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/cgroup.h"
47 #include "util/util.h"
48 #include "util/parse-options.h"
49 #include "util/parse-events.h"
50 #include "util/pmu.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/color.h"
56 #include "util/stat.h"
57 #include "util/header.h"
58 #include "util/cpumap.h"
59 #include "util/thread.h"
60 #include "util/thread_map.h"
61
62 #include <stdlib.h>
63 #include <sys/prctl.h>
64 #include <locale.h>
65
66 #define DEFAULT_SEPARATOR " "
67 #define CNTR_NOT_SUPPORTED "<not supported>"
68 #define CNTR_NOT_COUNTED "<not counted>"
69
70 static void print_stat(int argc, const char **argv);
71 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
72 static void print_counter(struct perf_evsel *counter, char *prefix);
73 static void print_aggr(char *prefix);
74
75 /* Default events used for perf stat -T */
76 static const char * const transaction_attrs[] = {
77 "task-clock",
78 "{"
79 "instructions,"
80 "cycles,"
81 "cpu/cycles-t/,"
82 "cpu/tx-start/,"
83 "cpu/el-start/,"
84 "cpu/cycles-ct/"
85 "}"
86 };
87
88 /* More limited version when the CPU does not have all events. */
89 static const char * const transaction_limited_attrs[] = {
90 "task-clock",
91 "{"
92 "instructions,"
93 "cycles,"
94 "cpu/cycles-t/,"
95 "cpu/tx-start/"
96 "}"
97 };
98
99 /* must match transaction_attrs and the beginning limited_attrs */
100 enum {
101 T_TASK_CLOCK,
102 T_INSTRUCTIONS,
103 T_CYCLES,
104 T_CYCLES_IN_TX,
105 T_TRANSACTION_START,
106 T_ELISION_START,
107 T_CYCLES_IN_TX_CP,
108 };
109
110 static struct perf_evlist *evsel_list;
111
112 static struct target target = {
113 .uid = UINT_MAX,
114 };
115
116 enum aggr_mode {
117 AGGR_NONE,
118 AGGR_GLOBAL,
119 AGGR_SOCKET,
120 AGGR_CORE,
121 };
122
123 static int run_count = 1;
124 static bool no_inherit = false;
125 static bool scale = true;
126 static enum aggr_mode aggr_mode = AGGR_GLOBAL;
127 static volatile pid_t child_pid = -1;
128 static bool null_run = false;
129 static int detailed_run = 0;
130 static bool transaction_run;
131 static bool big_num = true;
132 static int big_num_opt = -1;
133 static const char *csv_sep = NULL;
134 static bool csv_output = false;
135 static bool group = false;
136 static FILE *output = NULL;
137 static const char *pre_cmd = NULL;
138 static const char *post_cmd = NULL;
139 static bool sync_run = false;
140 static unsigned int interval = 0;
141 static unsigned int initial_delay = 0;
142 static unsigned int unit_width = 4; /* strlen("unit") */
143 static bool forever = false;
144 static struct timespec ref_time;
145 static struct cpu_map *aggr_map;
146 static int (*aggr_get_id)(struct cpu_map *m, int cpu);
147
148 static volatile int done = 0;
149
150 struct perf_stat {
151 struct stats res_stats[3];
152 };
153
diff_timespec(struct timespec * r,struct timespec * a,struct timespec * b)154 static inline void diff_timespec(struct timespec *r, struct timespec *a,
155 struct timespec *b)
156 {
157 r->tv_sec = a->tv_sec - b->tv_sec;
158 if (a->tv_nsec < b->tv_nsec) {
159 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
160 r->tv_sec--;
161 } else {
162 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
163 }
164 }
165
perf_evsel__cpus(struct perf_evsel * evsel)166 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
167 {
168 return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
169 }
170
perf_evsel__nr_cpus(struct perf_evsel * evsel)171 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
172 {
173 return perf_evsel__cpus(evsel)->nr;
174 }
175
perf_evsel__reset_stat_priv(struct perf_evsel * evsel)176 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
177 {
178 int i;
179 struct perf_stat *ps = evsel->priv;
180
181 for (i = 0; i < 3; i++)
182 init_stats(&ps->res_stats[i]);
183 }
184
perf_evsel__alloc_stat_priv(struct perf_evsel * evsel)185 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
186 {
187 evsel->priv = zalloc(sizeof(struct perf_stat));
188 if (evsel->priv == NULL)
189 return -ENOMEM;
190 perf_evsel__reset_stat_priv(evsel);
191 return 0;
192 }
193
perf_evsel__free_stat_priv(struct perf_evsel * evsel)194 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
195 {
196 zfree(&evsel->priv);
197 }
198
perf_evsel__alloc_prev_raw_counts(struct perf_evsel * evsel)199 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
200 {
201 void *addr;
202 size_t sz;
203
204 sz = sizeof(*evsel->counts) +
205 (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
206
207 addr = zalloc(sz);
208 if (!addr)
209 return -ENOMEM;
210
211 evsel->prev_raw_counts = addr;
212
213 return 0;
214 }
215
perf_evsel__free_prev_raw_counts(struct perf_evsel * evsel)216 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
217 {
218 zfree(&evsel->prev_raw_counts);
219 }
220
perf_evlist__free_stats(struct perf_evlist * evlist)221 static void perf_evlist__free_stats(struct perf_evlist *evlist)
222 {
223 struct perf_evsel *evsel;
224
225 evlist__for_each(evlist, evsel) {
226 perf_evsel__free_stat_priv(evsel);
227 perf_evsel__free_counts(evsel);
228 perf_evsel__free_prev_raw_counts(evsel);
229 }
230 }
231
perf_evlist__alloc_stats(struct perf_evlist * evlist,bool alloc_raw)232 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
233 {
234 struct perf_evsel *evsel;
235
236 evlist__for_each(evlist, evsel) {
237 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
238 perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
239 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
240 goto out_free;
241 }
242
243 return 0;
244
245 out_free:
246 perf_evlist__free_stats(evlist);
247 return -1;
248 }
249
250 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
251 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
252 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
253 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
254 static struct stats runtime_branches_stats[MAX_NR_CPUS];
255 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
256 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
257 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
258 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
259 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
260 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
261 static struct stats runtime_cycles_in_tx_stats[MAX_NR_CPUS];
262 static struct stats walltime_nsecs_stats;
263 static struct stats runtime_transaction_stats[MAX_NR_CPUS];
264 static struct stats runtime_elision_stats[MAX_NR_CPUS];
265
perf_stat__reset_stats(struct perf_evlist * evlist)266 static void perf_stat__reset_stats(struct perf_evlist *evlist)
267 {
268 struct perf_evsel *evsel;
269
270 evlist__for_each(evlist, evsel) {
271 perf_evsel__reset_stat_priv(evsel);
272 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
273 }
274
275 memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
276 memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
277 memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
278 memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
279 memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
280 memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
281 memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
282 memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
283 memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
284 memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
285 memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
286 memset(runtime_cycles_in_tx_stats, 0,
287 sizeof(runtime_cycles_in_tx_stats));
288 memset(runtime_transaction_stats, 0,
289 sizeof(runtime_transaction_stats));
290 memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
291 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
292 }
293
create_perf_stat_counter(struct perf_evsel * evsel)294 static int create_perf_stat_counter(struct perf_evsel *evsel)
295 {
296 struct perf_event_attr *attr = &evsel->attr;
297
298 if (scale)
299 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
300 PERF_FORMAT_TOTAL_TIME_RUNNING;
301
302 attr->inherit = !no_inherit;
303
304 if (target__has_cpu(&target))
305 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
306
307 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
308 attr->disabled = 1;
309 if (!initial_delay)
310 attr->enable_on_exec = 1;
311 }
312
313 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
314 }
315
316 /*
317 * Does the counter have nsecs as a unit?
318 */
nsec_counter(struct perf_evsel * evsel)319 static inline int nsec_counter(struct perf_evsel *evsel)
320 {
321 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
322 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
323 return 1;
324
325 return 0;
326 }
327
nth_evsel(int n)328 static struct perf_evsel *nth_evsel(int n)
329 {
330 static struct perf_evsel **array;
331 static int array_len;
332 struct perf_evsel *ev;
333 int j;
334
335 /* Assumes this only called when evsel_list does not change anymore. */
336 if (!array) {
337 evlist__for_each(evsel_list, ev)
338 array_len++;
339 array = malloc(array_len * sizeof(void *));
340 if (!array)
341 exit(ENOMEM);
342 j = 0;
343 evlist__for_each(evsel_list, ev)
344 array[j++] = ev;
345 }
346 if (n < array_len)
347 return array[n];
348 return NULL;
349 }
350
351 /*
352 * Update various tracking values we maintain to print
353 * more semantic information such as miss/hit ratios,
354 * instruction rates, etc:
355 */
update_shadow_stats(struct perf_evsel * counter,u64 * count)356 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
357 {
358 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
359 update_stats(&runtime_nsecs_stats[0], count[0]);
360 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
361 update_stats(&runtime_cycles_stats[0], count[0]);
362 else if (transaction_run &&
363 perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX)))
364 update_stats(&runtime_cycles_in_tx_stats[0], count[0]);
365 else if (transaction_run &&
366 perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START)))
367 update_stats(&runtime_transaction_stats[0], count[0]);
368 else if (transaction_run &&
369 perf_evsel__cmp(counter, nth_evsel(T_ELISION_START)))
370 update_stats(&runtime_elision_stats[0], count[0]);
371 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
372 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
373 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
374 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
375 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
376 update_stats(&runtime_branches_stats[0], count[0]);
377 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
378 update_stats(&runtime_cacherefs_stats[0], count[0]);
379 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
380 update_stats(&runtime_l1_dcache_stats[0], count[0]);
381 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
382 update_stats(&runtime_l1_icache_stats[0], count[0]);
383 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
384 update_stats(&runtime_ll_cache_stats[0], count[0]);
385 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
386 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
387 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
388 update_stats(&runtime_itlb_cache_stats[0], count[0]);
389 }
390
391 /*
392 * Read out the results of a single counter:
393 * aggregate counts across CPUs in system-wide mode
394 */
read_counter_aggr(struct perf_evsel * counter)395 static int read_counter_aggr(struct perf_evsel *counter)
396 {
397 struct perf_stat *ps = counter->priv;
398 u64 *count = counter->counts->aggr.values;
399 int i;
400
401 if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
402 thread_map__nr(evsel_list->threads), scale) < 0)
403 return -1;
404
405 for (i = 0; i < 3; i++)
406 update_stats(&ps->res_stats[i], count[i]);
407
408 if (verbose) {
409 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
410 perf_evsel__name(counter), count[0], count[1], count[2]);
411 }
412
413 /*
414 * Save the full runtime - to allow normalization during printout:
415 */
416 update_shadow_stats(counter, count);
417
418 return 0;
419 }
420
421 /*
422 * Read out the results of a single counter:
423 * do not aggregate counts across CPUs in system-wide mode
424 */
read_counter(struct perf_evsel * counter)425 static int read_counter(struct perf_evsel *counter)
426 {
427 u64 *count;
428 int cpu;
429
430 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
431 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
432 return -1;
433
434 count = counter->counts->cpu[cpu].values;
435
436 update_shadow_stats(counter, count);
437 }
438
439 return 0;
440 }
441
print_interval(void)442 static void print_interval(void)
443 {
444 static int num_print_interval;
445 struct perf_evsel *counter;
446 struct perf_stat *ps;
447 struct timespec ts, rs;
448 char prefix[64];
449
450 if (aggr_mode == AGGR_GLOBAL) {
451 evlist__for_each(evsel_list, counter) {
452 ps = counter->priv;
453 memset(ps->res_stats, 0, sizeof(ps->res_stats));
454 read_counter_aggr(counter);
455 }
456 } else {
457 evlist__for_each(evsel_list, counter) {
458 ps = counter->priv;
459 memset(ps->res_stats, 0, sizeof(ps->res_stats));
460 read_counter(counter);
461 }
462 }
463
464 clock_gettime(CLOCK_MONOTONIC, &ts);
465 diff_timespec(&rs, &ts, &ref_time);
466 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
467
468 if (num_print_interval == 0 && !csv_output) {
469 switch (aggr_mode) {
470 case AGGR_SOCKET:
471 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
472 break;
473 case AGGR_CORE:
474 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
475 break;
476 case AGGR_NONE:
477 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
478 break;
479 case AGGR_GLOBAL:
480 default:
481 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
482 }
483 }
484
485 if (++num_print_interval == 25)
486 num_print_interval = 0;
487
488 switch (aggr_mode) {
489 case AGGR_CORE:
490 case AGGR_SOCKET:
491 print_aggr(prefix);
492 break;
493 case AGGR_NONE:
494 evlist__for_each(evsel_list, counter)
495 print_counter(counter, prefix);
496 break;
497 case AGGR_GLOBAL:
498 default:
499 evlist__for_each(evsel_list, counter)
500 print_counter_aggr(counter, prefix);
501 }
502
503 fflush(output);
504 }
505
handle_initial_delay(void)506 static void handle_initial_delay(void)
507 {
508 struct perf_evsel *counter;
509
510 if (initial_delay) {
511 const int ncpus = cpu_map__nr(evsel_list->cpus),
512 nthreads = thread_map__nr(evsel_list->threads);
513
514 usleep(initial_delay * 1000);
515 evlist__for_each(evsel_list, counter)
516 perf_evsel__enable(counter, ncpus, nthreads);
517 }
518 }
519
520 static volatile int workload_exec_errno;
521
522 /*
523 * perf_evlist__prepare_workload will send a SIGUSR1
524 * if the fork fails, since we asked by setting its
525 * want_signal to true.
526 */
workload_exec_failed_signal(int signo __maybe_unused,siginfo_t * info,void * ucontext __maybe_unused)527 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
528 void *ucontext __maybe_unused)
529 {
530 workload_exec_errno = info->si_value.sival_int;
531 }
532
__run_perf_stat(int argc,const char ** argv)533 static int __run_perf_stat(int argc, const char **argv)
534 {
535 char msg[512];
536 unsigned long long t0, t1;
537 struct perf_evsel *counter;
538 struct timespec ts;
539 size_t l;
540 int status = 0;
541 const bool forks = (argc > 0);
542
543 if (interval) {
544 ts.tv_sec = interval / 1000;
545 ts.tv_nsec = (interval % 1000) * 1000000;
546 } else {
547 ts.tv_sec = 1;
548 ts.tv_nsec = 0;
549 }
550
551 if (forks) {
552 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
553 workload_exec_failed_signal) < 0) {
554 perror("failed to prepare workload");
555 return -1;
556 }
557 child_pid = evsel_list->workload.pid;
558 }
559
560 if (group)
561 perf_evlist__set_leader(evsel_list);
562
563 evlist__for_each(evsel_list, counter) {
564 if (create_perf_stat_counter(counter) < 0) {
565 /*
566 * PPC returns ENXIO for HW counters until 2.6.37
567 * (behavior changed with commit b0a873e).
568 */
569 if (errno == EINVAL || errno == ENOSYS ||
570 errno == ENOENT || errno == EOPNOTSUPP ||
571 errno == ENXIO) {
572 if (verbose)
573 ui__warning("%s event is not supported by the kernel.\n",
574 perf_evsel__name(counter));
575 counter->supported = false;
576 continue;
577 }
578
579 perf_evsel__open_strerror(counter, &target,
580 errno, msg, sizeof(msg));
581 ui__error("%s\n", msg);
582
583 if (child_pid != -1)
584 kill(child_pid, SIGTERM);
585
586 return -1;
587 }
588 counter->supported = true;
589
590 l = strlen(counter->unit);
591 if (l > unit_width)
592 unit_width = l;
593 }
594
595 if (perf_evlist__apply_filters(evsel_list)) {
596 error("failed to set filter with %d (%s)\n", errno,
597 strerror_r(errno, msg, sizeof(msg)));
598 return -1;
599 }
600
601 /*
602 * Enable counters and exec the command:
603 */
604 t0 = rdclock();
605 clock_gettime(CLOCK_MONOTONIC, &ref_time);
606
607 if (forks) {
608 perf_evlist__start_workload(evsel_list);
609 handle_initial_delay();
610
611 if (interval) {
612 while (!waitpid(child_pid, &status, WNOHANG)) {
613 nanosleep(&ts, NULL);
614 print_interval();
615 }
616 }
617 wait(&status);
618
619 if (workload_exec_errno) {
620 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
621 pr_err("Workload failed: %s\n", emsg);
622 return -1;
623 }
624
625 if (WIFSIGNALED(status))
626 psignal(WTERMSIG(status), argv[0]);
627 } else {
628 handle_initial_delay();
629 while (!done) {
630 nanosleep(&ts, NULL);
631 if (interval)
632 print_interval();
633 }
634 }
635
636 t1 = rdclock();
637
638 update_stats(&walltime_nsecs_stats, t1 - t0);
639
640 if (aggr_mode == AGGR_GLOBAL) {
641 evlist__for_each(evsel_list, counter) {
642 read_counter_aggr(counter);
643 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
644 thread_map__nr(evsel_list->threads));
645 }
646 } else {
647 evlist__for_each(evsel_list, counter) {
648 read_counter(counter);
649 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
650 }
651 }
652
653 return WEXITSTATUS(status);
654 }
655
run_perf_stat(int argc,const char ** argv)656 static int run_perf_stat(int argc, const char **argv)
657 {
658 int ret;
659
660 if (pre_cmd) {
661 ret = system(pre_cmd);
662 if (ret)
663 return ret;
664 }
665
666 if (sync_run)
667 sync();
668
669 ret = __run_perf_stat(argc, argv);
670 if (ret)
671 return ret;
672
673 if (post_cmd) {
674 ret = system(post_cmd);
675 if (ret)
676 return ret;
677 }
678
679 return ret;
680 }
681
print_noise_pct(double total,double avg)682 static void print_noise_pct(double total, double avg)
683 {
684 double pct = rel_stddev_stats(total, avg);
685
686 if (csv_output)
687 fprintf(output, "%s%.2f%%", csv_sep, pct);
688 else if (pct)
689 fprintf(output, " ( +-%6.2f%% )", pct);
690 }
691
print_noise(struct perf_evsel * evsel,double avg)692 static void print_noise(struct perf_evsel *evsel, double avg)
693 {
694 struct perf_stat *ps;
695
696 if (run_count == 1)
697 return;
698
699 ps = evsel->priv;
700 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
701 }
702
aggr_printout(struct perf_evsel * evsel,int id,int nr)703 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
704 {
705 switch (aggr_mode) {
706 case AGGR_CORE:
707 fprintf(output, "S%d-C%*d%s%*d%s",
708 cpu_map__id_to_socket(id),
709 csv_output ? 0 : -8,
710 cpu_map__id_to_cpu(id),
711 csv_sep,
712 csv_output ? 0 : 4,
713 nr,
714 csv_sep);
715 break;
716 case AGGR_SOCKET:
717 fprintf(output, "S%*d%s%*d%s",
718 csv_output ? 0 : -5,
719 id,
720 csv_sep,
721 csv_output ? 0 : 4,
722 nr,
723 csv_sep);
724 break;
725 case AGGR_NONE:
726 fprintf(output, "CPU%*d%s",
727 csv_output ? 0 : -4,
728 perf_evsel__cpus(evsel)->map[id], csv_sep);
729 break;
730 case AGGR_GLOBAL:
731 default:
732 break;
733 }
734 }
735
nsec_printout(int id,int nr,struct perf_evsel * evsel,double avg)736 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
737 {
738 double msecs = avg / 1e6;
739 const char *fmt_v, *fmt_n;
740 char name[25];
741
742 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
743 fmt_n = csv_output ? "%s" : "%-25s";
744
745 aggr_printout(evsel, id, nr);
746
747 scnprintf(name, sizeof(name), "%s%s",
748 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
749
750 fprintf(output, fmt_v, msecs, csv_sep);
751
752 if (csv_output)
753 fprintf(output, "%s%s", evsel->unit, csv_sep);
754 else
755 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
756
757 fprintf(output, fmt_n, name);
758
759 if (evsel->cgrp)
760 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
761
762 if (csv_output || interval)
763 return;
764
765 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
766 fprintf(output, " # %8.3f CPUs utilized ",
767 avg / avg_stats(&walltime_nsecs_stats));
768 else
769 fprintf(output, " ");
770 }
771
772 /* used for get_ratio_color() */
773 enum grc_type {
774 GRC_STALLED_CYCLES_FE,
775 GRC_STALLED_CYCLES_BE,
776 GRC_CACHE_MISSES,
777 GRC_MAX_NR
778 };
779
get_ratio_color(enum grc_type type,double ratio)780 static const char *get_ratio_color(enum grc_type type, double ratio)
781 {
782 static const double grc_table[GRC_MAX_NR][3] = {
783 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
784 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
785 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
786 };
787 const char *color = PERF_COLOR_NORMAL;
788
789 if (ratio > grc_table[type][0])
790 color = PERF_COLOR_RED;
791 else if (ratio > grc_table[type][1])
792 color = PERF_COLOR_MAGENTA;
793 else if (ratio > grc_table[type][2])
794 color = PERF_COLOR_YELLOW;
795
796 return color;
797 }
798
print_stalled_cycles_frontend(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)799 static void print_stalled_cycles_frontend(int cpu,
800 struct perf_evsel *evsel
801 __maybe_unused, double avg)
802 {
803 double total, ratio = 0.0;
804 const char *color;
805
806 total = avg_stats(&runtime_cycles_stats[cpu]);
807
808 if (total)
809 ratio = avg / total * 100.0;
810
811 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
812
813 fprintf(output, " # ");
814 color_fprintf(output, color, "%6.2f%%", ratio);
815 fprintf(output, " frontend cycles idle ");
816 }
817
print_stalled_cycles_backend(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)818 static void print_stalled_cycles_backend(int cpu,
819 struct perf_evsel *evsel
820 __maybe_unused, double avg)
821 {
822 double total, ratio = 0.0;
823 const char *color;
824
825 total = avg_stats(&runtime_cycles_stats[cpu]);
826
827 if (total)
828 ratio = avg / total * 100.0;
829
830 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
831
832 fprintf(output, " # ");
833 color_fprintf(output, color, "%6.2f%%", ratio);
834 fprintf(output, " backend cycles idle ");
835 }
836
print_branch_misses(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)837 static void print_branch_misses(int cpu,
838 struct perf_evsel *evsel __maybe_unused,
839 double avg)
840 {
841 double total, ratio = 0.0;
842 const char *color;
843
844 total = avg_stats(&runtime_branches_stats[cpu]);
845
846 if (total)
847 ratio = avg / total * 100.0;
848
849 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
850
851 fprintf(output, " # ");
852 color_fprintf(output, color, "%6.2f%%", ratio);
853 fprintf(output, " of all branches ");
854 }
855
print_l1_dcache_misses(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)856 static void print_l1_dcache_misses(int cpu,
857 struct perf_evsel *evsel __maybe_unused,
858 double avg)
859 {
860 double total, ratio = 0.0;
861 const char *color;
862
863 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
864
865 if (total)
866 ratio = avg / total * 100.0;
867
868 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
869
870 fprintf(output, " # ");
871 color_fprintf(output, color, "%6.2f%%", ratio);
872 fprintf(output, " of all L1-dcache hits ");
873 }
874
print_l1_icache_misses(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)875 static void print_l1_icache_misses(int cpu,
876 struct perf_evsel *evsel __maybe_unused,
877 double avg)
878 {
879 double total, ratio = 0.0;
880 const char *color;
881
882 total = avg_stats(&runtime_l1_icache_stats[cpu]);
883
884 if (total)
885 ratio = avg / total * 100.0;
886
887 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
888
889 fprintf(output, " # ");
890 color_fprintf(output, color, "%6.2f%%", ratio);
891 fprintf(output, " of all L1-icache hits ");
892 }
893
print_dtlb_cache_misses(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)894 static void print_dtlb_cache_misses(int cpu,
895 struct perf_evsel *evsel __maybe_unused,
896 double avg)
897 {
898 double total, ratio = 0.0;
899 const char *color;
900
901 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
902
903 if (total)
904 ratio = avg / total * 100.0;
905
906 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
907
908 fprintf(output, " # ");
909 color_fprintf(output, color, "%6.2f%%", ratio);
910 fprintf(output, " of all dTLB cache hits ");
911 }
912
print_itlb_cache_misses(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)913 static void print_itlb_cache_misses(int cpu,
914 struct perf_evsel *evsel __maybe_unused,
915 double avg)
916 {
917 double total, ratio = 0.0;
918 const char *color;
919
920 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
921
922 if (total)
923 ratio = avg / total * 100.0;
924
925 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
926
927 fprintf(output, " # ");
928 color_fprintf(output, color, "%6.2f%%", ratio);
929 fprintf(output, " of all iTLB cache hits ");
930 }
931
print_ll_cache_misses(int cpu,struct perf_evsel * evsel __maybe_unused,double avg)932 static void print_ll_cache_misses(int cpu,
933 struct perf_evsel *evsel __maybe_unused,
934 double avg)
935 {
936 double total, ratio = 0.0;
937 const char *color;
938
939 total = avg_stats(&runtime_ll_cache_stats[cpu]);
940
941 if (total)
942 ratio = avg / total * 100.0;
943
944 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
945
946 fprintf(output, " # ");
947 color_fprintf(output, color, "%6.2f%%", ratio);
948 fprintf(output, " of all LL-cache hits ");
949 }
950
abs_printout(int id,int nr,struct perf_evsel * evsel,double avg)951 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
952 {
953 double total, ratio = 0.0, total2;
954 double sc = evsel->scale;
955 const char *fmt;
956 int cpu = cpu_map__id_to_cpu(id);
957
958 if (csv_output) {
959 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
960 } else {
961 if (big_num)
962 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
963 else
964 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
965 }
966
967 aggr_printout(evsel, id, nr);
968
969 if (aggr_mode == AGGR_GLOBAL)
970 cpu = 0;
971
972 fprintf(output, fmt, avg, csv_sep);
973
974 if (evsel->unit)
975 fprintf(output, "%-*s%s",
976 csv_output ? 0 : unit_width,
977 evsel->unit, csv_sep);
978
979 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
980
981 if (evsel->cgrp)
982 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
983
984 if (csv_output || interval)
985 return;
986
987 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
988 total = avg_stats(&runtime_cycles_stats[cpu]);
989 if (total) {
990 ratio = avg / total;
991 fprintf(output, " # %5.2f insns per cycle ", ratio);
992 }
993 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
994 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
995
996 if (total && avg) {
997 ratio = total / avg;
998 fprintf(output, "\n");
999 if (aggr_mode == AGGR_NONE)
1000 fprintf(output, " ");
1001 fprintf(output, " # %5.2f stalled cycles per insn", ratio);
1002 }
1003
1004 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
1005 runtime_branches_stats[cpu].n != 0) {
1006 print_branch_misses(cpu, evsel, avg);
1007 } else if (
1008 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1009 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
1010 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1011 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1012 runtime_l1_dcache_stats[cpu].n != 0) {
1013 print_l1_dcache_misses(cpu, evsel, avg);
1014 } else if (
1015 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1016 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
1017 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1018 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1019 runtime_l1_icache_stats[cpu].n != 0) {
1020 print_l1_icache_misses(cpu, evsel, avg);
1021 } else if (
1022 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1023 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
1024 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1025 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1026 runtime_dtlb_cache_stats[cpu].n != 0) {
1027 print_dtlb_cache_misses(cpu, evsel, avg);
1028 } else if (
1029 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1030 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
1031 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1032 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1033 runtime_itlb_cache_stats[cpu].n != 0) {
1034 print_itlb_cache_misses(cpu, evsel, avg);
1035 } else if (
1036 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1037 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
1038 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1039 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1040 runtime_ll_cache_stats[cpu].n != 0) {
1041 print_ll_cache_misses(cpu, evsel, avg);
1042 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
1043 runtime_cacherefs_stats[cpu].n != 0) {
1044 total = avg_stats(&runtime_cacherefs_stats[cpu]);
1045
1046 if (total)
1047 ratio = avg * 100 / total;
1048
1049 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
1050
1051 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1052 print_stalled_cycles_frontend(cpu, evsel, avg);
1053 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1054 print_stalled_cycles_backend(cpu, evsel, avg);
1055 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1056 total = avg_stats(&runtime_nsecs_stats[cpu]);
1057
1058 if (total) {
1059 ratio = avg / total;
1060 fprintf(output, " # %8.3f GHz ", ratio);
1061 }
1062 } else if (transaction_run &&
1063 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1064 total = avg_stats(&runtime_cycles_stats[cpu]);
1065 if (total)
1066 fprintf(output,
1067 " # %5.2f%% transactional cycles ",
1068 100.0 * (avg / total));
1069 } else if (transaction_run &&
1070 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1071 total = avg_stats(&runtime_cycles_stats[cpu]);
1072 total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1073 if (total2 < avg)
1074 total2 = avg;
1075 if (total)
1076 fprintf(output,
1077 " # %5.2f%% aborted cycles ",
1078 100.0 * ((total2-avg) / total));
1079 } else if (transaction_run &&
1080 perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1081 avg > 0 &&
1082 runtime_cycles_in_tx_stats[cpu].n != 0) {
1083 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1084
1085 if (total)
1086 ratio = total / avg;
1087
1088 fprintf(output, " # %8.0f cycles / transaction ", ratio);
1089 } else if (transaction_run &&
1090 perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1091 avg > 0 &&
1092 runtime_cycles_in_tx_stats[cpu].n != 0) {
1093 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1094
1095 if (total)
1096 ratio = total / avg;
1097
1098 fprintf(output, " # %8.0f cycles / elision ", ratio);
1099 } else if (runtime_nsecs_stats[cpu].n != 0) {
1100 char unit = 'M';
1101
1102 total = avg_stats(&runtime_nsecs_stats[cpu]);
1103
1104 if (total)
1105 ratio = 1000.0 * avg / total;
1106 if (ratio < 0.001) {
1107 ratio *= 1000;
1108 unit = 'K';
1109 }
1110
1111 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
1112 } else {
1113 fprintf(output, " ");
1114 }
1115 }
1116
print_aggr(char * prefix)1117 static void print_aggr(char *prefix)
1118 {
1119 struct perf_evsel *counter;
1120 int cpu, s, s2, id, nr;
1121 double uval;
1122 u64 ena, run, val;
1123
1124 if (!(aggr_map || aggr_get_id))
1125 return;
1126
1127 for (s = 0; s < aggr_map->nr; s++) {
1128 id = aggr_map->map[s];
1129 evlist__for_each(evsel_list, counter) {
1130 val = ena = run = 0;
1131 nr = 0;
1132 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1133 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1134 if (s2 != id)
1135 continue;
1136 val += counter->counts->cpu[cpu].val;
1137 ena += counter->counts->cpu[cpu].ena;
1138 run += counter->counts->cpu[cpu].run;
1139 nr++;
1140 }
1141 if (prefix)
1142 fprintf(output, "%s", prefix);
1143
1144 if (run == 0 || ena == 0) {
1145 aggr_printout(counter, id, nr);
1146
1147 fprintf(output, "%*s%s",
1148 csv_output ? 0 : 18,
1149 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1150 csv_sep);
1151
1152 fprintf(output, "%-*s%s",
1153 csv_output ? 0 : unit_width,
1154 counter->unit, csv_sep);
1155
1156 fprintf(output, "%*s",
1157 csv_output ? 0 : -25,
1158 perf_evsel__name(counter));
1159
1160 if (counter->cgrp)
1161 fprintf(output, "%s%s",
1162 csv_sep, counter->cgrp->name);
1163
1164 fputc('\n', output);
1165 continue;
1166 }
1167 uval = val * counter->scale;
1168
1169 if (nsec_counter(counter))
1170 nsec_printout(id, nr, counter, uval);
1171 else
1172 abs_printout(id, nr, counter, uval);
1173
1174 if (!csv_output) {
1175 print_noise(counter, 1.0);
1176
1177 if (run != ena)
1178 fprintf(output, " (%.2f%%)",
1179 100.0 * run / ena);
1180 }
1181 fputc('\n', output);
1182 }
1183 }
1184 }
1185
1186 /*
1187 * Print out the results of a single counter:
1188 * aggregated counts in system-wide mode
1189 */
print_counter_aggr(struct perf_evsel * counter,char * prefix)1190 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1191 {
1192 struct perf_stat *ps = counter->priv;
1193 double avg = avg_stats(&ps->res_stats[0]);
1194 int scaled = counter->counts->scaled;
1195 double uval;
1196
1197 if (prefix)
1198 fprintf(output, "%s", prefix);
1199
1200 if (scaled == -1) {
1201 fprintf(output, "%*s%s",
1202 csv_output ? 0 : 18,
1203 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1204 csv_sep);
1205 fprintf(output, "%-*s%s",
1206 csv_output ? 0 : unit_width,
1207 counter->unit, csv_sep);
1208 fprintf(output, "%*s",
1209 csv_output ? 0 : -25,
1210 perf_evsel__name(counter));
1211
1212 if (counter->cgrp)
1213 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1214
1215 fputc('\n', output);
1216 return;
1217 }
1218
1219 uval = avg * counter->scale;
1220
1221 if (nsec_counter(counter))
1222 nsec_printout(-1, 0, counter, uval);
1223 else
1224 abs_printout(-1, 0, counter, uval);
1225
1226 print_noise(counter, avg);
1227
1228 if (csv_output) {
1229 fputc('\n', output);
1230 return;
1231 }
1232
1233 if (scaled) {
1234 double avg_enabled, avg_running;
1235
1236 avg_enabled = avg_stats(&ps->res_stats[1]);
1237 avg_running = avg_stats(&ps->res_stats[2]);
1238
1239 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1240 }
1241 fprintf(output, "\n");
1242 }
1243
1244 /*
1245 * Print out the results of a single counter:
1246 * does not use aggregated count in system-wide
1247 */
print_counter(struct perf_evsel * counter,char * prefix)1248 static void print_counter(struct perf_evsel *counter, char *prefix)
1249 {
1250 u64 ena, run, val;
1251 double uval;
1252 int cpu;
1253
1254 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1255 val = counter->counts->cpu[cpu].val;
1256 ena = counter->counts->cpu[cpu].ena;
1257 run = counter->counts->cpu[cpu].run;
1258
1259 if (prefix)
1260 fprintf(output, "%s", prefix);
1261
1262 if (run == 0 || ena == 0) {
1263 fprintf(output, "CPU%*d%s%*s%s",
1264 csv_output ? 0 : -4,
1265 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1266 csv_output ? 0 : 18,
1267 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1268 csv_sep);
1269
1270 fprintf(output, "%-*s%s",
1271 csv_output ? 0 : unit_width,
1272 counter->unit, csv_sep);
1273
1274 fprintf(output, "%*s",
1275 csv_output ? 0 : -25,
1276 perf_evsel__name(counter));
1277
1278 if (counter->cgrp)
1279 fprintf(output, "%s%s",
1280 csv_sep, counter->cgrp->name);
1281
1282 fputc('\n', output);
1283 continue;
1284 }
1285
1286 uval = val * counter->scale;
1287
1288 if (nsec_counter(counter))
1289 nsec_printout(cpu, 0, counter, uval);
1290 else
1291 abs_printout(cpu, 0, counter, uval);
1292
1293 if (!csv_output) {
1294 print_noise(counter, 1.0);
1295
1296 if (run != ena)
1297 fprintf(output, " (%.2f%%)",
1298 100.0 * run / ena);
1299 }
1300 fputc('\n', output);
1301 }
1302 }
1303
print_stat(int argc,const char ** argv)1304 static void print_stat(int argc, const char **argv)
1305 {
1306 struct perf_evsel *counter;
1307 int i;
1308
1309 fflush(stdout);
1310
1311 if (!csv_output) {
1312 fprintf(output, "\n");
1313 fprintf(output, " Performance counter stats for ");
1314 if (target.system_wide)
1315 fprintf(output, "\'system wide");
1316 else if (target.cpu_list)
1317 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1318 else if (!target__has_task(&target)) {
1319 fprintf(output, "\'%s", argv[0]);
1320 for (i = 1; i < argc; i++)
1321 fprintf(output, " %s", argv[i]);
1322 } else if (target.pid)
1323 fprintf(output, "process id \'%s", target.pid);
1324 else
1325 fprintf(output, "thread id \'%s", target.tid);
1326
1327 fprintf(output, "\'");
1328 if (run_count > 1)
1329 fprintf(output, " (%d runs)", run_count);
1330 fprintf(output, ":\n\n");
1331 }
1332
1333 switch (aggr_mode) {
1334 case AGGR_CORE:
1335 case AGGR_SOCKET:
1336 print_aggr(NULL);
1337 break;
1338 case AGGR_GLOBAL:
1339 evlist__for_each(evsel_list, counter)
1340 print_counter_aggr(counter, NULL);
1341 break;
1342 case AGGR_NONE:
1343 evlist__for_each(evsel_list, counter)
1344 print_counter(counter, NULL);
1345 break;
1346 default:
1347 break;
1348 }
1349
1350 if (!csv_output) {
1351 if (!null_run)
1352 fprintf(output, "\n");
1353 fprintf(output, " %17.9f seconds time elapsed",
1354 avg_stats(&walltime_nsecs_stats)/1e9);
1355 if (run_count > 1) {
1356 fprintf(output, " ");
1357 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1358 avg_stats(&walltime_nsecs_stats));
1359 }
1360 fprintf(output, "\n\n");
1361 }
1362 }
1363
1364 static volatile int signr = -1;
1365
skip_signal(int signo)1366 static void skip_signal(int signo)
1367 {
1368 if ((child_pid == -1) || interval)
1369 done = 1;
1370
1371 signr = signo;
1372 /*
1373 * render child_pid harmless
1374 * won't send SIGTERM to a random
1375 * process in case of race condition
1376 * and fast PID recycling
1377 */
1378 child_pid = -1;
1379 }
1380
sig_atexit(void)1381 static void sig_atexit(void)
1382 {
1383 sigset_t set, oset;
1384
1385 /*
1386 * avoid race condition with SIGCHLD handler
1387 * in skip_signal() which is modifying child_pid
1388 * goal is to avoid send SIGTERM to a random
1389 * process
1390 */
1391 sigemptyset(&set);
1392 sigaddset(&set, SIGCHLD);
1393 sigprocmask(SIG_BLOCK, &set, &oset);
1394
1395 if (child_pid != -1)
1396 kill(child_pid, SIGTERM);
1397
1398 sigprocmask(SIG_SETMASK, &oset, NULL);
1399
1400 if (signr == -1)
1401 return;
1402
1403 signal(signr, SIG_DFL);
1404 kill(getpid(), signr);
1405 }
1406
stat__set_big_num(const struct option * opt __maybe_unused,const char * s __maybe_unused,int unset)1407 static int stat__set_big_num(const struct option *opt __maybe_unused,
1408 const char *s __maybe_unused, int unset)
1409 {
1410 big_num_opt = unset ? 0 : 1;
1411 return 0;
1412 }
1413
perf_stat_init_aggr_mode(void)1414 static int perf_stat_init_aggr_mode(void)
1415 {
1416 switch (aggr_mode) {
1417 case AGGR_SOCKET:
1418 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1419 perror("cannot build socket map");
1420 return -1;
1421 }
1422 aggr_get_id = cpu_map__get_socket;
1423 break;
1424 case AGGR_CORE:
1425 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1426 perror("cannot build core map");
1427 return -1;
1428 }
1429 aggr_get_id = cpu_map__get_core;
1430 break;
1431 case AGGR_NONE:
1432 case AGGR_GLOBAL:
1433 default:
1434 break;
1435 }
1436 return 0;
1437 }
1438
setup_events(const char * const * attrs,unsigned len)1439 static int setup_events(const char * const *attrs, unsigned len)
1440 {
1441 unsigned i;
1442
1443 for (i = 0; i < len; i++) {
1444 if (parse_events(evsel_list, attrs[i]))
1445 return -1;
1446 }
1447 return 0;
1448 }
1449
1450 /*
1451 * Add default attributes, if there were no attributes specified or
1452 * if -d/--detailed, -d -d or -d -d -d is used:
1453 */
add_default_attributes(void)1454 static int add_default_attributes(void)
1455 {
1456 struct perf_event_attr default_attrs[] = {
1457
1458 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1459 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1460 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1461 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1462
1463 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1464 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1465 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1466 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1467 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1468 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1469
1470 };
1471
1472 /*
1473 * Detailed stats (-d), covering the L1 and last level data caches:
1474 */
1475 struct perf_event_attr detailed_attrs[] = {
1476
1477 { .type = PERF_TYPE_HW_CACHE,
1478 .config =
1479 PERF_COUNT_HW_CACHE_L1D << 0 |
1480 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1481 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1482
1483 { .type = PERF_TYPE_HW_CACHE,
1484 .config =
1485 PERF_COUNT_HW_CACHE_L1D << 0 |
1486 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1487 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1488
1489 { .type = PERF_TYPE_HW_CACHE,
1490 .config =
1491 PERF_COUNT_HW_CACHE_LL << 0 |
1492 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1493 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1494
1495 { .type = PERF_TYPE_HW_CACHE,
1496 .config =
1497 PERF_COUNT_HW_CACHE_LL << 0 |
1498 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1499 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1500 };
1501
1502 /*
1503 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1504 */
1505 struct perf_event_attr very_detailed_attrs[] = {
1506
1507 { .type = PERF_TYPE_HW_CACHE,
1508 .config =
1509 PERF_COUNT_HW_CACHE_L1I << 0 |
1510 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1511 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1512
1513 { .type = PERF_TYPE_HW_CACHE,
1514 .config =
1515 PERF_COUNT_HW_CACHE_L1I << 0 |
1516 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1517 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1518
1519 { .type = PERF_TYPE_HW_CACHE,
1520 .config =
1521 PERF_COUNT_HW_CACHE_DTLB << 0 |
1522 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1523 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1524
1525 { .type = PERF_TYPE_HW_CACHE,
1526 .config =
1527 PERF_COUNT_HW_CACHE_DTLB << 0 |
1528 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1529 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1530
1531 { .type = PERF_TYPE_HW_CACHE,
1532 .config =
1533 PERF_COUNT_HW_CACHE_ITLB << 0 |
1534 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1535 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1536
1537 { .type = PERF_TYPE_HW_CACHE,
1538 .config =
1539 PERF_COUNT_HW_CACHE_ITLB << 0 |
1540 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1541 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1542
1543 };
1544
1545 /*
1546 * Very, very detailed stats (-d -d -d), adding prefetch events:
1547 */
1548 struct perf_event_attr very_very_detailed_attrs[] = {
1549
1550 { .type = PERF_TYPE_HW_CACHE,
1551 .config =
1552 PERF_COUNT_HW_CACHE_L1D << 0 |
1553 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1554 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1555
1556 { .type = PERF_TYPE_HW_CACHE,
1557 .config =
1558 PERF_COUNT_HW_CACHE_L1D << 0 |
1559 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1560 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1561 };
1562
1563 /* Set attrs if no event is selected and !null_run: */
1564 if (null_run)
1565 return 0;
1566
1567 if (transaction_run) {
1568 int err;
1569 if (pmu_have_event("cpu", "cycles-ct") &&
1570 pmu_have_event("cpu", "el-start"))
1571 err = setup_events(transaction_attrs,
1572 ARRAY_SIZE(transaction_attrs));
1573 else
1574 err = setup_events(transaction_limited_attrs,
1575 ARRAY_SIZE(transaction_limited_attrs));
1576 if (err < 0) {
1577 fprintf(stderr, "Cannot set up transaction events\n");
1578 return -1;
1579 }
1580 return 0;
1581 }
1582
1583 if (!evsel_list->nr_entries) {
1584 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1585 return -1;
1586 }
1587
1588 /* Detailed events get appended to the event list: */
1589
1590 if (detailed_run < 1)
1591 return 0;
1592
1593 /* Append detailed run extra attributes: */
1594 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1595 return -1;
1596
1597 if (detailed_run < 2)
1598 return 0;
1599
1600 /* Append very detailed run extra attributes: */
1601 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1602 return -1;
1603
1604 if (detailed_run < 3)
1605 return 0;
1606
1607 /* Append very, very detailed run extra attributes: */
1608 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1609 }
1610
cmd_stat(int argc,const char ** argv,const char * prefix __maybe_unused)1611 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1612 {
1613 bool append_file = false;
1614 int output_fd = 0;
1615 const char *output_name = NULL;
1616 const struct option options[] = {
1617 OPT_BOOLEAN('T', "transaction", &transaction_run,
1618 "hardware transaction statistics"),
1619 OPT_CALLBACK('e', "event", &evsel_list, "event",
1620 "event selector. use 'perf list' to list available events",
1621 parse_events_option),
1622 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1623 "event filter", parse_filter),
1624 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1625 "child tasks do not inherit counters"),
1626 OPT_STRING('p', "pid", &target.pid, "pid",
1627 "stat events on existing process id"),
1628 OPT_STRING('t', "tid", &target.tid, "tid",
1629 "stat events on existing thread id"),
1630 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1631 "system-wide collection from all CPUs"),
1632 OPT_BOOLEAN('g', "group", &group,
1633 "put the counters into a counter group"),
1634 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1635 OPT_INCR('v', "verbose", &verbose,
1636 "be more verbose (show counter open errors, etc)"),
1637 OPT_INTEGER('r', "repeat", &run_count,
1638 "repeat command and print average + stddev (max: 100, forever: 0)"),
1639 OPT_BOOLEAN('n', "null", &null_run,
1640 "null run - dont start any counters"),
1641 OPT_INCR('d', "detailed", &detailed_run,
1642 "detailed run - start a lot of events"),
1643 OPT_BOOLEAN('S', "sync", &sync_run,
1644 "call sync() before starting a run"),
1645 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1646 "print large numbers with thousands\' separators",
1647 stat__set_big_num),
1648 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1649 "list of cpus to monitor in system-wide"),
1650 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1651 "disable CPU count aggregation", AGGR_NONE),
1652 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1653 "print counts with custom separator"),
1654 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1655 "monitor event in cgroup name only", parse_cgroups),
1656 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1657 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1658 OPT_INTEGER(0, "log-fd", &output_fd,
1659 "log output to fd, instead of stderr"),
1660 OPT_STRING(0, "pre", &pre_cmd, "command",
1661 "command to run prior to the measured command"),
1662 OPT_STRING(0, "post", &post_cmd, "command",
1663 "command to run after to the measured command"),
1664 OPT_UINTEGER('I', "interval-print", &interval,
1665 "print counts at regular interval in ms (>= 100)"),
1666 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1667 "aggregate counts per processor socket", AGGR_SOCKET),
1668 OPT_SET_UINT(0, "per-core", &aggr_mode,
1669 "aggregate counts per physical processor core", AGGR_CORE),
1670 OPT_UINTEGER('D', "delay", &initial_delay,
1671 "ms to wait before starting measurement after program start"),
1672 OPT_END()
1673 };
1674 const char * const stat_usage[] = {
1675 "perf stat [<options>] [<command>]",
1676 NULL
1677 };
1678 int status = -EINVAL, run_idx;
1679 const char *mode;
1680
1681 setlocale(LC_ALL, "");
1682
1683 evsel_list = perf_evlist__new();
1684 if (evsel_list == NULL)
1685 return -ENOMEM;
1686
1687 argc = parse_options(argc, argv, options, stat_usage,
1688 PARSE_OPT_STOP_AT_NON_OPTION);
1689
1690 output = stderr;
1691 if (output_name && strcmp(output_name, "-"))
1692 output = NULL;
1693
1694 if (output_name && output_fd) {
1695 fprintf(stderr, "cannot use both --output and --log-fd\n");
1696 parse_options_usage(stat_usage, options, "o", 1);
1697 parse_options_usage(NULL, options, "log-fd", 0);
1698 goto out;
1699 }
1700
1701 if (output_fd < 0) {
1702 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1703 parse_options_usage(stat_usage, options, "log-fd", 0);
1704 goto out;
1705 }
1706
1707 if (!output) {
1708 struct timespec tm;
1709 mode = append_file ? "a" : "w";
1710
1711 output = fopen(output_name, mode);
1712 if (!output) {
1713 perror("failed to create output file");
1714 return -1;
1715 }
1716 clock_gettime(CLOCK_REALTIME, &tm);
1717 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1718 } else if (output_fd > 0) {
1719 mode = append_file ? "a" : "w";
1720 output = fdopen(output_fd, mode);
1721 if (!output) {
1722 perror("Failed opening logfd");
1723 return -errno;
1724 }
1725 }
1726
1727 if (csv_sep) {
1728 csv_output = true;
1729 if (!strcmp(csv_sep, "\\t"))
1730 csv_sep = "\t";
1731 } else
1732 csv_sep = DEFAULT_SEPARATOR;
1733
1734 /*
1735 * let the spreadsheet do the pretty-printing
1736 */
1737 if (csv_output) {
1738 /* User explicitly passed -B? */
1739 if (big_num_opt == 1) {
1740 fprintf(stderr, "-B option not supported with -x\n");
1741 parse_options_usage(stat_usage, options, "B", 1);
1742 parse_options_usage(NULL, options, "x", 1);
1743 goto out;
1744 } else /* Nope, so disable big number formatting */
1745 big_num = false;
1746 } else if (big_num_opt == 0) /* User passed --no-big-num */
1747 big_num = false;
1748
1749 if (!argc && target__none(&target))
1750 usage_with_options(stat_usage, options);
1751
1752 if (run_count < 0) {
1753 pr_err("Run count must be a positive number\n");
1754 parse_options_usage(stat_usage, options, "r", 1);
1755 goto out;
1756 } else if (run_count == 0) {
1757 forever = true;
1758 run_count = 1;
1759 }
1760
1761 /* no_aggr, cgroup are for system-wide only */
1762 if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1763 !target__has_cpu(&target)) {
1764 fprintf(stderr, "both cgroup and no-aggregation "
1765 "modes only available in system-wide mode\n");
1766
1767 parse_options_usage(stat_usage, options, "G", 1);
1768 parse_options_usage(NULL, options, "A", 1);
1769 parse_options_usage(NULL, options, "a", 1);
1770 goto out;
1771 }
1772
1773 if (add_default_attributes())
1774 goto out;
1775
1776 target__validate(&target);
1777
1778 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1779 if (target__has_task(&target)) {
1780 pr_err("Problems finding threads of monitor\n");
1781 parse_options_usage(stat_usage, options, "p", 1);
1782 parse_options_usage(NULL, options, "t", 1);
1783 } else if (target__has_cpu(&target)) {
1784 perror("failed to parse CPUs map");
1785 parse_options_usage(stat_usage, options, "C", 1);
1786 parse_options_usage(NULL, options, "a", 1);
1787 }
1788 goto out;
1789 }
1790 if (interval && interval < 100) {
1791 pr_err("print interval must be >= 100ms\n");
1792 parse_options_usage(stat_usage, options, "I", 1);
1793 goto out;
1794 }
1795
1796 if (perf_evlist__alloc_stats(evsel_list, interval))
1797 goto out;
1798
1799 if (perf_stat_init_aggr_mode())
1800 goto out;
1801
1802 /*
1803 * We dont want to block the signals - that would cause
1804 * child tasks to inherit that and Ctrl-C would not work.
1805 * What we want is for Ctrl-C to work in the exec()-ed
1806 * task, but being ignored by perf stat itself:
1807 */
1808 atexit(sig_atexit);
1809 if (!forever)
1810 signal(SIGINT, skip_signal);
1811 signal(SIGCHLD, skip_signal);
1812 signal(SIGALRM, skip_signal);
1813 signal(SIGABRT, skip_signal);
1814
1815 status = 0;
1816 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1817 if (run_count != 1 && verbose)
1818 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1819 run_idx + 1);
1820
1821 status = run_perf_stat(argc, argv);
1822 if (forever && status != -1) {
1823 print_stat(argc, argv);
1824 perf_stat__reset_stats(evsel_list);
1825 }
1826 }
1827
1828 if (!forever && status != -1 && !interval)
1829 print_stat(argc, argv);
1830
1831 perf_evlist__free_stats(evsel_list);
1832 out:
1833 perf_evlist__delete(evsel_list);
1834 return status;
1835 }
1836