cpu].ena;
1367 run = counter->counts->cpu[cpu].run;
1370 fprintf(output, "%s", prefix);
1372 if (run == 0 || ena == 0) {
1373 fprintf(output, "CPU%*d%s%*s%s",
1374 csv_output ? 0 : -4,
1375 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1376 csv_output ? 0 : 18,
1377 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1380 fprintf(output, "%-*s%s",
1381 csv_output ? 0 : unit_width,
1382 counter->unit, csv_sep);
1384 fprintf(output, "%*s",
1385 csv_output ? 0 : -25,
1386 perf_evsel__name(counter));
1389 fprintf(output, "%s%s",
1390 csv_sep, counter->cgrp->name);
1392 print_running(run, ena);
1393 fputc('\n', output);
1397 uval = val * counter->scale;
1399 if (nsec_counter(counter))
1400 nsec_printout(cpu, 0, counter, uval);
1402 abs_printout(cpu, 0, counter, uval);
1405 print_noise(counter, 1.0);
1406 print_running(run, ena);
1408 fputc('\n', output);
1412 static void print_stat(int argc, const char **argv)
1414 struct perf_evsel *counter;
1420 fprintf(output, "\n");
1421 fprintf(output, " Performance counter stats for ");
1422 if (target.system_wide)
1423 fprintf(output, "\'system wide");
1424 else if (target.cpu_list)
1425 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1426 else if (!target__has_task(&target)) {
1427 fprintf(output, "\'%s", argv[0]);
1428 for (i = 1; i < argc; i++)
1429 fprintf(output, " %s", argv[i]);
1430 } else if (target.pid)
1431 fprintf(output, "process id \'%s", target.pid);
1433 fprintf(output, "thread id \'%s", target.tid);
1435 fprintf(output, "\'");
1437 fprintf(output, " (%d runs)", run_count);
1438 fprintf(output, ":\n\n");
1441 switch (aggr_mode) {
1447 evlist__for_each(evsel_list, counter)
1448 print_counter_aggr(counter, NULL);
1451 evlist__for_each(evsel_list, counter)
1452 print_counter(counter, NULL);
1460 fprintf(output, "\n");
1461 fprintf(output, " %17.9f seconds time elapsed",
1462 avg_stats(&walltime_nsecs_stats)/1e9);
1463 if (run_count > 1) {
1464 fprintf(output, " ");
1465 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1466 avg_stats(&walltime_nsecs_stats));
1468 fprintf(output, "\n\n");
1472 static volatile int signr = -1;
1474 static void skip_signal(int signo)
1476 if ((child_pid == -1) || interval)
1481 * render child_pid harmless
1482 * won't send SIGTERM to a random
1483 * process in case of race condition
1484 * and fast PID recycling
1489 static void sig_atexit(void)
1494 * avoid race condition with SIGCHLD handler
1495 * in skip_signal() which is modifying child_pid
1496 * goal is to avoid send SIGTERM to a random
1500 sigaddset(&set, SIGCHLD);
1501 sigprocmask(SIG_BLOCK, &set, &oset);
1503 if (child_pid != -1)
1504 kill(child_pid, SIGTERM);
1506 sigprocmask(SIG_SETMASK, &oset, NULL);
1511 signal(signr, SIG_DFL);
1512 kill(getpid(), signr);
1515 static int stat__set_big_num(const struct option *opt __maybe_unused,
1516 const char *s __maybe_unused, int unset)
1518 big_num_opt = unset ? 0 : 1;
1522 static int perf_stat_init_aggr_mode(void)
1524 switch (aggr_mode) {
1526 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1527 perror("cannot build socket map");
1530 aggr_get_id = cpu_map__get_socket;
1533 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1534 perror("cannot build core map");
1537 aggr_get_id = cpu_map__get_core;
1548 * Add default attributes, if there were no attributes specified or
1549 * if -d/--detailed, -d -d or -d -d -d is used:
1551 static int add_default_attributes(void)
1553 struct perf_event_attr default_attrs[] = {
1555 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1556 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1557 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1558 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1560 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1561 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1562 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1563 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1564 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1565 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1570 * Detailed stats (-d), covering the L1 and last level data caches:
1572 struct perf_event_attr detailed_attrs[] = {
1574 { .type = PERF_TYPE_HW_CACHE,
1576 PERF_COUNT_HW_CACHE_L1D << 0 |
1577 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1578 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1580 { .type = PERF_TYPE_HW_CACHE,
1582 PERF_COUNT_HW_CACHE_L1D << 0 |
1583 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1584 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1586 { .type = PERF_TYPE_HW_CACHE,
1588 PERF_COUNT_HW_CACHE_LL << 0 |
1589 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1590 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1592 { .type = PERF_TYPE_HW_CACHE,
1594 PERF_COUNT_HW_CACHE_LL << 0 |
1595 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1596 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1600 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1602 struct perf_event_attr very_detailed_attrs[] = {
1604 { .type = PERF_TYPE_HW_CACHE,
1606 PERF_COUNT_HW_CACHE_L1I << 0 |
1607 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1608 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1610 { .type = PERF_TYPE_HW_CACHE,
1612 PERF_COUNT_HW_CACHE_L1I << 0 |
1613 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1614 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1616 { .type = PERF_TYPE_HW_CACHE,
1618 PERF_COUNT_HW_CACHE_DTLB << 0 |
1619 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1620 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1622 { .type = PERF_TYPE_HW_CACHE,
1624 PERF_COUNT_HW_CACHE_DTLB << 0 |
1625 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1626 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1628 { .type = PERF_TYPE_HW_CACHE,
1630 PERF_COUNT_HW_CACHE_ITLB << 0 |
1631 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1632 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1634 { .type = PERF_TYPE_HW_CACHE,
1636 PERF_COUNT_HW_CACHE_ITLB << 0 |
1637 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1638 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1643 * Very, very detailed stats (-d -d -d), adding prefetch events:
1645 struct perf_event_attr very_very_detailed_attrs[] = {
1647 { .type = PERF_TYPE_HW_CACHE,
1649 PERF_COUNT_HW_CACHE_L1D << 0 |
1650 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1651 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1653 { .type = PERF_TYPE_HW_CACHE,
1655 PERF_COUNT_HW_CACHE_L1D << 0 |
1656 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1657 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1660 /* Set attrs if no event is selected and !null_run: */
1664 if (transaction_run) {
1666 if (pmu_have_event("cpu", "cycles-ct") &&
1667 pmu_have_event("cpu", "el-start"))
1668 err = parse_events(evsel_list, transaction_attrs, NULL);
1670 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1672 fprintf(stderr, "Cannot set up transaction events\n");
1678 if (!evsel_list->nr_entries) {
1679 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1683 /* Detailed events get appended to the event list: */
1685 if (detailed_run < 1)
1688 /* Append detailed run extra attributes: */
1689 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1692 if (detailed_run < 2)
1695 /* Append very detailed run extra attributes: */
1696 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1699 if (detailed_run < 3)
1702 /* Append very, very detailed run extra attributes: */
1703 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1706 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1708 bool append_file = false;
1710 const char *output_name = NULL;
1711 const struct option options[] = {
1712 OPT_BOOLEAN('T', "transaction", &transaction_run,
1713 "hardware transaction statistics"),
1714 OPT_CALLBACK('e', "event", &evsel_list, "event",
1715 "event selector. use 'perf list' to list available events",
1716 parse_events_option),
1717 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1718 "event filter", parse_filter),
1719 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1720 "child tasks do not inherit counters"),
1721 OPT_STRING('p', "pid", &target.pid, "pid",
1722 "stat events on existing process id"),
1723 OPT_STRING('t', "tid", &target.tid, "tid",
1724 "stat events on existing thread id"),
1725 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1726 "system-wide collection from all CPUs"),
1727 OPT_BOOLEAN('g', "group", &group,
1728 "put the counters into a counter group"),
1729 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1730 OPT_INCR('v', "verbose", &verbose,
1731 "be more verbose (show counter open errors, etc)"),
1732 OPT_INTEGER('r', "repeat", &run_count,
1733 "repeat command and print average + stddev (max: 100, forever: 0)"),
1734 OPT_BOOLEAN('n', "null", &null_run,
1735 "null run - dont start any counters"),
1736 OPT_INCR('d', "detailed", &detailed_run,
1737 "detailed run - start a lot of events"),
1738 OPT_BOOLEAN('S', "sync", &sync_run,
1739 "call sync() before starting a run"),
1740 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1741 "print large numbers with thousands\' separators",
1743 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1744 "list of cpus to monitor in system-wide"),
1745 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1746 "disable CPU count aggregation", AGGR_NONE),
1747 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1748 "print counts with custom separator"),
1749 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1750 "monitor event in cgroup name only", parse_cgroups),
1751 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1752 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1753 OPT_INTEGER(0, "log-fd", &output_fd,
1754 "log output to fd, instead of stderr"),
1755 OPT_STRING(0, "pre", &pre_cmd, "command",
1756 "command to run prior to the measured command"),
1757 OPT_STRING(0, "post", &post_cmd, "command",
1758 "command to run after to the measured command"),
1759 OPT_UINTEGER('I', "interval-print", &interval,
1760 "print counts at regular interval in ms (>= 100)"),
1761 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1762 "aggregate counts per processor socket", AGGR_SOCKET),
1763 OPT_SET_UINT(0, "per-core", &aggr_mode,
1764 "aggregate counts per physical processor core", AGGR_CORE),
1765 OPT_UINTEGER('D', "delay", &initial_delay,
1766 "ms to wait before starting measurement after program start"),
1769 const char * const stat_usage[] = {
1770 "perf stat [<options>] [<command>]",
1773 int status = -EINVAL, run_idx;
1776 setlocale(LC_ALL, "");
1778 evsel_list = perf_evlist__new();
1779 if (evsel_list == NULL)
1782 argc = parse_options(argc, argv, options, stat_usage,
1783 PARSE_OPT_STOP_AT_NON_OPTION);
1786 if (output_name && strcmp(output_name, "-"))
1789 if (output_name && output_fd) {
1790 fprintf(stderr, "cannot use both --output and --log-fd\n");
1791 parse_options_usage(stat_usage, options, "o", 1);
1792 parse_options_usage(NULL, options, "log-fd", 0);
1796 if (output_fd < 0) {
1797 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1798 parse_options_usage(stat_usage, options, "log-fd", 0);
1804 mode = append_file ? "a" : "w";
1806 output = fopen(output_name, mode);
1808 perror("failed to create output file");
1811 clock_gettime(CLOCK_REALTIME, &tm);
1812 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1813 } else if (output_fd > 0) {
1814 mode = append_file ? "a" : "w";
1815 output = fdopen(output_fd, mode);
1817 perror("Failed opening logfd");
1824 if (!strcmp(csv_sep, "\\t"))
1827 csv_sep = DEFAULT_SEPARATOR;
1830 * let the spreadsheet do the pretty-printing
1833 /* User explicitly passed -B? */
1834 if (big_num_opt == 1) {
1835 fprintf(stderr, "-B option not supported with -x\n");
1836 parse_options_usage(stat_usage, options, "B", 1);
1837 parse_options_usage(NULL, options, "x", 1);
1839 } else /* Nope, so disable big number formatting */
1841 } else if (big_num_opt == 0) /* User passed --no-big-num */
1844 if (!argc && target__none(&target))
1845 usage_with_options(stat_usage, options);
1847 if (run_count < 0) {
1848 pr_err("Run count must be a positive number\n");
1849 parse_options_usage(stat_usage, options, "r", 1);
1851 } else if (run_count == 0) {
1856 /* no_aggr, cgroup are for system-wide only */
1857 if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1858 !target__has_cpu(&target)) {
1859 fprintf(stderr, "both cgroup and no-aggregation "
1860 "modes only available in system-wide mode\n");
1862 parse_options_usage(stat_usage, options, "G", 1);
1863 parse_options_usage(NULL, options, "A", 1);
1864 parse_options_usage(NULL, options, "a", 1);
1868 if (add_default_attributes())
1871 target__validate(&target);
1873 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1874 if (target__has_task(&target)) {
1875 pr_err("Problems finding threads of monitor\n");
1876 parse_options_usage(stat_usage, options, "p", 1);
1877 parse_options_usage(NULL, options, "t", 1);
1878 } else if (target__has_cpu(&target)) {
1879 perror("failed to parse CPUs map");
1880 parse_options_usage(stat_usage, options, "C", 1);
1881 parse_options_usage(NULL, options, "a", 1);
1885 if (interval && interval < 100) {
1886 pr_err("print interval must be >= 100ms\n");
1887 parse_options_usage(stat_usage, options, "I", 1);
1891 if (perf_evlist__alloc_stats(evsel_list, interval))
1894 if (perf_stat_init_aggr_mode())
1898 * We dont want to block the signals - that would cause
1899 * child tasks to inherit that and Ctrl-C would not work.
1900 * What we want is for Ctrl-C to work in the exec()-ed
1901 * task, but being ignored by perf stat itself:
1905 signal(SIGINT, skip_signal);
1906 signal(SIGCHLD, skip_signal);
1907 signal(SIGALRM, skip_signal);
1908 signal(SIGABRT, skip_signal);
1911 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1912 if (run_count != 1 && verbose)
1913 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1916 status = run_perf_stat(argc, argv);
1917 if (forever && status != -1) {
1918 print_stat(argc, argv);
1919 perf_stat__reset_stats(evsel_list);
1923 if (!forever && status != -1 && !interval)
1924 print_stat(argc, argv);
1926 perf_evlist__free_stats(evsel_list);
1928 perf_evlist__delete(evsel_list);