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;
197 static volatile int done = 0;
205 struct stats res_stats[3];
208 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
210 evsel->priv = zalloc(sizeof(struct perf_stat));
211 return evsel->priv == NULL ? -ENOMEM : 0;
214 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
220 static void update_stats(struct stats *stats, u64 val)
225 delta = val - stats->mean;
226 stats->mean += delta / stats->n;
227 stats->M2 += delta*(val - stats->mean);
230 static double avg_stats(struct stats *stats)
236 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
238 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
239 * s^2 = -------------------------------
242 * http://en.wikipedia.org/wiki/Stddev
244 * The std dev of the mean is related to the std dev by:
251 static double stddev_stats(struct stats *stats)
253 double variance = stats->M2 / (stats->n - 1);
254 double variance_mean = variance / stats->n;
256 return sqrt(variance_mean);
259 struct stats runtime_nsecs_stats[MAX_NR_CPUS];
260 struct stats runtime_cycles_stats[MAX_NR_CPUS];
261 struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
262 struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
263 struct stats runtime_branches_stats[MAX_NR_CPUS];
264 struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
265 struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
266 struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
267 struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
268 struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
269 struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
270 struct stats walltime_nsecs_stats;
272 static int create_perf_stat_counter(struct perf_evsel *evsel)
274 struct perf_event_attr *attr = &evsel->attr;
277 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
278 PERF_FORMAT_TOTAL_TIME_RUNNING;
280 attr->inherit = !no_inherit;
283 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, false);
285 if (target_pid == -1 && target_tid == -1) {
287 attr->enable_on_exec = 1;
290 return perf_evsel__open_per_thread(evsel, evsel_list->threads, false);
294 * Does the counter have nsecs as a unit?
296 static inline int nsec_counter(struct perf_evsel *evsel)
298 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
299 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
306 * Update various tracking values we maintain to print
307 * more semantic information such as miss/hit ratios,
308 * instruction rates, etc:
310 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
312 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
313 update_stats(&runtime_nsecs_stats[0], count[0]);
314 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
315 update_stats(&runtime_cycles_stats[0], count[0]);
316 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
317 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
318 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
319 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
320 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
321 update_stats(&runtime_branches_stats[0], count[0]);
322 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
323 update_stats(&runtime_cacherefs_stats[0], count[0]);
324 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
325 update_stats(&runtime_l1_dcache_stats[0], count[0]);
326 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
327 update_stats(&runtime_l1_icache_stats[0], count[0]);
328 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
329 update_stats(&runtime_ll_cache_stats[0], count[0]);
330 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
331 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
332 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
333 update_stats(&runtime_itlb_cache_stats[0], count[0]);
337 * Read out the results of a single counter:
338 * aggregate counts across CPUs in system-wide mode
340 static int read_counter_aggr(struct perf_evsel *counter)
342 struct perf_stat *ps = counter->priv;
343 u64 *count = counter->counts->aggr.values;
346 if (__perf_evsel__read(counter, evsel_list->cpus->nr,
347 evsel_list->threads->nr, scale) < 0)
350 for (i = 0; i < 3; i++)
351 update_stats(&ps->res_stats[i], count[i]);
354 fprintf(stderr, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
355 event_name(counter), count[0], count[1], count[2]);
359 * Save the full runtime - to allow normalization during printout:
361 update_shadow_stats(counter, count);
367 * Read out the results of a single counter:
368 * do not aggregate counts across CPUs in system-wide mode
370 static int read_counter(struct perf_evsel *counter)
375 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
376 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
379 count = counter->counts->cpu[cpu].values;
381 update_shadow_stats(counter, count);
387 static int run_perf_stat(int argc __used, const char **argv)
389 unsigned long long t0, t1;
390 struct perf_evsel *counter;
392 int child_ready_pipe[2], go_pipe[2];
393 const bool forks = (argc > 0);
396 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
397 perror("failed to create pipes");
402 if ((child_pid = fork()) < 0)
403 perror("failed to fork");
406 close(child_ready_pipe[0]);
408 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
411 * Do a dummy execvp to get the PLT entry resolved,
412 * so we avoid the resolver overhead on the real
415 execvp("", (char **)argv);
418 * Tell the parent we're ready to go
420 close(child_ready_pipe[1]);
423 * Wait until the parent tells us to go.
425 if (read(go_pipe[0], &buf, 1) == -1)
426 perror("unable to read pipe");
428 execvp(argv[0], (char **)argv);
434 if (target_tid == -1 && target_pid == -1 && !system_wide)
435 evsel_list->threads->map[0] = child_pid;
438 * Wait for the child to be ready to exec.
440 close(child_ready_pipe[1]);
442 if (read(child_ready_pipe[0], &buf, 1) == -1)
443 perror("unable to read pipe");
444 close(child_ready_pipe[0]);
447 list_for_each_entry(counter, &evsel_list->entries, node) {
448 if (create_perf_stat_counter(counter) < 0) {
449 if (errno == EINVAL || errno == ENOSYS || errno == ENOENT) {
451 ui__warning("%s event is not supported by the kernel.\n",
452 event_name(counter));
453 counter->supported = false;
457 if (errno == EPERM || errno == EACCES) {
458 error("You may not have permission to collect %sstats.\n"
459 "\t Consider tweaking"
460 " /proc/sys/kernel/perf_event_paranoid or running as root.",
461 system_wide ? "system-wide " : "");
463 error("open_counter returned with %d (%s). "
464 "/bin/dmesg may provide additional information.\n",
465 errno, strerror(errno));
468 kill(child_pid, SIGTERM);
469 die("Not all events could be opened.\n");
472 counter->supported = true;
475 if (perf_evlist__set_filters(evsel_list)) {
476 error("failed to set filter with %d (%s)\n", errno,
482 * Enable counters and exec the command:
490 while(!done) sleep(1);
495 update_stats(&walltime_nsecs_stats, t1 - t0);
498 list_for_each_entry(counter, &evsel_list->entries, node) {
499 read_counter(counter);
500 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
503 list_for_each_entry(counter, &evsel_list->entries, node) {
504 read_counter_aggr(counter);
505 perf_evsel__close_fd(counter, evsel_list->cpus->nr,
506 evsel_list->threads->nr);
510 return WEXITSTATUS(status);
513 static void print_noise_pct(double total, double avg)
518 pct = 100.0*total/avg;
520 fprintf(stderr, " ( +-%6.2f%% )", pct);
523 static void print_noise(struct perf_evsel *evsel, double avg)
525 struct perf_stat *ps;
531 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
534 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
536 double msecs = avg / 1e6;
537 char cpustr[16] = { '\0', };
538 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
541 sprintf(cpustr, "CPU%*d%s",
543 evsel_list->cpus->map[cpu], csv_sep);
545 fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel));
548 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
553 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
554 fprintf(stderr, " # %8.3f CPUs utilized ", avg / avg_stats(&walltime_nsecs_stats));
557 static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg)
559 double total, ratio = 0.0;
562 total = avg_stats(&runtime_cycles_stats[cpu]);
565 ratio = avg / total * 100.0;
567 color = PERF_COLOR_NORMAL;
569 color = PERF_COLOR_RED;
570 else if (ratio > 30.0)
571 color = PERF_COLOR_MAGENTA;
572 else if (ratio > 10.0)
573 color = PERF_COLOR_YELLOW;
575 fprintf(stderr, " # ");
576 color_fprintf(stderr, color, "%6.2f%%", ratio);
577 fprintf(stderr, " frontend cycles idle ");
580 static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg)
582 double total, ratio = 0.0;
585 total = avg_stats(&runtime_cycles_stats[cpu]);
588 ratio = avg / total * 100.0;
590 color = PERF_COLOR_NORMAL;
592 color = PERF_COLOR_RED;
593 else if (ratio > 50.0)
594 color = PERF_COLOR_MAGENTA;
595 else if (ratio > 20.0)
596 color = PERF_COLOR_YELLOW;
598 fprintf(stderr, " # ");
599 color_fprintf(stderr, color, "%6.2f%%", ratio);
600 fprintf(stderr, " backend cycles idle ");
603 static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
605 double total, ratio = 0.0;
608 total = avg_stats(&runtime_branches_stats[cpu]);
611 ratio = avg / total * 100.0;
613 color = PERF_COLOR_NORMAL;
615 color = PERF_COLOR_RED;
616 else if (ratio > 10.0)
617 color = PERF_COLOR_MAGENTA;
618 else if (ratio > 5.0)
619 color = PERF_COLOR_YELLOW;
621 fprintf(stderr, " # ");
622 color_fprintf(stderr, color, "%6.2f%%", ratio);
623 fprintf(stderr, " of all branches ");
626 static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
628 double total, ratio = 0.0;
631 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
634 ratio = avg / total * 100.0;
636 color = PERF_COLOR_NORMAL;
638 color = PERF_COLOR_RED;
639 else if (ratio > 10.0)
640 color = PERF_COLOR_MAGENTA;
641 else if (ratio > 5.0)
642 color = PERF_COLOR_YELLOW;
644 fprintf(stderr, " # ");
645 color_fprintf(stderr, color, "%6.2f%%", ratio);
646 fprintf(stderr, " of all L1-dcache hits ");
649 static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
651 double total, ratio = 0.0;
654 total = avg_stats(&runtime_l1_icache_stats[cpu]);
657 ratio = avg / total * 100.0;
659 color = PERF_COLOR_NORMAL;
661 color = PERF_COLOR_RED;
662 else if (ratio > 10.0)
663 color = PERF_COLOR_MAGENTA;
664 else if (ratio > 5.0)
665 color = PERF_COLOR_YELLOW;
667 fprintf(stderr, " # ");
668 color_fprintf(stderr, color, "%6.2f%%", ratio);
669 fprintf(stderr, " of all L1-icache hits ");
672 static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
674 double total, ratio = 0.0;
677 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
680 ratio = avg / total * 100.0;
682 color = PERF_COLOR_NORMAL;
684 color = PERF_COLOR_RED;
685 else if (ratio > 10.0)
686 color = PERF_COLOR_MAGENTA;
687 else if (ratio > 5.0)
688 color = PERF_COLOR_YELLOW;
690 fprintf(stderr, " # ");
691 color_fprintf(stderr, color, "%6.2f%%", ratio);
692 fprintf(stderr, " of all dTLB cache hits ");
695 static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
697 double total, ratio = 0.0;
700 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
703 ratio = avg / total * 100.0;
705 color = PERF_COLOR_NORMAL;
707 color = PERF_COLOR_RED;
708 else if (ratio > 10.0)
709 color = PERF_COLOR_MAGENTA;
710 else if (ratio > 5.0)
711 color = PERF_COLOR_YELLOW;
713 fprintf(stderr, " # ");
714 color_fprintf(stderr, color, "%6.2f%%", ratio);
715 fprintf(stderr, " of all iTLB cache hits ");
718 static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
720 double total, ratio = 0.0;
723 total = avg_stats(&runtime_ll_cache_stats[cpu]);
726 ratio = avg / total * 100.0;
728 color = PERF_COLOR_NORMAL;
730 color = PERF_COLOR_RED;
731 else if (ratio > 10.0)
732 color = PERF_COLOR_MAGENTA;
733 else if (ratio > 5.0)
734 color = PERF_COLOR_YELLOW;
736 fprintf(stderr, " # ");
737 color_fprintf(stderr, color, "%6.2f%%", ratio);
738 fprintf(stderr, " of all LL-cache hits ");
741 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
743 double total, ratio = 0.0;
744 char cpustr[16] = { '\0', };
750 fmt = "%s%'18.0f%s%-25s";
752 fmt = "%s%18.0f%s%-25s";
755 sprintf(cpustr, "CPU%*d%s",
757 evsel_list->cpus->map[cpu], csv_sep);
761 fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel));
764 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
769 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
770 total = avg_stats(&runtime_cycles_stats[cpu]);
775 fprintf(stderr, " # %5.2f insns per cycle ", ratio);
777 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
778 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
782 fprintf(stderr, "\n # %5.2f stalled cycles per insn", ratio);
785 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
786 runtime_branches_stats[cpu].n != 0) {
787 print_branch_misses(cpu, evsel, avg);
789 evsel->attr.type == PERF_TYPE_HW_CACHE &&
790 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
791 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
792 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
793 runtime_l1_dcache_stats[cpu].n != 0) {
794 print_l1_dcache_misses(cpu, evsel, avg);
796 evsel->attr.type == PERF_TYPE_HW_CACHE &&
797 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
798 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
799 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
800 runtime_l1_icache_stats[cpu].n != 0) {
801 print_l1_icache_misses(cpu, evsel, avg);
803 evsel->attr.type == PERF_TYPE_HW_CACHE &&
804 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
805 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
806 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
807 runtime_dtlb_cache_stats[cpu].n != 0) {
808 print_dtlb_cache_misses(cpu, evsel, avg);
810 evsel->attr.type == PERF_TYPE_HW_CACHE &&
811 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
812 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
813 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
814 runtime_itlb_cache_stats[cpu].n != 0) {
815 print_itlb_cache_misses(cpu, evsel, avg);
817 evsel->attr.type == PERF_TYPE_HW_CACHE &&
818 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
819 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
820 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
821 runtime_ll_cache_stats[cpu].n != 0) {
822 print_ll_cache_misses(cpu, evsel, avg);
823 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
824 runtime_cacherefs_stats[cpu].n != 0) {
825 total = avg_stats(&runtime_cacherefs_stats[cpu]);
828 ratio = avg * 100 / total;
830 fprintf(stderr, " # %8.3f %% of all cache refs ", ratio);
832 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
833 print_stalled_cycles_frontend(cpu, evsel, avg);
834 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
835 print_stalled_cycles_backend(cpu, evsel, avg);
836 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
837 total = avg_stats(&runtime_nsecs_stats[cpu]);
840 ratio = 1.0 * avg / total;
842 fprintf(stderr, " # %8.3f GHz ", ratio);
843 } else if (runtime_nsecs_stats[cpu].n != 0) {
844 total = avg_stats(&runtime_nsecs_stats[cpu]);
847 ratio = 1000.0 * avg / total;
849 fprintf(stderr, " # %8.3f M/sec ", ratio);
851 fprintf(stderr, " ");
856 * Print out the results of a single counter:
857 * aggregated counts in system-wide mode
859 static void print_counter_aggr(struct perf_evsel *counter)
861 struct perf_stat *ps = counter->priv;
862 double avg = avg_stats(&ps->res_stats[0]);
863 int scaled = counter->counts->scaled;
866 fprintf(stderr, "%*s%s%*s",
868 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
870 csv_output ? 0 : -24,
871 event_name(counter));
874 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
880 if (nsec_counter(counter))
881 nsec_printout(-1, counter, avg);
883 abs_printout(-1, counter, avg);
890 print_noise(counter, avg);
893 double avg_enabled, avg_running;
895 avg_enabled = avg_stats(&ps->res_stats[1]);
896 avg_running = avg_stats(&ps->res_stats[2]);
898 fprintf(stderr, " [%5.2f%%]", 100 * avg_running / avg_enabled);
900 fprintf(stderr, "\n");
904 * Print out the results of a single counter:
905 * does not use aggregated count in system-wide
907 static void print_counter(struct perf_evsel *counter)
912 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
913 val = counter->counts->cpu[cpu].val;
914 ena = counter->counts->cpu[cpu].ena;
915 run = counter->counts->cpu[cpu].run;
916 if (run == 0 || ena == 0) {
917 fprintf(stderr, "CPU%*d%s%*s%s%*s",
919 evsel_list->cpus->map[cpu], csv_sep,
921 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
923 csv_output ? 0 : -24,
924 event_name(counter));
927 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
933 if (nsec_counter(counter))
934 nsec_printout(cpu, counter, val);
936 abs_printout(cpu, counter, val);
939 print_noise(counter, 1.0);
942 fprintf(stderr, " (%.2f%%)", 100.0 * run / ena);
948 static void print_stat(int argc, const char **argv)
950 struct perf_evsel *counter;
956 fprintf(stderr, "\n");
957 fprintf(stderr, " Performance counter stats for ");
958 if(target_pid == -1 && target_tid == -1) {
959 fprintf(stderr, "\'%s", argv[0]);
960 for (i = 1; i < argc; i++)
961 fprintf(stderr, " %s", argv[i]);
962 } else if (target_pid != -1)
963 fprintf(stderr, "process id \'%d", target_pid);
965 fprintf(stderr, "thread id \'%d", target_tid);
967 fprintf(stderr, "\'");
969 fprintf(stderr, " (%d runs)", run_count);
970 fprintf(stderr, ":\n\n");
974 list_for_each_entry(counter, &evsel_list->entries, node)
975 print_counter(counter);
977 list_for_each_entry(counter, &evsel_list->entries, node)
978 print_counter_aggr(counter);
983 fprintf(stderr, "\n");
984 fprintf(stderr, " %17.9f seconds time elapsed",
985 avg_stats(&walltime_nsecs_stats)/1e9);
987 fprintf(stderr, " ");
988 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
989 avg_stats(&walltime_nsecs_stats));
991 fprintf(stderr, "\n\n");
995 static volatile int signr = -1;
997 static void skip_signal(int signo)
1005 static void sig_atexit(void)
1007 if (child_pid != -1)
1008 kill(child_pid, SIGTERM);
1013 signal(signr, SIG_DFL);
1014 kill(getpid(), signr);
1017 static const char * const stat_usage[] = {
1018 "perf stat [<options>] [<command>]",
1022 static int stat__set_big_num(const struct option *opt __used,
1023 const char *s __used, int unset)
1025 big_num_opt = unset ? 0 : 1;
1029 static const struct option options[] = {
1030 OPT_CALLBACK('e', "event", &evsel_list, "event",
1031 "event selector. use 'perf list' to list available events",
1033 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1034 "event filter", parse_filter),
1035 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1036 "child tasks do not inherit counters"),
1037 OPT_INTEGER('p', "pid", &target_pid,
1038 "stat events on existing process id"),
1039 OPT_INTEGER('t', "tid", &target_tid,
1040 "stat events on existing thread id"),
1041 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1042 "system-wide collection from all CPUs"),
1043 OPT_BOOLEAN('c', "scale", &scale,
1044 "scale/normalize counters"),
1045 OPT_INCR('v', "verbose", &verbose,
1046 "be more verbose (show counter open errors, etc)"),
1047 OPT_INTEGER('r', "repeat", &run_count,
1048 "repeat command and print average + stddev (max: 100)"),
1049 OPT_BOOLEAN('n', "null", &null_run,
1050 "null run - dont start any counters"),
1051 OPT_INCR('d', "detailed", &detailed_run,
1052 "detailed run - start a lot of events"),
1053 OPT_BOOLEAN('S', "sync", &sync_run,
1054 "call sync() before starting a run"),
1055 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1056 "print large numbers with thousands\' separators",
1058 OPT_STRING('C', "cpu", &cpu_list, "cpu",
1059 "list of cpus to monitor in system-wide"),
1060 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1061 "disable CPU count aggregation"),
1062 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1063 "print counts with custom separator"),
1064 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1065 "monitor event in cgroup name only",
1071 * Add default attributes, if there were no attributes specified or
1072 * if -d/--detailed, -d -d or -d -d -d is used:
1074 static int add_default_attributes(void)
1076 struct perf_evsel *pos;
1080 /* Set attrs if no event is selected and !null_run: */
1084 if (!evsel_list->nr_entries) {
1085 for (c = 0; c < ARRAY_SIZE(default_attrs); c++) {
1086 pos = perf_evsel__new(default_attrs + c, c + attr_nr);
1089 perf_evlist__add(evsel_list, pos);
1094 /* Detailed events get appended to the event list: */
1096 if (detailed_run < 1)
1099 /* Append detailed run extra attributes: */
1100 for (c = 0; c < ARRAY_SIZE(detailed_attrs); c++) {
1101 pos = perf_evsel__new(detailed_attrs + c, c + attr_nr);
1104 perf_evlist__add(evsel_list, pos);
1108 if (detailed_run < 2)
1111 /* Append very detailed run extra attributes: */
1112 for (c = 0; c < ARRAY_SIZE(very_detailed_attrs); c++) {
1113 pos = perf_evsel__new(very_detailed_attrs + c, c + attr_nr);
1116 perf_evlist__add(evsel_list, pos);
1119 if (detailed_run < 3)
1122 /* Append very, very detailed run extra attributes: */
1123 for (c = 0; c < ARRAY_SIZE(very_very_detailed_attrs); c++) {
1124 pos = perf_evsel__new(very_very_detailed_attrs + c, c + attr_nr);
1127 perf_evlist__add(evsel_list, pos);
1134 int cmd_stat(int argc, const char **argv, const char *prefix __used)
1136 struct perf_evsel *pos;
1137 int status = -ENOMEM;
1139 setlocale(LC_ALL, "");
1141 evsel_list = perf_evlist__new(NULL, NULL);
1142 if (evsel_list == NULL)
1145 argc = parse_options(argc, argv, options, stat_usage,
1146 PARSE_OPT_STOP_AT_NON_OPTION);
1151 csv_sep = DEFAULT_SEPARATOR;
1154 * let the spreadsheet do the pretty-printing
1157 /* User explicitely passed -B? */
1158 if (big_num_opt == 1) {
1159 fprintf(stderr, "-B option not supported with -x\n");
1160 usage_with_options(stat_usage, options);
1161 } else /* Nope, so disable big number formatting */
1163 } else if (big_num_opt == 0) /* User passed --no-big-num */
1166 if (!argc && target_pid == -1 && target_tid == -1)
1167 usage_with_options(stat_usage, options);
1169 usage_with_options(stat_usage, options);
1171 /* no_aggr, cgroup are for system-wide only */
1172 if ((no_aggr || nr_cgroups) && !system_wide) {
1173 fprintf(stderr, "both cgroup and no-aggregation "
1174 "modes only available in system-wide mode\n");
1176 usage_with_options(stat_usage, options);
1179 if (add_default_attributes())
1182 if (target_pid != -1)
1183 target_tid = target_pid;
1185 evsel_list->threads = thread_map__new(target_pid, target_tid);
1186 if (evsel_list->threads == NULL) {
1187 pr_err("Problems finding threads of monitor\n");
1188 usage_with_options(stat_usage, options);
1192 evsel_list->cpus = cpu_map__new(cpu_list);
1194 evsel_list->cpus = cpu_map__dummy_new();
1196 if (evsel_list->cpus == NULL) {
1197 perror("failed to parse CPUs map");
1198 usage_with_options(stat_usage, options);
1202 list_for_each_entry(pos, &evsel_list->entries, node) {
1203 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1204 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 ||
1205 perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0)
1210 * We dont want to block the signals - that would cause
1211 * child tasks to inherit that and Ctrl-C would not work.
1212 * What we want is for Ctrl-C to work in the exec()-ed
1213 * task, but being ignored by perf stat itself:
1216 signal(SIGINT, skip_signal);
1217 signal(SIGALRM, skip_signal);
1218 signal(SIGABRT, skip_signal);
1221 for (run_idx = 0; run_idx < run_count; run_idx++) {
1222 if (run_count != 1 && verbose)
1223 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
1228 status = run_perf_stat(argc, argv);
1232 print_stat(argc, argv);
1234 list_for_each_entry(pos, &evsel_list->entries, node)
1235 perf_evsel__free_stat_priv(pos);
1236 perf_evlist__delete_maps(evsel_list);
1238 perf_evlist__delete(evsel_list);