?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge tag 'v3.3.7' into 3.3/master
[zen-stable.git] / tools / perf / builtin-stat.c
blobf5d2a63eba665c0d201500ff15a20042aec3afae
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 ||
467 errno == ENOENT || errno == EOPNOTSUPP) {
468 if (verbose)
469 ui__warning("%s event is not supported by the kernel.\n",
470 event_name(counter));
471 counter->supported = false;
472 continue;
473 }
474
475 if (errno == EPERM || errno == EACCES) {
476 error("You may not have permission to collect %sstats.\n"
477 "\t Consider tweaking"
478 " /proc/sys/kernel/perf_event_paranoid or running as root.",
479 system_wide ? "system-wide " : "");
480 } else {
481 error("open_counter returned with %d (%s). "
482 "/bin/dmesg may provide additional information.\n",
483 errno, strerror(errno));
484 }
485 if (child_pid != -1)
486 kill(child_pid, SIGTERM);
487 die("Not all events could be opened.\n");
488 return -1;
489 }
490 counter->supported = true;
491 }
492
493 if (perf_evlist__set_filters(evsel_list)) {
494 error("failed to set filter with %d (%s)\n", errno,
495 strerror(errno));
496 return -1;
497 }
498
499 /*
500 * Enable counters and exec the command:
501 */
502 t0 = rdclock();
503
504 if (forks) {
505 close(go_pipe[1]);
506 wait(&status);
507 if (WIFSIGNALED(status))
508 psignal(WTERMSIG(status), argv[0]);
509 } else {
510 while(!done) sleep(1);
511 }
512
513 t1 = rdclock();
514
515 update_stats(&walltime_nsecs_stats, t1 - t0);
516
517 if (no_aggr) {
518 list_for_each_entry(counter, &evsel_list->entries, node) {
519 read_counter(counter);
520 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
521 }
522 } else {
523 list_for_each_entry(counter, &evsel_list->entries, node) {
524 read_counter_aggr(counter);
525 perf_evsel__close_fd(counter, evsel_list->cpus->nr,
526 evsel_list->threads->nr);
527 }
528 }
529
530 return WEXITSTATUS(status);
531 }
532
533 static void print_noise_pct(double total, double avg)
534 {
535 double pct = 0.0;
536
537 if (avg)
538 pct = 100.0*total/avg;
539
540 if (csv_output)
541 fprintf(output, "%s%.2f%%", csv_sep, pct);
542 else if (pct)
543 fprintf(output, " ( +-%6.2f%% )", pct);
544 }
545
546 static void print_noise(struct perf_evsel *evsel, double avg)
547 {
548 struct perf_stat *ps;
549
550 if (run_count == 1)
551 return;
552
553 ps = evsel->priv;
554 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
555 }
556
557 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
558 {
559 double msecs = avg / 1e6;
560 char cpustr[16] = { '\0', };
561 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
562
563 if (no_aggr)
564 sprintf(cpustr, "CPU%*d%s",
565 csv_output ? 0 : -4,
566 evsel_list->cpus->map[cpu], csv_sep);
567
568 fprintf(output, fmt, cpustr, msecs, csv_sep, event_name(evsel));
569
570 if (evsel->cgrp)
571 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
572
573 if (csv_output)
574 return;
575
576 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
577 fprintf(output, " # %8.3f CPUs utilized ",
578 avg / avg_stats(&walltime_nsecs_stats));
579 }
580
581 /* used for get_ratio_color() */
582 enum grc_type {
583 GRC_STALLED_CYCLES_FE,
584 GRC_STALLED_CYCLES_BE,
585 GRC_CACHE_MISSES,
586 GRC_MAX_NR
587 };
588
589 static const char *get_ratio_color(enum grc_type type, double ratio)
590 {
591 static const double grc_table[GRC_MAX_NR][3] = {
592 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
593 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
594 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
595 };
596 const char *color = PERF_COLOR_NORMAL;
597
598 if (ratio > grc_table[type][0])
599 color = PERF_COLOR_RED;
600 else if (ratio > grc_table[type][1])
601 color = PERF_COLOR_MAGENTA;
602 else if (ratio > grc_table[type][2])
603 color = PERF_COLOR_YELLOW;
604
605 return color;
606 }
607
608 static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg)
609 {
610 double total, ratio = 0.0;
611 const char *color;
612
613 total = avg_stats(&runtime_cycles_stats[cpu]);
614
615 if (total)
616 ratio = avg / total * 100.0;
617
618 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
619
620 fprintf(output, " # ");
621 color_fprintf(output, color, "%6.2f%%", ratio);
622 fprintf(output, " frontend cycles idle ");
623 }
624
625 static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg)
626 {
627 double total, ratio = 0.0;
628 const char *color;
629
630 total = avg_stats(&runtime_cycles_stats[cpu]);
631
632 if (total)
633 ratio = avg / total * 100.0;
634
635 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
636
637 fprintf(output, " # ");
638 color_fprintf(output, color, "%6.2f%%", ratio);
639 fprintf(output, " backend cycles idle ");
640 }
641
642 static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
643 {
644 double total, ratio = 0.0;
645 const char *color;
646
647 total = avg_stats(&runtime_branches_stats[cpu]);
648
649 if (total)
650 ratio = avg / total * 100.0;
651
652 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
653
654 fprintf(output, " # ");
655 color_fprintf(output, color, "%6.2f%%", ratio);
656 fprintf(output, " of all branches ");
657 }
658
659 static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
660 {
661 double total, ratio = 0.0;
662 const char *color;
663
664 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
665
666 if (total)
667 ratio = avg / total * 100.0;
668
669 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
670
671 fprintf(output, " # ");
672 color_fprintf(output, color, "%6.2f%%", ratio);
673 fprintf(output, " of all L1-dcache hits ");
674 }
675
676 static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
677 {
678 double total, ratio = 0.0;
679 const char *color;
680
681 total = avg_stats(&runtime_l1_icache_stats[cpu]);
682
683 if (total)
684 ratio = avg / total * 100.0;
685
686 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
687
688 fprintf(output, " # ");
689 color_fprintf(output, color, "%6.2f%%", ratio);
690 fprintf(output, " of all L1-icache hits ");
691 }
692
693 static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
694 {
695 double total, ratio = 0.0;
696 const char *color;
697
698 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
699
700 if (total)
701 ratio = avg / total * 100.0;
702
703 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
704
705 fprintf(output, " # ");
706 color_fprintf(output, color, "%6.2f%%", ratio);
707 fprintf(output, " of all dTLB cache hits ");
708 }
709
710 static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
711 {
712 double total, ratio = 0.0;
713 const char *color;
714
715 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
716
717 if (total)
718 ratio = avg / total * 100.0;
719
720 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
721
722 fprintf(output, " # ");
723 color_fprintf(output, color, "%6.2f%%", ratio);
724 fprintf(output, " of all iTLB cache hits ");
725 }
726
727 static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
728 {
729 double total, ratio = 0.0;
730 const char *color;
731
732 total = avg_stats(&runtime_ll_cache_stats[cpu]);
733
734 if (total)
735 ratio = avg / total * 100.0;
736
737 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
738
739 fprintf(output, " # ");
740 color_fprintf(output, color, "%6.2f%%", ratio);
741 fprintf(output, " of all LL-cache hits ");
742 }
743
744 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
745 {
746 double total, ratio = 0.0;
747 char cpustr[16] = { '\0', };
748 const char *fmt;
749
750 if (csv_output)
751 fmt = "%s%.0f%s%s";
752 else if (big_num)
753 fmt = "%s%'18.0f%s%-25s";
754 else
755 fmt = "%s%18.0f%s%-25s";
756
757 if (no_aggr)
758 sprintf(cpustr, "CPU%*d%s",
759 csv_output ? 0 : -4,
760 evsel_list->cpus->map[cpu], csv_sep);
761 else
762 cpu = 0;
763
764 fprintf(output, fmt, cpustr, avg, csv_sep, event_name(evsel));
765
766 if (evsel->cgrp)
767 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
768
769 if (csv_output)
770 return;
771
772 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
773 total = avg_stats(&runtime_cycles_stats[cpu]);
774
775 if (total)
776 ratio = avg / total;
777
778 fprintf(output, " # %5.2f insns per cycle ", ratio);
779
780 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
781 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
782
783 if (total && avg) {
784 ratio = total / avg;
785 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
786 }
787
788 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
789 runtime_branches_stats[cpu].n != 0) {
790 print_branch_misses(cpu, evsel, avg);
791 } else if (
792 evsel->attr.type == PERF_TYPE_HW_CACHE &&
793 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
794 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
795 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
796 runtime_l1_dcache_stats[cpu].n != 0) {
797 print_l1_dcache_misses(cpu, evsel, avg);
798 } else if (
799 evsel->attr.type == PERF_TYPE_HW_CACHE &&
800 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
801 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
802 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
803 runtime_l1_icache_stats[cpu].n != 0) {
804 print_l1_icache_misses(cpu, evsel, avg);
805 } else if (
806 evsel->attr.type == PERF_TYPE_HW_CACHE &&
807 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
808 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
809 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
810 runtime_dtlb_cache_stats[cpu].n != 0) {
811 print_dtlb_cache_misses(cpu, evsel, avg);
812 } else if (
813 evsel->attr.type == PERF_TYPE_HW_CACHE &&
814 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
815 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
816 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
817 runtime_itlb_cache_stats[cpu].n != 0) {
818 print_itlb_cache_misses(cpu, evsel, avg);
819 } else if (
820 evsel->attr.type == PERF_TYPE_HW_CACHE &&
821 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
822 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
823 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
824 runtime_ll_cache_stats[cpu].n != 0) {
825 print_ll_cache_misses(cpu, evsel, avg);
826 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
827 runtime_cacherefs_stats[cpu].n != 0) {
828 total = avg_stats(&runtime_cacherefs_stats[cpu]);
829
830 if (total)
831 ratio = avg * 100 / total;
832
833 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
834
835 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
836 print_stalled_cycles_frontend(cpu, evsel, avg);
837 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
838 print_stalled_cycles_backend(cpu, evsel, avg);
839 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
840 total = avg_stats(&runtime_nsecs_stats[cpu]);
841
842 if (total)
843 ratio = 1.0 * avg / total;
844
845 fprintf(output, " # %8.3f GHz ", ratio);
846 } else if (runtime_nsecs_stats[cpu].n != 0) {
847 total = avg_stats(&runtime_nsecs_stats[cpu]);
848
849 if (total)
850 ratio = 1000.0 * avg / total;
851
852 fprintf(output, " # %8.3f M/sec ", ratio);
853 } else {
854 fprintf(output, " ");
855 }
856 }
857
858 /*
859 * Print out the results of a single counter:
860 * aggregated counts in system-wide mode
861 */
862 static void print_counter_aggr(struct perf_evsel *counter)
863 {
864 struct perf_stat *ps = counter->priv;
865 double avg = avg_stats(&ps->res_stats[0]);
866 int scaled = counter->counts->scaled;
867
868 if (scaled == -1) {
869 fprintf(output, "%*s%s%*s",
870 csv_output ? 0 : 18,
871 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
872 csv_sep,
873 csv_output ? 0 : -24,
874 event_name(counter));
875
876 if (counter->cgrp)
877 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
878
879 fputc('\n', output);
880 return;
881 }
882
883 if (nsec_counter(counter))
884 nsec_printout(-1, counter, avg);
885 else
886 abs_printout(-1, counter, avg);
887
888 print_noise(counter, avg);
889
890 if (csv_output) {
891 fputc('\n', output);
892 return;
893 }
894
895 if (scaled) {
896 double avg_enabled, avg_running;
897
898 avg_enabled = avg_stats(&ps->res_stats[1]);
899 avg_running = avg_stats(&ps->res_stats[2]);
900
901 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
902 }
903 fprintf(output, "\n");
904 }
905
906 /*
907 * Print out the results of a single counter:
908 * does not use aggregated count in system-wide
909 */
910 static void print_counter(struct perf_evsel *counter)
911 {
912 u64 ena, run, val;
913 int cpu;
914
915 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
916 val = counter->counts->cpu[cpu].val;
917 ena = counter->counts->cpu[cpu].ena;
918 run = counter->counts->cpu[cpu].run;
919 if (run == 0 || ena == 0) {
920 fprintf(output, "CPU%*d%s%*s%s%*s",
921 csv_output ? 0 : -4,
922 evsel_list->cpus->map[cpu], csv_sep,
923 csv_output ? 0 : 18,
924 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
925 csv_sep,
926 csv_output ? 0 : -24,
927 event_name(counter));
928
929 if (counter->cgrp)
930 fprintf(output, "%s%s",
931 csv_sep, counter->cgrp->name);
932
933 fputc('\n', output);
934 continue;
935 }
936
937 if (nsec_counter(counter))
938 nsec_printout(cpu, counter, val);
939 else
940 abs_printout(cpu, counter, val);
941
942 if (!csv_output) {
943 print_noise(counter, 1.0);
944
945 if (run != ena)
946 fprintf(output, " (%.2f%%)",
947 100.0 * run / ena);
948 }
949 fputc('\n', output);
950 }
951 }
952
953 static void print_stat(int argc, const char **argv)
954 {
955 struct perf_evsel *counter;
956 int i;
957
958 fflush(stdout);
959
960 if (!csv_output) {
961 fprintf(output, "\n");
962 fprintf(output, " Performance counter stats for ");
963 if(target_pid == -1 && target_tid == -1) {
964 fprintf(output, "\'%s", argv[0]);
965 for (i = 1; i < argc; i++)
966 fprintf(output, " %s", argv[i]);
967 } else if (target_pid != -1)
968 fprintf(output, "process id \'%d", target_pid);
969 else
970 fprintf(output, "thread id \'%d", target_tid);
971
972 fprintf(output, "\'");
973 if (run_count > 1)
974 fprintf(output, " (%d runs)", run_count);
975 fprintf(output, ":\n\n");
976 }
977
978 if (no_aggr) {
979 list_for_each_entry(counter, &evsel_list->entries, node)
980 print_counter(counter);
981 } else {
982 list_for_each_entry(counter, &evsel_list->entries, node)
983 print_counter_aggr(counter);
984 }
985
986 if (!csv_output) {
987 if (!null_run)
988 fprintf(output, "\n");
989 fprintf(output, " %17.9f seconds time elapsed",
990 avg_stats(&walltime_nsecs_stats)/1e9);
991 if (run_count > 1) {
992 fprintf(output, " ");
993 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
994 avg_stats(&walltime_nsecs_stats));
995 }
996 fprintf(output, "\n\n");
997 }
998 }
999
1000 static volatile int signr = -1;
1001
1002 static void skip_signal(int signo)
1003 {
1004 if(child_pid == -1)
1005 done = 1;
1006
1007 signr = signo;
1008 }
1009
1010 static void sig_atexit(void)
1011 {
1012 if (child_pid != -1)
1013 kill(child_pid, SIGTERM);
1014
1015 if (signr == -1)
1016 return;
1017
1018 signal(signr, SIG_DFL);
1019 kill(getpid(), signr);
1020 }
1021
1022 static const char * const stat_usage[] = {
1023 "perf stat [<options>] [<command>]",
1024 NULL
1025 };
1026
1027 static int stat__set_big_num(const struct option *opt __used,
1028 const char *s __used, int unset)
1029 {
1030 big_num_opt = unset ? 0 : 1;
1031 return 0;
1032 }
1033
1034 static bool append_file;
1035
1036 static const struct option options[] = {
1037 OPT_CALLBACK('e', "event", &evsel_list, "event",
1038 "event selector. use 'perf list' to list available events",
1039 parse_events_option),
1040 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1041 "event filter", parse_filter),
1042 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1043 "child tasks do not inherit counters"),
1044 OPT_INTEGER('p', "pid", &target_pid,
1045 "stat events on existing process id"),
1046 OPT_INTEGER('t', "tid", &target_tid,
1047 "stat events on existing thread id"),
1048 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1049 "system-wide collection from all CPUs"),
1050 OPT_BOOLEAN('g', "group", &group,
1051 "put the counters into a counter group"),
1052 OPT_BOOLEAN('c', "scale", &scale,
1053 "scale/normalize counters"),
1054 OPT_INCR('v', "verbose", &verbose,
1055 "be more verbose (show counter open errors, etc)"),
1056 OPT_INTEGER('r', "repeat", &run_count,
1057 "repeat command and print average + stddev (max: 100)"),
1058 OPT_BOOLEAN('n', "null", &null_run,
1059 "null run - dont start any counters"),
1060 OPT_INCR('d', "detailed", &detailed_run,
1061 "detailed run - start a lot of events"),
1062 OPT_BOOLEAN('S', "sync", &sync_run,
1063 "call sync() before starting a run"),
1064 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1065 "print large numbers with thousands\' separators",
1066 stat__set_big_num),
1067 OPT_STRING('C', "cpu", &cpu_list, "cpu",
1068 "list of cpus to monitor in system-wide"),
1069 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1070 "disable CPU count aggregation"),
1071 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1072 "print counts with custom separator"),
1073 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1074 "monitor event in cgroup name only",
1075 parse_cgroups),
1076 OPT_STRING('o', "output", &output_name, "file",
1077 "output file name"),
1078 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1079 OPT_INTEGER(0, "log-fd", &output_fd,
1080 "log output to fd, instead of stderr"),
1081 OPT_END()
1082 };
1083
1084 /*
1085 * Add default attributes, if there were no attributes specified or
1086 * if -d/--detailed, -d -d or -d -d -d is used:
1087 */
1088 static int add_default_attributes(void)
1089 {
1090 /* Set attrs if no event is selected and !null_run: */
1091 if (null_run)
1092 return 0;
1093
1094 if (!evsel_list->nr_entries) {
1095 if (perf_evlist__add_attrs_array(evsel_list, default_attrs) < 0)
1096 return -1;
1097 }
1098
1099 /* Detailed events get appended to the event list: */
1100
1101 if (detailed_run < 1)
1102 return 0;
1103
1104 /* Append detailed run extra attributes: */
1105 if (perf_evlist__add_attrs_array(evsel_list, detailed_attrs) < 0)
1106 return -1;
1107
1108 if (detailed_run < 2)
1109 return 0;
1110
1111 /* Append very detailed run extra attributes: */
1112 if (perf_evlist__add_attrs_array(evsel_list, very_detailed_attrs) < 0)
1113 return -1;
1114
1115 if (detailed_run < 3)
1116 return 0;
1117
1118 /* Append very, very detailed run extra attributes: */
1119 return perf_evlist__add_attrs_array(evsel_list, very_very_detailed_attrs);
1120 }
1121
1122 int cmd_stat(int argc, const char **argv, const char *prefix __used)
1123 {
1124 struct perf_evsel *pos;
1125 int status = -ENOMEM;
1126 const char *mode;
1127
1128 setlocale(LC_ALL, "");
1129
1130 evsel_list = perf_evlist__new(NULL, NULL);
1131 if (evsel_list == NULL)
1132 return -ENOMEM;
1133
1134 argc = parse_options(argc, argv, options, stat_usage,
1135 PARSE_OPT_STOP_AT_NON_OPTION);
1136
1137 output = stderr;
1138 if (output_name && strcmp(output_name, "-"))
1139 output = NULL;
1140
1141 if (output_name && output_fd) {
1142 fprintf(stderr, "cannot use both --output and --log-fd\n");
1143 usage_with_options(stat_usage, options);
1144 }
1145 if (!output) {
1146 struct timespec tm;
1147 mode = append_file ? "a" : "w";
1148
1149 output = fopen(output_name, mode);
1150 if (!output) {
1151 perror("failed to create output file");
1152 exit(-1);
1153 }
1154 clock_gettime(CLOCK_REALTIME, &tm);
1155 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1156 } else if (output_fd != 2) {
1157 mode = append_file ? "a" : "w";
1158 output = fdopen(output_fd, mode);
1159 if (!output) {
1160 perror("Failed opening logfd");
1161 return -errno;
1162 }
1163 }
1164
1165 if (csv_sep) {
1166 csv_output = true;
1167 if (!strcmp(csv_sep, "\\t"))
1168 csv_sep = "\t";
1169 } else
1170 csv_sep = DEFAULT_SEPARATOR;
1171
1172 /*
1173 * let the spreadsheet do the pretty-printing
1174 */
1175 if (csv_output) {
1176 /* User explicitly passed -B? */
1177 if (big_num_opt == 1) {
1178 fprintf(stderr, "-B option not supported with -x\n");
1179 usage_with_options(stat_usage, options);
1180 } else /* Nope, so disable big number formatting */
1181 big_num = false;
1182 } else if (big_num_opt == 0) /* User passed --no-big-num */
1183 big_num = false;
1184
1185 if (!argc && target_pid == -1 && target_tid == -1)
1186 usage_with_options(stat_usage, options);
1187 if (run_count <= 0)
1188 usage_with_options(stat_usage, options);
1189
1190 /* no_aggr, cgroup are for system-wide only */
1191 if ((no_aggr || nr_cgroups) && !system_wide) {
1192 fprintf(stderr, "both cgroup and no-aggregation "
1193 "modes only available in system-wide mode\n");
1194
1195 usage_with_options(stat_usage, options);
1196 }
1197
1198 if (add_default_attributes())
1199 goto out;
1200
1201 if (target_pid != -1)
1202 target_tid = target_pid;
1203
1204 evsel_list->threads = thread_map__new(target_pid, target_tid);
1205 if (evsel_list->threads == NULL) {
1206 pr_err("Problems finding threads of monitor\n");
1207 usage_with_options(stat_usage, options);
1208 }
1209
1210 if (system_wide)
1211 evsel_list->cpus = cpu_map__new(cpu_list);
1212 else
1213 evsel_list->cpus = cpu_map__dummy_new();
1214
1215 if (evsel_list->cpus == NULL) {
1216 perror("failed to parse CPUs map");
1217 usage_with_options(stat_usage, options);
1218 return -1;
1219 }
1220
1221 list_for_each_entry(pos, &evsel_list->entries, node) {
1222 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1223 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0)
1224 goto out_free_fd;
1225 }
1226
1227 /*
1228 * We dont want to block the signals - that would cause
1229 * child tasks to inherit that and Ctrl-C would not work.
1230 * What we want is for Ctrl-C to work in the exec()-ed
1231 * task, but being ignored by perf stat itself:
1232 */
1233 atexit(sig_atexit);
1234 signal(SIGINT, skip_signal);
1235 signal(SIGALRM, skip_signal);
1236 signal(SIGABRT, skip_signal);
1237
1238 status = 0;
1239 for (run_idx = 0; run_idx < run_count; run_idx++) {
1240 if (run_count != 1 && verbose)
1241 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1242 run_idx + 1);
1243
1244 if (sync_run)
1245 sync();
1246
1247 status = run_perf_stat(argc, argv);
1248 }
1249
1250 if (status != -1)
1251 print_stat(argc, argv);
1252 out_free_fd:
1253 list_for_each_entry(pos, &evsel_list->entries, node)
1254 perf_evsel__free_stat_priv(pos);
1255 perf_evlist__delete_maps(evsel_list);
1256 out:
1257 perf_evlist__delete(evsel_list);
1258 return status;
1259 }