?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
ARM: OMAP4: PM: Add power domain statistics support
[linux-2.6/btrfs-unstable.git] / tools / perf / builtin-stat.c
blob7d98676808d8722a39623de8983642f987cea7cc
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 struct perf_evlist *evsel_list;
177
178 static bool system_wide = false;
179 static int run_idx = 0;
180
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;
200
201 static volatile int done = 0;
202
203 struct stats
204 {
205 double n, mean, M2;
206 };
207
208 struct perf_stat {
209 struct stats res_stats[3];
210 };
211
212 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
213 {
214 evsel->priv = zalloc(sizeof(struct perf_stat));
215 return evsel->priv == NULL ? -ENOMEM : 0;
216 }
217
218 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
219 {
220 free(evsel->priv);
221 evsel->priv = NULL;
222 }
223
224 static void update_stats(struct stats *stats, u64 val)
225 {
226 double delta;
227
228 stats->n++;
229 delta = val - stats->mean;
230 stats->mean += delta / stats->n;
231 stats->M2 += delta*(val - stats->mean);
232 }
233
234 static double avg_stats(struct stats *stats)
235 {
236 return stats->mean;
237 }
238
239 /*
240 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
241 *
242 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
243 * s^2 = -------------------------------
244 * n - 1
245 *
246 * http://en.wikipedia.org/wiki/Stddev
247 *
248 * The std dev of the mean is related to the std dev by:
249 *
250 * s
251 * s_mean = -------
252 * sqrt(n)
253 *
254 */
255 static double stddev_stats(struct stats *stats)
256 {
257 double variance, variance_mean;
258
259 if (!stats->n)
260 return 0.0;
261
262 variance = stats->M2 / (stats->n - 1);
263 variance_mean = variance / stats->n;
264
265 return sqrt(variance_mean);
266 }
267
268 struct stats runtime_nsecs_stats[MAX_NR_CPUS];
269 struct stats runtime_cycles_stats[MAX_NR_CPUS];
270 struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
271 struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
272 struct stats runtime_branches_stats[MAX_NR_CPUS];
273 struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
274 struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
275 struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
276 struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
277 struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
278 struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
279 struct stats walltime_nsecs_stats;
280
281 static int create_perf_stat_counter(struct perf_evsel *evsel,
282 struct perf_evsel *first)
283 {
284 struct perf_event_attr *attr = &evsel->attr;
285 struct xyarray *group_fd = NULL;
286
287 if (group && evsel != first)
288 group_fd = first->fd;
289
290 if (scale)
291 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
292 PERF_FORMAT_TOTAL_TIME_RUNNING;
293
294 attr->inherit = !no_inherit;
295
296 if (system_wide)
297 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus,
298 group, group_fd);
299 if (target_pid == -1 && target_tid == -1) {
300 attr->disabled = 1;
301 attr->enable_on_exec = 1;
302 }
303
304 return perf_evsel__open_per_thread(evsel, evsel_list->threads,
305 group, group_fd);
306 }
307
308 /*
309 * Does the counter have nsecs as a unit?
310 */
311 static inline int nsec_counter(struct perf_evsel *evsel)
312 {
313 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
314 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
315 return 1;
316
317 return 0;
318 }
319
320 /*
321 * Update various tracking values we maintain to print
322 * more semantic information such as miss/hit ratios,
323 * instruction rates, etc:
324 */
325 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
326 {
327 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
328 update_stats(&runtime_nsecs_stats[0], count[0]);
329 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
330 update_stats(&runtime_cycles_stats[0], count[0]);
331 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
332 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
333 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
334 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
335 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
336 update_stats(&runtime_branches_stats[0], count[0]);
337 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
338 update_stats(&runtime_cacherefs_stats[0], count[0]);
339 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
340 update_stats(&runtime_l1_dcache_stats[0], count[0]);
341 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
342 update_stats(&runtime_l1_icache_stats[0], count[0]);
343 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
344 update_stats(&runtime_ll_cache_stats[0], count[0]);
345 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
346 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
347 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
348 update_stats(&runtime_itlb_cache_stats[0], count[0]);
349 }
350
351 /*
352 * Read out the results of a single counter:
353 * aggregate counts across CPUs in system-wide mode
354 */
355 static int read_counter_aggr(struct perf_evsel *counter)
356 {
357 struct perf_stat *ps = counter->priv;
358 u64 *count = counter->counts->aggr.values;
359 int i;
360
361 if (__perf_evsel__read(counter, evsel_list->cpus->nr,
362 evsel_list->threads->nr, scale) < 0)
363 return -1;
364
365 for (i = 0; i < 3; i++)
366 update_stats(&ps->res_stats[i], count[i]);
367
368 if (verbose) {
369 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
370 event_name(counter), count[0], count[1], count[2]);
371 }
372
373 /*
374 * Save the full runtime - to allow normalization during printout:
375 */
376 update_shadow_stats(counter, count);
377
378 return 0;
379 }
380
381 /*
382 * Read out the results of a single counter:
383 * do not aggregate counts across CPUs in system-wide mode
384 */
385 static int read_counter(struct perf_evsel *counter)
386 {
387 u64 *count;
388 int cpu;
389
390 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
391 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
392 return -1;
393
394 count = counter->counts->cpu[cpu].values;
395
396 update_shadow_stats(counter, count);
397 }
398
399 return 0;
400 }
401
402 static int run_perf_stat(int argc __used, const char **argv)
403 {
404 unsigned long long t0, t1;
405 struct perf_evsel *counter, *first;
406 int status = 0;
407 int child_ready_pipe[2], go_pipe[2];
408 const bool forks = (argc > 0);
409 char buf;
410
411 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
412 perror("failed to create pipes");
413 exit(1);
414 }
415
416 if (forks) {
417 if ((child_pid = fork()) < 0)
418 perror("failed to fork");
419
420 if (!child_pid) {
421 close(child_ready_pipe[0]);
422 close(go_pipe[1]);
423 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
424
425 /*
426 * Do a dummy execvp to get the PLT entry resolved,
427 * so we avoid the resolver overhead on the real
428 * execvp call.
429 */
430 execvp("", (char **)argv);
431
432 /*
433 * Tell the parent we're ready to go
434 */
435 close(child_ready_pipe[1]);
436
437 /*
438 * Wait until the parent tells us to go.
439 */
440 if (read(go_pipe[0], &buf, 1) == -1)
441 perror("unable to read pipe");
442
443 execvp(argv[0], (char **)argv);
444
445 perror(argv[0]);
446 exit(-1);
447 }
448
449 if (target_tid == -1 && target_pid == -1 && !system_wide)
450 evsel_list->threads->map[0] = child_pid;
451
452 /*
453 * Wait for the child to be ready to exec.
454 */
455 close(child_ready_pipe[1]);
456 close(go_pipe[0]);
457 if (read(child_ready_pipe[0], &buf, 1) == -1)
458 perror("unable to read pipe");
459 close(child_ready_pipe[0]);
460 }
461
462 first = list_entry(evsel_list->entries.next, struct perf_evsel, node);
463
464 list_for_each_entry(counter, &evsel_list->entries, node) {
465 if (create_perf_stat_counter(counter, first) < 0) {
466 if (errno == EINVAL || errno == ENOSYS || errno == ENOENT) {
467 if (verbose)
468 ui__warning("%s event is not supported by the kernel.\n",
469 event_name(counter));
470 counter->supported = false;
471 continue;
472 }
473
474 if (errno == EPERM || errno == EACCES) {
475 error("You may not have permission to collect %sstats.\n"
476 "\t Consider tweaking"
477 " /proc/sys/kernel/perf_event_paranoid or running as root.",
478 system_wide ? "system-wide " : "");
479 } else {
480 error("open_counter returned with %d (%s). "
481 "/bin/dmesg may provide additional information.\n",
482 errno, strerror(errno));
483 }
484 if (child_pid != -1)
485 kill(child_pid, SIGTERM);
486 die("Not all events could be opened.\n");
487 return -1;
488 }
489 counter->supported = true;
490 }
491
492 if (perf_evlist__set_filters(evsel_list)) {
493 error("failed to set filter with %d (%s)\n", errno,
494 strerror(errno));
495 return -1;
496 }
497
498 /*
499 * Enable counters and exec the command:
500 */
501 t0 = rdclock();
502
503 if (forks) {
504 close(go_pipe[1]);
505 wait(&status);
506 if (WIFSIGNALED(status))
507 psignal(WTERMSIG(status), argv[0]);
508 } else {
509 while(!done) sleep(1);
510 }
511
512 t1 = rdclock();
513
514 update_stats(&walltime_nsecs_stats, t1 - t0);
515
516 if (no_aggr) {
517 list_for_each_entry(counter, &evsel_list->entries, node) {
518 read_counter(counter);
519 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
520 }
521 } else {
522 list_for_each_entry(counter, &evsel_list->entries, node) {
523 read_counter_aggr(counter);
524 perf_evsel__close_fd(counter, evsel_list->cpus->nr,
525 evsel_list->threads->nr);
526 }
527 }
528
529 return WEXITSTATUS(status);
530 }
531
532 static void print_noise_pct(double total, double avg)
533 {
534 double pct = 0.0;
535
536 if (avg)
537 pct = 100.0*total/avg;
538
539 if (csv_output)
540 fprintf(output, "%s%.2f%%", csv_sep, pct);
541 else if (pct)
542 fprintf(output, " ( +-%6.2f%% )", pct);
543 }
544
545 static void print_noise(struct perf_evsel *evsel, double avg)
546 {
547 struct perf_stat *ps;
548
549 if (run_count == 1)
550 return;
551
552 ps = evsel->priv;
553 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
554 }
555
556 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
557 {
558 double msecs = avg / 1e6;
559 char cpustr[16] = { '\0', };
560 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
561
562 if (no_aggr)
563 sprintf(cpustr, "CPU%*d%s",
564 csv_output ? 0 : -4,
565 evsel_list->cpus->map[cpu], csv_sep);
566
567 fprintf(output, fmt, cpustr, msecs, csv_sep, event_name(evsel));
568
569 if (evsel->cgrp)
570 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
571
572 if (csv_output)
573 return;
574
575 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
576 fprintf(output, " # %8.3f CPUs utilized ",
577 avg / avg_stats(&walltime_nsecs_stats));
578 }
579
580 static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg)
581 {
582 double total, ratio = 0.0;
583 const char *color;
584
585 total = avg_stats(&runtime_cycles_stats[cpu]);
586
587 if (total)
588 ratio = avg / total * 100.0;
589
590 color = PERF_COLOR_NORMAL;
591 if (ratio > 50.0)
592 color = PERF_COLOR_RED;
593 else if (ratio > 30.0)
594 color = PERF_COLOR_MAGENTA;
595 else if (ratio > 10.0)
596 color = PERF_COLOR_YELLOW;
597
598 fprintf(output, " # ");
599 color_fprintf(output, color, "%6.2f%%", ratio);
600 fprintf(output, " frontend cycles idle ");
601 }
602
603 static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg)
604 {
605 double total, ratio = 0.0;
606 const char *color;
607
608 total = avg_stats(&runtime_cycles_stats[cpu]);
609
610 if (total)
611 ratio = avg / total * 100.0;
612
613 color = PERF_COLOR_NORMAL;
614 if (ratio > 75.0)
615 color = PERF_COLOR_RED;
616 else if (ratio > 50.0)
617 color = PERF_COLOR_MAGENTA;
618 else if (ratio > 20.0)
619 color = PERF_COLOR_YELLOW;
620
621 fprintf(output, " # ");
622 color_fprintf(output, color, "%6.2f%%", ratio);
623 fprintf(output, " backend cycles idle ");
624 }
625
626 static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
627 {
628 double total, ratio = 0.0;
629 const char *color;
630
631 total = avg_stats(&runtime_branches_stats[cpu]);
632
633 if (total)
634 ratio = avg / total * 100.0;
635
636 color = PERF_COLOR_NORMAL;
637 if (ratio > 20.0)
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;
643
644 fprintf(output, " # ");
645 color_fprintf(output, color, "%6.2f%%", ratio);
646 fprintf(output, " of all branches ");
647 }
648
649 static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
650 {
651 double total, ratio = 0.0;
652 const char *color;
653
654 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
655
656 if (total)
657 ratio = avg / total * 100.0;
658
659 color = PERF_COLOR_NORMAL;
660 if (ratio > 20.0)
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;
666
667 fprintf(output, " # ");
668 color_fprintf(output, color, "%6.2f%%", ratio);
669 fprintf(output, " of all L1-dcache hits ");
670 }
671
672 static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
673 {
674 double total, ratio = 0.0;
675 const char *color;
676
677 total = avg_stats(&runtime_l1_icache_stats[cpu]);
678
679 if (total)
680 ratio = avg / total * 100.0;
681
682 color = PERF_COLOR_NORMAL;
683 if (ratio > 20.0)
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;
689
690 fprintf(output, " # ");
691 color_fprintf(output, color, "%6.2f%%", ratio);
692 fprintf(output, " of all L1-icache hits ");
693 }
694
695 static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
696 {
697 double total, ratio = 0.0;
698 const char *color;
699
700 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
701
702 if (total)
703 ratio = avg / total * 100.0;
704
705 color = PERF_COLOR_NORMAL;
706 if (ratio > 20.0)
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;
712
713 fprintf(output, " # ");
714 color_fprintf(output, color, "%6.2f%%", ratio);
715 fprintf(output, " of all dTLB cache hits ");
716 }
717
718 static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
719 {
720 double total, ratio = 0.0;
721 const char *color;
722
723 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
724
725 if (total)
726 ratio = avg / total * 100.0;
727
728 color = PERF_COLOR_NORMAL;
729 if (ratio > 20.0)
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;
735
736 fprintf(output, " # ");
737 color_fprintf(output, color, "%6.2f%%", ratio);
738 fprintf(output, " of all iTLB cache hits ");
739 }
740
741 static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
742 {
743 double total, ratio = 0.0;
744 const char *color;
745
746 total = avg_stats(&runtime_ll_cache_stats[cpu]);
747
748 if (total)
749 ratio = avg / total * 100.0;
750
751 color = PERF_COLOR_NORMAL;
752 if (ratio > 20.0)
753 color = PERF_COLOR_RED;
754 else if (ratio > 10.0)
755 color = PERF_COLOR_MAGENTA;
756 else if (ratio > 5.0)
757 color = PERF_COLOR_YELLOW;
758
759 fprintf(output, " # ");
760 color_fprintf(output, color, "%6.2f%%", ratio);
761 fprintf(output, " of all LL-cache hits ");
762 }
763
764 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
765 {
766 double total, ratio = 0.0;
767 char cpustr[16] = { '\0', };
768 const char *fmt;
769
770 if (csv_output)
771 fmt = "%s%.0f%s%s";
772 else if (big_num)
773 fmt = "%s%'18.0f%s%-25s";
774 else
775 fmt = "%s%18.0f%s%-25s";
776
777 if (no_aggr)
778 sprintf(cpustr, "CPU%*d%s",
779 csv_output ? 0 : -4,
780 evsel_list->cpus->map[cpu], csv_sep);
781 else
782 cpu = 0;
783
784 fprintf(output, fmt, cpustr, avg, csv_sep, event_name(evsel));
785
786 if (evsel->cgrp)
787 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
788
789 if (csv_output)
790 return;
791
792 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
793 total = avg_stats(&runtime_cycles_stats[cpu]);
794
795 if (total)
796 ratio = avg / total;
797
798 fprintf(output, " # %5.2f insns per cycle ", ratio);
799
800 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
801 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
802
803 if (total && avg) {
804 ratio = total / avg;
805 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
806 }
807
808 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
809 runtime_branches_stats[cpu].n != 0) {
810 print_branch_misses(cpu, evsel, avg);
811 } else if (
812 evsel->attr.type == PERF_TYPE_HW_CACHE &&
813 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
814 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
815 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
816 runtime_l1_dcache_stats[cpu].n != 0) {
817 print_l1_dcache_misses(cpu, evsel, avg);
818 } else if (
819 evsel->attr.type == PERF_TYPE_HW_CACHE &&
820 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
821 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
822 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
823 runtime_l1_icache_stats[cpu].n != 0) {
824 print_l1_icache_misses(cpu, evsel, avg);
825 } else if (
826 evsel->attr.type == PERF_TYPE_HW_CACHE &&
827 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
828 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
829 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
830 runtime_dtlb_cache_stats[cpu].n != 0) {
831 print_dtlb_cache_misses(cpu, evsel, avg);
832 } else if (
833 evsel->attr.type == PERF_TYPE_HW_CACHE &&
834 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
835 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
836 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
837 runtime_itlb_cache_stats[cpu].n != 0) {
838 print_itlb_cache_misses(cpu, evsel, avg);
839 } else if (
840 evsel->attr.type == PERF_TYPE_HW_CACHE &&
841 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
842 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
843 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
844 runtime_ll_cache_stats[cpu].n != 0) {
845 print_ll_cache_misses(cpu, evsel, avg);
846 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
847 runtime_cacherefs_stats[cpu].n != 0) {
848 total = avg_stats(&runtime_cacherefs_stats[cpu]);
849
850 if (total)
851 ratio = avg * 100 / total;
852
853 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
854
855 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
856 print_stalled_cycles_frontend(cpu, evsel, avg);
857 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
858 print_stalled_cycles_backend(cpu, evsel, avg);
859 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
860 total = avg_stats(&runtime_nsecs_stats[cpu]);
861
862 if (total)
863 ratio = 1.0 * avg / total;
864
865 fprintf(output, " # %8.3f GHz ", ratio);
866 } else if (runtime_nsecs_stats[cpu].n != 0) {
867 total = avg_stats(&runtime_nsecs_stats[cpu]);
868
869 if (total)
870 ratio = 1000.0 * avg / total;
871
872 fprintf(output, " # %8.3f M/sec ", ratio);
873 } else {
874 fprintf(output, " ");
875 }
876 }
877
878 /*
879 * Print out the results of a single counter:
880 * aggregated counts in system-wide mode
881 */
882 static void print_counter_aggr(struct perf_evsel *counter)
883 {
884 struct perf_stat *ps = counter->priv;
885 double avg = avg_stats(&ps->res_stats[0]);
886 int scaled = counter->counts->scaled;
887
888 if (scaled == -1) {
889 fprintf(output, "%*s%s%*s",
890 csv_output ? 0 : 18,
891 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
892 csv_sep,
893 csv_output ? 0 : -24,
894 event_name(counter));
895
896 if (counter->cgrp)
897 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
898
899 fputc('\n', output);
900 return;
901 }
902
903 if (nsec_counter(counter))
904 nsec_printout(-1, counter, avg);
905 else
906 abs_printout(-1, counter, avg);
907
908 print_noise(counter, avg);
909
910 if (csv_output) {
911 fputc('\n', output);
912 return;
913 }
914
915 if (scaled) {
916 double avg_enabled, avg_running;
917
918 avg_enabled = avg_stats(&ps->res_stats[1]);
919 avg_running = avg_stats(&ps->res_stats[2]);
920
921 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
922 }
923 fprintf(output, "\n");
924 }
925
926 /*
927 * Print out the results of a single counter:
928 * does not use aggregated count in system-wide
929 */
930 static void print_counter(struct perf_evsel *counter)
931 {
932 u64 ena, run, val;
933 int cpu;
934
935 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
936 val = counter->counts->cpu[cpu].val;
937 ena = counter->counts->cpu[cpu].ena;
938 run = counter->counts->cpu[cpu].run;
939 if (run == 0 || ena == 0) {
940 fprintf(output, "CPU%*d%s%*s%s%*s",
941 csv_output ? 0 : -4,
942 evsel_list->cpus->map[cpu], csv_sep,
943 csv_output ? 0 : 18,
944 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
945 csv_sep,
946 csv_output ? 0 : -24,
947 event_name(counter));
948
949 if (counter->cgrp)
950 fprintf(output, "%s%s",
951 csv_sep, counter->cgrp->name);
952
953 fputc('\n', output);
954 continue;
955 }
956
957 if (nsec_counter(counter))
958 nsec_printout(cpu, counter, val);
959 else
960 abs_printout(cpu, counter, val);
961
962 if (!csv_output) {
963 print_noise(counter, 1.0);
964
965 if (run != ena)
966 fprintf(output, " (%.2f%%)",
967 100.0 * run / ena);
968 }
969 fputc('\n', output);
970 }
971 }
972
973 static void print_stat(int argc, const char **argv)
974 {
975 struct perf_evsel *counter;
976 int i;
977
978 fflush(stdout);
979
980 if (!csv_output) {
981 fprintf(output, "\n");
982 fprintf(output, " Performance counter stats for ");
983 if(target_pid == -1 && target_tid == -1) {
984 fprintf(output, "\'%s", argv[0]);
985 for (i = 1; i < argc; i++)
986 fprintf(output, " %s", argv[i]);
987 } else if (target_pid != -1)
988 fprintf(output, "process id \'%d", target_pid);
989 else
990 fprintf(output, "thread id \'%d", target_tid);
991
992 fprintf(output, "\'");
993 if (run_count > 1)
994 fprintf(output, " (%d runs)", run_count);
995 fprintf(output, ":\n\n");
996 }
997
998 if (no_aggr) {
999 list_for_each_entry(counter, &evsel_list->entries, node)
1000 print_counter(counter);
1001 } else {
1002 list_for_each_entry(counter, &evsel_list->entries, node)
1003 print_counter_aggr(counter);
1004 }
1005
1006 if (!csv_output) {
1007 if (!null_run)
1008 fprintf(output, "\n");
1009 fprintf(output, " %17.9f seconds time elapsed",
1010 avg_stats(&walltime_nsecs_stats)/1e9);
1011 if (run_count > 1) {
1012 fprintf(output, " ");
1013 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1014 avg_stats(&walltime_nsecs_stats));
1015 }
1016 fprintf(output, "\n\n");
1017 }
1018 }
1019
1020 static volatile int signr = -1;
1021
1022 static void skip_signal(int signo)
1023 {
1024 if(child_pid == -1)
1025 done = 1;
1026
1027 signr = signo;
1028 }
1029
1030 static void sig_atexit(void)
1031 {
1032 if (child_pid != -1)
1033 kill(child_pid, SIGTERM);
1034
1035 if (signr == -1)
1036 return;
1037
1038 signal(signr, SIG_DFL);
1039 kill(getpid(), signr);
1040 }
1041
1042 static const char * const stat_usage[] = {
1043 "perf stat [<options>] [<command>]",
1044 NULL
1045 };
1046
1047 static int stat__set_big_num(const struct option *opt __used,
1048 const char *s __used, int unset)
1049 {
1050 big_num_opt = unset ? 0 : 1;
1051 return 0;
1052 }
1053
1054 static bool append_file;
1055
1056 static const struct option options[] = {
1057 OPT_CALLBACK('e', "event", &evsel_list, "event",
1058 "event selector. use 'perf list' to list available events",
1059 parse_events_option),
1060 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1061 "event filter", parse_filter),
1062 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1063 "child tasks do not inherit counters"),
1064 OPT_INTEGER('p', "pid", &target_pid,
1065 "stat events on existing process id"),
1066 OPT_INTEGER('t', "tid", &target_tid,
1067 "stat events on existing thread id"),
1068 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1069 "system-wide collection from all CPUs"),
1070 OPT_BOOLEAN('g', "group", &group,
1071 "put the counters into a counter group"),
1072 OPT_BOOLEAN('c', "scale", &scale,
1073 "scale/normalize counters"),
1074 OPT_INCR('v', "verbose", &verbose,
1075 "be more verbose (show counter open errors, etc)"),
1076 OPT_INTEGER('r', "repeat", &run_count,
1077 "repeat command and print average + stddev (max: 100)"),
1078 OPT_BOOLEAN('n', "null", &null_run,
1079 "null run - dont start any counters"),
1080 OPT_INCR('d', "detailed", &detailed_run,
1081 "detailed run - start a lot of events"),
1082 OPT_BOOLEAN('S', "sync", &sync_run,
1083 "call sync() before starting a run"),
1084 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1085 "print large numbers with thousands\' separators",
1086 stat__set_big_num),
1087 OPT_STRING('C', "cpu", &cpu_list, "cpu",
1088 "list of cpus to monitor in system-wide"),
1089 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1090 "disable CPU count aggregation"),
1091 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1092 "print counts with custom separator"),
1093 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1094 "monitor event in cgroup name only",
1095 parse_cgroups),
1096 OPT_STRING('o', "output", &output_name, "file",
1097 "output file name"),
1098 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1099 OPT_INTEGER(0, "log-fd", &output_fd,
1100 "log output to fd, instead of stderr"),
1101 OPT_END()
1102 };
1103
1104 /*
1105 * Add default attributes, if there were no attributes specified or
1106 * if -d/--detailed, -d -d or -d -d -d is used:
1107 */
1108 static int add_default_attributes(void)
1109 {
1110 struct perf_evsel *pos;
1111 size_t attr_nr = 0;
1112 size_t c;
1113
1114 /* Set attrs if no event is selected and !null_run: */
1115 if (null_run)
1116 return 0;
1117
1118 if (!evsel_list->nr_entries) {
1119 for (c = 0; c < ARRAY_SIZE(default_attrs); c++) {
1120 pos = perf_evsel__new(default_attrs + c, c + attr_nr);
1121 if (pos == NULL)
1122 return -1;
1123 perf_evlist__add(evsel_list, pos);
1124 }
1125 attr_nr += c;
1126 }
1127
1128 /* Detailed events get appended to the event list: */
1129
1130 if (detailed_run < 1)
1131 return 0;
1132
1133 /* Append detailed run extra attributes: */
1134 for (c = 0; c < ARRAY_SIZE(detailed_attrs); c++) {
1135 pos = perf_evsel__new(detailed_attrs + c, c + attr_nr);
1136 if (pos == NULL)
1137 return -1;
1138 perf_evlist__add(evsel_list, pos);
1139 }
1140 attr_nr += c;
1141
1142 if (detailed_run < 2)
1143 return 0;
1144
1145 /* Append very detailed run extra attributes: */
1146 for (c = 0; c < ARRAY_SIZE(very_detailed_attrs); c++) {
1147 pos = perf_evsel__new(very_detailed_attrs + c, c + attr_nr);
1148 if (pos == NULL)
1149 return -1;
1150 perf_evlist__add(evsel_list, pos);
1151 }
1152
1153 if (detailed_run < 3)
1154 return 0;
1155
1156 /* Append very, very detailed run extra attributes: */
1157 for (c = 0; c < ARRAY_SIZE(very_very_detailed_attrs); c++) {
1158 pos = perf_evsel__new(very_very_detailed_attrs + c, c + attr_nr);
1159 if (pos == NULL)
1160 return -1;
1161 perf_evlist__add(evsel_list, pos);
1162 }
1163
1164
1165 return 0;
1166 }
1167
1168 int cmd_stat(int argc, const char **argv, const char *prefix __used)
1169 {
1170 struct perf_evsel *pos;
1171 int status = -ENOMEM;
1172 const char *mode;
1173
1174 setlocale(LC_ALL, "");
1175
1176 evsel_list = perf_evlist__new(NULL, NULL);
1177 if (evsel_list == NULL)
1178 return -ENOMEM;
1179
1180 argc = parse_options(argc, argv, options, stat_usage,
1181 PARSE_OPT_STOP_AT_NON_OPTION);
1182
1183 output = stderr;
1184 if (output_name && strcmp(output_name, "-"))
1185 output = NULL;
1186
1187 if (output_name && output_fd) {
1188 fprintf(stderr, "cannot use both --output and --log-fd\n");
1189 usage_with_options(stat_usage, options);
1190 }
1191 if (!output) {
1192 struct timespec tm;
1193 mode = append_file ? "a" : "w";
1194
1195 output = fopen(output_name, mode);
1196 if (!output) {
1197 perror("failed to create output file");
1198 exit(-1);
1199 }
1200 clock_gettime(CLOCK_REALTIME, &tm);
1201 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1202 } else if (output_fd != 2) {
1203 mode = append_file ? "a" : "w";
1204 output = fdopen(output_fd, mode);
1205 if (!output) {
1206 perror("Failed opening logfd");
1207 return -errno;
1208 }
1209 }
1210
1211 if (csv_sep) {
1212 csv_output = true;
1213 if (!strcmp(csv_sep, "\\t"))
1214 csv_sep = "\t";
1215 } else
1216 csv_sep = DEFAULT_SEPARATOR;
1217
1218 /*
1219 * let the spreadsheet do the pretty-printing
1220 */
1221 if (csv_output) {
1222 /* User explicitly passed -B? */
1223 if (big_num_opt == 1) {
1224 fprintf(stderr, "-B option not supported with -x\n");
1225 usage_with_options(stat_usage, options);
1226 } else /* Nope, so disable big number formatting */
1227 big_num = false;
1228 } else if (big_num_opt == 0) /* User passed --no-big-num */
1229 big_num = false;
1230
1231 if (!argc && target_pid == -1 && target_tid == -1)
1232 usage_with_options(stat_usage, options);
1233 if (run_count <= 0)
1234 usage_with_options(stat_usage, options);
1235
1236 /* no_aggr, cgroup are for system-wide only */
1237 if ((no_aggr || nr_cgroups) && !system_wide) {
1238 fprintf(stderr, "both cgroup and no-aggregation "
1239 "modes only available in system-wide mode\n");
1240
1241 usage_with_options(stat_usage, options);
1242 }
1243
1244 if (add_default_attributes())
1245 goto out;
1246
1247 if (target_pid != -1)
1248 target_tid = target_pid;
1249
1250 evsel_list->threads = thread_map__new(target_pid, target_tid);
1251 if (evsel_list->threads == NULL) {
1252 pr_err("Problems finding threads of monitor\n");
1253 usage_with_options(stat_usage, options);
1254 }
1255
1256 if (system_wide)
1257 evsel_list->cpus = cpu_map__new(cpu_list);
1258 else
1259 evsel_list->cpus = cpu_map__dummy_new();
1260
1261 if (evsel_list->cpus == NULL) {
1262 perror("failed to parse CPUs map");
1263 usage_with_options(stat_usage, options);
1264 return -1;
1265 }
1266
1267 list_for_each_entry(pos, &evsel_list->entries, node) {
1268 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1269 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 ||
1270 perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0)
1271 goto out_free_fd;
1272 }
1273
1274 /*
1275 * We dont want to block the signals - that would cause
1276 * child tasks to inherit that and Ctrl-C would not work.
1277 * What we want is for Ctrl-C to work in the exec()-ed
1278 * task, but being ignored by perf stat itself:
1279 */
1280 atexit(sig_atexit);
1281 signal(SIGINT, skip_signal);
1282 signal(SIGALRM, skip_signal);
1283 signal(SIGABRT, skip_signal);
1284
1285 status = 0;
1286 for (run_idx = 0; run_idx < run_count; run_idx++) {
1287 if (run_count != 1 && verbose)
1288 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1289 run_idx + 1);
1290
1291 if (sync_run)
1292 sync();
1293
1294 status = run_perf_stat(argc, argv);
1295 }
1296
1297 if (status != -1)
1298 print_stat(argc, argv);
1299 out_free_fd:
1300 list_for_each_entry(pos, &evsel_list->entries, node)
1301 perf_evsel__free_stat_priv(pos);
1302 perf_evlist__delete_maps(evsel_list);
1303 out:
1304 perf_evlist__delete(evsel_list);
1305 return status;
1306 }
(deprecated) Btrfs devel tree