4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ./hackbench 10
13 Performance counter stats for './hackbench 10':
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
27 0.154822978 seconds time elapsed
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 * Improvements and fixes by:
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
41 * Released under the GPL v2. (and only v2, not any later version)
46 #include "util/util.h"
47 #include "util/parse-options.h"
48 #include "util/parse-events.h"
49 #include "util/event.h"
50 #include "util/evlist.h"
51 #include "util/evsel.h"
52 #include "util/debug.h"
53 #include "util/color.h"
54 #include "util/stat.h"
55 #include "util/header.h"
56 #include "util/cpumap.h"
57 #include "util/thread.h"
58 #include "util/thread_map.h"
60 #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;
203 struct stats res_stats[3];
206 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
208 evsel->priv = zalloc(sizeof(struct perf_stat));
209 return evsel->priv == NULL ? -ENOMEM : 0;
212 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
218 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
219 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
220 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
221 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
222 static struct stats runtime_branches_stats[MAX_NR_CPUS];
223 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
224 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
225 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
226 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
227 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
228 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
229 static struct stats walltime_nsecs_stats;
231 static int create_perf_stat_counter(struct perf_evsel *evsel,
232 struct perf_evsel *first)
234 struct perf_event_attr *attr = &evsel->attr;
235 bool exclude_guest_missing = false;
239 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
240 PERF_FORMAT_TOTAL_TIME_RUNNING;
242 attr->inherit = !no_inherit;
245 if (exclude_guest_missing)
246 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
248 if (perf_target__has_cpu(&target)) {
249 ret = perf_evsel__open_per_cpu(evsel, evsel_list->cpus);
255 if (!perf_target__has_task(&target) && (!group || evsel == first)) {
257 attr->enable_on_exec = 1;
260 ret = perf_evsel__open_per_thread(evsel, evsel_list->threads);
265 if (ret && errno == EINVAL) {
266 if (!exclude_guest_missing &&
267 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
268 pr_debug("Old kernel, cannot exclude "
269 "guest or host samples.\n");
270 exclude_guest_missing = true;
278 * Does the counter have nsecs as a unit?
280 static inline int nsec_counter(struct perf_evsel *evsel)
282 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
283 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
290 * Update various tracking values we maintain to print
291 * more semantic information such as miss/hit ratios,
292 * instruction rates, etc:
294 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
296 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
297 update_stats(&runtime_nsecs_stats[0], count[0]);
298 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
299 update_stats(&runtime_cycles_stats[0], count[0]);
300 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
301 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
302 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
303 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
304 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
305 update_stats(&runtime_branches_stats[0], count[0]);
306 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
307 update_stats(&runtime_cacherefs_stats[0], count[0]);
308 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
309 update_stats(&runtime_l1_dcache_stats[0], count[0]);
310 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
311 update_stats(&runtime_l1_icache_stats[0], count[0]);
312 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
313 update_stats(&runtime_ll_cache_stats[0], count[0]);
314 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
315 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
316 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
317 update_stats(&runtime_itlb_cache_stats[0], count[0]);
321 * Read out the results of a single counter:
322 * aggregate counts across CPUs in system-wide mode
324 static int read_counter_aggr(struct perf_evsel *counter)
326 struct perf_stat *ps = counter->priv;
327 u64 *count = counter->counts->aggr.values;
330 if (__perf_evsel__read(counter, evsel_list->cpus->nr,
331 evsel_list->threads->nr, scale) < 0)
334 for (i = 0; i < 3; i++)
335 update_stats(&ps->res_stats[i], count[i]);
338 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
339 perf_evsel__name(counter), count[0], count[1], count[2]);
343 * Save the full runtime - to allow normalization during printout:
345 update_shadow_stats(counter, count);
351 * Read out the results of a single counter:
352 * do not aggregate counts across CPUs in system-wide mode
354 static int read_counter(struct perf_evsel *counter)
359 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
360 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
363 count = counter->counts->cpu[cpu].values;
365 update_shadow_stats(counter, count);
371 static int run_perf_stat(int argc __maybe_unused, const char **argv)
373 unsigned long long t0, t1;
374 struct perf_evsel *counter, *first;
376 int child_ready_pipe[2], go_pipe[2];
377 const bool forks = (argc > 0);
380 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
381 perror("failed to create pipes");
386 if ((child_pid = fork()) < 0)
387 perror("failed to fork");
390 close(child_ready_pipe[0]);
392 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
395 * Do a dummy execvp to get the PLT entry resolved,
396 * so we avoid the resolver overhead on the real
399 execvp("", (char **)argv);
402 * Tell the parent we're ready to go
404 close(child_ready_pipe[1]);
407 * Wait until the parent tells us to go.
409 if (read(go_pipe[0], &buf, 1) == -1)
410 perror("unable to read pipe");
412 execvp(argv[0], (char **)argv);
418 if (perf_target__none(&target))
419 evsel_list->threads->map[0] = child_pid;
422 * Wait for the child to be ready to exec.
424 close(child_ready_pipe[1]);
426 if (read(child_ready_pipe[0], &buf, 1) == -1)
427 perror("unable to read pipe");
428 close(child_ready_pipe[0]);
432 perf_evlist__set_leader(evsel_list);
434 first = perf_evlist__first(evsel_list);
436 list_for_each_entry(counter, &evsel_list->entries, node) {
437 if (create_perf_stat_counter(counter, first) < 0) {
439 * PPC returns ENXIO for HW counters until 2.6.37
440 * (behavior changed with commit b0a873e).
442 if (errno == EINVAL || errno == ENOSYS ||
443 errno == ENOENT || errno == EOPNOTSUPP ||
446 ui__warning("%s event is not supported by the kernel.\n",
447 perf_evsel__name(counter));
448 counter->supported = false;
452 if (errno == EPERM || errno == EACCES) {
453 error("You may not have permission to collect %sstats.\n"
454 "\t Consider tweaking"
455 " /proc/sys/kernel/perf_event_paranoid or running as root.",
456 target.system_wide ? "system-wide " : "");
458 error("open_counter returned with %d (%s). "
459 "/bin/dmesg may provide additional information.\n",
460 errno, strerror(errno));
463 kill(child_pid, SIGTERM);
465 pr_err("Not all events could be opened.\n");
468 counter->supported = true;
471 if (perf_evlist__set_filters(evsel_list)) {
472 error("failed to set filter with %d (%s)\n", errno,
478 * Enable counters and exec the command:
485 if (WIFSIGNALED(status))
486 psignal(WTERMSIG(status), argv[0]);
488 while(!done) sleep(1);
493 update_stats(&walltime_nsecs_stats, t1 - t0);
496 list_for_each_entry(counter, &evsel_list->entries, node) {
497 read_counter(counter);
498 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
501 list_for_each_entry(counter, &evsel_list->entries, node) {
502 read_counter_aggr(counter);
503 perf_evsel__close_fd(counter, evsel_list->cpus->nr,
504 evsel_list->threads->nr);
508 return WEXITSTATUS(status);
511 static void print_noise_pct(double total, double avg)
513 double pct = rel_stddev_stats(total, avg);
516 fprintf(output, "%s%.2f%%", csv_sep, pct);
518 fprintf(output, " ( +-%6.2f%% )", pct);
521 static void print_noise(struct perf_evsel *evsel, double avg)
523 struct perf_stat *ps;
529 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
532 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
534 double msecs = avg / 1e6;
535 char cpustr[16] = { '\0', };
536 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
539 sprintf(cpustr, "CPU%*d%s",
541 evsel_list->cpus->map[cpu], csv_sep);
543 fprintf(output, fmt, cpustr, msecs, csv_sep, perf_evsel__name(evsel));
546 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
551 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
552 fprintf(output, " # %8.3f CPUs utilized ",
553 avg / avg_stats(&walltime_nsecs_stats));
555 fprintf(output, " ");
558 /* used for get_ratio_color() */
560 GRC_STALLED_CYCLES_FE,
561 GRC_STALLED_CYCLES_BE,
566 static const char *get_ratio_color(enum grc_type type, double ratio)
568 static const double grc_table[GRC_MAX_NR][3] = {
569 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
570 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
571 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
573 const char *color = PERF_COLOR_NORMAL;
575 if (ratio > grc_table[type][0])
576 color = PERF_COLOR_RED;
577 else if (ratio > grc_table[type][1])
578 color = PERF_COLOR_MAGENTA;
579 else if (ratio > grc_table[type][2])
580 color = PERF_COLOR_YELLOW;
585 static void print_stalled_cycles_frontend(int cpu,
586 struct perf_evsel *evsel
587 __maybe_unused, double avg)
589 double total, ratio = 0.0;
592 total = avg_stats(&runtime_cycles_stats[cpu]);
595 ratio = avg / total * 100.0;
597 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
599 fprintf(output, " # ");
600 color_fprintf(output, color, "%6.2f%%", ratio);
601 fprintf(output, " frontend cycles idle ");
604 static void print_stalled_cycles_backend(int cpu,
605 struct perf_evsel *evsel
606 __maybe_unused, double avg)
608 double total, ratio = 0.0;
611 total = avg_stats(&runtime_cycles_stats[cpu]);
614 ratio = avg / total * 100.0;
616 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
618 fprintf(output, " # ");
619 color_fprintf(output, color, "%6.2f%%", ratio);
620 fprintf(output, " backend cycles idle ");
623 static void print_branch_misses(int cpu,
624 struct perf_evsel *evsel __maybe_unused,
627 double total, ratio = 0.0;
630 total = avg_stats(&runtime_branches_stats[cpu]);
633 ratio = avg / total * 100.0;
635 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
637 fprintf(output, " # ");
638 color_fprintf(output, color, "%6.2f%%", ratio);
639 fprintf(output, " of all branches ");
642 static void print_l1_dcache_misses(int cpu,
643 struct perf_evsel *evsel __maybe_unused,
646 double total, ratio = 0.0;
649 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
652 ratio = avg / total * 100.0;
654 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
656 fprintf(output, " # ");
657 color_fprintf(output, color, "%6.2f%%", ratio);
658 fprintf(output, " of all L1-dcache hits ");
661 static void print_l1_icache_misses(int cpu,
662 struct perf_evsel *evsel __maybe_unused,
665 double total, ratio = 0.0;
668 total = avg_stats(&runtime_l1_icache_stats[cpu]);
671 ratio = avg / total * 100.0;
673 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
675 fprintf(output, " # ");
676 color_fprintf(output, color, "%6.2f%%", ratio);
677 fprintf(output, " of all L1-icache hits ");
680 static void print_dtlb_cache_misses(int cpu,
681 struct perf_evsel *evsel __maybe_unused,
684 double total, ratio = 0.0;
687 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
690 ratio = avg / total * 100.0;
692 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
694 fprintf(output, " # ");
695 color_fprintf(output, color, "%6.2f%%", ratio);
696 fprintf(output, " of all dTLB cache hits ");
699 static void print_itlb_cache_misses(int cpu,
700 struct perf_evsel *evsel __maybe_unused,
703 double total, ratio = 0.0;
706 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
709 ratio = avg / total * 100.0;
711 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
713 fprintf(output, " # ");
714 color_fprintf(output, color, "%6.2f%%", ratio);
715 fprintf(output, " of all iTLB cache hits ");
718 static void print_ll_cache_misses(int cpu,
719 struct perf_evsel *evsel __maybe_unused,
722 double total, ratio = 0.0;
725 total = avg_stats(&runtime_ll_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 LL-cache hits ");
737 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
739 double total, ratio = 0.0;
740 char cpustr[16] = { '\0', };
746 fmt = "%s%'18.0f%s%-25s";
748 fmt = "%s%18.0f%s%-25s";
751 sprintf(cpustr, "CPU%*d%s",
753 evsel_list->cpus->map[cpu], csv_sep);
757 fprintf(output, fmt, cpustr, avg, csv_sep, perf_evsel__name(evsel));
760 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
765 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
766 total = avg_stats(&runtime_cycles_stats[cpu]);
771 fprintf(output, " # %5.2f insns per cycle ", ratio);
773 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
774 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
778 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
781 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
782 runtime_branches_stats[cpu].n != 0) {
783 print_branch_misses(cpu, evsel, avg);
785 evsel->attr.type == PERF_TYPE_HW_CACHE &&
786 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
787 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
788 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
789 runtime_l1_dcache_stats[cpu].n != 0) {
790 print_l1_dcache_misses(cpu, evsel, avg);
792 evsel->attr.type == PERF_TYPE_HW_CACHE &&
793 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
794 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
795 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
796 runtime_l1_icache_stats[cpu].n != 0) {
797 print_l1_icache_misses(cpu, evsel, avg);
799 evsel->attr.type == PERF_TYPE_HW_CACHE &&
800 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
801 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
802 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
803 runtime_dtlb_cache_stats[cpu].n != 0) {
804 print_dtlb_cache_misses(cpu, evsel, avg);
806 evsel->attr.type == PERF_TYPE_HW_CACHE &&
807 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
808 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
809 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
810 runtime_itlb_cache_stats[cpu].n != 0) {
811 print_itlb_cache_misses(cpu, evsel, avg);
813 evsel->attr.type == PERF_TYPE_HW_CACHE &&
814 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
815 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
816 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
817 runtime_ll_cache_stats[cpu].n != 0) {
818 print_ll_cache_misses(cpu, evsel, avg);
819 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
820 runtime_cacherefs_stats[cpu].n != 0) {
821 total = avg_stats(&runtime_cacherefs_stats[cpu]);
824 ratio = avg * 100 / total;
826 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
828 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
829 print_stalled_cycles_frontend(cpu, evsel, avg);
830 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
831 print_stalled_cycles_backend(cpu, evsel, avg);
832 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
833 total = avg_stats(&runtime_nsecs_stats[cpu]);
836 ratio = 1.0 * avg / total;
838 fprintf(output, " # %8.3f GHz ", ratio);
839 } else if (runtime_nsecs_stats[cpu].n != 0) {
842 total = avg_stats(&runtime_nsecs_stats[cpu]);
845 ratio = 1000.0 * avg / total;
851 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
853 fprintf(output, " ");
858 * Print out the results of a single counter:
859 * aggregated counts in system-wide mode
861 static void print_counter_aggr(struct perf_evsel *counter)
863 struct perf_stat *ps = counter->priv;
864 double avg = avg_stats(&ps->res_stats[0]);
865 int scaled = counter->counts->scaled;
868 fprintf(output, "%*s%s%*s",
870 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
872 csv_output ? 0 : -24,
873 perf_evsel__name(counter));
876 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
882 if (nsec_counter(counter))
883 nsec_printout(-1, counter, avg);
885 abs_printout(-1, counter, avg);
887 print_noise(counter, avg);
895 double avg_enabled, avg_running;
897 avg_enabled = avg_stats(&ps->res_stats[1]);
898 avg_running = avg_stats(&ps->res_stats[2]);
900 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
902 fprintf(output, "\n");
906 * Print out the results of a single counter:
907 * does not use aggregated count in system-wide
909 static void print_counter(struct perf_evsel *counter)
914 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
915 val = counter->counts->cpu[cpu].val;
916 ena = counter->counts->cpu[cpu].ena;
917 run = counter->counts->cpu[cpu].run;
918 if (run == 0 || ena == 0) {
919 fprintf(output, "CPU%*d%s%*s%s%*s",
921 evsel_list->cpus->map[cpu], csv_sep,
923 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
925 csv_output ? 0 : -24,
926 perf_evsel__name(counter));
929 fprintf(output, "%s%s",
930 csv_sep, counter->cgrp->name);
936 if (nsec_counter(counter))
937 nsec_printout(cpu, counter, val);
939 abs_printout(cpu, counter, val);
942 print_noise(counter, 1.0);
945 fprintf(output, " (%.2f%%)",
952 static void print_stat(int argc, const char **argv)
954 struct perf_evsel *counter;
960 fprintf(output, "\n");
961 fprintf(output, " Performance counter stats for ");
962 if (!perf_target__has_task(&target)) {
963 fprintf(output, "\'%s", argv[0]);
964 for (i = 1; i < argc; i++)
965 fprintf(output, " %s", argv[i]);
966 } else if (target.pid)
967 fprintf(output, "process id \'%s", target.pid);
969 fprintf(output, "thread id \'%s", target.tid);
971 fprintf(output, "\'");
973 fprintf(output, " (%d runs)", run_count);
974 fprintf(output, ":\n\n");
978 list_for_each_entry(counter, &evsel_list->entries, node)
979 print_counter(counter);
981 list_for_each_entry(counter, &evsel_list->entries, node)
982 print_counter_aggr(counter);
987 fprintf(output, "\n");
988 fprintf(output, " %17.9f seconds time elapsed",
989 avg_stats(&walltime_nsecs_stats)/1e9);
991 fprintf(output, " ");
992 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
993 avg_stats(&walltime_nsecs_stats));
995 fprintf(output, "\n\n");
999 static volatile int signr = -1;
1001 static void skip_signal(int signo)
1009 static void sig_atexit(void)
1011 if (child_pid != -1)
1012 kill(child_pid, SIGTERM);
1017 signal(signr, SIG_DFL);
1018 kill(getpid(), signr);
1021 static const char * const stat_usage[] = {
1022 "perf stat [<options>] [<command>]",
1026 static int stat__set_big_num(const struct option *opt __maybe_unused,
1027 const char *s __maybe_unused, int unset)
1029 big_num_opt = unset ? 0 : 1;
1033 static bool append_file;
1035 static const struct option options[] = {
1036 OPT_CALLBACK('e', "event", &evsel_list, "event",
1037 "event selector. use 'perf list' to list available events",
1038 parse_events_option),
1039 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1040 "event filter", parse_filter),
1041 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1042 "child tasks do not inherit counters"),
1043 OPT_STRING('p', "pid", &target.pid, "pid",
1044 "stat events on existing process id"),
1045 OPT_STRING('t', "tid", &target.tid, "tid",
1046 "stat events on existing thread id"),
1047 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1048 "system-wide collection from all CPUs"),
1049 OPT_BOOLEAN('g', "group", &group,
1050 "put the counters into a counter group"),
1051 OPT_BOOLEAN('c', "scale", &scale,
1052 "scale/normalize counters"),
1053 OPT_INCR('v', "verbose", &verbose,
1054 "be more verbose (show counter open errors, etc)"),
1055 OPT_INTEGER('r', "repeat", &run_count,
1056 "repeat command and print average + stddev (max: 100)"),
1057 OPT_BOOLEAN('n', "null", &null_run,
1058 "null run - dont start any counters"),
1059 OPT_INCR('d', "detailed", &detailed_run,
1060 "detailed run - start a lot of events"),
1061 OPT_BOOLEAN('S', "sync", &sync_run,
1062 "call sync() before starting a run"),
1063 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1064 "print large numbers with thousands\' separators",
1066 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1067 "list of cpus to monitor in system-wide"),
1068 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1069 "disable CPU count aggregation"),
1070 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1071 "print counts with custom separator"),
1072 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1073 "monitor event in cgroup name only",
1075 OPT_STRING('o', "output", &output_name, "file",
1076 "output file name"),
1077 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1078 OPT_INTEGER(0, "log-fd", &output_fd,
1079 "log output to fd, instead of stderr"),
1084 * Add default attributes, if there were no attributes specified or
1085 * if -d/--detailed, -d -d or -d -d -d is used:
1087 static int add_default_attributes(void)
1089 /* Set attrs if no event is selected and !null_run: */
1093 if (!evsel_list->nr_entries) {
1094 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1098 /* Detailed events get appended to the event list: */
1100 if (detailed_run < 1)
1103 /* Append detailed run extra attributes: */
1104 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1107 if (detailed_run < 2)
1110 /* Append very detailed run extra attributes: */
1111 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1114 if (detailed_run < 3)
1117 /* Append very, very detailed run extra attributes: */
1118 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1121 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1123 struct perf_evsel *pos;
1124 int status = -ENOMEM;
1127 setlocale(LC_ALL, "");
1129 evsel_list = perf_evlist__new(NULL, NULL);
1130 if (evsel_list == NULL)
1133 argc = parse_options(argc, argv, options, stat_usage,
1134 PARSE_OPT_STOP_AT_NON_OPTION);
1137 if (output_name && strcmp(output_name, "-"))
1140 if (output_name && output_fd) {
1141 fprintf(stderr, "cannot use both --output and --log-fd\n");
1142 usage_with_options(stat_usage, options);
1145 if (output_fd < 0) {
1146 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1147 usage_with_options(stat_usage, options);
1152 mode = append_file ? "a" : "w";
1154 output = fopen(output_name, mode);
1156 perror("failed to create output file");
1159 clock_gettime(CLOCK_REALTIME, &tm);
1160 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1161 } else if (output_fd > 0) {
1162 mode = append_file ? "a" : "w";
1163 output = fdopen(output_fd, mode);
1165 perror("Failed opening logfd");
1172 if (!strcmp(csv_sep, "\\t"))
1175 csv_sep = DEFAULT_SEPARATOR;
1178 * let the spreadsheet do the pretty-printing
1181 /* User explicitly passed -B? */
1182 if (big_num_opt == 1) {
1183 fprintf(stderr, "-B option not supported with -x\n");
1184 usage_with_options(stat_usage, options);
1185 } else /* Nope, so disable big number formatting */
1187 } else if (big_num_opt == 0) /* User passed --no-big-num */
1190 if (!argc && !perf_target__has_task(&target))
1191 usage_with_options(stat_usage, options);
1193 usage_with_options(stat_usage, options);
1195 /* no_aggr, cgroup are for system-wide only */
1196 if ((no_aggr || nr_cgroups) && !perf_target__has_cpu(&target)) {
1197 fprintf(stderr, "both cgroup and no-aggregation "
1198 "modes only available in system-wide mode\n");
1200 usage_with_options(stat_usage, options);
1203 if (add_default_attributes())
1206 perf_target__validate(&target);
1208 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1209 if (perf_target__has_task(&target))
1210 pr_err("Problems finding threads of monitor\n");
1211 if (perf_target__has_cpu(&target))
1212 perror("failed to parse CPUs map");
1214 usage_with_options(stat_usage, options);
1218 list_for_each_entry(pos, &evsel_list->entries, node) {
1219 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1220 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0)
1225 * We dont want to block the signals - that would cause
1226 * child tasks to inherit that and Ctrl-C would not work.
1227 * What we want is for Ctrl-C to work in the exec()-ed
1228 * task, but being ignored by perf stat itself:
1231 signal(SIGINT, skip_signal);
1232 signal(SIGALRM, skip_signal);
1233 signal(SIGABRT, skip_signal);
1236 for (run_idx = 0; run_idx < run_count; run_idx++) {
1237 if (run_count != 1 && verbose)
1238 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1244 status = run_perf_stat(argc, argv);
1248 print_stat(argc, argv);
1250 list_for_each_entry(pos, &evsel_list->entries, node)
1251 perf_evsel__free_stat_priv(pos);
1252 perf_evlist__delete_maps(evsel_list);
1254 perf_evlist__delete(evsel_list);