?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge branch 'linus'
[linux-2.6/x86.git] / tools / perf / builtin-stat.c
blobbec64a9e741c38d44378c1c90feeca3574929f05
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
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 = stats->M2 / (stats->n - 1);
257 double variance_mean = variance / stats->n;
258
259 return sqrt(variance_mean);
260 }
261
262 struct stats runtime_nsecs_stats[MAX_NR_CPUS];
263 struct stats runtime_cycles_stats[MAX_NR_CPUS];
264 struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
265 struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
266 struct stats runtime_branches_stats[MAX_NR_CPUS];
267 struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
268 struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
269 struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
270 struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
271 struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
272 struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
273 struct stats walltime_nsecs_stats;
274
275 static int create_perf_stat_counter(struct perf_evsel *evsel)
276 {
277 struct perf_event_attr *attr = &evsel->attr;
278
279 if (scale)
280 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
281 PERF_FORMAT_TOTAL_TIME_RUNNING;
282
283 attr->inherit = !no_inherit;
284
285 if (system_wide)
286 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, group);
287
288 if (target_pid == -1 && target_tid == -1) {
289 attr->disabled = 1;
290 attr->enable_on_exec = 1;
291 }
292
293 return perf_evsel__open_per_thread(evsel, evsel_list->threads, group);
294 }
295
296 /*
297 * Does the counter have nsecs as a unit?
298 */
299 static inline int nsec_counter(struct perf_evsel *evsel)
300 {
301 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
302 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
303 return 1;
304
305 return 0;
306 }
307
308 /*
309 * Update various tracking values we maintain to print
310 * more semantic information such as miss/hit ratios,
311 * instruction rates, etc:
312 */
313 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
314 {
315 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
316 update_stats(&runtime_nsecs_stats[0], count[0]);
317 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
318 update_stats(&runtime_cycles_stats[0], count[0]);
319 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
320 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
321 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
322 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
323 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
324 update_stats(&runtime_branches_stats[0], count[0]);
325 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
326 update_stats(&runtime_cacherefs_stats[0], count[0]);
327 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
328 update_stats(&runtime_l1_dcache_stats[0], count[0]);
329 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
330 update_stats(&runtime_l1_icache_stats[0], count[0]);
331 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
332 update_stats(&runtime_ll_cache_stats[0], count[0]);
333 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
334 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
335 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
336 update_stats(&runtime_itlb_cache_stats[0], count[0]);
337 }
338
339 /*
340 * Read out the results of a single counter:
341 * aggregate counts across CPUs in system-wide mode
342 */
343 static int read_counter_aggr(struct perf_evsel *counter)
344 {
345 struct perf_stat *ps = counter->priv;
346 u64 *count = counter->counts->aggr.values;
347 int i;
348
349 if (__perf_evsel__read(counter, evsel_list->cpus->nr,
350 evsel_list->threads->nr, scale) < 0)
351 return -1;
352
353 for (i = 0; i < 3; i++)
354 update_stats(&ps->res_stats[i], count[i]);
355
356 if (verbose) {
357 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
358 event_name(counter), count[0], count[1], count[2]);
359 }
360
361 /*
362 * Save the full runtime - to allow normalization during printout:
363 */
364 update_shadow_stats(counter, count);
365
366 return 0;
367 }
368
369 /*
370 * Read out the results of a single counter:
371 * do not aggregate counts across CPUs in system-wide mode
372 */
373 static int read_counter(struct perf_evsel *counter)
374 {
375 u64 *count;
376 int cpu;
377
378 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
379 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
380 return -1;
381
382 count = counter->counts->cpu[cpu].values;
383
384 update_shadow_stats(counter, count);
385 }
386
387 return 0;
388 }
389
390 static int run_perf_stat(int argc __used, const char **argv)
391 {
392 unsigned long long t0, t1;
393 struct perf_evsel *counter;
394 int status = 0;
395 int child_ready_pipe[2], go_pipe[2];
396 const bool forks = (argc > 0);
397 char buf;
398
399 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
400 perror("failed to create pipes");
401 exit(1);
402 }
403
404 if (forks) {
405 if ((child_pid = fork()) < 0)
406 perror("failed to fork");
407
408 if (!child_pid) {
409 close(child_ready_pipe[0]);
410 close(go_pipe[1]);
411 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
412
413 /*
414 * Do a dummy execvp to get the PLT entry resolved,
415 * so we avoid the resolver overhead on the real
416 * execvp call.
417 */
418 execvp("", (char **)argv);
419
420 /*
421 * Tell the parent we're ready to go
422 */
423 close(child_ready_pipe[1]);
424
425 /*
426 * Wait until the parent tells us to go.
427 */
428 if (read(go_pipe[0], &buf, 1) == -1)
429 perror("unable to read pipe");
430
431 execvp(argv[0], (char **)argv);
432
433 perror(argv[0]);
434 exit(-1);
435 }
436
437 if (target_tid == -1 && target_pid == -1 && !system_wide)
438 evsel_list->threads->map[0] = child_pid;
439
440 /*
441 * Wait for the child to be ready to exec.
442 */
443 close(child_ready_pipe[1]);
444 close(go_pipe[0]);
445 if (read(child_ready_pipe[0], &buf, 1) == -1)
446 perror("unable to read pipe");
447 close(child_ready_pipe[0]);
448 }
449
450 list_for_each_entry(counter, &evsel_list->entries, node) {
451 if (create_perf_stat_counter(counter) < 0) {
452 if (errno == EINVAL || errno == ENOSYS || errno == ENOENT) {
453 if (verbose)
454 ui__warning("%s event is not supported by the kernel.\n",
455 event_name(counter));
456 counter->supported = false;
457 continue;
458 }
459
460 if (errno == EPERM || errno == EACCES) {
461 error("You may not have permission to collect %sstats.\n"
462 "\t Consider tweaking"
463 " /proc/sys/kernel/perf_event_paranoid or running as root.",
464 system_wide ? "system-wide " : "");
465 } else {
466 error("open_counter returned with %d (%s). "
467 "/bin/dmesg may provide additional information.\n",
468 errno, strerror(errno));
469 }
470 if (child_pid != -1)
471 kill(child_pid, SIGTERM);
472 die("Not all events could be opened.\n");
473 return -1;
474 }
475 counter->supported = true;
476 }
477
478 if (perf_evlist__set_filters(evsel_list)) {
479 error("failed to set filter with %d (%s)\n", errno,
480 strerror(errno));
481 return -1;
482 }
483
484 /*
485 * Enable counters and exec the command:
486 */
487 t0 = rdclock();
488
489 if (forks) {
490 close(go_pipe[1]);
491 wait(&status);
492 } else {
493 while(!done) sleep(1);
494 }
495
496 t1 = rdclock();
497
498 update_stats(&walltime_nsecs_stats, t1 - t0);
499
500 if (no_aggr) {
501 list_for_each_entry(counter, &evsel_list->entries, node) {
502 read_counter(counter);
503 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
504 }
505 } else {
506 list_for_each_entry(counter, &evsel_list->entries, node) {
507 read_counter_aggr(counter);
508 perf_evsel__close_fd(counter, evsel_list->cpus->nr,
509 evsel_list->threads->nr);
510 }
511 }
512
513 return WEXITSTATUS(status);
514 }
515
516 static void print_noise_pct(double total, double avg)
517 {
518 double pct = 0.0;
519
520 if (avg)
521 pct = 100.0*total/avg;
522
523 if (csv_output)
524 fprintf(output, "%s%.2f%%", csv_sep, pct);
525 else
526 fprintf(output, " ( +-%6.2f%% )", pct);
527 }
528
529 static void print_noise(struct perf_evsel *evsel, double avg)
530 {
531 struct perf_stat *ps;
532
533 if (run_count == 1)
534 return;
535
536 ps = evsel->priv;
537 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
538 }
539
540 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
541 {
542 double msecs = avg / 1e6;
543 char cpustr[16] = { '\0', };
544 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
545
546 if (no_aggr)
547 sprintf(cpustr, "CPU%*d%s",
548 csv_output ? 0 : -4,
549 evsel_list->cpus->map[cpu], csv_sep);
550
551 fprintf(output, fmt, cpustr, msecs, csv_sep, event_name(evsel));
552
553 if (evsel->cgrp)
554 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
555
556 if (csv_output)
557 return;
558
559 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
560 fprintf(output, " # %8.3f CPUs utilized ",
561 avg / avg_stats(&walltime_nsecs_stats));
562 }
563
564 static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg)
565 {
566 double total, ratio = 0.0;
567 const char *color;
568
569 total = avg_stats(&runtime_cycles_stats[cpu]);
570
571 if (total)
572 ratio = avg / total * 100.0;
573
574 color = PERF_COLOR_NORMAL;
575 if (ratio > 50.0)
576 color = PERF_COLOR_RED;
577 else if (ratio > 30.0)
578 color = PERF_COLOR_MAGENTA;
579 else if (ratio > 10.0)
580 color = PERF_COLOR_YELLOW;
581
582 fprintf(output, " # ");
583 color_fprintf(output, color, "%6.2f%%", ratio);
584 fprintf(output, " frontend cycles idle ");
585 }
586
587 static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg)
588 {
589 double total, ratio = 0.0;
590 const char *color;
591
592 total = avg_stats(&runtime_cycles_stats[cpu]);
593
594 if (total)
595 ratio = avg / total * 100.0;
596
597 color = PERF_COLOR_NORMAL;
598 if (ratio > 75.0)
599 color = PERF_COLOR_RED;
600 else if (ratio > 50.0)
601 color = PERF_COLOR_MAGENTA;
602 else if (ratio > 20.0)
603 color = PERF_COLOR_YELLOW;
604
605 fprintf(output, " # ");
606 color_fprintf(output, color, "%6.2f%%", ratio);
607 fprintf(output, " backend cycles idle ");
608 }
609
610 static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
611 {
612 double total, ratio = 0.0;
613 const char *color;
614
615 total = avg_stats(&runtime_branches_stats[cpu]);
616
617 if (total)
618 ratio = avg / total * 100.0;
619
620 color = PERF_COLOR_NORMAL;
621 if (ratio > 20.0)
622 color = PERF_COLOR_RED;
623 else if (ratio > 10.0)
624 color = PERF_COLOR_MAGENTA;
625 else if (ratio > 5.0)
626 color = PERF_COLOR_YELLOW;
627
628 fprintf(output, " # ");
629 color_fprintf(output, color, "%6.2f%%", ratio);
630 fprintf(output, " of all branches ");
631 }
632
633 static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
634 {
635 double total, ratio = 0.0;
636 const char *color;
637
638 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
639
640 if (total)
641 ratio = avg / total * 100.0;
642
643 color = PERF_COLOR_NORMAL;
644 if (ratio > 20.0)
645 color = PERF_COLOR_RED;
646 else if (ratio > 10.0)
647 color = PERF_COLOR_MAGENTA;
648 else if (ratio > 5.0)
649 color = PERF_COLOR_YELLOW;
650
651 fprintf(output, " # ");
652 color_fprintf(output, color, "%6.2f%%", ratio);
653 fprintf(output, " of all L1-dcache hits ");
654 }
655
656 static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
657 {
658 double total, ratio = 0.0;
659 const char *color;
660
661 total = avg_stats(&runtime_l1_icache_stats[cpu]);
662
663 if (total)
664 ratio = avg / total * 100.0;
665
666 color = PERF_COLOR_NORMAL;
667 if (ratio > 20.0)
668 color = PERF_COLOR_RED;
669 else if (ratio > 10.0)
670 color = PERF_COLOR_MAGENTA;
671 else if (ratio > 5.0)
672 color = PERF_COLOR_YELLOW;
673
674 fprintf(output, " # ");
675 color_fprintf(output, color, "%6.2f%%", ratio);
676 fprintf(output, " of all L1-icache hits ");
677 }
678
679 static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
680 {
681 double total, ratio = 0.0;
682 const char *color;
683
684 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
685
686 if (total)
687 ratio = avg / total * 100.0;
688
689 color = PERF_COLOR_NORMAL;
690 if (ratio > 20.0)
691 color = PERF_COLOR_RED;
692 else if (ratio > 10.0)
693 color = PERF_COLOR_MAGENTA;
694 else if (ratio > 5.0)
695 color = PERF_COLOR_YELLOW;
696
697 fprintf(output, " # ");
698 color_fprintf(output, color, "%6.2f%%", ratio);
699 fprintf(output, " of all dTLB cache hits ");
700 }
701
702 static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
703 {
704 double total, ratio = 0.0;
705 const char *color;
706
707 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
708
709 if (total)
710 ratio = avg / total * 100.0;
711
712 color = PERF_COLOR_NORMAL;
713 if (ratio > 20.0)
714 color = PERF_COLOR_RED;
715 else if (ratio > 10.0)
716 color = PERF_COLOR_MAGENTA;
717 else if (ratio > 5.0)
718 color = PERF_COLOR_YELLOW;
719
720 fprintf(output, " # ");
721 color_fprintf(output, color, "%6.2f%%", ratio);
722 fprintf(output, " of all iTLB cache hits ");
723 }
724
725 static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
726 {
727 double total, ratio = 0.0;
728 const char *color;
729
730 total = avg_stats(&runtime_ll_cache_stats[cpu]);
731
732 if (total)
733 ratio = avg / total * 100.0;
734
735 color = PERF_COLOR_NORMAL;
736 if (ratio > 20.0)
737 color = PERF_COLOR_RED;
738 else if (ratio > 10.0)
739 color = PERF_COLOR_MAGENTA;
740 else if (ratio > 5.0)
741 color = PERF_COLOR_YELLOW;
742
743 fprintf(output, " # ");
744 color_fprintf(output, color, "%6.2f%%", ratio);
745 fprintf(output, " of all LL-cache hits ");
746 }
747
748 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
749 {
750 double total, ratio = 0.0;
751 char cpustr[16] = { '\0', };
752 const char *fmt;
753
754 if (csv_output)
755 fmt = "%s%.0f%s%s";
756 else if (big_num)
757 fmt = "%s%'18.0f%s%-25s";
758 else
759 fmt = "%s%18.0f%s%-25s";
760
761 if (no_aggr)
762 sprintf(cpustr, "CPU%*d%s",
763 csv_output ? 0 : -4,
764 evsel_list->cpus->map[cpu], csv_sep);
765 else
766 cpu = 0;
767
768 fprintf(output, fmt, cpustr, avg, csv_sep, event_name(evsel));
769
770 if (evsel->cgrp)
771 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
772
773 if (csv_output)
774 return;
775
776 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
777 total = avg_stats(&runtime_cycles_stats[cpu]);
778
779 if (total)
780 ratio = avg / total;
781
782 fprintf(output, " # %5.2f insns per cycle ", ratio);
783
784 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
785 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
786
787 if (total && avg) {
788 ratio = total / avg;
789 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
790 }
791
792 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
793 runtime_branches_stats[cpu].n != 0) {
794 print_branch_misses(cpu, evsel, avg);
795 } else if (
796 evsel->attr.type == PERF_TYPE_HW_CACHE &&
797 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
798 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
799 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
800 runtime_l1_dcache_stats[cpu].n != 0) {
801 print_l1_dcache_misses(cpu, evsel, avg);
802 } else if (
803 evsel->attr.type == PERF_TYPE_HW_CACHE &&
804 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
805 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
806 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
807 runtime_l1_icache_stats[cpu].n != 0) {
808 print_l1_icache_misses(cpu, evsel, avg);
809 } else if (
810 evsel->attr.type == PERF_TYPE_HW_CACHE &&
811 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
812 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
813 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
814 runtime_dtlb_cache_stats[cpu].n != 0) {
815 print_dtlb_cache_misses(cpu, evsel, avg);
816 } else if (
817 evsel->attr.type == PERF_TYPE_HW_CACHE &&
818 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
819 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
820 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
821 runtime_itlb_cache_stats[cpu].n != 0) {
822 print_itlb_cache_misses(cpu, evsel, avg);
823 } else if (
824 evsel->attr.type == PERF_TYPE_HW_CACHE &&
825 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
826 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
827 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
828 runtime_ll_cache_stats[cpu].n != 0) {
829 print_ll_cache_misses(cpu, evsel, avg);
830 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
831 runtime_cacherefs_stats[cpu].n != 0) {
832 total = avg_stats(&runtime_cacherefs_stats[cpu]);
833
834 if (total)
835 ratio = avg * 100 / total;
836
837 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
838
839 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
840 print_stalled_cycles_frontend(cpu, evsel, avg);
841 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
842 print_stalled_cycles_backend(cpu, evsel, avg);
843 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
844 total = avg_stats(&runtime_nsecs_stats[cpu]);
845
846 if (total)
847 ratio = 1.0 * avg / total;
848
849 fprintf(output, " # %8.3f GHz ", ratio);
850 } else if (runtime_nsecs_stats[cpu].n != 0) {
851 total = avg_stats(&runtime_nsecs_stats[cpu]);
852
853 if (total)
854 ratio = 1000.0 * avg / total;
855
856 fprintf(output, " # %8.3f M/sec ", ratio);
857 } else {
858 fprintf(output, " ");
859 }
860 }
861
862 /*
863 * Print out the results of a single counter:
864 * aggregated counts in system-wide mode
865 */
866 static void print_counter_aggr(struct perf_evsel *counter)
867 {
868 struct perf_stat *ps = counter->priv;
869 double avg = avg_stats(&ps->res_stats[0]);
870 int scaled = counter->counts->scaled;
871
872 if (scaled == -1) {
873 fprintf(output, "%*s%s%*s",
874 csv_output ? 0 : 18,
875 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
876 csv_sep,
877 csv_output ? 0 : -24,
878 event_name(counter));
879
880 if (counter->cgrp)
881 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
882
883 fputc('\n', output);
884 return;
885 }
886
887 if (nsec_counter(counter))
888 nsec_printout(-1, counter, avg);
889 else
890 abs_printout(-1, counter, avg);
891
892 print_noise(counter, avg);
893
894 if (csv_output) {
895 fputc('\n', output);
896 return;
897 }
898
899 if (scaled) {
900 double avg_enabled, avg_running;
901
902 avg_enabled = avg_stats(&ps->res_stats[1]);
903 avg_running = avg_stats(&ps->res_stats[2]);
904
905 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
906 }
907 fprintf(output, "\n");
908 }
909
910 /*
911 * Print out the results of a single counter:
912 * does not use aggregated count in system-wide
913 */
914 static void print_counter(struct perf_evsel *counter)
915 {
916 u64 ena, run, val;
917 int cpu;
918
919 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
920 val = counter->counts->cpu[cpu].val;
921 ena = counter->counts->cpu[cpu].ena;
922 run = counter->counts->cpu[cpu].run;
923 if (run == 0 || ena == 0) {
924 fprintf(output, "CPU%*d%s%*s%s%*s",
925 csv_output ? 0 : -4,
926 evsel_list->cpus->map[cpu], csv_sep,
927 csv_output ? 0 : 18,
928 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
929 csv_sep,
930 csv_output ? 0 : -24,
931 event_name(counter));
932
933 if (counter->cgrp)
934 fprintf(output, "%s%s",
935 csv_sep, counter->cgrp->name);
936
937 fputc('\n', output);
938 continue;
939 }
940
941 if (nsec_counter(counter))
942 nsec_printout(cpu, counter, val);
943 else
944 abs_printout(cpu, counter, val);
945
946 if (!csv_output) {
947 print_noise(counter, 1.0);
948
949 if (run != ena)
950 fprintf(output, " (%.2f%%)",
951 100.0 * run / ena);
952 }
953 fputc('\n', output);
954 }
955 }
956
957 static void print_stat(int argc, const char **argv)
958 {
959 struct perf_evsel *counter;
960 int i;
961
962 fflush(stdout);
963
964 if (!csv_output) {
965 fprintf(output, "\n");
966 fprintf(output, " Performance counter stats for ");
967 if(target_pid == -1 && target_tid == -1) {
968 fprintf(output, "\'%s", argv[0]);
969 for (i = 1; i < argc; i++)
970 fprintf(output, " %s", argv[i]);
971 } else if (target_pid != -1)
972 fprintf(output, "process id \'%d", target_pid);
973 else
974 fprintf(output, "thread id \'%d", target_tid);
975
976 fprintf(output, "\'");
977 if (run_count > 1)
978 fprintf(output, " (%d runs)", run_count);
979 fprintf(output, ":\n\n");
980 }
981
982 if (no_aggr) {
983 list_for_each_entry(counter, &evsel_list->entries, node)
984 print_counter(counter);
985 } else {
986 list_for_each_entry(counter, &evsel_list->entries, node)
987 print_counter_aggr(counter);
988 }
989
990 if (!csv_output) {
991 if (!null_run)
992 fprintf(output, "\n");
993 fprintf(output, " %17.9f seconds time elapsed",
994 avg_stats(&walltime_nsecs_stats)/1e9);
995 if (run_count > 1) {
996 fprintf(output, " ");
997 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
998 avg_stats(&walltime_nsecs_stats));
999 }
1000 fprintf(output, "\n\n");
1001 }
1002 }
1003
1004 static volatile int signr = -1;
1005
1006 static void skip_signal(int signo)
1007 {
1008 if(child_pid == -1)
1009 done = 1;
1010
1011 signr = signo;
1012 }
1013
1014 static void sig_atexit(void)
1015 {
1016 if (child_pid != -1)
1017 kill(child_pid, SIGTERM);
1018
1019 if (signr == -1)
1020 return;
1021
1022 signal(signr, SIG_DFL);
1023 kill(getpid(), signr);
1024 }
1025
1026 static const char * const stat_usage[] = {
1027 "perf stat [<options>] [<command>]",
1028 NULL
1029 };
1030
1031 static int stat__set_big_num(const struct option *opt __used,
1032 const char *s __used, int unset)
1033 {
1034 big_num_opt = unset ? 0 : 1;
1035 return 0;
1036 }
1037
1038 static bool append_file;
1039
1040 static const struct option options[] = {
1041 OPT_CALLBACK('e', "event", &evsel_list, "event",
1042 "event selector. use 'perf list' to list available events",
1043 parse_events_option),
1044 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1045 "event filter", parse_filter),
1046 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1047 "child tasks do not inherit counters"),
1048 OPT_INTEGER('p', "pid", &target_pid,
1049 "stat events on existing process id"),
1050 OPT_INTEGER('t', "tid", &target_tid,
1051 "stat events on existing thread id"),
1052 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1053 "system-wide collection from all CPUs"),
1054 OPT_BOOLEAN('g', "group", &group,
1055 "put the counters into a counter group"),
1056 OPT_BOOLEAN('c', "scale", &scale,
1057 "scale/normalize counters"),
1058 OPT_INCR('v', "verbose", &verbose,
1059 "be more verbose (show counter open errors, etc)"),
1060 OPT_INTEGER('r', "repeat", &run_count,
1061 "repeat command and print average + stddev (max: 100)"),
1062 OPT_BOOLEAN('n', "null", &null_run,
1063 "null run - dont start any counters"),
1064 OPT_INCR('d', "detailed", &detailed_run,
1065 "detailed run - start a lot of events"),
1066 OPT_BOOLEAN('S', "sync", &sync_run,
1067 "call sync() before starting a run"),
1068 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1069 "print large numbers with thousands\' separators",
1070 stat__set_big_num),
1071 OPT_STRING('C', "cpu", &cpu_list, "cpu",
1072 "list of cpus to monitor in system-wide"),
1073 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1074 "disable CPU count aggregation"),
1075 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1076 "print counts with custom separator"),
1077 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1078 "monitor event in cgroup name only",
1079 parse_cgroups),
1080 OPT_STRING('o', "output", &output_name, "file",
1081 "output file name"),
1082 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1083 OPT_END()
1084 };
1085
1086 /*
1087 * Add default attributes, if there were no attributes specified or
1088 * if -d/--detailed, -d -d or -d -d -d is used:
1089 */
1090 static int add_default_attributes(void)
1091 {
1092 struct perf_evsel *pos;
1093 size_t attr_nr = 0;
1094 size_t c;
1095
1096 /* Set attrs if no event is selected and !null_run: */
1097 if (null_run)
1098 return 0;
1099
1100 if (!evsel_list->nr_entries) {
1101 for (c = 0; c < ARRAY_SIZE(default_attrs); c++) {
1102 pos = perf_evsel__new(default_attrs + c, c + attr_nr);
1103 if (pos == NULL)
1104 return -1;
1105 perf_evlist__add(evsel_list, pos);
1106 }
1107 attr_nr += c;
1108 }
1109
1110 /* Detailed events get appended to the event list: */
1111
1112 if (detailed_run < 1)
1113 return 0;
1114
1115 /* Append detailed run extra attributes: */
1116 for (c = 0; c < ARRAY_SIZE(detailed_attrs); c++) {
1117 pos = perf_evsel__new(detailed_attrs + c, c + attr_nr);
1118 if (pos == NULL)
1119 return -1;
1120 perf_evlist__add(evsel_list, pos);
1121 }
1122 attr_nr += c;
1123
1124 if (detailed_run < 2)
1125 return 0;
1126
1127 /* Append very detailed run extra attributes: */
1128 for (c = 0; c < ARRAY_SIZE(very_detailed_attrs); c++) {
1129 pos = perf_evsel__new(very_detailed_attrs + c, c + attr_nr);
1130 if (pos == NULL)
1131 return -1;
1132 perf_evlist__add(evsel_list, pos);
1133 }
1134
1135 if (detailed_run < 3)
1136 return 0;
1137
1138 /* Append very, very detailed run extra attributes: */
1139 for (c = 0; c < ARRAY_SIZE(very_very_detailed_attrs); c++) {
1140 pos = perf_evsel__new(very_very_detailed_attrs + c, c + attr_nr);
1141 if (pos == NULL)
1142 return -1;
1143 perf_evlist__add(evsel_list, pos);
1144 }
1145
1146
1147 return 0;
1148 }
1149
1150 int cmd_stat(int argc, const char **argv, const char *prefix __used)
1151 {
1152 struct perf_evsel *pos;
1153 int status = -ENOMEM;
1154 const char *mode;
1155
1156 setlocale(LC_ALL, "");
1157
1158 evsel_list = perf_evlist__new(NULL, NULL);
1159 if (evsel_list == NULL)
1160 return -ENOMEM;
1161
1162 argc = parse_options(argc, argv, options, stat_usage,
1163 PARSE_OPT_STOP_AT_NON_OPTION);
1164
1165 output = stderr;
1166 if (output_name && strcmp(output_name, "-"))
1167 output = NULL;
1168
1169 if (!output) {
1170 struct timespec tm;
1171 mode = append_file ? "a" : "w";
1172
1173 output = fopen(output_name, mode);
1174 if (!output) {
1175 perror("failed to create output file");
1176 exit(-1);
1177 }
1178 clock_gettime(CLOCK_REALTIME, &tm);
1179 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1180 }
1181
1182 if (csv_sep)
1183 csv_output = true;
1184 else
1185 csv_sep = DEFAULT_SEPARATOR;
1186
1187 /*
1188 * let the spreadsheet do the pretty-printing
1189 */
1190 if (csv_output) {
1191 /* User explicitely passed -B? */
1192 if (big_num_opt == 1) {
1193 fprintf(stderr, "-B option not supported with -x\n");
1194 usage_with_options(stat_usage, options);
1195 } else /* Nope, so disable big number formatting */
1196 big_num = false;
1197 } else if (big_num_opt == 0) /* User passed --no-big-num */
1198 big_num = false;
1199
1200 if (!argc && target_pid == -1 && target_tid == -1)
1201 usage_with_options(stat_usage, options);
1202 if (run_count <= 0)
1203 usage_with_options(stat_usage, options);
1204
1205 /* no_aggr, cgroup are for system-wide only */
1206 if ((no_aggr || nr_cgroups) && !system_wide) {
1207 fprintf(stderr, "both cgroup and no-aggregation "
1208 "modes only available in system-wide mode\n");
1209
1210 usage_with_options(stat_usage, options);
1211 }
1212
1213 if (add_default_attributes())
1214 goto out;
1215
1216 if (target_pid != -1)
1217 target_tid = target_pid;
1218
1219 evsel_list->threads = thread_map__new(target_pid, target_tid);
1220 if (evsel_list->threads == NULL) {
1221 pr_err("Problems finding threads of monitor\n");
1222 usage_with_options(stat_usage, options);
1223 }
1224
1225 if (system_wide)
1226 evsel_list->cpus = cpu_map__new(cpu_list);
1227 else
1228 evsel_list->cpus = cpu_map__dummy_new();
1229
1230 if (evsel_list->cpus == NULL) {
1231 perror("failed to parse CPUs map");
1232 usage_with_options(stat_usage, options);
1233 return -1;
1234 }
1235
1236 list_for_each_entry(pos, &evsel_list->entries, node) {
1237 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1238 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 ||
1239 perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0)
1240 goto out_free_fd;
1241 }
1242
1243 /*
1244 * We dont want to block the signals - that would cause
1245 * child tasks to inherit that and Ctrl-C would not work.
1246 * What we want is for Ctrl-C to work in the exec()-ed
1247 * task, but being ignored by perf stat itself:
1248 */
1249 atexit(sig_atexit);
1250 signal(SIGINT, skip_signal);
1251 signal(SIGALRM, skip_signal);
1252 signal(SIGABRT, skip_signal);
1253
1254 status = 0;
1255 for (run_idx = 0; run_idx < run_count; run_idx++) {
1256 if (run_count != 1 && verbose)
1257 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1258 run_idx + 1);
1259
1260 if (sync_run)
1261 sync();
1262
1263 status = run_perf_stat(argc, argv);
1264 }
1265
1266 if (status != -1)
1267 print_stat(argc, argv);
1268 out_free_fd:
1269 list_for_each_entry(pos, &evsel_list->entries, node)
1270 perf_evsel__free_stat_priv(pos);
1271 perf_evlist__delete_maps(evsel_list);
1272 out:
1273 perf_evlist__delete(evsel_list);
1274 return status;
1275 }
x86 "subsystem" repository