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 struct perf_evlist *evsel_list;
178 static bool system_wide = false;
179 static int run_idx = 0;
181 static int run_count = 1;
182 static bool no_inherit = false;
183 static bool scale = true;
184 static bool no_aggr = false;
185 static const char *target_pid;
186 static const char *target_tid;
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 *cpu_list;
194 static const char *csv_sep = NULL;
195 static bool csv_output = false;
196 static bool group = false;
197 static const char *output_name = NULL;
198 static FILE *output = NULL;
199 static int output_fd;
201 static volatile int done = 0;
209 struct stats res_stats[3];
212 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
214 evsel->priv = zalloc(sizeof(struct perf_stat));
215 return evsel->priv == NULL ? -ENOMEM : 0;
218 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
224 static void update_stats(struct stats *stats, u64 val)
229 delta = val - stats->mean;
230 stats->mean += delta / stats->n;
231 stats->M2 += delta*(val - stats->mean);
234 static double avg_stats(struct stats *stats)
240 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
242 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
243 * s^2 = -------------------------------
246 * http://en.wikipedia.org/wiki/Stddev
248 * The std dev of the mean is related to the std dev by:
255 static double stddev_stats(struct stats *stats)
257 double variance, variance_mean;
262 variance = stats->M2 / (stats->n - 1);
263 variance_mean = variance / stats->n;
265 return sqrt(variance_mean);
268 struct stats runtime_nsecs_stats[MAX_NR_CPUS];
269 struct stats runtime_cycles_stats[MAX_NR_CPUS];
270 struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
271 struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
272 struct stats runtime_branches_stats[MAX_NR_CPUS];
273 struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
274 struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
275 struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
276 struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
277 struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
278 struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
279 struct stats walltime_nsecs_stats;
281 static int create_perf_stat_counter(struct perf_evsel *evsel,
282 struct perf_evsel *first)
284 struct perf_event_attr *attr = &evsel->attr;
285 struct xyarray *group_fd = NULL;
286 bool exclude_guest_missing = false;
289 if (group && evsel != first)
290 group_fd = first->fd;
293 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
294 PERF_FORMAT_TOTAL_TIME_RUNNING;
296 attr->inherit = !no_inherit;
299 if (exclude_guest_missing)
300 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
303 ret = perf_evsel__open_per_cpu(evsel, evsel_list->cpus,
310 if (!target_pid && !target_tid && (!group || evsel == first)) {
312 attr->enable_on_exec = 1;
315 ret = perf_evsel__open_per_thread(evsel, evsel_list->threads,
321 if (ret && errno == EINVAL) {
322 if (!exclude_guest_missing &&
323 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
324 pr_debug("Old kernel, cannot exclude "
325 "guest or host samples.\n");
326 exclude_guest_missing = true;
334 * Does the counter have nsecs as a unit?
336 static inline int nsec_counter(struct perf_evsel *evsel)
338 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
339 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
346 * Update various tracking values we maintain to print
347 * more semantic information such as miss/hit ratios,
348 * instruction rates, etc:
350 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
352 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
353 update_stats(&runtime_nsecs_stats[0], count[0]);
354 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
355 update_stats(&runtime_cycles_stats[0], count[0]);
356 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
357 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
358 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
359 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
360 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
361 update_stats(&runtime_branches_stats[0], count[0]);
362 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
363 update_stats(&runtime_cacherefs_stats[0], count[0]);
364 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
365 update_stats(&runtime_l1_dcache_stats[0], count[0]);
366 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
367 update_stats(&runtime_l1_icache_stats[0], count[0]);
368 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
369 update_stats(&runtime_ll_cache_stats[0], count[0]);
370 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
371 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
372 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
373 update_stats(&runtime_itlb_cache_stats[0], count[0]);
377 * Read out the results of a single counter:
378 * aggregate counts across CPUs in system-wide mode
380 static int read_counter_aggr(struct perf_evsel *counter)
382 struct perf_stat *ps = counter->priv;
383 u64 *count = counter->counts->aggr.values;
386 if (__perf_evsel__read(counter, evsel_list->cpus->nr,
387 evsel_list->threads->nr, scale) < 0)
390 for (i = 0; i < 3; i++)
391 update_stats(&ps->res_stats[i], count[i]);
394 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
395 event_name(counter), count[0], count[1], count[2]);
399 * Save the full runtime - to allow normalization during printout:
401 update_shadow_stats(counter, count);
407 * Read out the results of a single counter:
408 * do not aggregate counts across CPUs in system-wide mode
410 static int read_counter(struct perf_evsel *counter)
415 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
416 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
419 count = counter->counts->cpu[cpu].values;
421 update_shadow_stats(counter, count);
427 static int run_perf_stat(int argc __used, const char **argv)
429 unsigned long long t0, t1;
430 struct perf_evsel *counter, *first;
432 int child_ready_pipe[2], go_pipe[2];
433 const bool forks = (argc > 0);
436 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
437 perror("failed to create pipes");
442 if ((child_pid = fork()) < 0)
443 perror("failed to fork");
446 close(child_ready_pipe[0]);
448 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
451 * Do a dummy execvp to get the PLT entry resolved,
452 * so we avoid the resolver overhead on the real
455 execvp("", (char **)argv);
458 * Tell the parent we're ready to go
460 close(child_ready_pipe[1]);
463 * Wait until the parent tells us to go.
465 if (read(go_pipe[0], &buf, 1) == -1)
466 perror("unable to read pipe");
468 execvp(argv[0], (char **)argv);
474 if (!target_tid && !target_pid && !system_wide)
475 evsel_list->threads->map[0] = child_pid;
478 * Wait for the child to be ready to exec.
480 close(child_ready_pipe[1]);
482 if (read(child_ready_pipe[0], &buf, 1) == -1)
483 perror("unable to read pipe");
484 close(child_ready_pipe[0]);
487 first = list_entry(evsel_list->entries.next, struct perf_evsel, node);
489 list_for_each_entry(counter, &evsel_list->entries, node) {
490 if (create_perf_stat_counter(counter, first) < 0) {
491 if (errno == EINVAL || errno == ENOSYS ||
492 errno == ENOENT || errno == EOPNOTSUPP) {
494 ui__warning("%s event is not supported by the kernel.\n",
495 event_name(counter));
496 counter->supported = false;
500 if (errno == EPERM || errno == EACCES) {
501 error("You may not have permission to collect %sstats.\n"
502 "\t Consider tweaking"
503 " /proc/sys/kernel/perf_event_paranoid or running as root.",
504 system_wide ? "system-wide " : "");
506 error("open_counter returned with %d (%s). "
507 "/bin/dmesg may provide additional information.\n",
508 errno, strerror(errno));
511 kill(child_pid, SIGTERM);
512 die("Not all events could be opened.\n");
515 counter->supported = true;
518 if (perf_evlist__set_filters(evsel_list)) {
519 error("failed to set filter with %d (%s)\n", errno,
525 * Enable counters and exec the command:
532 if (WIFSIGNALED(status))
533 psignal(WTERMSIG(status), argv[0]);
535 while(!done) sleep(1);
540 update_stats(&walltime_nsecs_stats, t1 - t0);
543 list_for_each_entry(counter, &evsel_list->entries, node) {
544 read_counter(counter);
545 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
548 list_for_each_entry(counter, &evsel_list->entries, node) {
549 read_counter_aggr(counter);
550 perf_evsel__close_fd(counter, evsel_list->cpus->nr,
551 evsel_list->threads->nr);
555 return WEXITSTATUS(status);
558 static void print_noise_pct(double total, double avg)
563 pct = 100.0*total/avg;
566 fprintf(output, "%s%.2f%%", csv_sep, pct);
568 fprintf(output, " ( +-%6.2f%% )", pct);
571 static void print_noise(struct perf_evsel *evsel, double avg)
573 struct perf_stat *ps;
579 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
582 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
584 double msecs = avg / 1e6;
585 char cpustr[16] = { '\0', };
586 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
589 sprintf(cpustr, "CPU%*d%s",
591 evsel_list->cpus->map[cpu], csv_sep);
593 fprintf(output, fmt, cpustr, msecs, csv_sep, event_name(evsel));
596 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
601 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
602 fprintf(output, " # %8.3f CPUs utilized ",
603 avg / avg_stats(&walltime_nsecs_stats));
605 fprintf(output, " ");
608 /* used for get_ratio_color() */
610 GRC_STALLED_CYCLES_FE,
611 GRC_STALLED_CYCLES_BE,
616 static const char *get_ratio_color(enum grc_type type, double ratio)
618 static const double grc_table[GRC_MAX_NR][3] = {
619 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
620 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
621 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
623 const char *color = PERF_COLOR_NORMAL;
625 if (ratio > grc_table[type][0])
626 color = PERF_COLOR_RED;
627 else if (ratio > grc_table[type][1])
628 color = PERF_COLOR_MAGENTA;
629 else if (ratio > grc_table[type][2])
630 color = PERF_COLOR_YELLOW;
635 static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg)
637 double total, ratio = 0.0;
640 total = avg_stats(&runtime_cycles_stats[cpu]);
643 ratio = avg / total * 100.0;
645 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
647 fprintf(output, " # ");
648 color_fprintf(output, color, "%6.2f%%", ratio);
649 fprintf(output, " frontend cycles idle ");
652 static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg)
654 double total, ratio = 0.0;
657 total = avg_stats(&runtime_cycles_stats[cpu]);
660 ratio = avg / total * 100.0;
662 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
664 fprintf(output, " # ");
665 color_fprintf(output, color, "%6.2f%%", ratio);
666 fprintf(output, " backend cycles idle ");
669 static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
671 double total, ratio = 0.0;
674 total = avg_stats(&runtime_branches_stats[cpu]);
677 ratio = avg / total * 100.0;
679 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
681 fprintf(output, " # ");
682 color_fprintf(output, color, "%6.2f%%", ratio);
683 fprintf(output, " of all branches ");
686 static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
688 double total, ratio = 0.0;
691 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
694 ratio = avg / total * 100.0;
696 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
698 fprintf(output, " # ");
699 color_fprintf(output, color, "%6.2f%%", ratio);
700 fprintf(output, " of all L1-dcache hits ");
703 static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
705 double total, ratio = 0.0;
708 total = avg_stats(&runtime_l1_icache_stats[cpu]);
711 ratio = avg / total * 100.0;
713 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
715 fprintf(output, " # ");
716 color_fprintf(output, color, "%6.2f%%", ratio);
717 fprintf(output, " of all L1-icache hits ");
720 static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
722 double total, ratio = 0.0;
725 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
728 ratio = avg / total * 100.0;
730 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
732 fprintf(output, " # ");
733 color_fprintf(output, color, "%6.2f%%", ratio);
734 fprintf(output, " of all dTLB cache hits ");
737 static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
739 double total, ratio = 0.0;
742 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
745 ratio = avg / total * 100.0;
747 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
749 fprintf(output, " # ");
750 color_fprintf(output, color, "%6.2f%%", ratio);
751 fprintf(output, " of all iTLB cache hits ");
754 static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
756 double total, ratio = 0.0;
759 total = avg_stats(&runtime_ll_cache_stats[cpu]);
762 ratio = avg / total * 100.0;
764 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
766 fprintf(output, " # ");
767 color_fprintf(output, color, "%6.2f%%", ratio);
768 fprintf(output, " of all LL-cache hits ");
771 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
773 double total, ratio = 0.0;
774 char cpustr[16] = { '\0', };
780 fmt = "%s%'18.0f%s%-25s";
782 fmt = "%s%18.0f%s%-25s";
785 sprintf(cpustr, "CPU%*d%s",
787 evsel_list->cpus->map[cpu], csv_sep);
791 fprintf(output, fmt, cpustr, avg, csv_sep, event_name(evsel));
794 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
799 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
800 total = avg_stats(&runtime_cycles_stats[cpu]);
805 fprintf(output, " # %5.2f insns per cycle ", ratio);
807 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
808 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
812 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
815 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
816 runtime_branches_stats[cpu].n != 0) {
817 print_branch_misses(cpu, evsel, avg);
819 evsel->attr.type == PERF_TYPE_HW_CACHE &&
820 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
821 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
822 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
823 runtime_l1_dcache_stats[cpu].n != 0) {
824 print_l1_dcache_misses(cpu, evsel, avg);
826 evsel->attr.type == PERF_TYPE_HW_CACHE &&
827 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
828 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
829 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
830 runtime_l1_icache_stats[cpu].n != 0) {
831 print_l1_icache_misses(cpu, evsel, avg);
833 evsel->attr.type == PERF_TYPE_HW_CACHE &&
834 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
835 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
836 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
837 runtime_dtlb_cache_stats[cpu].n != 0) {
838 print_dtlb_cache_misses(cpu, evsel, avg);
840 evsel->attr.type == PERF_TYPE_HW_CACHE &&
841 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
842 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
843 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
844 runtime_itlb_cache_stats[cpu].n != 0) {
845 print_itlb_cache_misses(cpu, evsel, avg);
847 evsel->attr.type == PERF_TYPE_HW_CACHE &&
848 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
849 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
850 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
851 runtime_ll_cache_stats[cpu].n != 0) {
852 print_ll_cache_misses(cpu, evsel, avg);
853 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
854 runtime_cacherefs_stats[cpu].n != 0) {
855 total = avg_stats(&runtime_cacherefs_stats[cpu]);
858 ratio = avg * 100 / total;
860 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
862 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
863 print_stalled_cycles_frontend(cpu, evsel, avg);
864 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
865 print_stalled_cycles_backend(cpu, evsel, avg);
866 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
867 total = avg_stats(&runtime_nsecs_stats[cpu]);
870 ratio = 1.0 * avg / total;
872 fprintf(output, " # %8.3f GHz ", ratio);
873 } else if (runtime_nsecs_stats[cpu].n != 0) {
876 total = avg_stats(&runtime_nsecs_stats[cpu]);
879 ratio = 1000.0 * avg / total;
885 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
887 fprintf(output, " ");
892 * Print out the results of a single counter:
893 * aggregated counts in system-wide mode
895 static void print_counter_aggr(struct perf_evsel *counter)
897 struct perf_stat *ps = counter->priv;
898 double avg = avg_stats(&ps->res_stats[0]);
899 int scaled = counter->counts->scaled;
902 fprintf(output, "%*s%s%*s",
904 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
906 csv_output ? 0 : -24,
907 event_name(counter));
910 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
916 if (nsec_counter(counter))
917 nsec_printout(-1, counter, avg);
919 abs_printout(-1, counter, avg);
921 print_noise(counter, avg);
929 double avg_enabled, avg_running;
931 avg_enabled = avg_stats(&ps->res_stats[1]);
932 avg_running = avg_stats(&ps->res_stats[2]);
934 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
936 fprintf(output, "\n");
940 * Print out the results of a single counter:
941 * does not use aggregated count in system-wide
943 static void print_counter(struct perf_evsel *counter)
948 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
949 val = counter->counts->cpu[cpu].val;
950 ena = counter->counts->cpu[cpu].ena;
951 run = counter->counts->cpu[cpu].run;
952 if (run == 0 || ena == 0) {
953 fprintf(output, "CPU%*d%s%*s%s%*s",
955 evsel_list->cpus->map[cpu], csv_sep,
957 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
959 csv_output ? 0 : -24,
960 event_name(counter));
963 fprintf(output, "%s%s",
964 csv_sep, counter->cgrp->name);
970 if (nsec_counter(counter))
971 nsec_printout(cpu, counter, val);
973 abs_printout(cpu, counter, val);
976 print_noise(counter, 1.0);
979 fprintf(output, " (%.2f%%)",
986 static void print_stat(int argc, const char **argv)
988 struct perf_evsel *counter;
994 fprintf(output, "\n");
995 fprintf(output, " Performance counter stats for ");
996 if (!target_pid && !target_tid) {
997 fprintf(output, "\'%s", argv[0]);
998 for (i = 1; i < argc; i++)
999 fprintf(output, " %s", argv[i]);
1000 } else if (target_pid)
1001 fprintf(output, "process id \'%s", target_pid);
1003 fprintf(output, "thread id \'%s", target_tid);
1005 fprintf(output, "\'");
1007 fprintf(output, " (%d runs)", run_count);
1008 fprintf(output, ":\n\n");
1012 list_for_each_entry(counter, &evsel_list->entries, node)
1013 print_counter(counter);
1015 list_for_each_entry(counter, &evsel_list->entries, node)
1016 print_counter_aggr(counter);
1021 fprintf(output, "\n");
1022 fprintf(output, " %17.9f seconds time elapsed",
1023 avg_stats(&walltime_nsecs_stats)/1e9);
1024 if (run_count > 1) {
1025 fprintf(output, " ");
1026 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1027 avg_stats(&walltime_nsecs_stats));
1029 fprintf(output, "\n\n");
1033 static volatile int signr = -1;
1035 static void skip_signal(int signo)
1043 static void sig_atexit(void)
1045 if (child_pid != -1)
1046 kill(child_pid, SIGTERM);
1051 signal(signr, SIG_DFL);
1052 kill(getpid(), signr);
1055 static const char * const stat_usage[] = {
1056 "perf stat [<options>] [<command>]",
1060 static int stat__set_big_num(const struct option *opt __used,
1061 const char *s __used, int unset)
1063 big_num_opt = unset ? 0 : 1;
1067 static bool append_file;
1069 static const struct option options[] = {
1070 OPT_CALLBACK('e', "event", &evsel_list, "event",
1071 "event selector. use 'perf list' to list available events",
1072 parse_events_option),
1073 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1074 "event filter", parse_filter),
1075 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1076 "child tasks do not inherit counters"),
1077 OPT_STRING('p', "pid", &target_pid, "pid",
1078 "stat events on existing process id"),
1079 OPT_STRING('t', "tid", &target_tid, "tid",
1080 "stat events on existing thread id"),
1081 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1082 "system-wide collection from all CPUs"),
1083 OPT_BOOLEAN('g', "group", &group,
1084 "put the counters into a counter group"),
1085 OPT_BOOLEAN('c', "scale", &scale,
1086 "scale/normalize counters"),
1087 OPT_INCR('v', "verbose", &verbose,
1088 "be more verbose (show counter open errors, etc)"),
1089 OPT_INTEGER('r', "repeat", &run_count,
1090 "repeat command and print average + stddev (max: 100)"),
1091 OPT_BOOLEAN('n', "null", &null_run,
1092 "null run - dont start any counters"),
1093 OPT_INCR('d', "detailed", &detailed_run,
1094 "detailed run - start a lot of events"),
1095 OPT_BOOLEAN('S', "sync", &sync_run,
1096 "call sync() before starting a run"),
1097 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1098 "print large numbers with thousands\' separators",
1100 OPT_STRING('C', "cpu", &cpu_list, "cpu",
1101 "list of cpus to monitor in system-wide"),
1102 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1103 "disable CPU count aggregation"),
1104 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1105 "print counts with custom separator"),
1106 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1107 "monitor event in cgroup name only",
1109 OPT_STRING('o', "output", &output_name, "file",
1110 "output file name"),
1111 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1112 OPT_INTEGER(0, "log-fd", &output_fd,
1113 "log output to fd, instead of stderr"),
1118 * Add default attributes, if there were no attributes specified or
1119 * if -d/--detailed, -d -d or -d -d -d is used:
1121 static int add_default_attributes(void)
1123 /* Set attrs if no event is selected and !null_run: */
1127 if (!evsel_list->nr_entries) {
1128 if (perf_evlist__add_attrs_array(evsel_list, default_attrs) < 0)
1132 /* Detailed events get appended to the event list: */
1134 if (detailed_run < 1)
1137 /* Append detailed run extra attributes: */
1138 if (perf_evlist__add_attrs_array(evsel_list, detailed_attrs) < 0)
1141 if (detailed_run < 2)
1144 /* Append very detailed run extra attributes: */
1145 if (perf_evlist__add_attrs_array(evsel_list, very_detailed_attrs) < 0)
1148 if (detailed_run < 3)
1151 /* Append very, very detailed run extra attributes: */
1152 return perf_evlist__add_attrs_array(evsel_list, very_very_detailed_attrs);
1155 int cmd_stat(int argc, const char **argv, const char *prefix __used)
1157 struct perf_evsel *pos;
1158 int status = -ENOMEM;
1161 setlocale(LC_ALL, "");
1163 evsel_list = perf_evlist__new(NULL, NULL);
1164 if (evsel_list == NULL)
1167 argc = parse_options(argc, argv, options, stat_usage,
1168 PARSE_OPT_STOP_AT_NON_OPTION);
1171 if (output_name && strcmp(output_name, "-"))
1174 if (output_name && output_fd) {
1175 fprintf(stderr, "cannot use both --output and --log-fd\n");
1176 usage_with_options(stat_usage, options);
1180 mode = append_file ? "a" : "w";
1182 output = fopen(output_name, mode);
1184 perror("failed to create output file");
1187 clock_gettime(CLOCK_REALTIME, &tm);
1188 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1189 } else if (output_fd != 2) {
1190 mode = append_file ? "a" : "w";
1191 output = fdopen(output_fd, mode);
1193 perror("Failed opening logfd");
1200 if (!strcmp(csv_sep, "\\t"))
1203 csv_sep = DEFAULT_SEPARATOR;
1206 * let the spreadsheet do the pretty-printing
1209 /* User explicitly passed -B? */
1210 if (big_num_opt == 1) {
1211 fprintf(stderr, "-B option not supported with -x\n");
1212 usage_with_options(stat_usage, options);
1213 } else /* Nope, so disable big number formatting */
1215 } else if (big_num_opt == 0) /* User passed --no-big-num */
1218 if (!argc && !target_pid && !target_tid)
1219 usage_with_options(stat_usage, options);
1221 usage_with_options(stat_usage, options);
1223 /* no_aggr, cgroup are for system-wide only */
1224 if ((no_aggr || nr_cgroups) && !system_wide) {
1225 fprintf(stderr, "both cgroup and no-aggregation "
1226 "modes only available in system-wide mode\n");
1228 usage_with_options(stat_usage, options);
1231 if (add_default_attributes())
1235 target_tid = target_pid;
1237 evsel_list->threads = thread_map__new_str(target_pid,
1238 target_tid, UINT_MAX);
1239 if (evsel_list->threads == NULL) {
1240 pr_err("Problems finding threads of monitor\n");
1241 usage_with_options(stat_usage, options);
1245 evsel_list->cpus = cpu_map__new(cpu_list);
1247 evsel_list->cpus = cpu_map__dummy_new();
1249 if (evsel_list->cpus == NULL) {
1250 perror("failed to parse CPUs map");
1251 usage_with_options(stat_usage, options);
1255 list_for_each_entry(pos, &evsel_list->entries, node) {
1256 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1257 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0)
1262 * We dont want to block the signals - that would cause
1263 * child tasks to inherit that and Ctrl-C would not work.
1264 * What we want is for Ctrl-C to work in the exec()-ed
1265 * task, but being ignored by perf stat itself:
1268 signal(SIGINT, skip_signal);
1269 signal(SIGALRM, skip_signal);
1270 signal(SIGABRT, skip_signal);
1273 for (run_idx = 0; run_idx < run_count; run_idx++) {
1274 if (run_count != 1 && verbose)
1275 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1281 status = run_perf_stat(argc, argv);
1285 print_stat(argc, argv);
1287 list_for_each_entry(pos, &evsel_list->entries, node)
1288 perf_evsel__free_stat_priv(pos);
1289 perf_evlist__delete_maps(evsel_list);
1291 perf_evlist__delete(evsel_list);