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/header.h"
55 #include "util/cpumap.h"
56 #include "util/thread.h"
57 #include "util/thread_map.h"
59 #include <sys/prctl.h>
63 #define DEFAULT_SEPARATOR " "
64 #define CNTR_NOT_SUPPORTED "<not supported>"
65 #define CNTR_NOT_COUNTED "<not counted>"
67 static struct perf_event_attr default_attrs[] = {
69 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
70 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
71 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
72 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
74 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
75 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
76 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
77 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
78 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
79 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
84 * Detailed stats (-d), covering the L1 and last level data caches:
86 static struct perf_event_attr detailed_attrs[] = {
88 { .type = PERF_TYPE_HW_CACHE,
90 PERF_COUNT_HW_CACHE_L1D << 0 |
91 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
92 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
94 { .type = PERF_TYPE_HW_CACHE,
96 PERF_COUNT_HW_CACHE_L1D << 0 |
97 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
98 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
100 { .type = PERF_TYPE_HW_CACHE,
102 PERF_COUNT_HW_CACHE_LL << 0 |
103 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
104 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
106 { .type = PERF_TYPE_HW_CACHE,
108 PERF_COUNT_HW_CACHE_LL << 0 |
109 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
110 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
114 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
116 static struct perf_event_attr very_detailed_attrs[] = {
118 { .type = PERF_TYPE_HW_CACHE,
120 PERF_COUNT_HW_CACHE_L1I << 0 |
121 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
122 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
124 { .type = PERF_TYPE_HW_CACHE,
126 PERF_COUNT_HW_CACHE_L1I << 0 |
127 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
128 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
130 { .type = PERF_TYPE_HW_CACHE,
132 PERF_COUNT_HW_CACHE_DTLB << 0 |
133 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
134 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
136 { .type = PERF_TYPE_HW_CACHE,
138 PERF_COUNT_HW_CACHE_DTLB << 0 |
139 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
140 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
142 { .type = PERF_TYPE_HW_CACHE,
144 PERF_COUNT_HW_CACHE_ITLB << 0 |
145 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
146 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
148 { .type = PERF_TYPE_HW_CACHE,
150 PERF_COUNT_HW_CACHE_ITLB << 0 |
151 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
152 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
157 * Very, very detailed stats (-d -d -d), adding prefetch events:
159 static struct perf_event_attr very_very_detailed_attrs[] = {
161 { .type = PERF_TYPE_HW_CACHE,
163 PERF_COUNT_HW_CACHE_L1D << 0 |
164 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
165 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
167 { .type = PERF_TYPE_HW_CACHE,
169 PERF_COUNT_HW_CACHE_L1D << 0 |
170 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
171 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
176 static struct perf_evlist *evsel_list;
178 static struct perf_target target = {
182 static int run_idx = 0;
183 static int run_count = 1;
184 static bool no_inherit = false;
185 static bool scale = true;
186 static bool no_aggr = false;
187 static pid_t child_pid = -1;
188 static bool null_run = false;
189 static int detailed_run = 0;
190 static bool sync_run = false;
191 static bool big_num = true;
192 static int big_num_opt = -1;
193 static const char *csv_sep = NULL;
194 static bool csv_output = false;
195 static bool group = false;
196 static const char *output_name = NULL;
197 static FILE *output = NULL;
198 static int output_fd;
200 static volatile int done = 0;
208 struct stats res_stats[3];
211 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
213 evsel->priv = zalloc(sizeof(struct perf_stat));
214 return evsel->priv == NULL ? -ENOMEM : 0;
217 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
223 static void update_stats(struct stats *stats, u64 val)
228 delta = val - stats->mean;
229 stats->mean += delta / stats->n;
230 stats->M2 += delta*(val - stats->mean);
233 static double avg_stats(struct stats *stats)
239 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
241 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
242 * s^2 = -------------------------------
245 * http://en.wikipedia.org/wiki/Stddev
247 * The std dev of the mean is related to the std dev by:
254 static double stddev_stats(struct stats *stats)
256 double variance, variance_mean;
261 variance = stats->M2 / (stats->n - 1);
262 variance_mean = variance / stats->n;
264 return sqrt(variance_mean);
267 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
268 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
269 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
270 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
271 static struct stats runtime_branches_stats[MAX_NR_CPUS];
272 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
273 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
274 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
275 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
276 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
277 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
278 static struct stats walltime_nsecs_stats;
280 static int create_perf_stat_counter(struct perf_evsel *evsel,
281 struct perf_evsel *first)
283 struct perf_event_attr *attr = &evsel->attr;
284 bool exclude_guest_missing = false;
288 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
289 PERF_FORMAT_TOTAL_TIME_RUNNING;
291 attr->inherit = !no_inherit;
294 if (exclude_guest_missing)
295 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
297 if (perf_target__has_cpu(&target)) {
298 ret = perf_evsel__open_per_cpu(evsel, evsel_list->cpus);
304 if (!perf_target__has_task(&target) && (!group || evsel == first)) {
306 attr->enable_on_exec = 1;
309 ret = perf_evsel__open_per_thread(evsel, evsel_list->threads);
314 if (ret && errno == EINVAL) {
315 if (!exclude_guest_missing &&
316 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
317 pr_debug("Old kernel, cannot exclude "
318 "guest or host samples.\n");
319 exclude_guest_missing = true;
327 * Does the counter have nsecs as a unit?
329 static inline int nsec_counter(struct perf_evsel *evsel)
331 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
332 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
339 * Update various tracking values we maintain to print
340 * more semantic information such as miss/hit ratios,
341 * instruction rates, etc:
343 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
345 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
346 update_stats(&runtime_nsecs_stats[0], count[0]);
347 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
348 update_stats(&runtime_cycles_stats[0], count[0]);
349 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
350 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
351 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
352 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
353 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
354 update_stats(&runtime_branches_stats[0], count[0]);
355 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
356 update_stats(&runtime_cacherefs_stats[0], count[0]);
357 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
358 update_stats(&runtime_l1_dcache_stats[0], count[0]);
359 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
360 update_stats(&runtime_l1_icache_stats[0], count[0]);
361 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
362 update_stats(&runtime_ll_cache_stats[0], count[0]);
363 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
364 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
365 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
366 update_stats(&runtime_itlb_cache_stats[0], count[0]);
370 * Read out the results of a single counter:
371 * aggregate counts across CPUs in system-wide mode
373 static int read_counter_aggr(struct perf_evsel *counter)
375 struct perf_stat *ps = counter->priv;
376 u64 *count = counter->counts->aggr.values;
379 if (__perf_evsel__read(counter, evsel_list->cpus->nr,
380 evsel_list->threads->nr, scale) < 0)
383 for (i = 0; i < 3; i++)
384 update_stats(&ps->res_stats[i], count[i]);
387 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
388 perf_evsel__name(counter), count[0], count[1], count[2]);
392 * Save the full runtime - to allow normalization during printout:
394 update_shadow_stats(counter, count);
400 * Read out the results of a single counter:
401 * do not aggregate counts across CPUs in system-wide mode
403 static int read_counter(struct perf_evsel *counter)
408 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
409 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
412 count = counter->counts->cpu[cpu].values;
414 update_shadow_stats(counter, count);
420 static int run_perf_stat(int argc __maybe_unused, const char **argv)
422 unsigned long long t0, t1;
423 struct perf_evsel *counter, *first;
425 int child_ready_pipe[2], go_pipe[2];
426 const bool forks = (argc > 0);
429 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
430 perror("failed to create pipes");
435 if ((child_pid = fork()) < 0)
436 perror("failed to fork");
439 close(child_ready_pipe[0]);
441 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
444 * Do a dummy execvp to get the PLT entry resolved,
445 * so we avoid the resolver overhead on the real
448 execvp("", (char **)argv);
451 * Tell the parent we're ready to go
453 close(child_ready_pipe[1]);
456 * Wait until the parent tells us to go.
458 if (read(go_pipe[0], &buf, 1) == -1)
459 perror("unable to read pipe");
461 execvp(argv[0], (char **)argv);
467 if (perf_target__none(&target))
468 evsel_list->threads->map[0] = child_pid;
471 * Wait for the child to be ready to exec.
473 close(child_ready_pipe[1]);
475 if (read(child_ready_pipe[0], &buf, 1) == -1)
476 perror("unable to read pipe");
477 close(child_ready_pipe[0]);
481 perf_evlist__set_leader(evsel_list);
483 first = perf_evlist__first(evsel_list);
485 list_for_each_entry(counter, &evsel_list->entries, node) {
486 if (create_perf_stat_counter(counter, first) < 0) {
488 * PPC returns ENXIO for HW counters until 2.6.37
489 * (behavior changed with commit b0a873e).
491 if (errno == EINVAL || errno == ENOSYS ||
492 errno == ENOENT || errno == EOPNOTSUPP ||
495 ui__warning("%s event is not supported by the kernel.\n",
496 perf_evsel__name(counter));
497 counter->supported = false;
501 if (errno == EPERM || errno == EACCES) {
502 error("You may not have permission to collect %sstats.\n"
503 "\t Consider tweaking"
504 " /proc/sys/kernel/perf_event_paranoid or running as root.",
505 target.system_wide ? "system-wide " : "");
507 error("open_counter returned with %d (%s). "
508 "/bin/dmesg may provide additional information.\n",
509 errno, strerror(errno));
512 kill(child_pid, SIGTERM);
514 pr_err("Not all events could be opened.\n");
517 counter->supported = true;
520 if (perf_evlist__set_filters(evsel_list)) {
521 error("failed to set filter with %d (%s)\n", errno,
527 * Enable counters and exec the command:
534 if (WIFSIGNALED(status))
535 psignal(WTERMSIG(status), argv[0]);
537 while(!done) sleep(1);
542 update_stats(&walltime_nsecs_stats, t1 - t0);
545 list_for_each_entry(counter, &evsel_list->entries, node) {
546 read_counter(counter);
547 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
550 list_for_each_entry(counter, &evsel_list->entries, node) {
551 read_counter_aggr(counter);
552 perf_evsel__close_fd(counter, evsel_list->cpus->nr,
553 evsel_list->threads->nr);
557 return WEXITSTATUS(status);
560 static void print_noise_pct(double total, double avg)
565 pct = 100.0*total/avg;
568 fprintf(output, "%s%.2f%%", csv_sep, pct);
570 fprintf(output, " ( +-%6.2f%% )", pct);
573 static void print_noise(struct perf_evsel *evsel, double avg)
575 struct perf_stat *ps;
581 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
584 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
586 double msecs = avg / 1e6;
587 char cpustr[16] = { '\0', };
588 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
591 sprintf(cpustr, "CPU%*d%s",
593 evsel_list->cpus->map[cpu], csv_sep);
595 fprintf(output, fmt, cpustr, msecs, csv_sep, perf_evsel__name(evsel));
598 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
603 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
604 fprintf(output, " # %8.3f CPUs utilized ",
605 avg / avg_stats(&walltime_nsecs_stats));
607 fprintf(output, " ");
610 /* used for get_ratio_color() */
612 GRC_STALLED_CYCLES_FE,
613 GRC_STALLED_CYCLES_BE,
618 static const char *get_ratio_color(enum grc_type type, double ratio)
620 static const double grc_table[GRC_MAX_NR][3] = {
621 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
622 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
623 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
625 const char *color = PERF_COLOR_NORMAL;
627 if (ratio > grc_table[type][0])
628 color = PERF_COLOR_RED;
629 else if (ratio > grc_table[type][1])
630 color = PERF_COLOR_MAGENTA;
631 else if (ratio > grc_table[type][2])
632 color = PERF_COLOR_YELLOW;
637 static void print_stalled_cycles_frontend(int cpu,
638 struct perf_evsel *evsel
639 __maybe_unused, double avg)
641 double total, ratio = 0.0;
644 total = avg_stats(&runtime_cycles_stats[cpu]);
647 ratio = avg / total * 100.0;
649 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
651 fprintf(output, " # ");
652 color_fprintf(output, color, "%6.2f%%", ratio);
653 fprintf(output, " frontend cycles idle ");
656 static void print_stalled_cycles_backend(int cpu,
657 struct perf_evsel *evsel
658 __maybe_unused, double avg)
660 double total, ratio = 0.0;
663 total = avg_stats(&runtime_cycles_stats[cpu]);
666 ratio = avg / total * 100.0;
668 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
670 fprintf(output, " # ");
671 color_fprintf(output, color, "%6.2f%%", ratio);
672 fprintf(output, " backend cycles idle ");
675 static void print_branch_misses(int cpu,
676 struct perf_evsel *evsel __maybe_unused,
679 double total, ratio = 0.0;
682 total = avg_stats(&runtime_branches_stats[cpu]);
685 ratio = avg / total * 100.0;
687 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
689 fprintf(output, " # ");
690 color_fprintf(output, color, "%6.2f%%", ratio);
691 fprintf(output, " of all branches ");
694 static void print_l1_dcache_misses(int cpu,
695 struct perf_evsel *evsel __maybe_unused,
698 double total, ratio = 0.0;
701 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
704 ratio = avg / total * 100.0;
706 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
708 fprintf(output, " # ");
709 color_fprintf(output, color, "%6.2f%%", ratio);
710 fprintf(output, " of all L1-dcache hits ");
713 static void print_l1_icache_misses(int cpu,
714 struct perf_evsel *evsel __maybe_unused,
717 double total, ratio = 0.0;
720 total = avg_stats(&runtime_l1_icache_stats[cpu]);
723 ratio = avg / total * 100.0;
725 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
727 fprintf(output, " # ");
728 color_fprintf(output, color, "%6.2f%%", ratio);
729 fprintf(output, " of all L1-icache hits ");
732 static void print_dtlb_cache_misses(int cpu,
733 struct perf_evsel *evsel __maybe_unused,
736 double total, ratio = 0.0;
739 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
742 ratio = avg / total * 100.0;
744 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
746 fprintf(output, " # ");
747 color_fprintf(output, color, "%6.2f%%", ratio);
748 fprintf(output, " of all dTLB cache hits ");
751 static void print_itlb_cache_misses(int cpu,
752 struct perf_evsel *evsel __maybe_unused,
755 double total, ratio = 0.0;
758 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
761 ratio = avg / total * 100.0;
763 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
765 fprintf(output, " # ");
766 color_fprintf(output, color, "%6.2f%%", ratio);
767 fprintf(output, " of all iTLB cache hits ");
770 static void print_ll_cache_misses(int cpu,
771 struct perf_evsel *evsel __maybe_unused,
774 double total, ratio = 0.0;
777 total = avg_stats(&runtime_ll_cache_stats[cpu]);
780 ratio = avg / total * 100.0;
782 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
784 fprintf(output, " # ");
785 color_fprintf(output, color, "%6.2f%%", ratio);
786 fprintf(output, " of all LL-cache hits ");
789 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
791 double total, ratio = 0.0;
792 char cpustr[16] = { '\0', };
798 fmt = "%s%'18.0f%s%-25s";
800 fmt = "%s%18.0f%s%-25s";
803 sprintf(cpustr, "CPU%*d%s",
805 evsel_list->cpus->map[cpu], csv_sep);
809 fprintf(output, fmt, cpustr, avg, csv_sep, perf_evsel__name(evsel));
812 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
817 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
818 total = avg_stats(&runtime_cycles_stats[cpu]);
823 fprintf(output, " # %5.2f insns per cycle ", ratio);
825 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
826 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
830 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
833 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
834 runtime_branches_stats[cpu].n != 0) {
835 print_branch_misses(cpu, evsel, avg);
837 evsel->attr.type == PERF_TYPE_HW_CACHE &&
838 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
839 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
840 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
841 runtime_l1_dcache_stats[cpu].n != 0) {
842 print_l1_dcache_misses(cpu, evsel, avg);
844 evsel->attr.type == PERF_TYPE_HW_CACHE &&
845 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
846 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
847 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
848 runtime_l1_icache_stats[cpu].n != 0) {
849 print_l1_icache_misses(cpu, evsel, avg);
851 evsel->attr.type == PERF_TYPE_HW_CACHE &&
852 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
853 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
854 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
855 runtime_dtlb_cache_stats[cpu].n != 0) {
856 print_dtlb_cache_misses(cpu, evsel, avg);
858 evsel->attr.type == PERF_TYPE_HW_CACHE &&
859 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
860 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
861 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
862 runtime_itlb_cache_stats[cpu].n != 0) {
863 print_itlb_cache_misses(cpu, evsel, avg);
865 evsel->attr.type == PERF_TYPE_HW_CACHE &&
866 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
867 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
868 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
869 runtime_ll_cache_stats[cpu].n != 0) {
870 print_ll_cache_misses(cpu, evsel, avg);
871 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
872 runtime_cacherefs_stats[cpu].n != 0) {
873 total = avg_stats(&runtime_cacherefs_stats[cpu]);
876 ratio = avg * 100 / total;
878 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
880 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
881 print_stalled_cycles_frontend(cpu, evsel, avg);
882 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
883 print_stalled_cycles_backend(cpu, evsel, avg);
884 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
885 total = avg_stats(&runtime_nsecs_stats[cpu]);
888 ratio = 1.0 * avg / total;
890 fprintf(output, " # %8.3f GHz ", ratio);
891 } else if (runtime_nsecs_stats[cpu].n != 0) {
894 total = avg_stats(&runtime_nsecs_stats[cpu]);
897 ratio = 1000.0 * avg / total;
903 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
905 fprintf(output, " ");
910 * Print out the results of a single counter:
911 * aggregated counts in system-wide mode
913 static void print_counter_aggr(struct perf_evsel *counter)
915 struct perf_stat *ps = counter->priv;
916 double avg = avg_stats(&ps->res_stats[0]);
917 int scaled = counter->counts->scaled;
920 fprintf(output, "%*s%s%*s",
922 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
924 csv_output ? 0 : -24,
925 perf_evsel__name(counter));
928 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
934 if (nsec_counter(counter))
935 nsec_printout(-1, counter, avg);
937 abs_printout(-1, counter, avg);
939 print_noise(counter, avg);
947 double avg_enabled, avg_running;
949 avg_enabled = avg_stats(&ps->res_stats[1]);
950 avg_running = avg_stats(&ps->res_stats[2]);
952 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
954 fprintf(output, "\n");
958 * Print out the results of a single counter:
959 * does not use aggregated count in system-wide
961 static void print_counter(struct perf_evsel *counter)
966 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
967 val = counter->counts->cpu[cpu].val;
968 ena = counter->counts->cpu[cpu].ena;
969 run = counter->counts->cpu[cpu].run;
970 if (run == 0 || ena == 0) {
971 fprintf(output, "CPU%*d%s%*s%s%*s",
973 evsel_list->cpus->map[cpu], csv_sep,
975 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
977 csv_output ? 0 : -24,
978 perf_evsel__name(counter));
981 fprintf(output, "%s%s",
982 csv_sep, counter->cgrp->name);
988 if (nsec_counter(counter))
989 nsec_printout(cpu, counter, val);
991 abs_printout(cpu, counter, val);
994 print_noise(counter, 1.0);
997 fprintf(output, " (%.2f%%)",
1000 fputc('\n', output);
1004 static void print_stat(int argc, const char **argv)
1006 struct perf_evsel *counter;
1012 fprintf(output, "\n");
1013 fprintf(output, " Performance counter stats for ");
1014 if (!perf_target__has_task(&target)) {
1015 fprintf(output, "\'%s", argv[0]);
1016 for (i = 1; i < argc; i++)
1017 fprintf(output, " %s", argv[i]);
1018 } else if (target.pid)
1019 fprintf(output, "process id \'%s", target.pid);
1021 fprintf(output, "thread id \'%s", target.tid);
1023 fprintf(output, "\'");
1025 fprintf(output, " (%d runs)", run_count);
1026 fprintf(output, ":\n\n");
1030 list_for_each_entry(counter, &evsel_list->entries, node)
1031 print_counter(counter);
1033 list_for_each_entry(counter, &evsel_list->entries, node)
1034 print_counter_aggr(counter);
1039 fprintf(output, "\n");
1040 fprintf(output, " %17.9f seconds time elapsed",
1041 avg_stats(&walltime_nsecs_stats)/1e9);
1042 if (run_count > 1) {
1043 fprintf(output, " ");
1044 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1045 avg_stats(&walltime_nsecs_stats));
1047 fprintf(output, "\n\n");
1051 static volatile int signr = -1;
1053 static void skip_signal(int signo)
1061 static void sig_atexit(void)
1063 if (child_pid != -1)
1064 kill(child_pid, SIGTERM);
1069 signal(signr, SIG_DFL);
1070 kill(getpid(), signr);
1073 static const char * const stat_usage[] = {
1074 "perf stat [<options>] [<command>]",
1078 static int stat__set_big_num(const struct option *opt __maybe_unused,
1079 const char *s __maybe_unused, int unset)
1081 big_num_opt = unset ? 0 : 1;
1085 static bool append_file;
1087 static const struct option options[] = {
1088 OPT_CALLBACK('e', "event", &evsel_list, "event",
1089 "event selector. use 'perf list' to list available events",
1090 parse_events_option),
1091 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1092 "event filter", parse_filter),
1093 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1094 "child tasks do not inherit counters"),
1095 OPT_STRING('p', "pid", &target.pid, "pid",
1096 "stat events on existing process id"),
1097 OPT_STRING('t', "tid", &target.tid, "tid",
1098 "stat events on existing thread id"),
1099 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1100 "system-wide collection from all CPUs"),
1101 OPT_BOOLEAN('g', "group", &group,
1102 "put the counters into a counter group"),
1103 OPT_BOOLEAN('c', "scale", &scale,
1104 "scale/normalize counters"),
1105 OPT_INCR('v', "verbose", &verbose,
1106 "be more verbose (show counter open errors, etc)"),
1107 OPT_INTEGER('r', "repeat", &run_count,
1108 "repeat command and print average + stddev (max: 100)"),
1109 OPT_BOOLEAN('n', "null", &null_run,
1110 "null run - dont start any counters"),
1111 OPT_INCR('d', "detailed", &detailed_run,
1112 "detailed run - start a lot of events"),
1113 OPT_BOOLEAN('S', "sync", &sync_run,
1114 "call sync() before starting a run"),
1115 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1116 "print large numbers with thousands\' separators",
1118 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1119 "list of cpus to monitor in system-wide"),
1120 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1121 "disable CPU count aggregation"),
1122 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1123 "print counts with custom separator"),
1124 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1125 "monitor event in cgroup name only",
1127 OPT_STRING('o', "output", &output_name, "file",
1128 "output file name"),
1129 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1130 OPT_INTEGER(0, "log-fd", &output_fd,
1131 "log output to fd, instead of stderr"),
1136 * Add default attributes, if there were no attributes specified or
1137 * if -d/--detailed, -d -d or -d -d -d is used:
1139 static int add_default_attributes(void)
1141 /* Set attrs if no event is selected and !null_run: */
1145 if (!evsel_list->nr_entries) {
1146 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1150 /* Detailed events get appended to the event list: */
1152 if (detailed_run < 1)
1155 /* Append detailed run extra attributes: */
1156 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1159 if (detailed_run < 2)
1162 /* Append very detailed run extra attributes: */
1163 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1166 if (detailed_run < 3)
1169 /* Append very, very detailed run extra attributes: */
1170 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1173 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1175 struct perf_evsel *pos;
1176 int status = -ENOMEM;
1179 setlocale(LC_ALL, "");
1181 evsel_list = perf_evlist__new(NULL, NULL);
1182 if (evsel_list == NULL)
1185 argc = parse_options(argc, argv, options, stat_usage,
1186 PARSE_OPT_STOP_AT_NON_OPTION);
1189 if (output_name && strcmp(output_name, "-"))
1192 if (output_name && output_fd) {
1193 fprintf(stderr, "cannot use both --output and --log-fd\n");
1194 usage_with_options(stat_usage, options);
1197 if (output_fd < 0) {
1198 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1199 usage_with_options(stat_usage, options);
1204 mode = append_file ? "a" : "w";
1206 output = fopen(output_name, mode);
1208 perror("failed to create output file");
1211 clock_gettime(CLOCK_REALTIME, &tm);
1212 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1213 } else if (output_fd > 0) {
1214 mode = append_file ? "a" : "w";
1215 output = fdopen(output_fd, mode);
1217 perror("Failed opening logfd");
1224 if (!strcmp(csv_sep, "\\t"))
1227 csv_sep = DEFAULT_SEPARATOR;
1230 * let the spreadsheet do the pretty-printing
1233 /* User explicitly passed -B? */
1234 if (big_num_opt == 1) {
1235 fprintf(stderr, "-B option not supported with -x\n");
1236 usage_with_options(stat_usage, options);
1237 } else /* Nope, so disable big number formatting */
1239 } else if (big_num_opt == 0) /* User passed --no-big-num */
1242 if (!argc && !perf_target__has_task(&target))
1243 usage_with_options(stat_usage, options);
1245 usage_with_options(stat_usage, options);
1247 /* no_aggr, cgroup are for system-wide only */
1248 if ((no_aggr || nr_cgroups) && !perf_target__has_cpu(&target)) {
1249 fprintf(stderr, "both cgroup and no-aggregation "
1250 "modes only available in system-wide mode\n");
1252 usage_with_options(stat_usage, options);
1255 if (add_default_attributes())
1258 perf_target__validate(&target);
1260 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1261 if (perf_target__has_task(&target))
1262 pr_err("Problems finding threads of monitor\n");
1263 if (perf_target__has_cpu(&target))
1264 perror("failed to parse CPUs map");
1266 usage_with_options(stat_usage, options);
1270 list_for_each_entry(pos, &evsel_list->entries, node) {
1271 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1272 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0)
1277 * We dont want to block the signals - that would cause
1278 * child tasks to inherit that and Ctrl-C would not work.
1279 * What we want is for Ctrl-C to work in the exec()-ed
1280 * task, but being ignored by perf stat itself:
1283 signal(SIGINT, skip_signal);
1284 signal(SIGALRM, skip_signal);
1285 signal(SIGABRT, skip_signal);
1288 for (run_idx = 0; run_idx < run_count; run_idx++) {
1289 if (run_count != 1 && verbose)
1290 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1296 status = run_perf_stat(argc, argv);
1300 print_stat(argc, argv);
1302 list_for_each_entry(pos, &evsel_list->entries, node)
1303 perf_evsel__free_stat_priv(pos);
1304 perf_evlist__delete_maps(evsel_list);
1306 perf_evlist__delete(evsel_list);