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 pid_t target_pid = -1;
186 static pid_t target_tid = -1;
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 = stats->M2 / (stats->n - 1);
258 double variance_mean = variance / stats->n;
260 return sqrt(variance_mean);
263 struct stats runtime_nsecs_stats[MAX_NR_CPUS];
264 struct stats runtime_cycles_stats[MAX_NR_CPUS];
265 struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
266 struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
267 struct stats runtime_branches_stats[MAX_NR_CPUS];
268 struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
269 struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
270 struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
271 struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
272 struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
273 struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
274 struct stats walltime_nsecs_stats;
276 static int create_perf_stat_counter(struct perf_evsel *evsel)
278 struct perf_event_attr *attr = &evsel->attr;
281 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
282 PERF_FORMAT_TOTAL_TIME_RUNNING;
284 attr->inherit = !no_inherit;
287 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, group);
289 if (target_pid == -1 && target_tid == -1) {
291 attr->enable_on_exec = 1;
294 return perf_evsel__open_per_thread(evsel, evsel_list->threads, group);
298 * Does the counter have nsecs as a unit?
300 static inline int nsec_counter(struct perf_evsel *evsel)
302 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
303 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
310 * Update various tracking values we maintain to print
311 * more semantic information such as miss/hit ratios,
312 * instruction rates, etc:
314 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
316 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
317 update_stats(&runtime_nsecs_stats[0], count[0]);
318 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
319 update_stats(&runtime_cycles_stats[0], count[0]);
320 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
321 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
322 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
323 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
324 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
325 update_stats(&runtime_branches_stats[0], count[0]);
326 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
327 update_stats(&runtime_cacherefs_stats[0], count[0]);
328 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
329 update_stats(&runtime_l1_dcache_stats[0], count[0]);
330 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
331 update_stats(&runtime_l1_icache_stats[0], count[0]);
332 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
333 update_stats(&runtime_ll_cache_stats[0], count[0]);
334 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
335 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
336 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
337 update_stats(&runtime_itlb_cache_stats[0], count[0]);
341 * Read out the results of a single counter:
342 * aggregate counts across CPUs in system-wide mode
344 static int read_counter_aggr(struct perf_evsel *counter)
346 struct perf_stat *ps = counter->priv;
347 u64 *count = counter->counts->aggr.values;
350 if (__perf_evsel__read(counter, evsel_list->cpus->nr,
351 evsel_list->threads->nr, scale) < 0)
354 for (i = 0; i < 3; i++)
355 update_stats(&ps->res_stats[i], count[i]);
358 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
359 event_name(counter), count[0], count[1], count[2]);
363 * Save the full runtime - to allow normalization during printout:
365 update_shadow_stats(counter, count);
371 * Read out the results of a single counter:
372 * do not aggregate counts across CPUs in system-wide mode
374 static int read_counter(struct perf_evsel *counter)
379 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
380 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
383 count = counter->counts->cpu[cpu].values;
385 update_shadow_stats(counter, count);
391 static int run_perf_stat(int argc __used, const char **argv)
393 unsigned long long t0, t1;
394 struct perf_evsel *counter;
396 int child_ready_pipe[2], go_pipe[2];
397 const bool forks = (argc > 0);
400 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
401 perror("failed to create pipes");
406 if ((child_pid = fork()) < 0)
407 perror("failed to fork");
410 close(child_ready_pipe[0]);
412 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
415 * Do a dummy execvp to get the PLT entry resolved,
416 * so we avoid the resolver overhead on the real
419 execvp("", (char **)argv);
422 * Tell the parent we're ready to go
424 close(child_ready_pipe[1]);
427 * Wait until the parent tells us to go.
429 if (read(go_pipe[0], &buf, 1) == -1)
430 perror("unable to read pipe");
432 execvp(argv[0], (char **)argv);
438 if (target_tid == -1 && target_pid == -1 && !system_wide)
439 evsel_list->threads->map[0] = child_pid;
442 * Wait for the child to be ready to exec.
444 close(child_ready_pipe[1]);
446 if (read(child_ready_pipe[0], &buf, 1) == -1)
447 perror("unable to read pipe");
448 close(child_ready_pipe[0]);
451 list_for_each_entry(counter, &evsel_list->entries, node) {
452 if (create_perf_stat_counter(counter) < 0) {
453 if (errno == EINVAL || errno == ENOSYS || errno == ENOENT) {
455 ui__warning("%s event is not supported by the kernel.\n",
456 event_name(counter));
457 counter->supported = false;
461 if (errno == EPERM || errno == EACCES) {
462 error("You may not have permission to collect %sstats.\n"
463 "\t Consider tweaking"
464 " /proc/sys/kernel/perf_event_paranoid or running as root.",
465 system_wide ? "system-wide " : "");
467 error("open_counter returned with %d (%s). "
468 "/bin/dmesg may provide additional information.\n",
469 errno, strerror(errno));
472 kill(child_pid, SIGTERM);
473 die("Not all events could be opened.\n");
476 counter->supported = true;
479 if (perf_evlist__set_filters(evsel_list)) {
480 error("failed to set filter with %d (%s)\n", errno,
486 * Enable counters and exec the command:
494 while(!done) sleep(1);
499 update_stats(&walltime_nsecs_stats, t1 - t0);
502 list_for_each_entry(counter, &evsel_list->entries, node) {
503 read_counter(counter);
504 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
507 list_for_each_entry(counter, &evsel_list->entries, node) {
508 read_counter_aggr(counter);
509 perf_evsel__close_fd(counter, evsel_list->cpus->nr,
510 evsel_list->threads->nr);
514 return WEXITSTATUS(status);
517 static void print_noise_pct(double total, double avg)
522 pct = 100.0*total/avg;
525 fprintf(output, "%s%.2f%%", csv_sep, pct);
527 fprintf(output, " ( +-%6.2f%% )", pct);
530 static void print_noise(struct perf_evsel *evsel, double avg)
532 struct perf_stat *ps;
538 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
541 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
543 double msecs = avg / 1e6;
544 char cpustr[16] = { '\0', };
545 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
548 sprintf(cpustr, "CPU%*d%s",
550 evsel_list->cpus->map[cpu], csv_sep);
552 fprintf(output, fmt, cpustr, msecs, csv_sep, event_name(evsel));
555 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
560 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
561 fprintf(output, " # %8.3f CPUs utilized ",
562 avg / avg_stats(&walltime_nsecs_stats));
565 static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg)
567 double total, ratio = 0.0;
570 total = avg_stats(&runtime_cycles_stats[cpu]);
573 ratio = avg / total * 100.0;
575 color = PERF_COLOR_NORMAL;
577 color = PERF_COLOR_RED;
578 else if (ratio > 30.0)
579 color = PERF_COLOR_MAGENTA;
580 else if (ratio > 10.0)
581 color = PERF_COLOR_YELLOW;
583 fprintf(output, " # ");
584 color_fprintf(output, color, "%6.2f%%", ratio);
585 fprintf(output, " frontend cycles idle ");
588 static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg)
590 double total, ratio = 0.0;
593 total = avg_stats(&runtime_cycles_stats[cpu]);
596 ratio = avg / total * 100.0;
598 color = PERF_COLOR_NORMAL;
600 color = PERF_COLOR_RED;
601 else if (ratio > 50.0)
602 color = PERF_COLOR_MAGENTA;
603 else if (ratio > 20.0)
604 color = PERF_COLOR_YELLOW;
606 fprintf(output, " # ");
607 color_fprintf(output, color, "%6.2f%%", ratio);
608 fprintf(output, " backend cycles idle ");
611 static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
613 double total, ratio = 0.0;
616 total = avg_stats(&runtime_branches_stats[cpu]);
619 ratio = avg / total * 100.0;
621 color = PERF_COLOR_NORMAL;
623 color = PERF_COLOR_RED;
624 else if (ratio > 10.0)
625 color = PERF_COLOR_MAGENTA;
626 else if (ratio > 5.0)
627 color = PERF_COLOR_YELLOW;
629 fprintf(output, " # ");
630 color_fprintf(output, color, "%6.2f%%", ratio);
631 fprintf(output, " of all branches ");
634 static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
636 double total, ratio = 0.0;
639 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
642 ratio = avg / total * 100.0;
644 color = PERF_COLOR_NORMAL;
646 color = PERF_COLOR_RED;
647 else if (ratio > 10.0)
648 color = PERF_COLOR_MAGENTA;
649 else if (ratio > 5.0)
650 color = PERF_COLOR_YELLOW;
652 fprintf(output, " # ");
653 color_fprintf(output, color, "%6.2f%%", ratio);
654 fprintf(output, " of all L1-dcache hits ");
657 static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
659 double total, ratio = 0.0;
662 total = avg_stats(&runtime_l1_icache_stats[cpu]);
665 ratio = avg / total * 100.0;
667 color = PERF_COLOR_NORMAL;
669 color = PERF_COLOR_RED;
670 else if (ratio > 10.0)
671 color = PERF_COLOR_MAGENTA;
672 else if (ratio > 5.0)
673 color = PERF_COLOR_YELLOW;
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, struct perf_evsel *evsel __used, double avg)
682 double total, ratio = 0.0;
685 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
688 ratio = avg / total * 100.0;
690 color = PERF_COLOR_NORMAL;
692 color = PERF_COLOR_RED;
693 else if (ratio > 10.0)
694 color = PERF_COLOR_MAGENTA;
695 else if (ratio > 5.0)
696 color = PERF_COLOR_YELLOW;
698 fprintf(output, " # ");
699 color_fprintf(output, color, "%6.2f%%", ratio);
700 fprintf(output, " of all dTLB cache hits ");
703 static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
705 double total, ratio = 0.0;
708 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
711 ratio = avg / total * 100.0;
713 color = PERF_COLOR_NORMAL;
715 color = PERF_COLOR_RED;
716 else if (ratio > 10.0)
717 color = PERF_COLOR_MAGENTA;
718 else if (ratio > 5.0)
719 color = PERF_COLOR_YELLOW;
721 fprintf(output, " # ");
722 color_fprintf(output, color, "%6.2f%%", ratio);
723 fprintf(output, " of all iTLB cache hits ");
726 static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
728 double total, ratio = 0.0;
731 total = avg_stats(&runtime_ll_cache_stats[cpu]);
734 ratio = avg / total * 100.0;
736 color = PERF_COLOR_NORMAL;
738 color = PERF_COLOR_RED;
739 else if (ratio > 10.0)
740 color = PERF_COLOR_MAGENTA;
741 else if (ratio > 5.0)
742 color = PERF_COLOR_YELLOW;
744 fprintf(output, " # ");
745 color_fprintf(output, color, "%6.2f%%", ratio);
746 fprintf(output, " of all LL-cache hits ");
749 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
751 double total, ratio = 0.0;
752 char cpustr[16] = { '\0', };
758 fmt = "%s%'18.0f%s%-25s";
760 fmt = "%s%18.0f%s%-25s";
763 sprintf(cpustr, "CPU%*d%s",
765 evsel_list->cpus->map[cpu], csv_sep);
769 fprintf(output, fmt, cpustr, avg, csv_sep, event_name(evsel));
772 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
777 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
778 total = avg_stats(&runtime_cycles_stats[cpu]);
783 fprintf(output, " # %5.2f insns per cycle ", ratio);
785 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
786 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
790 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
793 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
794 runtime_branches_stats[cpu].n != 0) {
795 print_branch_misses(cpu, evsel, avg);
797 evsel->attr.type == PERF_TYPE_HW_CACHE &&
798 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
799 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
800 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
801 runtime_l1_dcache_stats[cpu].n != 0) {
802 print_l1_dcache_misses(cpu, evsel, avg);
804 evsel->attr.type == PERF_TYPE_HW_CACHE &&
805 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
806 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
807 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
808 runtime_l1_icache_stats[cpu].n != 0) {
809 print_l1_icache_misses(cpu, evsel, avg);
811 evsel->attr.type == PERF_TYPE_HW_CACHE &&
812 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
813 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
814 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
815 runtime_dtlb_cache_stats[cpu].n != 0) {
816 print_dtlb_cache_misses(cpu, evsel, avg);
818 evsel->attr.type == PERF_TYPE_HW_CACHE &&
819 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
820 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
821 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
822 runtime_itlb_cache_stats[cpu].n != 0) {
823 print_itlb_cache_misses(cpu, evsel, avg);
825 evsel->attr.type == PERF_TYPE_HW_CACHE &&
826 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
827 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
828 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
829 runtime_ll_cache_stats[cpu].n != 0) {
830 print_ll_cache_misses(cpu, evsel, avg);
831 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
832 runtime_cacherefs_stats[cpu].n != 0) {
833 total = avg_stats(&runtime_cacherefs_stats[cpu]);
836 ratio = avg * 100 / total;
838 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
840 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
841 print_stalled_cycles_frontend(cpu, evsel, avg);
842 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
843 print_stalled_cycles_backend(cpu, evsel, avg);
844 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
845 total = avg_stats(&runtime_nsecs_stats[cpu]);
848 ratio = 1.0 * avg / total;
850 fprintf(output, " # %8.3f GHz ", ratio);
851 } else if (runtime_nsecs_stats[cpu].n != 0) {
852 total = avg_stats(&runtime_nsecs_stats[cpu]);
855 ratio = 1000.0 * avg / total;
857 fprintf(output, " # %8.3f M/sec ", ratio);
859 fprintf(output, " ");
864 * Print out the results of a single counter:
865 * aggregated counts in system-wide mode
867 static void print_counter_aggr(struct perf_evsel *counter)
869 struct perf_stat *ps = counter->priv;
870 double avg = avg_stats(&ps->res_stats[0]);
871 int scaled = counter->counts->scaled;
874 fprintf(output, "%*s%s%*s",
876 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
878 csv_output ? 0 : -24,
879 event_name(counter));
882 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
888 if (nsec_counter(counter))
889 nsec_printout(-1, counter, avg);
891 abs_printout(-1, counter, avg);
893 print_noise(counter, avg);
901 double avg_enabled, avg_running;
903 avg_enabled = avg_stats(&ps->res_stats[1]);
904 avg_running = avg_stats(&ps->res_stats[2]);
906 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
908 fprintf(output, "\n");
912 * Print out the results of a single counter:
913 * does not use aggregated count in system-wide
915 static void print_counter(struct perf_evsel *counter)
920 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
921 val = counter->counts->cpu[cpu].val;
922 ena = counter->counts->cpu[cpu].ena;
923 run = counter->counts->cpu[cpu].run;
924 if (run == 0 || ena == 0) {
925 fprintf(output, "CPU%*d%s%*s%s%*s",
927 evsel_list->cpus->map[cpu], csv_sep,
929 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
931 csv_output ? 0 : -24,
932 event_name(counter));
935 fprintf(output, "%s%s",
936 csv_sep, counter->cgrp->name);
942 if (nsec_counter(counter))
943 nsec_printout(cpu, counter, val);
945 abs_printout(cpu, counter, val);
948 print_noise(counter, 1.0);
951 fprintf(output, " (%.2f%%)",
958 static void print_stat(int argc, const char **argv)
960 struct perf_evsel *counter;
966 fprintf(output, "\n");
967 fprintf(output, " Performance counter stats for ");
968 if(target_pid == -1 && target_tid == -1) {
969 fprintf(output, "\'%s", argv[0]);
970 for (i = 1; i < argc; i++)
971 fprintf(output, " %s", argv[i]);
972 } else if (target_pid != -1)
973 fprintf(output, "process id \'%d", target_pid);
975 fprintf(output, "thread id \'%d", target_tid);
977 fprintf(output, "\'");
979 fprintf(output, " (%d runs)", run_count);
980 fprintf(output, ":\n\n");
984 list_for_each_entry(counter, &evsel_list->entries, node)
985 print_counter(counter);
987 list_for_each_entry(counter, &evsel_list->entries, node)
988 print_counter_aggr(counter);
993 fprintf(output, "\n");
994 fprintf(output, " %17.9f seconds time elapsed",
995 avg_stats(&walltime_nsecs_stats)/1e9);
997 fprintf(output, " ");
998 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
999 avg_stats(&walltime_nsecs_stats));
1001 fprintf(output, "\n\n");
1005 static volatile int signr = -1;
1007 static void skip_signal(int signo)
1015 static void sig_atexit(void)
1017 if (child_pid != -1)
1018 kill(child_pid, SIGTERM);
1023 signal(signr, SIG_DFL);
1024 kill(getpid(), signr);
1027 static const char * const stat_usage[] = {
1028 "perf stat [<options>] [<command>]",
1032 static int stat__set_big_num(const struct option *opt __used,
1033 const char *s __used, int unset)
1035 big_num_opt = unset ? 0 : 1;
1039 static bool append_file;
1041 static const struct option options[] = {
1042 OPT_CALLBACK('e', "event", &evsel_list, "event",
1043 "event selector. use 'perf list' to list available events",
1044 parse_events_option),
1045 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1046 "event filter", parse_filter),
1047 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1048 "child tasks do not inherit counters"),
1049 OPT_INTEGER('p', "pid", &target_pid,
1050 "stat events on existing process id"),
1051 OPT_INTEGER('t', "tid", &target_tid,
1052 "stat events on existing thread id"),
1053 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1054 "system-wide collection from all CPUs"),
1055 OPT_BOOLEAN('g', "group", &group,
1056 "put the counters into a counter group"),
1057 OPT_BOOLEAN('c', "scale", &scale,
1058 "scale/normalize counters"),
1059 OPT_INCR('v', "verbose", &verbose,
1060 "be more verbose (show counter open errors, etc)"),
1061 OPT_INTEGER('r', "repeat", &run_count,
1062 "repeat command and print average + stddev (max: 100)"),
1063 OPT_BOOLEAN('n', "null", &null_run,
1064 "null run - dont start any counters"),
1065 OPT_INCR('d', "detailed", &detailed_run,
1066 "detailed run - start a lot of events"),
1067 OPT_BOOLEAN('S', "sync", &sync_run,
1068 "call sync() before starting a run"),
1069 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1070 "print large numbers with thousands\' separators",
1072 OPT_STRING('C', "cpu", &cpu_list, "cpu",
1073 "list of cpus to monitor in system-wide"),
1074 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1075 "disable CPU count aggregation"),
1076 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1077 "print counts with custom separator"),
1078 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1079 "monitor event in cgroup name only",
1081 OPT_STRING('o', "output", &output_name, "file",
1082 "output file name"),
1083 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1084 OPT_INTEGER(0, "log-fd", &output_fd,
1085 "log output to fd, instead of stderr"),
1090 * Add default attributes, if there were no attributes specified or
1091 * if -d/--detailed, -d -d or -d -d -d is used:
1093 static int add_default_attributes(void)
1095 struct perf_evsel *pos;
1099 /* Set attrs if no event is selected and !null_run: */
1103 if (!evsel_list->nr_entries) {
1104 for (c = 0; c < ARRAY_SIZE(default_attrs); c++) {
1105 pos = perf_evsel__new(default_attrs + c, c + attr_nr);
1108 perf_evlist__add(evsel_list, pos);
1113 /* Detailed events get appended to the event list: */
1115 if (detailed_run < 1)
1118 /* Append detailed run extra attributes: */
1119 for (c = 0; c < ARRAY_SIZE(detailed_attrs); c++) {
1120 pos = perf_evsel__new(detailed_attrs + c, c + attr_nr);
1123 perf_evlist__add(evsel_list, pos);
1127 if (detailed_run < 2)
1130 /* Append very detailed run extra attributes: */
1131 for (c = 0; c < ARRAY_SIZE(very_detailed_attrs); c++) {
1132 pos = perf_evsel__new(very_detailed_attrs + c, c + attr_nr);
1135 perf_evlist__add(evsel_list, pos);
1138 if (detailed_run < 3)
1141 /* Append very, very detailed run extra attributes: */
1142 for (c = 0; c < ARRAY_SIZE(very_very_detailed_attrs); c++) {
1143 pos = perf_evsel__new(very_very_detailed_attrs + c, c + attr_nr);
1146 perf_evlist__add(evsel_list, pos);
1153 int cmd_stat(int argc, const char **argv, const char *prefix __used)
1155 struct perf_evsel *pos;
1156 int status = -ENOMEM;
1159 setlocale(LC_ALL, "");
1161 evsel_list = perf_evlist__new(NULL, NULL);
1162 if (evsel_list == NULL)
1165 argc = parse_options(argc, argv, options, stat_usage,
1166 PARSE_OPT_STOP_AT_NON_OPTION);
1169 if (output_name && strcmp(output_name, "-"))
1172 if (output_name && output_fd) {
1173 fprintf(stderr, "cannot use both --output and --log-fd\n");
1174 usage_with_options(stat_usage, options);
1178 mode = append_file ? "a" : "w";
1180 output = fopen(output_name, mode);
1182 perror("failed to create output file");
1185 clock_gettime(CLOCK_REALTIME, &tm);
1186 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1187 } else if (output_fd != 2) {
1188 mode = append_file ? "a" : "w";
1189 output = fdopen(output_fd, mode);
1191 perror("Failed opening logfd");
1199 csv_sep = DEFAULT_SEPARATOR;
1202 * let the spreadsheet do the pretty-printing
1205 /* User explicitely passed -B? */
1206 if (big_num_opt == 1) {
1207 fprintf(stderr, "-B option not supported with -x\n");
1208 usage_with_options(stat_usage, options);
1209 } else /* Nope, so disable big number formatting */
1211 } else if (big_num_opt == 0) /* User passed --no-big-num */
1214 if (!argc && target_pid == -1 && target_tid == -1)
1215 usage_with_options(stat_usage, options);
1217 usage_with_options(stat_usage, options);
1219 /* no_aggr, cgroup are for system-wide only */
1220 if ((no_aggr || nr_cgroups) && !system_wide) {
1221 fprintf(stderr, "both cgroup and no-aggregation "
1222 "modes only available in system-wide mode\n");
1224 usage_with_options(stat_usage, options);
1227 if (add_default_attributes())
1230 if (target_pid != -1)
1231 target_tid = target_pid;
1233 evsel_list->threads = thread_map__new(target_pid, target_tid);
1234 if (evsel_list->threads == NULL) {
1235 pr_err("Problems finding threads of monitor\n");
1236 usage_with_options(stat_usage, options);
1240 evsel_list->cpus = cpu_map__new(cpu_list);
1242 evsel_list->cpus = cpu_map__dummy_new();
1244 if (evsel_list->cpus == NULL) {
1245 perror("failed to parse CPUs map");
1246 usage_with_options(stat_usage, options);
1250 list_for_each_entry(pos, &evsel_list->entries, node) {
1251 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1252 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 ||
1253 perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0)
1258 * We dont want to block the signals - that would cause
1259 * child tasks to inherit that and Ctrl-C would not work.
1260 * What we want is for Ctrl-C to work in the exec()-ed
1261 * task, but being ignored by perf stat itself:
1264 signal(SIGINT, skip_signal);
1265 signal(SIGALRM, skip_signal);
1266 signal(SIGABRT, skip_signal);
1269 for (run_idx = 0; run_idx < run_count; run_idx++) {
1270 if (run_count != 1 && verbose)
1271 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1277 status = run_perf_stat(argc, argv);
1281 print_stat(argc, argv);
1283 list_for_each_entry(pos, &evsel_list->entries, node)
1284 perf_evsel__free_stat_priv(pos);
1285 perf_evlist__delete_maps(evsel_list);
1287 perf_evlist__delete(evsel_list);