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/stat.h"
55 #include "util/header.h"
56 #include "util/cpumap.h"
57 #include "util/thread.h"
58 #include "util/thread_map.h"
59
60 #include <stdlib.h>
61 #include <sys/prctl.h>
62 #include <locale.h>
63
64 #define DEFAULT_SEPARATOR " "
65 #define CNTR_NOT_SUPPORTED "<not supported>"
66 #define CNTR_NOT_COUNTED "<not counted>"
67
68 static void print_stat(int argc, const char **argv);
69 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
70 static void print_counter(struct perf_evsel *counter, char *prefix);
71 static void print_aggr(char *prefix);
72
73 static struct perf_evlist *evsel_list;
74
75 static struct perf_target target = {
76 .uid = UINT_MAX,
77 };
78
79 enum aggr_mode {
80 AGGR_NONE,
81 AGGR_GLOBAL,
82 AGGR_SOCKET,
83 AGGR_CORE,
84 };
85
86 static int run_count = 1;
87 static bool no_inherit = false;
88 static bool scale = true;
89 static enum aggr_mode aggr_mode = AGGR_GLOBAL;
90 static volatile pid_t child_pid = -1;
91 static bool null_run = false;
92 static int detailed_run = 0;
93 static bool big_num = true;
94 static int big_num_opt = -1;
95 static const char *csv_sep = NULL;
96 static bool csv_output = false;
97 static bool group = false;
98 static FILE *output = NULL;
99 static const char *pre_cmd = NULL;
100 static const char *post_cmd = NULL;
101 static bool sync_run = false;
102 static unsigned int interval = 0;
103 static unsigned int initial_delay = 0;
104 static bool forever = false;
105 static struct timespec ref_time;
106 static struct cpu_map *aggr_map;
107 static int (*aggr_get_id)(struct cpu_map *m, int cpu);
108
109 static volatile int done = 0;
110
111 struct perf_stat {
112 struct stats res_stats[3];
113 };
114
diff_timespec(struct timespec * r,struct timespec * a,struct timespec * b)115 static inline void diff_timespec(struct timespec *r, struct timespec *a,
116 struct timespec *b)
117 {
118 r->tv_sec = a->tv_sec - b->tv_sec;
119 if (a->tv_nsec < b->tv_nsec) {
120 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
121 r->tv_sec--;
122 } else {
123 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
124 }
125 }
126
perf_evsel__cpus(struct perf_evsel * evsel)127 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
128 {
129 return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
130 }
131
perf_evsel__nr_cpus(struct perf_evsel * evsel)132 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
133 {
134 return perf_evsel__cpus(evsel)->nr;
135 }
136
perf_evsel__reset_stat_priv(struct perf_evsel * evsel)137 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
138 {
139 memset(evsel->priv, 0, sizeof(struct perf_stat));
140 }
141
perf_evsel__alloc_stat_priv(struct perf_evsel * evsel)142 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
143 {
144 evsel->priv = zalloc(sizeof(struct perf_stat));
145 return evsel->priv == NULL ? -ENOMEM : 0;
146 }
147
perf_evsel__free_stat_priv(struct perf_evsel * evsel)148 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
149 {
150 free(evsel->priv);
151 evsel->priv = NULL;
152 }
153
perf_evsel__alloc_prev_raw_counts(struct perf_evsel * evsel)154 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
155 {
156 void *addr;
157 size_t sz;
158
159 sz = sizeof(*evsel->counts) +
160 (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
161
162 addr = zalloc(sz);
163 if (!addr)
164 return -ENOMEM;
165
166 evsel->prev_raw_counts = addr;
167
168 return 0;
169 }
170
perf_evsel__free_prev_raw_counts(struct perf_evsel * evsel)171 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
172 {
173 free(evsel->prev_raw_counts);
174 evsel->prev_raw_counts = NULL;
175 }
176
perf_evlist__free_stats(struct perf_evlist * evlist)177 static void perf_evlist__free_stats(struct perf_evlist *evlist)
178 {
179 struct perf_evsel *evsel;
180
181 list_for_each_entry(evsel, &evlist->entries, node) {
182 perf_evsel__free_stat_priv(evsel);
183 perf_evsel__free_counts(evsel);
184 perf_evsel__free_prev_raw_counts(evsel);
185 }
186 }
187
perf_evlist__alloc_stats(struct perf_evlist * evlist,bool alloc_raw)188 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
189 {
190 struct perf_evsel *evsel;
191
192 list_for_each_entry(evsel, &evlist->entries, node) {
193 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
194 perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
195 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
196 goto out_free;
197 }
198
199 return 0;
200
201 out_free:
202 perf_evlist__free_stats(evlist);
203 return -1;
204 }
205
206 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
207 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
208 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
209 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
210 static struct stats runtime_branches_stats[MAX_NR_CPUS];
211 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
212 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
213 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
214 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
215 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
216 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
217 static struct stats walltime_nsecs_stats;
218
perf_stat__reset_stats(struct perf_evlist * evlist)219 static void perf_stat__reset_stats(struct perf_evlist *evlist)
220 {
221 struct perf_evsel *evsel;
222
223 list_for_each_entry(evsel, &evlist->entries, node) {
224 perf_evsel__reset_stat_priv(evsel);
225 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
226 }
227
228 memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
229 memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
230 memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
231 memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
232 memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
233 memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
234 memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
235 memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
236 memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
237 memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
238 memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
239 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
240 }
241
create_perf_stat_counter(struct perf_evsel * evsel)242 static int create_perf_stat_counter(struct perf_evsel *evsel)
243 {
244 struct perf_event_attr *attr = &evsel->attr;
245
246 if (scale)
247 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
248 PERF_FORMAT_TOTAL_TIME_RUNNING;
249
250 attr->inherit = !no_inherit;
251
252 if (perf_target__has_cpu(&target))
253 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
254
255 if (!perf_target__has_task(&target) &&
256 perf_evsel__is_group_leader(evsel)) {
257 attr->disabled = 1;
258 if (!initial_delay)
259 attr->enable_on_exec = 1;
260 }
261
262 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
263 }
264
265 /*
266 * Does the counter have nsecs as a unit?
267 */
nsec_counter(struct perf_evsel * evsel)268 static inline int nsec_counter(struct perf_evsel *evsel)
269 {
270 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
271 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
272 return 1;
273
274 return 0;
275 }
276
277 /*
278 * Update various tracking values we maintain to print
279 * more semantic information such as miss/hit ratios,
280 * instruction rates, etc:
281 */
update_shadow_stats(struct perf_evsel * counter,u64 * count)282 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
283 {
284 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
285 update_stats(&runtime_nsecs_stats[0], count[0]);
286 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
287 update_stats(&runtime_cycles_stats[0], count[0]);
288 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
289 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
290 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
291 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
292 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
293 update_stats(&runtime_branches_stats[0], count[0]);
294 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
295 update_stats(&runtime_cacherefs_stats[0], count[0]);
296 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
297 update_stats(&runtime_l1_dcache_stats[0], count[0]);
298 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
299 update_stats(&runtime_l1_icache_stats[0], count[0]);
300 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
301 update_stats(&runtime_ll_cache_stats[0], count[0]);
302 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
303 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
304 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
305 update_stats(&runtime_itlb_cache_stats[0], count[0]);
306 }
307
308 /*
309 * Read out the results of a single counter:
310 * aggregate counts across CPUs in system-wide mode
311 */
read_counter_aggr(struct perf_evsel * counter)312 static int read_counter_aggr(struct perf_evsel *counter)
313 {
314 struct perf_stat *ps = counter->priv;
315 u64 *count = counter->counts->aggr.values;
316 int i;
317
318 if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
319 thread_map__nr(evsel_list->threads), scale) < 0)
320 return -1;
321
322 for (i = 0; i < 3; i++)
323 update_stats(&ps->res_stats[i], count[i]);
324
325 if (verbose) {
326 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
327 perf_evsel__name(counter), count[0], count[1], count[2]);
328 }
329
330 /*
331 * Save the full runtime - to allow normalization during printout:
332 */
333 update_shadow_stats(counter, count);
334
335 return 0;
336 }
337
338 /*
339 * Read out the results of a single counter:
340 * do not aggregate counts across CPUs in system-wide mode
341 */
read_counter(struct perf_evsel * counter)342 static int read_counter(struct perf_evsel *counter)
343 {
344 u64 *count;
345 int cpu;
346
347 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
348 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
349 return -1;
350
351 count = counter->counts->cpu[cpu].values;
352
353 update_shadow_stats(counter, count);
354 }
355
356 return 0;
357 }
358
print_interval(void)359 static void print_interval(void)
360 {
361 #ifndef __APPLE__
362 static int num_print_interval;
363 struct perf_evsel *counter;
364 struct perf_stat *ps;
365 struct timespec ts, rs;
366 char prefix[64];
367
368 if (aggr_mode == AGGR_GLOBAL) {
369 list_for_each_entry(counter, &evsel_list->entries, node) {
370 ps = counter->priv;
371 memset(ps->res_stats, 0, sizeof(ps->res_stats));
372 read_counter_aggr(counter);
373 }
374 } else {
375 list_for_each_entry(counter, &evsel_list->entries, node) {
376 ps = counter->priv;
377 memset(ps->res_stats, 0, sizeof(ps->res_stats));
378 read_counter(counter);
379 }
380 }
381
382 clock_gettime(CLOCK_MONOTONIC, &ts);
383 diff_timespec(&rs, &ts, &ref_time);
384 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
385
386 if (num_print_interval == 0 && !csv_output) {
387 switch (aggr_mode) {
388 case AGGR_SOCKET:
389 fprintf(output, "# time socket cpus counts events\n");
390 break;
391 case AGGR_CORE:
392 fprintf(output, "# time core cpus counts events\n");
393 break;
394 case AGGR_NONE:
395 fprintf(output, "# time CPU counts events\n");
396 break;
397 case AGGR_GLOBAL:
398 default:
399 fprintf(output, "# time counts events\n");
400 }
401 }
402
403 if (++num_print_interval == 25)
404 num_print_interval = 0;
405
406 switch (aggr_mode) {
407 case AGGR_CORE:
408 case AGGR_SOCKET:
409 print_aggr(prefix);
410 break;
411 case AGGR_NONE:
412 list_for_each_entry(counter, &evsel_list->entries, node)
413 print_counter(counter, prefix);
414 break;
415 case AGGR_GLOBAL:
416 default:
417 list_for_each_entry(counter, &evsel_list->entries, node)
418 print_counter_aggr(counter, prefix);
419 }
420
421 fflush(output);
422 #else
423 perror("print_interval not supported on MacOS");
424 #endif
425 }
426
handle_initial_delay(void)427 static void handle_initial_delay(void)
428 {
429 struct perf_evsel *counter;
430
431 if (initial_delay) {
432 const int ncpus = cpu_map__nr(evsel_list->cpus),
433 nthreads = thread_map__nr(evsel_list->threads);
434
435 usleep(initial_delay * 1000);
436 list_for_each_entry(counter, &evsel_list->entries, node)
437 perf_evsel__enable(counter, ncpus, nthreads);
438 }
439 }
440
__run_perf_stat(int argc,const char ** argv)441 static int __run_perf_stat(int argc, const char **argv)
442 {
443 #ifndef __APPLE__
444 char msg[512];
445 unsigned long long t0, t1;
446 struct perf_evsel *counter;
447 struct timespec ts;
448 int status = 0;
449 const bool forks = (argc > 0);
450
451 if (interval) {
452 ts.tv_sec = interval / 1000;
453 ts.tv_nsec = (interval % 1000) * 1000000;
454 } else {
455 ts.tv_sec = 1;
456 ts.tv_nsec = 0;
457 }
458
459 if (forks) {
460 if (perf_evlist__prepare_workload(evsel_list, &target, argv,
461 false, false) < 0) {
462 perror("failed to prepare workload");
463 return -1;
464 }
465 child_pid = evsel_list->workload.pid;
466 }
467
468 if (group)
469 perf_evlist__set_leader(evsel_list);
470
471 list_for_each_entry(counter, &evsel_list->entries, node) {
472 if (create_perf_stat_counter(counter) < 0) {
473 /*
474 * PPC returns ENXIO for HW counters until 2.6.37
475 * (behavior changed with commit b0a873e).
476 */
477 if (errno == EINVAL || errno == ENOSYS ||
478 errno == ENOENT || errno == EOPNOTSUPP ||
479 errno == ENXIO) {
480 if (verbose)
481 ui__warning("%s event is not supported by the kernel.\n",
482 perf_evsel__name(counter));
483 counter->supported = false;
484 continue;
485 }
486
487 perf_evsel__open_strerror(counter, &target,
488 errno, msg, sizeof(msg));
489 ui__error("%s\n", msg);
490
491 if (child_pid != -1)
492 kill(child_pid, SIGTERM);
493
494 return -1;
495 }
496 counter->supported = true;
497 }
498
499 if (perf_evlist__apply_filters(evsel_list)) {
500 error("failed to set filter with %d (%s)\n", errno,
501 strerror(errno));
502 return -1;
503 }
504
505 /*
506 * Enable counters and exec the command:
507 */
508 t0 = rdclock();
509 clock_gettime(CLOCK_MONOTONIC, &ref_time);
510
511 if (forks) {
512 perf_evlist__start_workload(evsel_list);
513 handle_initial_delay();
514
515 if (interval) {
516 while (!waitpid(child_pid, &status, WNOHANG)) {
517 nanosleep(&ts, NULL);
518 print_interval();
519 }
520 }
521 wait(&status);
522 if (WIFSIGNALED(status))
523 psignal(WTERMSIG(status), argv[0]);
524 } else {
525 handle_initial_delay();
526 while (!done) {
527 nanosleep(&ts, NULL);
528 if (interval)
529 print_interval();
530 }
531 }
532
533 t1 = rdclock();
534
535 update_stats(&walltime_nsecs_stats, t1 - t0);
536
537 if (aggr_mode == AGGR_GLOBAL) {
538 list_for_each_entry(counter, &evsel_list->entries, node) {
539 read_counter_aggr(counter);
540 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
541 thread_map__nr(evsel_list->threads));
542 }
543 } else {
544 list_for_each_entry(counter, &evsel_list->entries, node) {
545 read_counter(counter);
546 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
547 }
548 }
549
550 return WEXITSTATUS(status);
551 #else
552 return -1;
553 #endif
554 }
555
run_perf_stat(int argc __maybe_unused,const char ** argv)556 static int run_perf_stat(int argc __maybe_unused, const char **argv)
557 {
558 int ret;
559
560 if (pre_cmd) {
561 ret = system(pre_cmd);
562 if (ret)
563 return ret;
564 }
565
566 if (sync_run)
567 sync();
568
569 ret = __run_perf_stat(argc, argv);
570 if (ret)
571 return ret;
572
573 if (post_cmd) {
574 ret = system(post_cmd);
575 if (ret)
576 return ret;
577 }
578
579 return ret;
580 }
581
print_noise_pct(double total,double avg)582 static void print_noise_pct(double total, double avg)
583 {
584 double pct = rel_stddev_stats(total, avg);
585
586 if (csv_output)
587 fprintf(output, "%s%.2f%%", csv_sep, pct);
588 else if (pct)
589 fprintf(output, " ( +-%6.2f%% )", pct);
590 }
591
print_noise(struct perf_evsel * evsel,double avg)592 static void print_noise(struct perf_evsel *evsel, double avg)
593 {
594 struct perf_stat *ps;
595
596 if (run_count == 1)
597 return;
598
599 ps = evsel->priv;
600 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
601 }
602
aggr_printout(struct perf_evsel * evsel,int id,int nr)603 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
604 {
605 switch (aggr_mode) {
606 case AGGR_CORE:
607 fprintf(output, "S%d-C%*d%s%*d%s",
608 cpu_map__id_to_socket(id),
609 csv_output ? 0 : -8,
610 cpu_map__id_to_cpu(id),
611 csv_sep,
612 csv_output ? 0 : 4,
613 nr,
614 csv_sep);
615 break;
616 case AGGR_SOCKET:
617 fprintf(output, "S%*d%s%*d%s",
618 csv_output ? 0 : -5,
619 id,
620 csv_sep,
621 csv_output ? 0 : 4,
622 nr,
623 csv_sep);
624 break;
625 case AGGR_NONE:
626 fprintf(output, "CPU%*d%s",
627 csv_output ? 0 : -4,
628 perf_evsel__cpus(evsel)->map[id], csv_sep);
629 break;
630 case AGGR_GLOBAL:
631 default:
632 break;
633 }
634 }
635
nsec_printout(int cpu,int nr,struct perf_evsel * evsel,double avg)636 static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
637 {
638 double msecs = avg / 1e6;
639 const char *fmt = csv_output ? "%.6f%s%s" : "%18.6f%s%-25s";
640
641 aggr_printout(evsel, cpu, nr);
642
643 fprintf(output, fmt, msecs, csv_sep, perf_evsel__name(evsel));
644
645 if (evsel->cgrp)
646 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
647
648 if (csv_output || interval)
649 return;
650
651 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
652 fprintf(output, " # %8.3f CPUs utilized ",
653 avg / avg_stats(&walltime_nsecs_stats));
654 else
655 fprintf(output, " ");
656 }
657
658 /* used for get_ratio_color() */
659 enum grc_type {
660 GRC_STALLED_CYCLES_FE,
661 GRC_STALLED_CYCLES_BE,
662 GRC_CACHE_MISSES,
663 GRC_MAX_NR
664 };
665
get_ratio_color(enum grc_type type,double ratio)666 static const char *get_ratio_color(enum grc_type type, double ratio)
667 {
668 static const double grc_table[GRC_MAX_NR][3] = {
669 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
670 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
671 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
672 };
673 const char *color = PERF_COLOR_NORMAL;
674
675 if (ratio > grc_table[type][0])
676 color = PERF_COLOR_RED;
677 else if (ratio > grc_table[type][1])
678 color = PERF_COLOR_MAGENTA;
679 else if (ratio > grc_table[type][2])
680 color = PERF_COLOR_YELLOW;
681
682 return color;
683 }
684
print_stalled_cycles_frontend(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)685 static void print_stalled_cycles_frontend(int cpu,
686 struct perf_evsel *evsel
687 __maybe_unused, double avg)
688 {
689 double total, ratio = 0.0;
690 const char *color;
691
692 total = avg_stats(&runtime_cycles_stats[cpu]);
693
694 if (total)
695 ratio = avg / total * 100.0;
696
697 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
698
699 fprintf(output, " # ");
700 color_fprintf(output, color, "%6.2f%%", ratio);
701 fprintf(output, " frontend cycles idle ");
702 }
703
print_stalled_cycles_backend(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)704 static void print_stalled_cycles_backend(int cpu,
705 struct perf_evsel *evsel
706 __maybe_unused, double avg)
707 {
708 double total, ratio = 0.0;
709 const char *color;
710
711 total = avg_stats(&runtime_cycles_stats[cpu]);
712
713 if (total)
714 ratio = avg / total * 100.0;
715
716 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
717
718 fprintf(output, " # ");
719 color_fprintf(output, color, "%6.2f%%", ratio);
720 fprintf(output, " backend cycles idle ");
721 }
722
print_branch_misses(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)723 static void print_branch_misses(int cpu,
724 struct perf_evsel *evsel __maybe_unused,
725 double avg)
726 {
727 double total, ratio = 0.0;
728 const char *color;
729
730 total = avg_stats(&runtime_branches_stats[cpu]);
731
732 if (total)
733 ratio = avg / total * 100.0;
734
735 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
736
737 fprintf(output, " # ");
738 color_fprintf(output, color, "%6.2f%%", ratio);
739 fprintf(output, " of all branches ");
740 }
741
print_l1_dcache_misses(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)742 static void print_l1_dcache_misses(int cpu,
743 struct perf_evsel *evsel __maybe_unused,
744 double avg)
745 {
746 double total, ratio = 0.0;
747 const char *color;
748
749 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
750
751 if (total)
752 ratio = avg / total * 100.0;
753
754 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
755
756 fprintf(output, " # ");
757 color_fprintf(output, color, "%6.2f%%", ratio);
758 fprintf(output, " of all L1-dcache hits ");
759 }
760
print_l1_icache_misses(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)761 static void print_l1_icache_misses(int cpu,
762 struct perf_evsel *evsel __maybe_unused,
763 double avg)
764 {
765 double total, ratio = 0.0;
766 const char *color;
767
768 total = avg_stats(&runtime_l1_icache_stats[cpu]);
769
770 if (total)
771 ratio = avg / total * 100.0;
772
773 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
774
775 fprintf(output, " # ");
776 color_fprintf(output, color, "%6.2f%%", ratio);
777 fprintf(output, " of all L1-icache hits ");
778 }
779
print_dtlb_cache_misses(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)780 static void print_dtlb_cache_misses(int cpu,
781 struct perf_evsel *evsel __maybe_unused,
782 double avg)
783 {
784 double total, ratio = 0.0;
785 const char *color;
786
787 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
788
789 if (total)
790 ratio = avg / total * 100.0;
791
792 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
793
794 fprintf(output, " # ");
795 color_fprintf(output, color, "%6.2f%%", ratio);
796 fprintf(output, " of all dTLB cache hits ");
797 }
798
print_itlb_cache_misses(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)799 static void print_itlb_cache_misses(int cpu,
800 struct perf_evsel *evsel __maybe_unused,
801 double avg)
802 {
803 double total, ratio = 0.0;
804 const char *color;
805
806 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
807
808 if (total)
809 ratio = avg / total * 100.0;
810
811 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
812
813 fprintf(output, " # ");
814 color_fprintf(output, color, "%6.2f%%", ratio);
815 fprintf(output, " of all iTLB cache hits ");
816 }
817
print_ll_cache_misses(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)818 static void print_ll_cache_misses(int cpu,
819 struct perf_evsel *evsel __maybe_unused,
820 double avg)
821 {
822 double total, ratio = 0.0;
823 const char *color;
824
825 total = avg_stats(&runtime_ll_cache_stats[cpu]);
826
827 if (total)
828 ratio = avg / total * 100.0;
829
830 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
831
832 fprintf(output, " # ");
833 color_fprintf(output, color, "%6.2f%%", ratio);
834 fprintf(output, " of all LL-cache hits ");
835 }
836
abs_printout(int cpu,int nr,struct perf_evsel * evsel,double avg)837 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
838 {
839 double total, ratio = 0.0;
840 const char *fmt;
841
842 if (csv_output)
843 fmt = "%.0f%s%s";
844 else if (big_num)
845 fmt = "%'18.0f%s%-25s";
846 else
847 fmt = "%18.0f%s%-25s";
848
849 aggr_printout(evsel, cpu, nr);
850
851 if (aggr_mode == AGGR_GLOBAL)
852 cpu = 0;
853
854 fprintf(output, fmt, avg, csv_sep, perf_evsel__name(evsel));
855
856 if (evsel->cgrp)
857 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
858
859 if (csv_output || interval)
860 return;
861
862 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
863 total = avg_stats(&runtime_cycles_stats[cpu]);
864 if (total)
865 ratio = avg / total;
866
867 fprintf(output, " # %5.2f insns per cycle ", ratio);
868
869 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
870 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
871
872 if (total && avg) {
873 ratio = total / avg;
874 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
875 }
876
877 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
878 runtime_branches_stats[cpu].n != 0) {
879 print_branch_misses(cpu, evsel, avg);
880 } else if (
881 evsel->attr.type == PERF_TYPE_HW_CACHE &&
882 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
883 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
884 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
885 runtime_l1_dcache_stats[cpu].n != 0) {
886 print_l1_dcache_misses(cpu, evsel, avg);
887 } else if (
888 evsel->attr.type == PERF_TYPE_HW_CACHE &&
889 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
890 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
891 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
892 runtime_l1_icache_stats[cpu].n != 0) {
893 print_l1_icache_misses(cpu, evsel, avg);
894 } else if (
895 evsel->attr.type == PERF_TYPE_HW_CACHE &&
896 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
897 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
898 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
899 runtime_dtlb_cache_stats[cpu].n != 0) {
900 print_dtlb_cache_misses(cpu, evsel, avg);
901 } else if (
902 evsel->attr.type == PERF_TYPE_HW_CACHE &&
903 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
904 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
905 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
906 runtime_itlb_cache_stats[cpu].n != 0) {
907 print_itlb_cache_misses(cpu, evsel, avg);
908 } else if (
909 evsel->attr.type == PERF_TYPE_HW_CACHE &&
910 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
911 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
912 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
913 runtime_ll_cache_stats[cpu].n != 0) {
914 print_ll_cache_misses(cpu, evsel, avg);
915 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
916 runtime_cacherefs_stats[cpu].n != 0) {
917 total = avg_stats(&runtime_cacherefs_stats[cpu]);
918
919 if (total)
920 ratio = avg * 100 / total;
921
922 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
923
924 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
925 print_stalled_cycles_frontend(cpu, evsel, avg);
926 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
927 print_stalled_cycles_backend(cpu, evsel, avg);
928 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
929 total = avg_stats(&runtime_nsecs_stats[cpu]);
930
931 if (total)
932 ratio = 1.0 * avg / total;
933
934 fprintf(output, " # %8.3f GHz ", ratio);
935 } else if (runtime_nsecs_stats[cpu].n != 0) {
936 char unit = 'M';
937
938 total = avg_stats(&runtime_nsecs_stats[cpu]);
939
940 if (total)
941 ratio = 1000.0 * avg / total;
942 if (ratio < 0.001) {
943 ratio *= 1000;
944 unit = 'K';
945 }
946
947 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
948 } else {
949 fprintf(output, " ");
950 }
951 }
952
print_aggr(char * prefix)953 static void print_aggr(char *prefix)
954 {
955 struct perf_evsel *counter;
956 int cpu, cpu2, s, s2, id, nr;
957 u64 ena, run, val;
958
959 if (!(aggr_map || aggr_get_id))
960 return;
961
962 for (s = 0; s < aggr_map->nr; s++) {
963 id = aggr_map->map[s];
964 list_for_each_entry(counter, &evsel_list->entries, node) {
965 val = ena = run = 0;
966 nr = 0;
967 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
968 cpu2 = perf_evsel__cpus(counter)->map[cpu];
969 s2 = aggr_get_id(evsel_list->cpus, cpu2);
970 if (s2 != id)
971 continue;
972 val += counter->counts->cpu[cpu].val;
973 ena += counter->counts->cpu[cpu].ena;
974 run += counter->counts->cpu[cpu].run;
975 nr++;
976 }
977 if (prefix)
978 fprintf(output, "%s", prefix);
979
980 if (run == 0 || ena == 0) {
981 aggr_printout(counter, id, nr);
982
983 fprintf(output, "%*s%s%*s",
984 csv_output ? 0 : 18,
985 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
986 csv_sep,
987 csv_output ? 0 : -24,
988 perf_evsel__name(counter));
989
990 if (counter->cgrp)
991 fprintf(output, "%s%s",
992 csv_sep, counter->cgrp->name);
993
994 fputc('\n', output);
995 continue;
996 }
997
998 if (nsec_counter(counter))
999 nsec_printout(id, nr, counter, val);
1000 else
1001 abs_printout(id, nr, counter, val);
1002
1003 if (!csv_output) {
1004 print_noise(counter, 1.0);
1005
1006 if (run != ena)
1007 fprintf(output, " (%.2f%%)",
1008 100.0 * run / ena);
1009 }
1010 fputc('\n', output);
1011 }
1012 }
1013 }
1014
1015 /*
1016 * Print out the results of a single counter:
1017 * aggregated counts in system-wide mode
1018 */
print_counter_aggr(struct perf_evsel * counter,char * prefix)1019 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1020 {
1021 struct perf_stat *ps = counter->priv;
1022 double avg = avg_stats(&ps->res_stats[0]);
1023 int scaled = counter->counts->scaled;
1024
1025 if (prefix)
1026 fprintf(output, "%s", prefix);
1027
1028 if (scaled == -1) {
1029 fprintf(output, "%*s%s%*s",
1030 csv_output ? 0 : 18,
1031 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1032 csv_sep,
1033 csv_output ? 0 : -24,
1034 perf_evsel__name(counter));
1035
1036 if (counter->cgrp)
1037 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1038
1039 fputc('\n', output);
1040 return;
1041 }
1042
1043 if (nsec_counter(counter))
1044 nsec_printout(-1, 0, counter, avg);
1045 else
1046 abs_printout(-1, 0, counter, avg);
1047
1048 print_noise(counter, avg);
1049
1050 if (csv_output) {
1051 fputc('\n', output);
1052 return;
1053 }
1054
1055 if (scaled) {
1056 double avg_enabled, avg_running;
1057
1058 avg_enabled = avg_stats(&ps->res_stats[1]);
1059 avg_running = avg_stats(&ps->res_stats[2]);
1060
1061 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1062 }
1063 fprintf(output, "\n");
1064 }
1065
1066 /*
1067 * Print out the results of a single counter:
1068 * does not use aggregated count in system-wide
1069 */
print_counter(struct perf_evsel * counter,char * prefix)1070 static void print_counter(struct perf_evsel *counter, char *prefix)
1071 {
1072 u64 ena, run, val;
1073 int cpu;
1074
1075 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1076 val = counter->counts->cpu[cpu].val;
1077 ena = counter->counts->cpu[cpu].ena;
1078 run = counter->counts->cpu[cpu].run;
1079
1080 if (prefix)
1081 fprintf(output, "%s", prefix);
1082
1083 if (run == 0 || ena == 0) {
1084 fprintf(output, "CPU%*d%s%*s%s%*s",
1085 csv_output ? 0 : -4,
1086 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1087 csv_output ? 0 : 18,
1088 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1089 csv_sep,
1090 csv_output ? 0 : -24,
1091 perf_evsel__name(counter));
1092
1093 if (counter->cgrp)
1094 fprintf(output, "%s%s",
1095 csv_sep, counter->cgrp->name);
1096
1097 fputc('\n', output);
1098 continue;
1099 }
1100
1101 if (nsec_counter(counter))
1102 nsec_printout(cpu, 0, counter, val);
1103 else
1104 abs_printout(cpu, 0, counter, val);
1105
1106 if (!csv_output) {
1107 print_noise(counter, 1.0);
1108
1109 if (run != ena)
1110 fprintf(output, " (%.2f%%)",
1111 100.0 * run / ena);
1112 }
1113 fputc('\n', output);
1114 }
1115 }
1116
print_stat(int argc,const char ** argv)1117 static void print_stat(int argc, const char **argv)
1118 {
1119 struct perf_evsel *counter;
1120 int i;
1121
1122 fflush(stdout);
1123
1124 if (!csv_output) {
1125 fprintf(output, "\n");
1126 fprintf(output, " Performance counter stats for ");
1127 if (!perf_target__has_task(&target)) {
1128 fprintf(output, "\'%s", argv[0]);
1129 for (i = 1; i < argc; i++)
1130 fprintf(output, " %s", argv[i]);
1131 } else if (target.pid)
1132 fprintf(output, "process id \'%s", target.pid);
1133 else
1134 fprintf(output, "thread id \'%s", target.tid);
1135
1136 fprintf(output, "\'");
1137 if (run_count > 1)
1138 fprintf(output, " (%d runs)", run_count);
1139 fprintf(output, ":\n\n");
1140 }
1141
1142 switch (aggr_mode) {
1143 case AGGR_CORE:
1144 case AGGR_SOCKET:
1145 print_aggr(NULL);
1146 break;
1147 case AGGR_GLOBAL:
1148 list_for_each_entry(counter, &evsel_list->entries, node)
1149 print_counter_aggr(counter, NULL);
1150 break;
1151 case AGGR_NONE:
1152 list_for_each_entry(counter, &evsel_list->entries, node)
1153 print_counter(counter, NULL);
1154 break;
1155 default:
1156 break;
1157 }
1158
1159 if (!csv_output) {
1160 if (!null_run)
1161 fprintf(output, "\n");
1162 fprintf(output, " %17.9f seconds time elapsed",
1163 avg_stats(&walltime_nsecs_stats)/1e9);
1164 if (run_count > 1) {
1165 fprintf(output, " ");
1166 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1167 avg_stats(&walltime_nsecs_stats));
1168 }
1169 fprintf(output, "\n\n");
1170 }
1171 }
1172
1173 static volatile int signr = -1;
1174
skip_signal(int signo)1175 static void skip_signal(int signo)
1176 {
1177 if ((child_pid == -1) || interval)
1178 done = 1;
1179
1180 signr = signo;
1181 /*
1182 * render child_pid harmless
1183 * won't send SIGTERM to a random
1184 * process in case of race condition
1185 * and fast PID recycling
1186 */
1187 child_pid = -1;
1188 }
1189
sig_atexit(void)1190 static void sig_atexit(void)
1191 {
1192 sigset_t set, oset;
1193
1194 /*
1195 * avoid race condition with SIGCHLD handler
1196 * in skip_signal() which is modifying child_pid
1197 * goal is to avoid send SIGTERM to a random
1198 * process
1199 */
1200 sigemptyset(&set);
1201 sigaddset(&set, SIGCHLD);
1202 sigprocmask(SIG_BLOCK, &set, &oset);
1203
1204 if (child_pid != -1)
1205 kill(child_pid, SIGTERM);
1206
1207 sigprocmask(SIG_SETMASK, &oset, NULL);
1208
1209 if (signr == -1)
1210 return;
1211
1212 signal(signr, SIG_DFL);
1213 kill(getpid(), signr);
1214 }
1215
stat__set_big_num(const struct option * opt __maybe_unused,const char * s __maybe_unused,int unset)1216 static int stat__set_big_num(const struct option *opt __maybe_unused,
1217 const char *s __maybe_unused, int unset)
1218 {
1219 big_num_opt = unset ? 0 : 1;
1220 return 0;
1221 }
1222
perf_stat_init_aggr_mode(void)1223 static int perf_stat_init_aggr_mode(void)
1224 {
1225 switch (aggr_mode) {
1226 case AGGR_SOCKET:
1227 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1228 perror("cannot build socket map");
1229 return -1;
1230 }
1231 aggr_get_id = cpu_map__get_socket;
1232 break;
1233 case AGGR_CORE:
1234 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1235 perror("cannot build core map");
1236 return -1;
1237 }
1238 aggr_get_id = cpu_map__get_core;
1239 break;
1240 case AGGR_NONE:
1241 case AGGR_GLOBAL:
1242 default:
1243 break;
1244 }
1245 return 0;
1246 }
1247
1248
1249 /*
1250 * Add default attributes, if there were no attributes specified or
1251 * if -d/--detailed, -d -d or -d -d -d is used:
1252 */
add_default_attributes(void)1253 static int add_default_attributes(void)
1254 {
1255 struct perf_event_attr default_attrs[] = {
1256
1257 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1258 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1259 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1260 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1261
1262 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1263 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1264 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1265 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1266 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1267 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1268
1269 };
1270
1271 /*
1272 * Detailed stats (-d), covering the L1 and last level data caches:
1273 */
1274 struct perf_event_attr detailed_attrs[] = {
1275
1276 { .type = PERF_TYPE_HW_CACHE,
1277 .config =
1278 PERF_COUNT_HW_CACHE_L1D << 0 |
1279 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1280 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1281
1282 { .type = PERF_TYPE_HW_CACHE,
1283 .config =
1284 PERF_COUNT_HW_CACHE_L1D << 0 |
1285 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1286 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1287
1288 { .type = PERF_TYPE_HW_CACHE,
1289 .config =
1290 PERF_COUNT_HW_CACHE_LL << 0 |
1291 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1292 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1293
1294 { .type = PERF_TYPE_HW_CACHE,
1295 .config =
1296 PERF_COUNT_HW_CACHE_LL << 0 |
1297 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1298 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1299 };
1300
1301 /*
1302 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1303 */
1304 struct perf_event_attr very_detailed_attrs[] = {
1305
1306 { .type = PERF_TYPE_HW_CACHE,
1307 .config =
1308 PERF_COUNT_HW_CACHE_L1I << 0 |
1309 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1310 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1311
1312 { .type = PERF_TYPE_HW_CACHE,
1313 .config =
1314 PERF_COUNT_HW_CACHE_L1I << 0 |
1315 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1316 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1317
1318 { .type = PERF_TYPE_HW_CACHE,
1319 .config =
1320 PERF_COUNT_HW_CACHE_DTLB << 0 |
1321 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1322 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1323
1324 { .type = PERF_TYPE_HW_CACHE,
1325 .config =
1326 PERF_COUNT_HW_CACHE_DTLB << 0 |
1327 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1328 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1329
1330 { .type = PERF_TYPE_HW_CACHE,
1331 .config =
1332 PERF_COUNT_HW_CACHE_ITLB << 0 |
1333 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1334 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1335
1336 { .type = PERF_TYPE_HW_CACHE,
1337 .config =
1338 PERF_COUNT_HW_CACHE_ITLB << 0 |
1339 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1340 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1341
1342 };
1343
1344 /*
1345 * Very, very detailed stats (-d -d -d), adding prefetch events:
1346 */
1347 struct perf_event_attr very_very_detailed_attrs[] = {
1348
1349 { .type = PERF_TYPE_HW_CACHE,
1350 .config =
1351 PERF_COUNT_HW_CACHE_L1D << 0 |
1352 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1353 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1354
1355 { .type = PERF_TYPE_HW_CACHE,
1356 .config =
1357 PERF_COUNT_HW_CACHE_L1D << 0 |
1358 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1359 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1360 };
1361
1362 /* Set attrs if no event is selected and !null_run: */
1363 if (null_run)
1364 return 0;
1365
1366 if (!evsel_list->nr_entries) {
1367 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1368 return -1;
1369 }
1370
1371 /* Detailed events get appended to the event list: */
1372
1373 if (detailed_run < 1)
1374 return 0;
1375
1376 /* Append detailed run extra attributes: */
1377 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1378 return -1;
1379
1380 if (detailed_run < 2)
1381 return 0;
1382
1383 /* Append very detailed run extra attributes: */
1384 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1385 return -1;
1386
1387 if (detailed_run < 3)
1388 return 0;
1389
1390 /* Append very, very detailed run extra attributes: */
1391 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1392 }
1393
cmd_stat(int argc,const char ** argv,const char * prefix __maybe_unused)1394 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1395 {
1396 #ifndef __APPLE__
1397 bool append_file = false;
1398 int output_fd = 0;
1399 const char *output_name = NULL;
1400 const struct option options[] = {
1401 OPT_CALLBACK('e', "event", &evsel_list, "event",
1402 "event selector. use 'perf list' to list available events",
1403 parse_events_option),
1404 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1405 "event filter", parse_filter),
1406 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1407 "child tasks do not inherit counters"),
1408 OPT_STRING('p', "pid", &target.pid, "pid",
1409 "stat events on existing process id"),
1410 OPT_STRING('t', "tid", &target.tid, "tid",
1411 "stat events on existing thread id"),
1412 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1413 "system-wide collection from all CPUs"),
1414 OPT_BOOLEAN('g', "group", &group,
1415 "put the counters into a counter group"),
1416 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1417 OPT_INCR('v', "verbose", &verbose,
1418 "be more verbose (show counter open errors, etc)"),
1419 OPT_INTEGER('r', "repeat", &run_count,
1420 "repeat command and print average + stddev (max: 100, forever: 0)"),
1421 OPT_BOOLEAN('n', "null", &null_run,
1422 "null run - dont start any counters"),
1423 OPT_INCR('d', "detailed", &detailed_run,
1424 "detailed run - start a lot of events"),
1425 OPT_BOOLEAN('S', "sync", &sync_run,
1426 "call sync() before starting a run"),
1427 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1428 "print large numbers with thousands\' separators",
1429 stat__set_big_num),
1430 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1431 "list of cpus to monitor in system-wide"),
1432 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1433 "disable CPU count aggregation", AGGR_NONE),
1434 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1435 "print counts with custom separator"),
1436 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1437 "monitor event in cgroup name only", parse_cgroups),
1438 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1439 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1440 OPT_INTEGER(0, "log-fd", &output_fd,
1441 "log output to fd, instead of stderr"),
1442 OPT_STRING(0, "pre", &pre_cmd, "command",
1443 "command to run prior to the measured command"),
1444 OPT_STRING(0, "post", &post_cmd, "command",
1445 "command to run after to the measured command"),
1446 OPT_UINTEGER('I', "interval-print", &interval,
1447 "print counts at regular interval in ms (>= 100)"),
1448 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1449 "aggregate counts per processor socket", AGGR_SOCKET),
1450 OPT_SET_UINT(0, "per-core", &aggr_mode,
1451 "aggregate counts per physical processor core", AGGR_CORE),
1452 OPT_UINTEGER('D', "delay", &initial_delay,
1453 "ms to wait before starting measurement after program start"),
1454 OPT_END()
1455 };
1456 const char * const stat_usage[] = {
1457 "perf stat [<options>] [<command>]",
1458 NULL
1459 };
1460 int status = -ENOMEM, run_idx;
1461 const char *mode;
1462
1463 setlocale(LC_ALL, "");
1464
1465 evsel_list = perf_evlist__new();
1466 if (evsel_list == NULL)
1467 return -ENOMEM;
1468
1469 argc = parse_options(argc, argv, options, stat_usage,
1470 PARSE_OPT_STOP_AT_NON_OPTION);
1471
1472 output = stderr;
1473 if (output_name && strcmp(output_name, "-"))
1474 output = NULL;
1475
1476 if (output_name && output_fd) {
1477 fprintf(stderr, "cannot use both --output and --log-fd\n");
1478 usage_with_options(stat_usage, options);
1479 }
1480
1481 if (output_fd < 0) {
1482 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1483 usage_with_options(stat_usage, options);
1484 }
1485
1486 if (!output) {
1487 struct timespec tm;
1488 mode = append_file ? "a" : "w";
1489
1490 output = fopen(output_name, mode);
1491 if (!output) {
1492 perror("failed to create output file");
1493 return -1;
1494 }
1495 clock_gettime(CLOCK_REALTIME, &tm);
1496 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1497 } else if (output_fd > 0) {
1498 mode = append_file ? "a" : "w";
1499 output = fdopen(output_fd, mode);
1500 if (!output) {
1501 perror("Failed opening logfd");
1502 return -errno;
1503 }
1504 }
1505
1506 if (csv_sep) {
1507 csv_output = true;
1508 if (!strcmp(csv_sep, "\\t"))
1509 csv_sep = "\t";
1510 } else
1511 csv_sep = DEFAULT_SEPARATOR;
1512
1513 /*
1514 * let the spreadsheet do the pretty-printing
1515 */
1516 if (csv_output) {
1517 /* User explicitly passed -B? */
1518 if (big_num_opt == 1) {
1519 fprintf(stderr, "-B option not supported with -x\n");
1520 usage_with_options(stat_usage, options);
1521 } else /* Nope, so disable big number formatting */
1522 big_num = false;
1523 } else if (big_num_opt == 0) /* User passed --no-big-num */
1524 big_num = false;
1525
1526 if (!argc && !perf_target__has_task(&target))
1527 usage_with_options(stat_usage, options);
1528 if (run_count < 0) {
1529 usage_with_options(stat_usage, options);
1530 } else if (run_count == 0) {
1531 forever = true;
1532 run_count = 1;
1533 }
1534
1535 /* no_aggr, cgroup are for system-wide only */
1536 if ((aggr_mode != AGGR_GLOBAL || nr_cgroups)
1537 && !perf_target__has_cpu(&target)) {
1538 fprintf(stderr, "both cgroup and no-aggregation "
1539 "modes only available in system-wide mode\n");
1540
1541 usage_with_options(stat_usage, options);
1542 return -1;
1543 }
1544
1545 if (add_default_attributes())
1546 goto out;
1547
1548 perf_target__validate(&target);
1549
1550 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1551 if (perf_target__has_task(&target))
1552 pr_err("Problems finding threads of monitor\n");
1553 if (perf_target__has_cpu(&target))
1554 perror("failed to parse CPUs map");
1555
1556 usage_with_options(stat_usage, options);
1557 return -1;
1558 }
1559 if (interval && interval < 100) {
1560 pr_err("print interval must be >= 100ms\n");
1561 usage_with_options(stat_usage, options);
1562 return -1;
1563 }
1564
1565 if (perf_evlist__alloc_stats(evsel_list, interval))
1566 goto out_free_maps;
1567
1568 if (perf_stat_init_aggr_mode())
1569 goto out;
1570
1571 /*
1572 * We dont want to block the signals - that would cause
1573 * child tasks to inherit that and Ctrl-C would not work.
1574 * What we want is for Ctrl-C to work in the exec()-ed
1575 * task, but being ignored by perf stat itself:
1576 */
1577 atexit(sig_atexit);
1578 if (!forever)
1579 signal(SIGINT, skip_signal);
1580 signal(SIGCHLD, skip_signal);
1581 signal(SIGALRM, skip_signal);
1582 signal(SIGABRT, skip_signal);
1583
1584 status = 0;
1585 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1586 if (run_count != 1 && verbose)
1587 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1588 run_idx + 1);
1589
1590 status = run_perf_stat(argc, argv);
1591 if (forever && status != -1) {
1592 print_stat(argc, argv);
1593 perf_stat__reset_stats(evsel_list);
1594 }
1595 }
1596
1597 if (!forever && status != -1 && !interval)
1598 print_stat(argc, argv);
1599
1600 perf_evlist__free_stats(evsel_list);
1601 out_free_maps:
1602 perf_evlist__delete_maps(evsel_list);
1603 out:
1604 perf_evlist__delete(evsel_list);
1605 return status;
1606 #else
1607 perror("cmd_stat not supported on MacOS");
1608 return -1;
1609 #endif
1610 }
1611