?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
perf tools: Enable grouping logic for parsed events
[linux-2.6/btrfs-unstable.git] / tools / perf / builtin-stat.c
blob23908a85bba9f14b35cad0258878d5ffb3aabafe
1 /*
2 * builtin-stat.c
3 *
4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
6 *
7 * Sample output:
8
9 $ perf stat ./hackbench 10
10
11 Time: 0.118
12
13 Performance counter stats for './hackbench 10':
14
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
26
27 0.154822978 seconds time elapsed
28
29 *
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
31 *
32 * Improvements and fixes by:
33 *
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>
40 *
41 * Released under the GPL v2. (and only v2, not any later version)
42 */
43
44 #include "perf.h"
45 #include "builtin.h"
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"
58
59 #include <sys/prctl.h>
60 #include <math.h>
61 #include <locale.h>
62
63 #define DEFAULT_SEPARATOR " "
64 #define CNTR_NOT_SUPPORTED "<not supported>"
65 #define CNTR_NOT_COUNTED "<not counted>"
66
67 static struct perf_event_attr default_attrs[] = {
68
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 },
73
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 },
80
81 };
82
83 /*
84 * Detailed stats (-d), covering the L1 and last level data caches:
85 */
86 static struct perf_event_attr detailed_attrs[] = {
87
88 { .type = PERF_TYPE_HW_CACHE,
89 .config =
90 PERF_COUNT_HW_CACHE_L1D << 0 |
91 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
92 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
93
94 { .type = PERF_TYPE_HW_CACHE,
95 .config =
96 PERF_COUNT_HW_CACHE_L1D << 0 |
97 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
98 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
99
100 { .type = PERF_TYPE_HW_CACHE,
101 .config =
102 PERF_COUNT_HW_CACHE_LL << 0 |
103 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
104 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
105
106 { .type = PERF_TYPE_HW_CACHE,
107 .config =
108 PERF_COUNT_HW_CACHE_LL << 0 |
109 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
110 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
111 };
112
113 /*
114 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
115 */
116 static struct perf_event_attr very_detailed_attrs[] = {
117
118 { .type = PERF_TYPE_HW_CACHE,
119 .config =
120 PERF_COUNT_HW_CACHE_L1I << 0 |
121 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
122 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
123
124 { .type = PERF_TYPE_HW_CACHE,
125 .config =
126 PERF_COUNT_HW_CACHE_L1I << 0 |
127 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
128 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
129
130 { .type = PERF_TYPE_HW_CACHE,
131 .config =
132 PERF_COUNT_HW_CACHE_DTLB << 0 |
133 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
134 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
135
136 { .type = PERF_TYPE_HW_CACHE,
137 .config =
138 PERF_COUNT_HW_CACHE_DTLB << 0 |
139 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
140 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
141
142 { .type = PERF_TYPE_HW_CACHE,
143 .config =
144 PERF_COUNT_HW_CACHE_ITLB << 0 |
145 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
146 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
147
148 { .type = PERF_TYPE_HW_CACHE,
149 .config =
150 PERF_COUNT_HW_CACHE_ITLB << 0 |
151 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
152 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
153
154 };
155
156 /*
157 * Very, very detailed stats (-d -d -d), adding prefetch events:
158 */
159 static struct perf_event_attr very_very_detailed_attrs[] = {
160
161 { .type = PERF_TYPE_HW_CACHE,
162 .config =
163 PERF_COUNT_HW_CACHE_L1D << 0 |
164 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
165 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
166
167 { .type = PERF_TYPE_HW_CACHE,
168 .config =
169 PERF_COUNT_HW_CACHE_L1D << 0 |
170 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
171 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
172 };
173
174
175
176 static struct perf_evlist *evsel_list;
177
178 static struct perf_target target = {
179 .uid = UINT_MAX,
180 };
181
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;
199
200 static volatile int done = 0;
201
202 struct stats
203 {
204 double n, mean, M2;
205 };
206
207 struct perf_stat {
208 struct stats res_stats[3];
209 };
210
211 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
212 {
213 evsel->priv = zalloc(sizeof(struct perf_stat));
214 return evsel->priv == NULL ? -ENOMEM : 0;
215 }
216
217 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
218 {
219 free(evsel->priv);
220 evsel->priv = NULL;
221 }
222
223 static void update_stats(struct stats *stats, u64 val)
224 {
225 double delta;
226
227 stats->n++;
228 delta = val - stats->mean;
229 stats->mean += delta / stats->n;
230 stats->M2 += delta*(val - stats->mean);
231 }
232
233 static double avg_stats(struct stats *stats)
234 {
235 return stats->mean;
236 }
237
238 /*
239 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
240 *
241 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
242 * s^2 = -------------------------------
243 * n - 1
244 *
245 * http://en.wikipedia.org/wiki/Stddev
246 *
247 * The std dev of the mean is related to the std dev by:
248 *
249 * s
250 * s_mean = -------
251 * sqrt(n)
252 *
253 */
254 static double stddev_stats(struct stats *stats)
255 {
256 double variance, variance_mean;
257
258 if (!stats->n)
259 return 0.0;
260
261 variance = stats->M2 / (stats->n - 1);
262 variance_mean = variance / stats->n;
263
264 return sqrt(variance_mean);
265 }
266
267 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
268 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
269 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
270 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
271 static struct stats runtime_branches_stats[MAX_NR_CPUS];
272 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
273 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
274 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
275 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
276 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
277 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
278 static struct stats walltime_nsecs_stats;
279
280 static int create_perf_stat_counter(struct perf_evsel *evsel,
281 struct perf_evsel *first)
282 {
283 struct perf_event_attr *attr = &evsel->attr;
284 bool exclude_guest_missing = false;
285 int ret;
286
287 if (scale)
288 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
289 PERF_FORMAT_TOTAL_TIME_RUNNING;
290
291 attr->inherit = !no_inherit;
292
293 retry:
294 if (exclude_guest_missing)
295 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
296
297 if (perf_target__has_cpu(&target)) {
298 ret = perf_evsel__open_per_cpu(evsel, evsel_list->cpus);
299 if (ret)
300 goto check_ret;
301 return 0;
302 }
303
304 if (!perf_target__has_task(&target) && (!group || evsel == first)) {
305 attr->disabled = 1;
306 attr->enable_on_exec = 1;
307 }
308
309 ret = perf_evsel__open_per_thread(evsel, evsel_list->threads);
310 if (!ret)
311 return 0;
312 /* fall through */
313 check_ret:
314 if (ret && errno == EINVAL) {
315 if (!exclude_guest_missing &&
316 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
317 pr_debug("Old kernel, cannot exclude "
318 "guest or host samples.\n");
319 exclude_guest_missing = true;
320 goto retry;
321 }
322 }
323 return ret;
324 }
325
326 /*
327 * Does the counter have nsecs as a unit?
328 */
329 static inline int nsec_counter(struct perf_evsel *evsel)
330 {
331 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
332 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
333 return 1;
334
335 return 0;
336 }
337
338 /*
339 * Update various tracking values we maintain to print
340 * more semantic information such as miss/hit ratios,
341 * instruction rates, etc:
342 */
343 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
344 {
345 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
346 update_stats(&runtime_nsecs_stats[0], count[0]);
347 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
348 update_stats(&runtime_cycles_stats[0], count[0]);
349 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
350 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
351 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
352 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
353 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
354 update_stats(&runtime_branches_stats[0], count[0]);
355 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
356 update_stats(&runtime_cacherefs_stats[0], count[0]);
357 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
358 update_stats(&runtime_l1_dcache_stats[0], count[0]);
359 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
360 update_stats(&runtime_l1_icache_stats[0], count[0]);
361 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
362 update_stats(&runtime_ll_cache_stats[0], count[0]);
363 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
364 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
365 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
366 update_stats(&runtime_itlb_cache_stats[0], count[0]);
367 }
368
369 /*
370 * Read out the results of a single counter:
371 * aggregate counts across CPUs in system-wide mode
372 */
373 static int read_counter_aggr(struct perf_evsel *counter)
374 {
375 struct perf_stat *ps = counter->priv;
376 u64 *count = counter->counts->aggr.values;
377 int i;
378
379 if (__perf_evsel__read(counter, evsel_list->cpus->nr,
380 evsel_list->threads->nr, scale) < 0)
381 return -1;
382
383 for (i = 0; i < 3; i++)
384 update_stats(&ps->res_stats[i], count[i]);
385
386 if (verbose) {
387 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
388 perf_evsel__name(counter), count[0], count[1], count[2]);
389 }
390
391 /*
392 * Save the full runtime - to allow normalization during printout:
393 */
394 update_shadow_stats(counter, count);
395
396 return 0;
397 }
398
399 /*
400 * Read out the results of a single counter:
401 * do not aggregate counts across CPUs in system-wide mode
402 */
403 static int read_counter(struct perf_evsel *counter)
404 {
405 u64 *count;
406 int cpu;
407
408 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
409 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
410 return -1;
411
412 count = counter->counts->cpu[cpu].values;
413
414 update_shadow_stats(counter, count);
415 }
416
417 return 0;
418 }
419
420 static int run_perf_stat(int argc __used, const char **argv)
421 {
422 unsigned long long t0, t1;
423 struct perf_evsel *counter, *first;
424 int status = 0;
425 int child_ready_pipe[2], go_pipe[2];
426 const bool forks = (argc > 0);
427 char buf;
428
429 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
430 perror("failed to create pipes");
431 exit(1);
432 }
433
434 if (forks) {
435 if ((child_pid = fork()) < 0)
436 perror("failed to fork");
437
438 if (!child_pid) {
439 close(child_ready_pipe[0]);
440 close(go_pipe[1]);
441 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
442
443 /*
444 * Do a dummy execvp to get the PLT entry resolved,
445 * so we avoid the resolver overhead on the real
446 * execvp call.
447 */
448 execvp("", (char **)argv);
449
450 /*
451 * Tell the parent we're ready to go
452 */
453 close(child_ready_pipe[1]);
454
455 /*
456 * Wait until the parent tells us to go.
457 */
458 if (read(go_pipe[0], &buf, 1) == -1)
459 perror("unable to read pipe");
460
461 execvp(argv[0], (char **)argv);
462
463 perror(argv[0]);
464 exit(-1);
465 }
466
467 if (perf_target__none(&target))
468 evsel_list->threads->map[0] = child_pid;
469
470 /*
471 * Wait for the child to be ready to exec.
472 */
473 close(child_ready_pipe[1]);
474 close(go_pipe[0]);
475 if (read(child_ready_pipe[0], &buf, 1) == -1)
476 perror("unable to read pipe");
477 close(child_ready_pipe[0]);
478 }
479
480 if (group)
481 perf_evlist__group(evsel_list);
482
483 first = list_entry(evsel_list->entries.next, struct perf_evsel, node);
484
485 list_for_each_entry(counter, &evsel_list->entries, node) {
486 if (create_perf_stat_counter(counter, first) < 0) {
487 /*
488 * PPC returns ENXIO for HW counters until 2.6.37
489 * (behavior changed with commit b0a873e).
490 */
491 if (errno == EINVAL || errno == ENOSYS ||
492 errno == ENOENT || errno == EOPNOTSUPP ||
493 errno == ENXIO) {
494 if (verbose)
495 ui__warning("%s event is not supported by the kernel.\n",
496 perf_evsel__name(counter));
497 counter->supported = false;
498 continue;
499 }
500
501 if (errno == EPERM || errno == EACCES) {
502 error("You may not have permission to collect %sstats.\n"
503 "\t Consider tweaking"
504 " /proc/sys/kernel/perf_event_paranoid or running as root.",
505 target.system_wide ? "system-wide " : "");
506 } else {
507 error("open_counter returned with %d (%s). "
508 "/bin/dmesg may provide additional information.\n",
509 errno, strerror(errno));
510 }
511 if (child_pid != -1)
512 kill(child_pid, SIGTERM);
513 die("Not all events could be opened.\n");
514 return -1;
515 }
516 counter->supported = true;
517 }
518
519 if (perf_evlist__set_filters(evsel_list)) {
520 error("failed to set filter with %d (%s)\n", errno,
521 strerror(errno));
522 return -1;
523 }
524
525 /*
526 * Enable counters and exec the command:
527 */
528 t0 = rdclock();
529
530 if (forks) {
531 close(go_pipe[1]);
532 wait(&status);
533 if (WIFSIGNALED(status))
534 psignal(WTERMSIG(status), argv[0]);
535 } else {
536 while(!done) sleep(1);
537 }
538
539 t1 = rdclock();
540
541 update_stats(&walltime_nsecs_stats, t1 - t0);
542
543 if (no_aggr) {
544 list_for_each_entry(counter, &evsel_list->entries, node) {
545 read_counter(counter);
546 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
547 }
548 } else {
549 list_for_each_entry(counter, &evsel_list->entries, node) {
550 read_counter_aggr(counter);
551 perf_evsel__close_fd(counter, evsel_list->cpus->nr,
552 evsel_list->threads->nr);
553 }
554 }
555
556 return WEXITSTATUS(status);
557 }
558
559 static void print_noise_pct(double total, double avg)
560 {
561 double pct = 0.0;
562
563 if (avg)
564 pct = 100.0*total/avg;
565
566 if (csv_output)
567 fprintf(output, "%s%.2f%%", csv_sep, pct);
568 else if (pct)
569 fprintf(output, " ( +-%6.2f%% )", pct);
570 }
571
572 static void print_noise(struct perf_evsel *evsel, double avg)
573 {
574 struct perf_stat *ps;
575
576 if (run_count == 1)
577 return;
578
579 ps = evsel->priv;
580 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
581 }
582
583 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
584 {
585 double msecs = avg / 1e6;
586 char cpustr[16] = { '\0', };
587 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
588
589 if (no_aggr)
590 sprintf(cpustr, "CPU%*d%s",
591 csv_output ? 0 : -4,
592 evsel_list->cpus->map[cpu], csv_sep);
593
594 fprintf(output, fmt, cpustr, msecs, csv_sep, perf_evsel__name(evsel));
595
596 if (evsel->cgrp)
597 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
598
599 if (csv_output)
600 return;
601
602 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
603 fprintf(output, " # %8.3f CPUs utilized ",
604 avg / avg_stats(&walltime_nsecs_stats));
605 else
606 fprintf(output, " ");
607 }
608
609 /* used for get_ratio_color() */
610 enum grc_type {
611 GRC_STALLED_CYCLES_FE,
612 GRC_STALLED_CYCLES_BE,
613 GRC_CACHE_MISSES,
614 GRC_MAX_NR
615 };
616
617 static const char *get_ratio_color(enum grc_type type, double ratio)
618 {
619 static const double grc_table[GRC_MAX_NR][3] = {
620 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
621 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
622 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
623 };
624 const char *color = PERF_COLOR_NORMAL;
625
626 if (ratio > grc_table[type][0])
627 color = PERF_COLOR_RED;
628 else if (ratio > grc_table[type][1])
629 color = PERF_COLOR_MAGENTA;
630 else if (ratio > grc_table[type][2])
631 color = PERF_COLOR_YELLOW;
632
633 return color;
634 }
635
636 static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg)
637 {
638 double total, ratio = 0.0;
639 const char *color;
640
641 total = avg_stats(&runtime_cycles_stats[cpu]);
642
643 if (total)
644 ratio = avg / total * 100.0;
645
646 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
647
648 fprintf(output, " # ");
649 color_fprintf(output, color, "%6.2f%%", ratio);
650 fprintf(output, " frontend cycles idle ");
651 }
652
653 static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg)
654 {
655 double total, ratio = 0.0;
656 const char *color;
657
658 total = avg_stats(&runtime_cycles_stats[cpu]);
659
660 if (total)
661 ratio = avg / total * 100.0;
662
663 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
664
665 fprintf(output, " # ");
666 color_fprintf(output, color, "%6.2f%%", ratio);
667 fprintf(output, " backend cycles idle ");
668 }
669
670 static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
671 {
672 double total, ratio = 0.0;
673 const char *color;
674
675 total = avg_stats(&runtime_branches_stats[cpu]);
676
677 if (total)
678 ratio = avg / total * 100.0;
679
680 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
681
682 fprintf(output, " # ");
683 color_fprintf(output, color, "%6.2f%%", ratio);
684 fprintf(output, " of all branches ");
685 }
686
687 static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
688 {
689 double total, ratio = 0.0;
690 const char *color;
691
692 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
693
694 if (total)
695 ratio = avg / total * 100.0;
696
697 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
698
699 fprintf(output, " # ");
700 color_fprintf(output, color, "%6.2f%%", ratio);
701 fprintf(output, " of all L1-dcache hits ");
702 }
703
704 static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
705 {
706 double total, ratio = 0.0;
707 const char *color;
708
709 total = avg_stats(&runtime_l1_icache_stats[cpu]);
710
711 if (total)
712 ratio = avg / total * 100.0;
713
714 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
715
716 fprintf(output, " # ");
717 color_fprintf(output, color, "%6.2f%%", ratio);
718 fprintf(output, " of all L1-icache hits ");
719 }
720
721 static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
722 {
723 double total, ratio = 0.0;
724 const char *color;
725
726 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
727
728 if (total)
729 ratio = avg / total * 100.0;
730
731 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
732
733 fprintf(output, " # ");
734 color_fprintf(output, color, "%6.2f%%", ratio);
735 fprintf(output, " of all dTLB cache hits ");
736 }
737
738 static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
739 {
740 double total, ratio = 0.0;
741 const char *color;
742
743 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
744
745 if (total)
746 ratio = avg / total * 100.0;
747
748 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
749
750 fprintf(output, " # ");
751 color_fprintf(output, color, "%6.2f%%", ratio);
752 fprintf(output, " of all iTLB cache hits ");
753 }
754
755 static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
756 {
757 double total, ratio = 0.0;
758 const char *color;
759
760 total = avg_stats(&runtime_ll_cache_stats[cpu]);
761
762 if (total)
763 ratio = avg / total * 100.0;
764
765 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
766
767 fprintf(output, " # ");
768 color_fprintf(output, color, "%6.2f%%", ratio);
769 fprintf(output, " of all LL-cache hits ");
770 }
771
772 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
773 {
774 double total, ratio = 0.0;
775 char cpustr[16] = { '\0', };
776 const char *fmt;
777
778 if (csv_output)
779 fmt = "%s%.0f%s%s";
780 else if (big_num)
781 fmt = "%s%'18.0f%s%-25s";
782 else
783 fmt = "%s%18.0f%s%-25s";
784
785 if (no_aggr)
786 sprintf(cpustr, "CPU%*d%s",
787 csv_output ? 0 : -4,
788 evsel_list->cpus->map[cpu], csv_sep);
789 else
790 cpu = 0;
791
792 fprintf(output, fmt, cpustr, avg, csv_sep, perf_evsel__name(evsel));
793
794 if (evsel->cgrp)
795 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
796
797 if (csv_output)
798 return;
799
800 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
801 total = avg_stats(&runtime_cycles_stats[cpu]);
802
803 if (total)
804 ratio = avg / total;
805
806 fprintf(output, " # %5.2f insns per cycle ", ratio);
807
808 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
809 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
810
811 if (total && avg) {
812 ratio = total / avg;
813 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
814 }
815
816 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
817 runtime_branches_stats[cpu].n != 0) {
818 print_branch_misses(cpu, evsel, avg);
819 } else if (
820 evsel->attr.type == PERF_TYPE_HW_CACHE &&
821 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
822 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
823 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
824 runtime_l1_dcache_stats[cpu].n != 0) {
825 print_l1_dcache_misses(cpu, evsel, avg);
826 } else if (
827 evsel->attr.type == PERF_TYPE_HW_CACHE &&
828 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
829 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
830 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
831 runtime_l1_icache_stats[cpu].n != 0) {
832 print_l1_icache_misses(cpu, evsel, avg);
833 } else if (
834 evsel->attr.type == PERF_TYPE_HW_CACHE &&
835 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
836 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
837 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
838 runtime_dtlb_cache_stats[cpu].n != 0) {
839 print_dtlb_cache_misses(cpu, evsel, avg);
840 } else if (
841 evsel->attr.type == PERF_TYPE_HW_CACHE &&
842 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
843 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
844 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
845 runtime_itlb_cache_stats[cpu].n != 0) {
846 print_itlb_cache_misses(cpu, evsel, avg);
847 } else if (
848 evsel->attr.type == PERF_TYPE_HW_CACHE &&
849 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
850 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
851 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
852 runtime_ll_cache_stats[cpu].n != 0) {
853 print_ll_cache_misses(cpu, evsel, avg);
854 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
855 runtime_cacherefs_stats[cpu].n != 0) {
856 total = avg_stats(&runtime_cacherefs_stats[cpu]);
857
858 if (total)
859 ratio = avg * 100 / total;
860
861 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
862
863 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
864 print_stalled_cycles_frontend(cpu, evsel, avg);
865 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
866 print_stalled_cycles_backend(cpu, evsel, avg);
867 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
868 total = avg_stats(&runtime_nsecs_stats[cpu]);
869
870 if (total)
871 ratio = 1.0 * avg / total;
872
873 fprintf(output, " # %8.3f GHz ", ratio);
874 } else if (runtime_nsecs_stats[cpu].n != 0) {
875 char unit = 'M';
876
877 total = avg_stats(&runtime_nsecs_stats[cpu]);
878
879 if (total)
880 ratio = 1000.0 * avg / total;
881 if (ratio < 0.001) {
882 ratio *= 1000;
883 unit = 'K';
884 }
885
886 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
887 } else {
888 fprintf(output, " ");
889 }
890 }
891
892 /*
893 * Print out the results of a single counter:
894 * aggregated counts in system-wide mode
895 */
896 static void print_counter_aggr(struct perf_evsel *counter)
897 {
898 struct perf_stat *ps = counter->priv;
899 double avg = avg_stats(&ps->res_stats[0]);
900 int scaled = counter->counts->scaled;
901
902 if (scaled == -1) {
903 fprintf(output, "%*s%s%*s",
904 csv_output ? 0 : 18,
905 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
906 csv_sep,
907 csv_output ? 0 : -24,
908 perf_evsel__name(counter));
909
910 if (counter->cgrp)
911 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
912
913 fputc('\n', output);
914 return;
915 }
916
917 if (nsec_counter(counter))
918 nsec_printout(-1, counter, avg);
919 else
920 abs_printout(-1, counter, avg);
921
922 print_noise(counter, avg);
923
924 if (csv_output) {
925 fputc('\n', output);
926 return;
927 }
928
929 if (scaled) {
930 double avg_enabled, avg_running;
931
932 avg_enabled = avg_stats(&ps->res_stats[1]);
933 avg_running = avg_stats(&ps->res_stats[2]);
934
935 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
936 }
937 fprintf(output, "\n");
938 }
939
940 /*
941 * Print out the results of a single counter:
942 * does not use aggregated count in system-wide
943 */
944 static void print_counter(struct perf_evsel *counter)
945 {
946 u64 ena, run, val;
947 int cpu;
948
949 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
950 val = counter->counts->cpu[cpu].val;
951 ena = counter->counts->cpu[cpu].ena;
952 run = counter->counts->cpu[cpu].run;
953 if (run == 0 || ena == 0) {
954 fprintf(output, "CPU%*d%s%*s%s%*s",
955 csv_output ? 0 : -4,
956 evsel_list->cpus->map[cpu], csv_sep,
957 csv_output ? 0 : 18,
958 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
959 csv_sep,
960 csv_output ? 0 : -24,
961 perf_evsel__name(counter));
962
963 if (counter->cgrp)
964 fprintf(output, "%s%s",
965 csv_sep, counter->cgrp->name);
966
967 fputc('\n', output);
968 continue;
969 }
970
971 if (nsec_counter(counter))
972 nsec_printout(cpu, counter, val);
973 else
974 abs_printout(cpu, counter, val);
975
976 if (!csv_output) {
977 print_noise(counter, 1.0);
978
979 if (run != ena)
980 fprintf(output, " (%.2f%%)",
981 100.0 * run / ena);
982 }
983 fputc('\n', output);
984 }
985 }
986
987 static void print_stat(int argc, const char **argv)
988 {
989 struct perf_evsel *counter;
990 int i;
991
992 fflush(stdout);
993
994 if (!csv_output) {
995 fprintf(output, "\n");
996 fprintf(output, " Performance counter stats for ");
997 if (!perf_target__has_task(&target)) {
998 fprintf(output, "\'%s", argv[0]);
999 for (i = 1; i < argc; i++)
1000 fprintf(output, " %s", argv[i]);
1001 } else if (target.pid)
1002 fprintf(output, "process id \'%s", target.pid);
1003 else
1004 fprintf(output, "thread id \'%s", target.tid);
1005
1006 fprintf(output, "\'");
1007 if (run_count > 1)
1008 fprintf(output, " (%d runs)", run_count);
1009 fprintf(output, ":\n\n");
1010 }
1011
1012 if (no_aggr) {
1013 list_for_each_entry(counter, &evsel_list->entries, node)
1014 print_counter(counter);
1015 } else {
1016 list_for_each_entry(counter, &evsel_list->entries, node)
1017 print_counter_aggr(counter);
1018 }
1019
1020 if (!csv_output) {
1021 if (!null_run)
1022 fprintf(output, "\n");
1023 fprintf(output, " %17.9f seconds time elapsed",
1024 avg_stats(&walltime_nsecs_stats)/1e9);
1025 if (run_count > 1) {
1026 fprintf(output, " ");
1027 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1028 avg_stats(&walltime_nsecs_stats));
1029 }
1030 fprintf(output, "\n\n");
1031 }
1032 }
1033
1034 static volatile int signr = -1;
1035
1036 static void skip_signal(int signo)
1037 {
1038 if(child_pid == -1)
1039 done = 1;
1040
1041 signr = signo;
1042 }
1043
1044 static void sig_atexit(void)
1045 {
1046 if (child_pid != -1)
1047 kill(child_pid, SIGTERM);
1048
1049 if (signr == -1)
1050 return;
1051
1052 signal(signr, SIG_DFL);
1053 kill(getpid(), signr);
1054 }
1055
1056 static const char * const stat_usage[] = {
1057 "perf stat [<options>] [<command>]",
1058 NULL
1059 };
1060
1061 static int stat__set_big_num(const struct option *opt __used,
1062 const char *s __used, int unset)
1063 {
1064 big_num_opt = unset ? 0 : 1;
1065 return 0;
1066 }
1067
1068 static bool append_file;
1069
1070 static const struct option options[] = {
1071 OPT_CALLBACK('e', "event", &evsel_list, "event",
1072 "event selector. use 'perf list' to list available events",
1073 parse_events_option),
1074 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1075 "event filter", parse_filter),
1076 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1077 "child tasks do not inherit counters"),
1078 OPT_STRING('p', "pid", &target.pid, "pid",
1079 "stat events on existing process id"),
1080 OPT_STRING('t', "tid", &target.tid, "tid",
1081 "stat events on existing thread id"),
1082 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1083 "system-wide collection from all CPUs"),
1084 OPT_BOOLEAN('g', "group", &group,
1085 "put the counters into a counter group"),
1086 OPT_BOOLEAN('c', "scale", &scale,
1087 "scale/normalize counters"),
1088 OPT_INCR('v', "verbose", &verbose,
1089 "be more verbose (show counter open errors, etc)"),
1090 OPT_INTEGER('r', "repeat", &run_count,
1091 "repeat command and print average + stddev (max: 100)"),
1092 OPT_BOOLEAN('n', "null", &null_run,
1093 "null run - dont start any counters"),
1094 OPT_INCR('d', "detailed", &detailed_run,
1095 "detailed run - start a lot of events"),
1096 OPT_BOOLEAN('S', "sync", &sync_run,
1097 "call sync() before starting a run"),
1098 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1099 "print large numbers with thousands\' separators",
1100 stat__set_big_num),
1101 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1102 "list of cpus to monitor in system-wide"),
1103 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1104 "disable CPU count aggregation"),
1105 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1106 "print counts with custom separator"),
1107 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1108 "monitor event in cgroup name only",
1109 parse_cgroups),
1110 OPT_STRING('o', "output", &output_name, "file",
1111 "output file name"),
1112 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1113 OPT_INTEGER(0, "log-fd", &output_fd,
1114 "log output to fd, instead of stderr"),
1115 OPT_END()
1116 };
1117
1118 /*
1119 * Add default attributes, if there were no attributes specified or
1120 * if -d/--detailed, -d -d or -d -d -d is used:
1121 */
1122 static int add_default_attributes(void)
1123 {
1124 /* Set attrs if no event is selected and !null_run: */
1125 if (null_run)
1126 return 0;
1127
1128 if (!evsel_list->nr_entries) {
1129 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1130 return -1;
1131 }
1132
1133 /* Detailed events get appended to the event list: */
1134
1135 if (detailed_run < 1)
1136 return 0;
1137
1138 /* Append detailed run extra attributes: */
1139 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1140 return -1;
1141
1142 if (detailed_run < 2)
1143 return 0;
1144
1145 /* Append very detailed run extra attributes: */
1146 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1147 return -1;
1148
1149 if (detailed_run < 3)
1150 return 0;
1151
1152 /* Append very, very detailed run extra attributes: */
1153 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1154 }
1155
1156 int cmd_stat(int argc, const char **argv, const char *prefix __used)
1157 {
1158 struct perf_evsel *pos;
1159 int status = -ENOMEM;
1160 const char *mode;
1161
1162 setlocale(LC_ALL, "");
1163
1164 evsel_list = perf_evlist__new(NULL, NULL);
1165 if (evsel_list == NULL)
1166 return -ENOMEM;
1167
1168 argc = parse_options(argc, argv, options, stat_usage,
1169 PARSE_OPT_STOP_AT_NON_OPTION);
1170
1171 output = stderr;
1172 if (output_name && strcmp(output_name, "-"))
1173 output = NULL;
1174
1175 if (output_name && output_fd) {
1176 fprintf(stderr, "cannot use both --output and --log-fd\n");
1177 usage_with_options(stat_usage, options);
1178 }
1179
1180 if (output_fd < 0) {
1181 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1182 usage_with_options(stat_usage, options);
1183 }
1184
1185 if (!output) {
1186 struct timespec tm;
1187 mode = append_file ? "a" : "w";
1188
1189 output = fopen(output_name, mode);
1190 if (!output) {
1191 perror("failed to create output file");
1192 exit(-1);
1193 }
1194 clock_gettime(CLOCK_REALTIME, &tm);
1195 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1196 } else if (output_fd > 0) {
1197 mode = append_file ? "a" : "w";
1198 output = fdopen(output_fd, mode);
1199 if (!output) {
1200 perror("Failed opening logfd");
1201 return -errno;
1202 }
1203 }
1204
1205 if (csv_sep) {
1206 csv_output = true;
1207 if (!strcmp(csv_sep, "\\t"))
1208 csv_sep = "\t";
1209 } else
1210 csv_sep = DEFAULT_SEPARATOR;
1211
1212 /*
1213 * let the spreadsheet do the pretty-printing
1214 */
1215 if (csv_output) {
1216 /* User explicitly passed -B? */
1217 if (big_num_opt == 1) {
1218 fprintf(stderr, "-B option not supported with -x\n");
1219 usage_with_options(stat_usage, options);
1220 } else /* Nope, so disable big number formatting */
1221 big_num = false;
1222 } else if (big_num_opt == 0) /* User passed --no-big-num */
1223 big_num = false;
1224
1225 if (!argc && !perf_target__has_task(&target))
1226 usage_with_options(stat_usage, options);
1227 if (run_count <= 0)
1228 usage_with_options(stat_usage, options);
1229
1230 /* no_aggr, cgroup are for system-wide only */
1231 if ((no_aggr || nr_cgroups) && !perf_target__has_cpu(&target)) {
1232 fprintf(stderr, "both cgroup and no-aggregation "
1233 "modes only available in system-wide mode\n");
1234
1235 usage_with_options(stat_usage, options);
1236 }
1237
1238 if (add_default_attributes())
1239 goto out;
1240
1241 perf_target__validate(&target);
1242
1243 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1244 if (perf_target__has_task(&target))
1245 pr_err("Problems finding threads of monitor\n");
1246 if (perf_target__has_cpu(&target))
1247 perror("failed to parse CPUs map");
1248
1249 usage_with_options(stat_usage, options);
1250 return -1;
1251 }
1252
1253 list_for_each_entry(pos, &evsel_list->entries, node) {
1254 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1255 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0)
1256 goto out_free_fd;
1257 }
1258
1259 /*
1260 * We dont want to block the signals - that would cause
1261 * child tasks to inherit that and Ctrl-C would not work.
1262 * What we want is for Ctrl-C to work in the exec()-ed
1263 * task, but being ignored by perf stat itself:
1264 */
1265 atexit(sig_atexit);
1266 signal(SIGINT, skip_signal);
1267 signal(SIGALRM, skip_signal);
1268 signal(SIGABRT, skip_signal);
1269
1270 status = 0;
1271 for (run_idx = 0; run_idx < run_count; run_idx++) {
1272 if (run_count != 1 && verbose)
1273 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1274 run_idx + 1);
1275
1276 if (sync_run)
1277 sync();
1278
1279 status = run_perf_stat(argc, argv);
1280 }
1281
1282 if (status != -1)
1283 print_stat(argc, argv);
1284 out_free_fd:
1285 list_for_each_entry(pos, &evsel_list->entries, node)
1286 perf_evsel__free_stat_priv(pos);
1287 perf_evlist__delete_maps(evsel_list);
1288 out:
1289 perf_evlist__delete(evsel_list);
1290 return status;
1291 }
(deprecated) Btrfs devel tree