Loading...
1/*
2 * builtin-stat.c
3 *
4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
6 *
7 * Sample output:
8
9 $ perf stat ./hackbench 10
10
11 Time: 0.118
12
13 Performance counter stats for './hackbench 10':
14
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
26
27 0.154822978 seconds time elapsed
28
29 *
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
31 *
32 * Improvements and fixes by:
33 *
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
40 *
41 * Released under the GPL v2. (and only v2, not any later version)
42 */
43
44#include "perf.h"
45#include "builtin.h"
46#include "util/util.h"
47#include "util/parse-options.h"
48#include "util/parse-events.h"
49#include "util/event.h"
50#include "util/evlist.h"
51#include "util/evsel.h"
52#include "util/debug.h"
53#include "util/color.h"
54#include "util/header.h"
55#include "util/cpumap.h"
56#include "util/thread.h"
57#include "util/thread_map.h"
58
59#include <sys/prctl.h>
60#include <math.h>
61#include <locale.h>
62
63#define DEFAULT_SEPARATOR " "
64#define CNTR_NOT_SUPPORTED "<not supported>"
65#define CNTR_NOT_COUNTED "<not counted>"
66
67static struct perf_event_attr default_attrs[] = {
68
69 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
70 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
71 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
72 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
73
74 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
75 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
76 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
77 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
78 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
79 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
80
81};
82
83/*
84 * Detailed stats (-d), covering the L1 and last level data caches:
85 */
86static struct perf_event_attr detailed_attrs[] = {
87
88 { .type = PERF_TYPE_HW_CACHE,
89 .config =
90 PERF_COUNT_HW_CACHE_L1D << 0 |
91 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
92 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
93
94 { .type = PERF_TYPE_HW_CACHE,
95 .config =
96 PERF_COUNT_HW_CACHE_L1D << 0 |
97 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
98 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
99
100 { .type = PERF_TYPE_HW_CACHE,
101 .config =
102 PERF_COUNT_HW_CACHE_LL << 0 |
103 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
104 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
105
106 { .type = PERF_TYPE_HW_CACHE,
107 .config =
108 PERF_COUNT_HW_CACHE_LL << 0 |
109 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
110 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
111};
112
113/*
114 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
115 */
116static struct perf_event_attr very_detailed_attrs[] = {
117
118 { .type = PERF_TYPE_HW_CACHE,
119 .config =
120 PERF_COUNT_HW_CACHE_L1I << 0 |
121 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
122 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
123
124 { .type = PERF_TYPE_HW_CACHE,
125 .config =
126 PERF_COUNT_HW_CACHE_L1I << 0 |
127 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
128 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
129
130 { .type = PERF_TYPE_HW_CACHE,
131 .config =
132 PERF_COUNT_HW_CACHE_DTLB << 0 |
133 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
134 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
135
136 { .type = PERF_TYPE_HW_CACHE,
137 .config =
138 PERF_COUNT_HW_CACHE_DTLB << 0 |
139 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
140 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
141
142 { .type = PERF_TYPE_HW_CACHE,
143 .config =
144 PERF_COUNT_HW_CACHE_ITLB << 0 |
145 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
146 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
147
148 { .type = PERF_TYPE_HW_CACHE,
149 .config =
150 PERF_COUNT_HW_CACHE_ITLB << 0 |
151 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
152 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
153
154};
155
156/*
157 * Very, very detailed stats (-d -d -d), adding prefetch events:
158 */
159static struct perf_event_attr very_very_detailed_attrs[] = {
160
161 { .type = PERF_TYPE_HW_CACHE,
162 .config =
163 PERF_COUNT_HW_CACHE_L1D << 0 |
164 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
165 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
166
167 { .type = PERF_TYPE_HW_CACHE,
168 .config =
169 PERF_COUNT_HW_CACHE_L1D << 0 |
170 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
171 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
172};
173
174
175
176struct perf_evlist *evsel_list;
177
178static bool system_wide = false;
179static int run_idx = 0;
180
181static int run_count = 1;
182static bool no_inherit = false;
183static bool scale = true;
184static bool no_aggr = false;
185static pid_t target_pid = -1;
186static pid_t target_tid = -1;
187static pid_t child_pid = -1;
188static bool null_run = false;
189static int detailed_run = 0;
190static bool sync_run = false;
191static bool big_num = true;
192static int big_num_opt = -1;
193static const char *cpu_list;
194static const char *csv_sep = NULL;
195static bool csv_output = false;
196static bool group = false;
197
198static volatile int done = 0;
199
200struct stats
201{
202 double n, mean, M2;
203};
204
205struct perf_stat {
206 struct stats res_stats[3];
207};
208
209static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
210{
211 evsel->priv = zalloc(sizeof(struct perf_stat));
212 return evsel->priv == NULL ? -ENOMEM : 0;
213}
214
215static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
216{
217 free(evsel->priv);
218 evsel->priv = NULL;
219}
220
221static void update_stats(struct stats *stats, u64 val)
222{
223 double delta;
224
225 stats->n++;
226 delta = val - stats->mean;
227 stats->mean += delta / stats->n;
228 stats->M2 += delta*(val - stats->mean);
229}
230
231static double avg_stats(struct stats *stats)
232{
233 return stats->mean;
234}
235
236/*
237 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
238 *
239 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
240 * s^2 = -------------------------------
241 * n - 1
242 *
243 * http://en.wikipedia.org/wiki/Stddev
244 *
245 * The std dev of the mean is related to the std dev by:
246 *
247 * s
248 * s_mean = -------
249 * sqrt(n)
250 *
251 */
252static double stddev_stats(struct stats *stats)
253{
254 double variance = stats->M2 / (stats->n - 1);
255 double variance_mean = variance / stats->n;
256
257 return sqrt(variance_mean);
258}
259
260struct stats runtime_nsecs_stats[MAX_NR_CPUS];
261struct stats runtime_cycles_stats[MAX_NR_CPUS];
262struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
263struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
264struct stats runtime_branches_stats[MAX_NR_CPUS];
265struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
266struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
267struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
268struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
269struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
270struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
271struct stats walltime_nsecs_stats;
272
273static int create_perf_stat_counter(struct perf_evsel *evsel)
274{
275 struct perf_event_attr *attr = &evsel->attr;
276
277 if (scale)
278 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
279 PERF_FORMAT_TOTAL_TIME_RUNNING;
280
281 attr->inherit = !no_inherit;
282
283 if (system_wide)
284 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, group);
285
286 if (target_pid == -1 && target_tid == -1) {
287 attr->disabled = 1;
288 attr->enable_on_exec = 1;
289 }
290
291 return perf_evsel__open_per_thread(evsel, evsel_list->threads, group);
292}
293
294/*
295 * Does the counter have nsecs as a unit?
296 */
297static inline int nsec_counter(struct perf_evsel *evsel)
298{
299 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
300 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
301 return 1;
302
303 return 0;
304}
305
306/*
307 * Update various tracking values we maintain to print
308 * more semantic information such as miss/hit ratios,
309 * instruction rates, etc:
310 */
311static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
312{
313 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
314 update_stats(&runtime_nsecs_stats[0], count[0]);
315 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
316 update_stats(&runtime_cycles_stats[0], count[0]);
317 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
318 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
319 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
320 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
321 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
322 update_stats(&runtime_branches_stats[0], count[0]);
323 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
324 update_stats(&runtime_cacherefs_stats[0], count[0]);
325 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
326 update_stats(&runtime_l1_dcache_stats[0], count[0]);
327 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
328 update_stats(&runtime_l1_icache_stats[0], count[0]);
329 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
330 update_stats(&runtime_ll_cache_stats[0], count[0]);
331 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
332 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
333 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
334 update_stats(&runtime_itlb_cache_stats[0], count[0]);
335}
336
337/*
338 * Read out the results of a single counter:
339 * aggregate counts across CPUs in system-wide mode
340 */
341static int read_counter_aggr(struct perf_evsel *counter)
342{
343 struct perf_stat *ps = counter->priv;
344 u64 *count = counter->counts->aggr.values;
345 int i;
346
347 if (__perf_evsel__read(counter, evsel_list->cpus->nr,
348 evsel_list->threads->nr, scale) < 0)
349 return -1;
350
351 for (i = 0; i < 3; i++)
352 update_stats(&ps->res_stats[i], count[i]);
353
354 if (verbose) {
355 fprintf(stderr, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
356 event_name(counter), count[0], count[1], count[2]);
357 }
358
359 /*
360 * Save the full runtime - to allow normalization during printout:
361 */
362 update_shadow_stats(counter, count);
363
364 return 0;
365}
366
367/*
368 * Read out the results of a single counter:
369 * do not aggregate counts across CPUs in system-wide mode
370 */
371static int read_counter(struct perf_evsel *counter)
372{
373 u64 *count;
374 int cpu;
375
376 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
377 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
378 return -1;
379
380 count = counter->counts->cpu[cpu].values;
381
382 update_shadow_stats(counter, count);
383 }
384
385 return 0;
386}
387
388static int run_perf_stat(int argc __used, const char **argv)
389{
390 unsigned long long t0, t1;
391 struct perf_evsel *counter;
392 int status = 0;
393 int child_ready_pipe[2], go_pipe[2];
394 const bool forks = (argc > 0);
395 char buf;
396
397 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
398 perror("failed to create pipes");
399 exit(1);
400 }
401
402 if (forks) {
403 if ((child_pid = fork()) < 0)
404 perror("failed to fork");
405
406 if (!child_pid) {
407 close(child_ready_pipe[0]);
408 close(go_pipe[1]);
409 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
410
411 /*
412 * Do a dummy execvp to get the PLT entry resolved,
413 * so we avoid the resolver overhead on the real
414 * execvp call.
415 */
416 execvp("", (char **)argv);
417
418 /*
419 * Tell the parent we're ready to go
420 */
421 close(child_ready_pipe[1]);
422
423 /*
424 * Wait until the parent tells us to go.
425 */
426 if (read(go_pipe[0], &buf, 1) == -1)
427 perror("unable to read pipe");
428
429 execvp(argv[0], (char **)argv);
430
431 perror(argv[0]);
432 exit(-1);
433 }
434
435 if (target_tid == -1 && target_pid == -1 && !system_wide)
436 evsel_list->threads->map[0] = child_pid;
437
438 /*
439 * Wait for the child to be ready to exec.
440 */
441 close(child_ready_pipe[1]);
442 close(go_pipe[0]);
443 if (read(child_ready_pipe[0], &buf, 1) == -1)
444 perror("unable to read pipe");
445 close(child_ready_pipe[0]);
446 }
447
448 list_for_each_entry(counter, &evsel_list->entries, node) {
449 if (create_perf_stat_counter(counter) < 0) {
450 if (errno == EINVAL || errno == ENOSYS || errno == ENOENT) {
451 if (verbose)
452 ui__warning("%s event is not supported by the kernel.\n",
453 event_name(counter));
454 counter->supported = false;
455 continue;
456 }
457
458 if (errno == EPERM || errno == EACCES) {
459 error("You may not have permission to collect %sstats.\n"
460 "\t Consider tweaking"
461 " /proc/sys/kernel/perf_event_paranoid or running as root.",
462 system_wide ? "system-wide " : "");
463 } else {
464 error("open_counter returned with %d (%s). "
465 "/bin/dmesg may provide additional information.\n",
466 errno, strerror(errno));
467 }
468 if (child_pid != -1)
469 kill(child_pid, SIGTERM);
470 die("Not all events could be opened.\n");
471 return -1;
472 }
473 counter->supported = true;
474 }
475
476 if (perf_evlist__set_filters(evsel_list)) {
477 error("failed to set filter with %d (%s)\n", errno,
478 strerror(errno));
479 return -1;
480 }
481
482 /*
483 * Enable counters and exec the command:
484 */
485 t0 = rdclock();
486
487 if (forks) {
488 close(go_pipe[1]);
489 wait(&status);
490 } else {
491 while(!done) sleep(1);
492 }
493
494 t1 = rdclock();
495
496 update_stats(&walltime_nsecs_stats, t1 - t0);
497
498 if (no_aggr) {
499 list_for_each_entry(counter, &evsel_list->entries, node) {
500 read_counter(counter);
501 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
502 }
503 } else {
504 list_for_each_entry(counter, &evsel_list->entries, node) {
505 read_counter_aggr(counter);
506 perf_evsel__close_fd(counter, evsel_list->cpus->nr,
507 evsel_list->threads->nr);
508 }
509 }
510
511 return WEXITSTATUS(status);
512}
513
514static void print_noise_pct(double total, double avg)
515{
516 double pct = 0.0;
517
518 if (avg)
519 pct = 100.0*total/avg;
520
521 if (csv_output)
522 fprintf(stderr, "%s%.2f%%", csv_sep, pct);
523 else
524 fprintf(stderr, " ( +-%6.2f%% )", pct);
525}
526
527static void print_noise(struct perf_evsel *evsel, double avg)
528{
529 struct perf_stat *ps;
530
531 if (run_count == 1)
532 return;
533
534 ps = evsel->priv;
535 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
536}
537
538static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
539{
540 double msecs = avg / 1e6;
541 char cpustr[16] = { '\0', };
542 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
543
544 if (no_aggr)
545 sprintf(cpustr, "CPU%*d%s",
546 csv_output ? 0 : -4,
547 evsel_list->cpus->map[cpu], csv_sep);
548
549 fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel));
550
551 if (evsel->cgrp)
552 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
553
554 if (csv_output)
555 return;
556
557 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
558 fprintf(stderr, " # %8.3f CPUs utilized ", avg / avg_stats(&walltime_nsecs_stats));
559}
560
561static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg)
562{
563 double total, ratio = 0.0;
564 const char *color;
565
566 total = avg_stats(&runtime_cycles_stats[cpu]);
567
568 if (total)
569 ratio = avg / total * 100.0;
570
571 color = PERF_COLOR_NORMAL;
572 if (ratio > 50.0)
573 color = PERF_COLOR_RED;
574 else if (ratio > 30.0)
575 color = PERF_COLOR_MAGENTA;
576 else if (ratio > 10.0)
577 color = PERF_COLOR_YELLOW;
578
579 fprintf(stderr, " # ");
580 color_fprintf(stderr, color, "%6.2f%%", ratio);
581 fprintf(stderr, " frontend cycles idle ");
582}
583
584static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg)
585{
586 double total, ratio = 0.0;
587 const char *color;
588
589 total = avg_stats(&runtime_cycles_stats[cpu]);
590
591 if (total)
592 ratio = avg / total * 100.0;
593
594 color = PERF_COLOR_NORMAL;
595 if (ratio > 75.0)
596 color = PERF_COLOR_RED;
597 else if (ratio > 50.0)
598 color = PERF_COLOR_MAGENTA;
599 else if (ratio > 20.0)
600 color = PERF_COLOR_YELLOW;
601
602 fprintf(stderr, " # ");
603 color_fprintf(stderr, color, "%6.2f%%", ratio);
604 fprintf(stderr, " backend cycles idle ");
605}
606
607static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
608{
609 double total, ratio = 0.0;
610 const char *color;
611
612 total = avg_stats(&runtime_branches_stats[cpu]);
613
614 if (total)
615 ratio = avg / total * 100.0;
616
617 color = PERF_COLOR_NORMAL;
618 if (ratio > 20.0)
619 color = PERF_COLOR_RED;
620 else if (ratio > 10.0)
621 color = PERF_COLOR_MAGENTA;
622 else if (ratio > 5.0)
623 color = PERF_COLOR_YELLOW;
624
625 fprintf(stderr, " # ");
626 color_fprintf(stderr, color, "%6.2f%%", ratio);
627 fprintf(stderr, " of all branches ");
628}
629
630static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
631{
632 double total, ratio = 0.0;
633 const char *color;
634
635 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
636
637 if (total)
638 ratio = avg / total * 100.0;
639
640 color = PERF_COLOR_NORMAL;
641 if (ratio > 20.0)
642 color = PERF_COLOR_RED;
643 else if (ratio > 10.0)
644 color = PERF_COLOR_MAGENTA;
645 else if (ratio > 5.0)
646 color = PERF_COLOR_YELLOW;
647
648 fprintf(stderr, " # ");
649 color_fprintf(stderr, color, "%6.2f%%", ratio);
650 fprintf(stderr, " of all L1-dcache hits ");
651}
652
653static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
654{
655 double total, ratio = 0.0;
656 const char *color;
657
658 total = avg_stats(&runtime_l1_icache_stats[cpu]);
659
660 if (total)
661 ratio = avg / total * 100.0;
662
663 color = PERF_COLOR_NORMAL;
664 if (ratio > 20.0)
665 color = PERF_COLOR_RED;
666 else if (ratio > 10.0)
667 color = PERF_COLOR_MAGENTA;
668 else if (ratio > 5.0)
669 color = PERF_COLOR_YELLOW;
670
671 fprintf(stderr, " # ");
672 color_fprintf(stderr, color, "%6.2f%%", ratio);
673 fprintf(stderr, " of all L1-icache hits ");
674}
675
676static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
677{
678 double total, ratio = 0.0;
679 const char *color;
680
681 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
682
683 if (total)
684 ratio = avg / total * 100.0;
685
686 color = PERF_COLOR_NORMAL;
687 if (ratio > 20.0)
688 color = PERF_COLOR_RED;
689 else if (ratio > 10.0)
690 color = PERF_COLOR_MAGENTA;
691 else if (ratio > 5.0)
692 color = PERF_COLOR_YELLOW;
693
694 fprintf(stderr, " # ");
695 color_fprintf(stderr, color, "%6.2f%%", ratio);
696 fprintf(stderr, " of all dTLB cache hits ");
697}
698
699static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
700{
701 double total, ratio = 0.0;
702 const char *color;
703
704 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
705
706 if (total)
707 ratio = avg / total * 100.0;
708
709 color = PERF_COLOR_NORMAL;
710 if (ratio > 20.0)
711 color = PERF_COLOR_RED;
712 else if (ratio > 10.0)
713 color = PERF_COLOR_MAGENTA;
714 else if (ratio > 5.0)
715 color = PERF_COLOR_YELLOW;
716
717 fprintf(stderr, " # ");
718 color_fprintf(stderr, color, "%6.2f%%", ratio);
719 fprintf(stderr, " of all iTLB cache hits ");
720}
721
722static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
723{
724 double total, ratio = 0.0;
725 const char *color;
726
727 total = avg_stats(&runtime_ll_cache_stats[cpu]);
728
729 if (total)
730 ratio = avg / total * 100.0;
731
732 color = PERF_COLOR_NORMAL;
733 if (ratio > 20.0)
734 color = PERF_COLOR_RED;
735 else if (ratio > 10.0)
736 color = PERF_COLOR_MAGENTA;
737 else if (ratio > 5.0)
738 color = PERF_COLOR_YELLOW;
739
740 fprintf(stderr, " # ");
741 color_fprintf(stderr, color, "%6.2f%%", ratio);
742 fprintf(stderr, " of all LL-cache hits ");
743}
744
745static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
746{
747 double total, ratio = 0.0;
748 char cpustr[16] = { '\0', };
749 const char *fmt;
750
751 if (csv_output)
752 fmt = "%s%.0f%s%s";
753 else if (big_num)
754 fmt = "%s%'18.0f%s%-25s";
755 else
756 fmt = "%s%18.0f%s%-25s";
757
758 if (no_aggr)
759 sprintf(cpustr, "CPU%*d%s",
760 csv_output ? 0 : -4,
761 evsel_list->cpus->map[cpu], csv_sep);
762 else
763 cpu = 0;
764
765 fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel));
766
767 if (evsel->cgrp)
768 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
769
770 if (csv_output)
771 return;
772
773 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
774 total = avg_stats(&runtime_cycles_stats[cpu]);
775
776 if (total)
777 ratio = avg / total;
778
779 fprintf(stderr, " # %5.2f insns per cycle ", ratio);
780
781 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
782 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
783
784 if (total && avg) {
785 ratio = total / avg;
786 fprintf(stderr, "\n # %5.2f stalled cycles per insn", ratio);
787 }
788
789 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
790 runtime_branches_stats[cpu].n != 0) {
791 print_branch_misses(cpu, evsel, avg);
792 } else if (
793 evsel->attr.type == PERF_TYPE_HW_CACHE &&
794 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
795 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
796 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
797 runtime_l1_dcache_stats[cpu].n != 0) {
798 print_l1_dcache_misses(cpu, evsel, avg);
799 } else if (
800 evsel->attr.type == PERF_TYPE_HW_CACHE &&
801 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
802 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
803 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
804 runtime_l1_icache_stats[cpu].n != 0) {
805 print_l1_icache_misses(cpu, evsel, avg);
806 } else if (
807 evsel->attr.type == PERF_TYPE_HW_CACHE &&
808 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
809 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
810 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
811 runtime_dtlb_cache_stats[cpu].n != 0) {
812 print_dtlb_cache_misses(cpu, evsel, avg);
813 } else if (
814 evsel->attr.type == PERF_TYPE_HW_CACHE &&
815 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
816 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
817 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
818 runtime_itlb_cache_stats[cpu].n != 0) {
819 print_itlb_cache_misses(cpu, evsel, avg);
820 } else if (
821 evsel->attr.type == PERF_TYPE_HW_CACHE &&
822 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
823 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
824 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
825 runtime_ll_cache_stats[cpu].n != 0) {
826 print_ll_cache_misses(cpu, evsel, avg);
827 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
828 runtime_cacherefs_stats[cpu].n != 0) {
829 total = avg_stats(&runtime_cacherefs_stats[cpu]);
830
831 if (total)
832 ratio = avg * 100 / total;
833
834 fprintf(stderr, " # %8.3f %% of all cache refs ", ratio);
835
836 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
837 print_stalled_cycles_frontend(cpu, evsel, avg);
838 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
839 print_stalled_cycles_backend(cpu, evsel, avg);
840 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
841 total = avg_stats(&runtime_nsecs_stats[cpu]);
842
843 if (total)
844 ratio = 1.0 * avg / total;
845
846 fprintf(stderr, " # %8.3f GHz ", ratio);
847 } else if (runtime_nsecs_stats[cpu].n != 0) {
848 total = avg_stats(&runtime_nsecs_stats[cpu]);
849
850 if (total)
851 ratio = 1000.0 * avg / total;
852
853 fprintf(stderr, " # %8.3f M/sec ", ratio);
854 } else {
855 fprintf(stderr, " ");
856 }
857}
858
859/*
860 * Print out the results of a single counter:
861 * aggregated counts in system-wide mode
862 */
863static void print_counter_aggr(struct perf_evsel *counter)
864{
865 struct perf_stat *ps = counter->priv;
866 double avg = avg_stats(&ps->res_stats[0]);
867 int scaled = counter->counts->scaled;
868
869 if (scaled == -1) {
870 fprintf(stderr, "%*s%s%*s",
871 csv_output ? 0 : 18,
872 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
873 csv_sep,
874 csv_output ? 0 : -24,
875 event_name(counter));
876
877 if (counter->cgrp)
878 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
879
880 fputc('\n', stderr);
881 return;
882 }
883
884 if (nsec_counter(counter))
885 nsec_printout(-1, counter, avg);
886 else
887 abs_printout(-1, counter, avg);
888
889 print_noise(counter, avg);
890
891 if (csv_output) {
892 fputc('\n', stderr);
893 return;
894 }
895
896 if (scaled) {
897 double avg_enabled, avg_running;
898
899 avg_enabled = avg_stats(&ps->res_stats[1]);
900 avg_running = avg_stats(&ps->res_stats[2]);
901
902 fprintf(stderr, " [%5.2f%%]", 100 * avg_running / avg_enabled);
903 }
904 fprintf(stderr, "\n");
905}
906
907/*
908 * Print out the results of a single counter:
909 * does not use aggregated count in system-wide
910 */
911static void print_counter(struct perf_evsel *counter)
912{
913 u64 ena, run, val;
914 int cpu;
915
916 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
917 val = counter->counts->cpu[cpu].val;
918 ena = counter->counts->cpu[cpu].ena;
919 run = counter->counts->cpu[cpu].run;
920 if (run == 0 || ena == 0) {
921 fprintf(stderr, "CPU%*d%s%*s%s%*s",
922 csv_output ? 0 : -4,
923 evsel_list->cpus->map[cpu], csv_sep,
924 csv_output ? 0 : 18,
925 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
926 csv_sep,
927 csv_output ? 0 : -24,
928 event_name(counter));
929
930 if (counter->cgrp)
931 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
932
933 fputc('\n', stderr);
934 continue;
935 }
936
937 if (nsec_counter(counter))
938 nsec_printout(cpu, counter, val);
939 else
940 abs_printout(cpu, counter, val);
941
942 if (!csv_output) {
943 print_noise(counter, 1.0);
944
945 if (run != ena)
946 fprintf(stderr, " (%.2f%%)", 100.0 * run / ena);
947 }
948 fputc('\n', stderr);
949 }
950}
951
952static void print_stat(int argc, const char **argv)
953{
954 struct perf_evsel *counter;
955 int i;
956
957 fflush(stdout);
958
959 if (!csv_output) {
960 fprintf(stderr, "\n");
961 fprintf(stderr, " Performance counter stats for ");
962 if(target_pid == -1 && target_tid == -1) {
963 fprintf(stderr, "\'%s", argv[0]);
964 for (i = 1; i < argc; i++)
965 fprintf(stderr, " %s", argv[i]);
966 } else if (target_pid != -1)
967 fprintf(stderr, "process id \'%d", target_pid);
968 else
969 fprintf(stderr, "thread id \'%d", target_tid);
970
971 fprintf(stderr, "\'");
972 if (run_count > 1)
973 fprintf(stderr, " (%d runs)", run_count);
974 fprintf(stderr, ":\n\n");
975 }
976
977 if (no_aggr) {
978 list_for_each_entry(counter, &evsel_list->entries, node)
979 print_counter(counter);
980 } else {
981 list_for_each_entry(counter, &evsel_list->entries, node)
982 print_counter_aggr(counter);
983 }
984
985 if (!csv_output) {
986 if (!null_run)
987 fprintf(stderr, "\n");
988 fprintf(stderr, " %17.9f seconds time elapsed",
989 avg_stats(&walltime_nsecs_stats)/1e9);
990 if (run_count > 1) {
991 fprintf(stderr, " ");
992 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
993 avg_stats(&walltime_nsecs_stats));
994 }
995 fprintf(stderr, "\n\n");
996 }
997}
998
999static volatile int signr = -1;
1000
1001static void skip_signal(int signo)
1002{
1003 if(child_pid == -1)
1004 done = 1;
1005
1006 signr = signo;
1007}
1008
1009static void sig_atexit(void)
1010{
1011 if (child_pid != -1)
1012 kill(child_pid, SIGTERM);
1013
1014 if (signr == -1)
1015 return;
1016
1017 signal(signr, SIG_DFL);
1018 kill(getpid(), signr);
1019}
1020
1021static const char * const stat_usage[] = {
1022 "perf stat [<options>] [<command>]",
1023 NULL
1024};
1025
1026static int stat__set_big_num(const struct option *opt __used,
1027 const char *s __used, int unset)
1028{
1029 big_num_opt = unset ? 0 : 1;
1030 return 0;
1031}
1032
1033static const struct option options[] = {
1034 OPT_CALLBACK('e', "event", &evsel_list, "event",
1035 "event selector. use 'perf list' to list available events",
1036 parse_events_option),
1037 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1038 "event filter", parse_filter),
1039 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1040 "child tasks do not inherit counters"),
1041 OPT_INTEGER('p', "pid", &target_pid,
1042 "stat events on existing process id"),
1043 OPT_INTEGER('t', "tid", &target_tid,
1044 "stat events on existing thread id"),
1045 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1046 "system-wide collection from all CPUs"),
1047 OPT_BOOLEAN('g', "group", &group,
1048 "put the counters into a counter group"),
1049 OPT_BOOLEAN('c', "scale", &scale,
1050 "scale/normalize counters"),
1051 OPT_INCR('v', "verbose", &verbose,
1052 "be more verbose (show counter open errors, etc)"),
1053 OPT_INTEGER('r', "repeat", &run_count,
1054 "repeat command and print average + stddev (max: 100)"),
1055 OPT_BOOLEAN('n', "null", &null_run,
1056 "null run - dont start any counters"),
1057 OPT_INCR('d', "detailed", &detailed_run,
1058 "detailed run - start a lot of events"),
1059 OPT_BOOLEAN('S', "sync", &sync_run,
1060 "call sync() before starting a run"),
1061 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1062 "print large numbers with thousands\' separators",
1063 stat__set_big_num),
1064 OPT_STRING('C', "cpu", &cpu_list, "cpu",
1065 "list of cpus to monitor in system-wide"),
1066 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1067 "disable CPU count aggregation"),
1068 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1069 "print counts with custom separator"),
1070 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1071 "monitor event in cgroup name only",
1072 parse_cgroups),
1073 OPT_END()
1074};
1075
1076/*
1077 * Add default attributes, if there were no attributes specified or
1078 * if -d/--detailed, -d -d or -d -d -d is used:
1079 */
1080static int add_default_attributes(void)
1081{
1082 struct perf_evsel *pos;
1083 size_t attr_nr = 0;
1084 size_t c;
1085
1086 /* Set attrs if no event is selected and !null_run: */
1087 if (null_run)
1088 return 0;
1089
1090 if (!evsel_list->nr_entries) {
1091 for (c = 0; c < ARRAY_SIZE(default_attrs); c++) {
1092 pos = perf_evsel__new(default_attrs + c, c + attr_nr);
1093 if (pos == NULL)
1094 return -1;
1095 perf_evlist__add(evsel_list, pos);
1096 }
1097 attr_nr += c;
1098 }
1099
1100 /* Detailed events get appended to the event list: */
1101
1102 if (detailed_run < 1)
1103 return 0;
1104
1105 /* Append detailed run extra attributes: */
1106 for (c = 0; c < ARRAY_SIZE(detailed_attrs); c++) {
1107 pos = perf_evsel__new(detailed_attrs + c, c + attr_nr);
1108 if (pos == NULL)
1109 return -1;
1110 perf_evlist__add(evsel_list, pos);
1111 }
1112 attr_nr += c;
1113
1114 if (detailed_run < 2)
1115 return 0;
1116
1117 /* Append very detailed run extra attributes: */
1118 for (c = 0; c < ARRAY_SIZE(very_detailed_attrs); c++) {
1119 pos = perf_evsel__new(very_detailed_attrs + c, c + attr_nr);
1120 if (pos == NULL)
1121 return -1;
1122 perf_evlist__add(evsel_list, pos);
1123 }
1124
1125 if (detailed_run < 3)
1126 return 0;
1127
1128 /* Append very, very detailed run extra attributes: */
1129 for (c = 0; c < ARRAY_SIZE(very_very_detailed_attrs); c++) {
1130 pos = perf_evsel__new(very_very_detailed_attrs + c, c + attr_nr);
1131 if (pos == NULL)
1132 return -1;
1133 perf_evlist__add(evsel_list, pos);
1134 }
1135
1136
1137 return 0;
1138}
1139
1140int cmd_stat(int argc, const char **argv, const char *prefix __used)
1141{
1142 struct perf_evsel *pos;
1143 int status = -ENOMEM;
1144
1145 setlocale(LC_ALL, "");
1146
1147 evsel_list = perf_evlist__new(NULL, NULL);
1148 if (evsel_list == NULL)
1149 return -ENOMEM;
1150
1151 argc = parse_options(argc, argv, options, stat_usage,
1152 PARSE_OPT_STOP_AT_NON_OPTION);
1153
1154 if (csv_sep)
1155 csv_output = true;
1156 else
1157 csv_sep = DEFAULT_SEPARATOR;
1158
1159 /*
1160 * let the spreadsheet do the pretty-printing
1161 */
1162 if (csv_output) {
1163 /* User explicitely passed -B? */
1164 if (big_num_opt == 1) {
1165 fprintf(stderr, "-B option not supported with -x\n");
1166 usage_with_options(stat_usage, options);
1167 } else /* Nope, so disable big number formatting */
1168 big_num = false;
1169 } else if (big_num_opt == 0) /* User passed --no-big-num */
1170 big_num = false;
1171
1172 if (!argc && target_pid == -1 && target_tid == -1)
1173 usage_with_options(stat_usage, options);
1174 if (run_count <= 0)
1175 usage_with_options(stat_usage, options);
1176
1177 /* no_aggr, cgroup are for system-wide only */
1178 if ((no_aggr || nr_cgroups) && !system_wide) {
1179 fprintf(stderr, "both cgroup and no-aggregation "
1180 "modes only available in system-wide mode\n");
1181
1182 usage_with_options(stat_usage, options);
1183 }
1184
1185 if (add_default_attributes())
1186 goto out;
1187
1188 if (target_pid != -1)
1189 target_tid = target_pid;
1190
1191 evsel_list->threads = thread_map__new(target_pid, target_tid);
1192 if (evsel_list->threads == NULL) {
1193 pr_err("Problems finding threads of monitor\n");
1194 usage_with_options(stat_usage, options);
1195 }
1196
1197 if (system_wide)
1198 evsel_list->cpus = cpu_map__new(cpu_list);
1199 else
1200 evsel_list->cpus = cpu_map__dummy_new();
1201
1202 if (evsel_list->cpus == NULL) {
1203 perror("failed to parse CPUs map");
1204 usage_with_options(stat_usage, options);
1205 return -1;
1206 }
1207
1208 list_for_each_entry(pos, &evsel_list->entries, node) {
1209 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1210 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 ||
1211 perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0)
1212 goto out_free_fd;
1213 }
1214
1215 /*
1216 * We dont want to block the signals - that would cause
1217 * child tasks to inherit that and Ctrl-C would not work.
1218 * What we want is for Ctrl-C to work in the exec()-ed
1219 * task, but being ignored by perf stat itself:
1220 */
1221 atexit(sig_atexit);
1222 signal(SIGINT, skip_signal);
1223 signal(SIGALRM, skip_signal);
1224 signal(SIGABRT, skip_signal);
1225
1226 status = 0;
1227 for (run_idx = 0; run_idx < run_count; run_idx++) {
1228 if (run_count != 1 && verbose)
1229 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
1230
1231 if (sync_run)
1232 sync();
1233
1234 status = run_perf_stat(argc, argv);
1235 }
1236
1237 if (status != -1)
1238 print_stat(argc, argv);
1239out_free_fd:
1240 list_for_each_entry(pos, &evsel_list->entries, node)
1241 perf_evsel__free_stat_priv(pos);
1242 perf_evlist__delete_maps(evsel_list);
1243out:
1244 perf_evlist__delete(evsel_list);
1245 return status;
1246}
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 <subcmd/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/drv_configs.h"
56#include "util/color.h"
57#include "util/stat.h"
58#include "util/header.h"
59#include "util/cpumap.h"
60#include "util/thread.h"
61#include "util/thread_map.h"
62#include "util/counts.h"
63#include "util/group.h"
64#include "util/session.h"
65#include "util/tool.h"
66#include "util/group.h"
67#include "asm/bug.h"
68
69#include <linux/time64.h>
70#include <api/fs/fs.h>
71#include <stdlib.h>
72#include <sys/prctl.h>
73#include <locale.h>
74#include <math.h>
75
76#define DEFAULT_SEPARATOR " "
77#define CNTR_NOT_SUPPORTED "<not supported>"
78#define CNTR_NOT_COUNTED "<not counted>"
79
80static void print_counters(struct timespec *ts, int argc, const char **argv);
81
82/* Default events used for perf stat -T */
83static const char *transaction_attrs = {
84 "task-clock,"
85 "{"
86 "instructions,"
87 "cycles,"
88 "cpu/cycles-t/,"
89 "cpu/tx-start/,"
90 "cpu/el-start/,"
91 "cpu/cycles-ct/"
92 "}"
93};
94
95/* More limited version when the CPU does not have all events. */
96static const char * transaction_limited_attrs = {
97 "task-clock,"
98 "{"
99 "instructions,"
100 "cycles,"
101 "cpu/cycles-t/,"
102 "cpu/tx-start/"
103 "}"
104};
105
106static const char * topdown_attrs[] = {
107 "topdown-total-slots",
108 "topdown-slots-retired",
109 "topdown-recovery-bubbles",
110 "topdown-fetch-bubbles",
111 "topdown-slots-issued",
112 NULL,
113};
114
115static struct perf_evlist *evsel_list;
116
117static struct target target = {
118 .uid = UINT_MAX,
119};
120
121typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
122
123static int run_count = 1;
124static bool no_inherit = false;
125static volatile pid_t child_pid = -1;
126static bool null_run = false;
127static int detailed_run = 0;
128static bool transaction_run;
129static bool topdown_run = false;
130static bool big_num = true;
131static int big_num_opt = -1;
132static const char *csv_sep = NULL;
133static bool csv_output = false;
134static bool group = false;
135static const char *pre_cmd = NULL;
136static const char *post_cmd = NULL;
137static bool sync_run = false;
138static unsigned int initial_delay = 0;
139static unsigned int unit_width = 4; /* strlen("unit") */
140static bool forever = false;
141static bool metric_only = false;
142static bool force_metric_only = false;
143static struct timespec ref_time;
144static struct cpu_map *aggr_map;
145static aggr_get_id_t aggr_get_id;
146static bool append_file;
147static const char *output_name;
148static int output_fd;
149
150struct perf_stat {
151 bool record;
152 struct perf_data_file file;
153 struct perf_session *session;
154 u64 bytes_written;
155 struct perf_tool tool;
156 bool maps_allocated;
157 struct cpu_map *cpus;
158 struct thread_map *threads;
159 enum aggr_mode aggr_mode;
160};
161
162static struct perf_stat perf_stat;
163#define STAT_RECORD perf_stat.record
164
165static volatile int done = 0;
166
167static struct perf_stat_config stat_config = {
168 .aggr_mode = AGGR_GLOBAL,
169 .scale = true,
170};
171
172static inline void diff_timespec(struct timespec *r, struct timespec *a,
173 struct timespec *b)
174{
175 r->tv_sec = a->tv_sec - b->tv_sec;
176 if (a->tv_nsec < b->tv_nsec) {
177 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
178 r->tv_sec--;
179 } else {
180 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
181 }
182}
183
184static void perf_stat__reset_stats(void)
185{
186 perf_evlist__reset_stats(evsel_list);
187 perf_stat__reset_shadow_stats();
188}
189
190static int create_perf_stat_counter(struct perf_evsel *evsel)
191{
192 struct perf_event_attr *attr = &evsel->attr;
193
194 if (stat_config.scale)
195 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
196 PERF_FORMAT_TOTAL_TIME_RUNNING;
197
198 attr->inherit = !no_inherit;
199
200 /*
201 * Some events get initialized with sample_(period/type) set,
202 * like tracepoints. Clear it up for counting.
203 */
204 attr->sample_period = 0;
205
206 /*
207 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
208 * while avoiding that older tools show confusing messages.
209 *
210 * However for pipe sessions we need to keep it zero,
211 * because script's perf_evsel__check_attr is triggered
212 * by attr->sample_type != 0, and we can't run it on
213 * stat sessions.
214 */
215 if (!(STAT_RECORD && perf_stat.file.is_pipe))
216 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
217
218 /*
219 * Disabling all counters initially, they will be enabled
220 * either manually by us or by kernel via enable_on_exec
221 * set later.
222 */
223 if (perf_evsel__is_group_leader(evsel)) {
224 attr->disabled = 1;
225
226 /*
227 * In case of initial_delay we enable tracee
228 * events manually.
229 */
230 if (target__none(&target) && !initial_delay)
231 attr->enable_on_exec = 1;
232 }
233
234 if (target__has_cpu(&target))
235 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
236
237 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
238}
239
240/*
241 * Does the counter have nsecs as a unit?
242 */
243static inline int nsec_counter(struct perf_evsel *evsel)
244{
245 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
246 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
247 return 1;
248
249 return 0;
250}
251
252static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
253 union perf_event *event,
254 struct perf_sample *sample __maybe_unused,
255 struct machine *machine __maybe_unused)
256{
257 if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
258 pr_err("failed to write perf data, error: %m\n");
259 return -1;
260 }
261
262 perf_stat.bytes_written += event->header.size;
263 return 0;
264}
265
266static int write_stat_round_event(u64 tm, u64 type)
267{
268 return perf_event__synthesize_stat_round(NULL, tm, type,
269 process_synthesized_event,
270 NULL);
271}
272
273#define WRITE_STAT_ROUND_EVENT(time, interval) \
274 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
275
276#define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
277
278static int
279perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
280 struct perf_counts_values *count)
281{
282 struct perf_sample_id *sid = SID(counter, cpu, thread);
283
284 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
285 process_synthesized_event, NULL);
286}
287
288/*
289 * Read out the results of a single counter:
290 * do not aggregate counts across CPUs in system-wide mode
291 */
292static int read_counter(struct perf_evsel *counter)
293{
294 int nthreads = thread_map__nr(evsel_list->threads);
295 int ncpus, cpu, thread;
296
297 if (target__has_cpu(&target))
298 ncpus = perf_evsel__nr_cpus(counter);
299 else
300 ncpus = 1;
301
302 if (!counter->supported)
303 return -ENOENT;
304
305 if (counter->system_wide)
306 nthreads = 1;
307
308 for (thread = 0; thread < nthreads; thread++) {
309 for (cpu = 0; cpu < ncpus; cpu++) {
310 struct perf_counts_values *count;
311
312 count = perf_counts(counter->counts, cpu, thread);
313 if (perf_evsel__read(counter, cpu, thread, count))
314 return -1;
315
316 if (STAT_RECORD) {
317 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
318 pr_err("failed to write stat event\n");
319 return -1;
320 }
321 }
322
323 if (verbose > 1) {
324 fprintf(stat_config.output,
325 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
326 perf_evsel__name(counter),
327 cpu,
328 count->val, count->ena, count->run);
329 }
330 }
331 }
332
333 return 0;
334}
335
336static void read_counters(void)
337{
338 struct perf_evsel *counter;
339
340 evlist__for_each_entry(evsel_list, counter) {
341 if (read_counter(counter))
342 pr_debug("failed to read counter %s\n", counter->name);
343
344 if (perf_stat_process_counter(&stat_config, counter))
345 pr_warning("failed to process counter %s\n", counter->name);
346 }
347}
348
349static void process_interval(void)
350{
351 struct timespec ts, rs;
352
353 read_counters();
354
355 clock_gettime(CLOCK_MONOTONIC, &ts);
356 diff_timespec(&rs, &ts, &ref_time);
357
358 if (STAT_RECORD) {
359 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
360 pr_err("failed to write stat round event\n");
361 }
362
363 print_counters(&rs, 0, NULL);
364}
365
366static void enable_counters(void)
367{
368 if (initial_delay)
369 usleep(initial_delay * USEC_PER_MSEC);
370
371 /*
372 * We need to enable counters only if:
373 * - we don't have tracee (attaching to task or cpu)
374 * - we have initial delay configured
375 */
376 if (!target__none(&target) || initial_delay)
377 perf_evlist__enable(evsel_list);
378}
379
380static void disable_counters(void)
381{
382 /*
383 * If we don't have tracee (attaching to task or cpu), counters may
384 * still be running. To get accurate group ratios, we must stop groups
385 * from counting before reading their constituent counters.
386 */
387 if (!target__none(&target))
388 perf_evlist__disable(evsel_list);
389}
390
391static volatile int workload_exec_errno;
392
393/*
394 * perf_evlist__prepare_workload will send a SIGUSR1
395 * if the fork fails, since we asked by setting its
396 * want_signal to true.
397 */
398static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
399 void *ucontext __maybe_unused)
400{
401 workload_exec_errno = info->si_value.sival_int;
402}
403
404static bool has_unit(struct perf_evsel *counter)
405{
406 return counter->unit && *counter->unit;
407}
408
409static bool has_scale(struct perf_evsel *counter)
410{
411 return counter->scale != 1;
412}
413
414static int perf_stat_synthesize_config(bool is_pipe)
415{
416 struct perf_evsel *counter;
417 int err;
418
419 if (is_pipe) {
420 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
421 process_synthesized_event);
422 if (err < 0) {
423 pr_err("Couldn't synthesize attrs.\n");
424 return err;
425 }
426 }
427
428 /*
429 * Synthesize other events stuff not carried within
430 * attr event - unit, scale, name
431 */
432 evlist__for_each_entry(evsel_list, counter) {
433 if (!counter->supported)
434 continue;
435
436 /*
437 * Synthesize unit and scale only if it's defined.
438 */
439 if (has_unit(counter)) {
440 err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
441 if (err < 0) {
442 pr_err("Couldn't synthesize evsel unit.\n");
443 return err;
444 }
445 }
446
447 if (has_scale(counter)) {
448 err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
449 if (err < 0) {
450 pr_err("Couldn't synthesize evsel scale.\n");
451 return err;
452 }
453 }
454
455 if (counter->own_cpus) {
456 err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
457 if (err < 0) {
458 pr_err("Couldn't synthesize evsel scale.\n");
459 return err;
460 }
461 }
462
463 /*
464 * Name is needed only for pipe output,
465 * perf.data carries event names.
466 */
467 if (is_pipe) {
468 err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
469 if (err < 0) {
470 pr_err("Couldn't synthesize evsel name.\n");
471 return err;
472 }
473 }
474 }
475
476 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
477 process_synthesized_event,
478 NULL);
479 if (err < 0) {
480 pr_err("Couldn't synthesize thread map.\n");
481 return err;
482 }
483
484 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
485 process_synthesized_event, NULL);
486 if (err < 0) {
487 pr_err("Couldn't synthesize thread map.\n");
488 return err;
489 }
490
491 err = perf_event__synthesize_stat_config(NULL, &stat_config,
492 process_synthesized_event, NULL);
493 if (err < 0) {
494 pr_err("Couldn't synthesize config.\n");
495 return err;
496 }
497
498 return 0;
499}
500
501#define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
502
503static int __store_counter_ids(struct perf_evsel *counter,
504 struct cpu_map *cpus,
505 struct thread_map *threads)
506{
507 int cpu, thread;
508
509 for (cpu = 0; cpu < cpus->nr; cpu++) {
510 for (thread = 0; thread < threads->nr; thread++) {
511 int fd = FD(counter, cpu, thread);
512
513 if (perf_evlist__id_add_fd(evsel_list, counter,
514 cpu, thread, fd) < 0)
515 return -1;
516 }
517 }
518
519 return 0;
520}
521
522static int store_counter_ids(struct perf_evsel *counter)
523{
524 struct cpu_map *cpus = counter->cpus;
525 struct thread_map *threads = counter->threads;
526
527 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
528 return -ENOMEM;
529
530 return __store_counter_ids(counter, cpus, threads);
531}
532
533static int __run_perf_stat(int argc, const char **argv)
534{
535 int interval = stat_config.interval;
536 char msg[512];
537 unsigned long long t0, t1;
538 struct perf_evsel *counter;
539 struct timespec ts;
540 size_t l;
541 int status = 0;
542 const bool forks = (argc > 0);
543 bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
544 struct perf_evsel_config_term *err_term;
545
546 if (interval) {
547 ts.tv_sec = interval / USEC_PER_MSEC;
548 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
549 } else {
550 ts.tv_sec = 1;
551 ts.tv_nsec = 0;
552 }
553
554 if (forks) {
555 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
556 workload_exec_failed_signal) < 0) {
557 perror("failed to prepare workload");
558 return -1;
559 }
560 child_pid = evsel_list->workload.pid;
561 }
562
563 if (group)
564 perf_evlist__set_leader(evsel_list);
565
566 evlist__for_each_entry(evsel_list, counter) {
567try_again:
568 if (create_perf_stat_counter(counter) < 0) {
569 /*
570 * PPC returns ENXIO for HW counters until 2.6.37
571 * (behavior changed with commit b0a873e).
572 */
573 if (errno == EINVAL || errno == ENOSYS ||
574 errno == ENOENT || errno == EOPNOTSUPP ||
575 errno == ENXIO) {
576 if (verbose)
577 ui__warning("%s event is not supported by the kernel.\n",
578 perf_evsel__name(counter));
579 counter->supported = false;
580
581 if ((counter->leader != counter) ||
582 !(counter->leader->nr_members > 1))
583 continue;
584 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
585 if (verbose)
586 ui__warning("%s\n", msg);
587 goto try_again;
588 }
589
590 perf_evsel__open_strerror(counter, &target,
591 errno, msg, sizeof(msg));
592 ui__error("%s\n", msg);
593
594 if (child_pid != -1)
595 kill(child_pid, SIGTERM);
596
597 return -1;
598 }
599 counter->supported = true;
600
601 l = strlen(counter->unit);
602 if (l > unit_width)
603 unit_width = l;
604
605 if (STAT_RECORD && store_counter_ids(counter))
606 return -1;
607 }
608
609 if (perf_evlist__apply_filters(evsel_list, &counter)) {
610 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
611 counter->filter, perf_evsel__name(counter), errno,
612 str_error_r(errno, msg, sizeof(msg)));
613 return -1;
614 }
615
616 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
617 error("failed to set config \"%s\" on event %s with %d (%s)\n",
618 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
619 str_error_r(errno, msg, sizeof(msg)));
620 return -1;
621 }
622
623 if (STAT_RECORD) {
624 int err, fd = perf_data_file__fd(&perf_stat.file);
625
626 if (is_pipe) {
627 err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
628 } else {
629 err = perf_session__write_header(perf_stat.session, evsel_list,
630 fd, false);
631 }
632
633 if (err < 0)
634 return err;
635
636 err = perf_stat_synthesize_config(is_pipe);
637 if (err < 0)
638 return err;
639 }
640
641 /*
642 * Enable counters and exec the command:
643 */
644 t0 = rdclock();
645 clock_gettime(CLOCK_MONOTONIC, &ref_time);
646
647 if (forks) {
648 perf_evlist__start_workload(evsel_list);
649 enable_counters();
650
651 if (interval) {
652 while (!waitpid(child_pid, &status, WNOHANG)) {
653 nanosleep(&ts, NULL);
654 process_interval();
655 }
656 }
657 wait(&status);
658
659 if (workload_exec_errno) {
660 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
661 pr_err("Workload failed: %s\n", emsg);
662 return -1;
663 }
664
665 if (WIFSIGNALED(status))
666 psignal(WTERMSIG(status), argv[0]);
667 } else {
668 enable_counters();
669 while (!done) {
670 nanosleep(&ts, NULL);
671 if (interval)
672 process_interval();
673 }
674 }
675
676 disable_counters();
677
678 t1 = rdclock();
679
680 update_stats(&walltime_nsecs_stats, t1 - t0);
681
682 /*
683 * Closing a group leader splits the group, and as we only disable
684 * group leaders, results in remaining events becoming enabled. To
685 * avoid arbitrary skew, we must read all counters before closing any
686 * group leaders.
687 */
688 read_counters();
689 perf_evlist__close(evsel_list);
690
691 return WEXITSTATUS(status);
692}
693
694static int run_perf_stat(int argc, const char **argv)
695{
696 int ret;
697
698 if (pre_cmd) {
699 ret = system(pre_cmd);
700 if (ret)
701 return ret;
702 }
703
704 if (sync_run)
705 sync();
706
707 ret = __run_perf_stat(argc, argv);
708 if (ret)
709 return ret;
710
711 if (post_cmd) {
712 ret = system(post_cmd);
713 if (ret)
714 return ret;
715 }
716
717 return ret;
718}
719
720static void print_running(u64 run, u64 ena)
721{
722 if (csv_output) {
723 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
724 csv_sep,
725 run,
726 csv_sep,
727 ena ? 100.0 * run / ena : 100.0);
728 } else if (run != ena) {
729 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
730 }
731}
732
733static void print_noise_pct(double total, double avg)
734{
735 double pct = rel_stddev_stats(total, avg);
736
737 if (csv_output)
738 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
739 else if (pct)
740 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
741}
742
743static void print_noise(struct perf_evsel *evsel, double avg)
744{
745 struct perf_stat_evsel *ps;
746
747 if (run_count == 1)
748 return;
749
750 ps = evsel->priv;
751 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
752}
753
754static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
755{
756 switch (stat_config.aggr_mode) {
757 case AGGR_CORE:
758 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
759 cpu_map__id_to_socket(id),
760 csv_output ? 0 : -8,
761 cpu_map__id_to_cpu(id),
762 csv_sep,
763 csv_output ? 0 : 4,
764 nr,
765 csv_sep);
766 break;
767 case AGGR_SOCKET:
768 fprintf(stat_config.output, "S%*d%s%*d%s",
769 csv_output ? 0 : -5,
770 id,
771 csv_sep,
772 csv_output ? 0 : 4,
773 nr,
774 csv_sep);
775 break;
776 case AGGR_NONE:
777 fprintf(stat_config.output, "CPU%*d%s",
778 csv_output ? 0 : -4,
779 perf_evsel__cpus(evsel)->map[id], csv_sep);
780 break;
781 case AGGR_THREAD:
782 fprintf(stat_config.output, "%*s-%*d%s",
783 csv_output ? 0 : 16,
784 thread_map__comm(evsel->threads, id),
785 csv_output ? 0 : -8,
786 thread_map__pid(evsel->threads, id),
787 csv_sep);
788 break;
789 case AGGR_GLOBAL:
790 case AGGR_UNSET:
791 default:
792 break;
793 }
794}
795
796struct outstate {
797 FILE *fh;
798 bool newline;
799 const char *prefix;
800 int nfields;
801 int id, nr;
802 struct perf_evsel *evsel;
803};
804
805#define METRIC_LEN 35
806
807static void new_line_std(void *ctx)
808{
809 struct outstate *os = ctx;
810
811 os->newline = true;
812}
813
814static void do_new_line_std(struct outstate *os)
815{
816 fputc('\n', os->fh);
817 fputs(os->prefix, os->fh);
818 aggr_printout(os->evsel, os->id, os->nr);
819 if (stat_config.aggr_mode == AGGR_NONE)
820 fprintf(os->fh, " ");
821 fprintf(os->fh, " ");
822}
823
824static void print_metric_std(void *ctx, const char *color, const char *fmt,
825 const char *unit, double val)
826{
827 struct outstate *os = ctx;
828 FILE *out = os->fh;
829 int n;
830 bool newline = os->newline;
831
832 os->newline = false;
833
834 if (unit == NULL || fmt == NULL) {
835 fprintf(out, "%-*s", METRIC_LEN, "");
836 return;
837 }
838
839 if (newline)
840 do_new_line_std(os);
841
842 n = fprintf(out, " # ");
843 if (color)
844 n += color_fprintf(out, color, fmt, val);
845 else
846 n += fprintf(out, fmt, val);
847 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
848}
849
850static void new_line_csv(void *ctx)
851{
852 struct outstate *os = ctx;
853 int i;
854
855 fputc('\n', os->fh);
856 if (os->prefix)
857 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
858 aggr_printout(os->evsel, os->id, os->nr);
859 for (i = 0; i < os->nfields; i++)
860 fputs(csv_sep, os->fh);
861}
862
863static void print_metric_csv(void *ctx,
864 const char *color __maybe_unused,
865 const char *fmt, const char *unit, double val)
866{
867 struct outstate *os = ctx;
868 FILE *out = os->fh;
869 char buf[64], *vals, *ends;
870
871 if (unit == NULL || fmt == NULL) {
872 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
873 return;
874 }
875 snprintf(buf, sizeof(buf), fmt, val);
876 vals = buf;
877 while (isspace(*vals))
878 vals++;
879 ends = vals;
880 while (isdigit(*ends) || *ends == '.')
881 ends++;
882 *ends = 0;
883 while (isspace(*unit))
884 unit++;
885 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
886}
887
888#define METRIC_ONLY_LEN 20
889
890/* Filter out some columns that don't work well in metrics only mode */
891
892static bool valid_only_metric(const char *unit)
893{
894 if (!unit)
895 return false;
896 if (strstr(unit, "/sec") ||
897 strstr(unit, "hz") ||
898 strstr(unit, "Hz") ||
899 strstr(unit, "CPUs utilized"))
900 return false;
901 return true;
902}
903
904static const char *fixunit(char *buf, struct perf_evsel *evsel,
905 const char *unit)
906{
907 if (!strncmp(unit, "of all", 6)) {
908 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
909 unit);
910 return buf;
911 }
912 return unit;
913}
914
915static void print_metric_only(void *ctx, const char *color, const char *fmt,
916 const char *unit, double val)
917{
918 struct outstate *os = ctx;
919 FILE *out = os->fh;
920 int n;
921 char buf[1024];
922 unsigned mlen = METRIC_ONLY_LEN;
923
924 if (!valid_only_metric(unit))
925 return;
926 unit = fixunit(buf, os->evsel, unit);
927 if (color)
928 n = color_fprintf(out, color, fmt, val);
929 else
930 n = fprintf(out, fmt, val);
931 if (n > METRIC_ONLY_LEN)
932 n = METRIC_ONLY_LEN;
933 if (mlen < strlen(unit))
934 mlen = strlen(unit) + 1;
935 fprintf(out, "%*s", mlen - n, "");
936}
937
938static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
939 const char *fmt,
940 const char *unit, double val)
941{
942 struct outstate *os = ctx;
943 FILE *out = os->fh;
944 char buf[64], *vals, *ends;
945 char tbuf[1024];
946
947 if (!valid_only_metric(unit))
948 return;
949 unit = fixunit(tbuf, os->evsel, unit);
950 snprintf(buf, sizeof buf, fmt, val);
951 vals = buf;
952 while (isspace(*vals))
953 vals++;
954 ends = vals;
955 while (isdigit(*ends) || *ends == '.')
956 ends++;
957 *ends = 0;
958 fprintf(out, "%s%s", vals, csv_sep);
959}
960
961static void new_line_metric(void *ctx __maybe_unused)
962{
963}
964
965static void print_metric_header(void *ctx, const char *color __maybe_unused,
966 const char *fmt __maybe_unused,
967 const char *unit, double val __maybe_unused)
968{
969 struct outstate *os = ctx;
970 char tbuf[1024];
971
972 if (!valid_only_metric(unit))
973 return;
974 unit = fixunit(tbuf, os->evsel, unit);
975 if (csv_output)
976 fprintf(os->fh, "%s%s", unit, csv_sep);
977 else
978 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
979}
980
981static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
982{
983 FILE *output = stat_config.output;
984 double msecs = avg / NSEC_PER_MSEC;
985 const char *fmt_v, *fmt_n;
986 char name[25];
987
988 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
989 fmt_n = csv_output ? "%s" : "%-25s";
990
991 aggr_printout(evsel, id, nr);
992
993 scnprintf(name, sizeof(name), "%s%s",
994 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
995
996 fprintf(output, fmt_v, msecs, csv_sep);
997
998 if (csv_output)
999 fprintf(output, "%s%s", evsel->unit, csv_sep);
1000 else
1001 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
1002
1003 fprintf(output, fmt_n, name);
1004
1005 if (evsel->cgrp)
1006 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1007}
1008
1009static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1010{
1011 int i;
1012
1013 if (!aggr_get_id)
1014 return 0;
1015
1016 if (stat_config.aggr_mode == AGGR_NONE)
1017 return id;
1018
1019 if (stat_config.aggr_mode == AGGR_GLOBAL)
1020 return 0;
1021
1022 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1023 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1024
1025 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1026 return cpu2;
1027 }
1028 return 0;
1029}
1030
1031static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1032{
1033 FILE *output = stat_config.output;
1034 double sc = evsel->scale;
1035 const char *fmt;
1036
1037 if (csv_output) {
1038 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
1039 } else {
1040 if (big_num)
1041 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1042 else
1043 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1044 }
1045
1046 aggr_printout(evsel, id, nr);
1047
1048 fprintf(output, fmt, avg, csv_sep);
1049
1050 if (evsel->unit)
1051 fprintf(output, "%-*s%s",
1052 csv_output ? 0 : unit_width,
1053 evsel->unit, csv_sep);
1054
1055 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1056
1057 if (evsel->cgrp)
1058 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1059}
1060
1061static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1062 char *prefix, u64 run, u64 ena, double noise)
1063{
1064 struct perf_stat_output_ctx out;
1065 struct outstate os = {
1066 .fh = stat_config.output,
1067 .prefix = prefix ? prefix : "",
1068 .id = id,
1069 .nr = nr,
1070 .evsel = counter,
1071 };
1072 print_metric_t pm = print_metric_std;
1073 void (*nl)(void *);
1074
1075 if (metric_only) {
1076 nl = new_line_metric;
1077 if (csv_output)
1078 pm = print_metric_only_csv;
1079 else
1080 pm = print_metric_only;
1081 } else
1082 nl = new_line_std;
1083
1084 if (csv_output && !metric_only) {
1085 static int aggr_fields[] = {
1086 [AGGR_GLOBAL] = 0,
1087 [AGGR_THREAD] = 1,
1088 [AGGR_NONE] = 1,
1089 [AGGR_SOCKET] = 2,
1090 [AGGR_CORE] = 2,
1091 };
1092
1093 pm = print_metric_csv;
1094 nl = new_line_csv;
1095 os.nfields = 3;
1096 os.nfields += aggr_fields[stat_config.aggr_mode];
1097 if (counter->cgrp)
1098 os.nfields++;
1099 }
1100 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1101 if (metric_only) {
1102 pm(&os, NULL, "", "", 0);
1103 return;
1104 }
1105 aggr_printout(counter, id, nr);
1106
1107 fprintf(stat_config.output, "%*s%s",
1108 csv_output ? 0 : 18,
1109 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1110 csv_sep);
1111
1112 fprintf(stat_config.output, "%-*s%s",
1113 csv_output ? 0 : unit_width,
1114 counter->unit, csv_sep);
1115
1116 fprintf(stat_config.output, "%*s",
1117 csv_output ? 0 : -25,
1118 perf_evsel__name(counter));
1119
1120 if (counter->cgrp)
1121 fprintf(stat_config.output, "%s%s",
1122 csv_sep, counter->cgrp->name);
1123
1124 if (!csv_output)
1125 pm(&os, NULL, NULL, "", 0);
1126 print_noise(counter, noise);
1127 print_running(run, ena);
1128 if (csv_output)
1129 pm(&os, NULL, NULL, "", 0);
1130 return;
1131 }
1132
1133 if (metric_only)
1134 /* nothing */;
1135 else if (nsec_counter(counter))
1136 nsec_printout(id, nr, counter, uval);
1137 else
1138 abs_printout(id, nr, counter, uval);
1139
1140 out.print_metric = pm;
1141 out.new_line = nl;
1142 out.ctx = &os;
1143
1144 if (csv_output && !metric_only) {
1145 print_noise(counter, noise);
1146 print_running(run, ena);
1147 }
1148
1149 perf_stat__print_shadow_stats(counter, uval,
1150 first_shadow_cpu(counter, id),
1151 &out);
1152 if (!csv_output && !metric_only) {
1153 print_noise(counter, noise);
1154 print_running(run, ena);
1155 }
1156}
1157
1158static void aggr_update_shadow(void)
1159{
1160 int cpu, s2, id, s;
1161 u64 val;
1162 struct perf_evsel *counter;
1163
1164 for (s = 0; s < aggr_map->nr; s++) {
1165 id = aggr_map->map[s];
1166 evlist__for_each_entry(evsel_list, counter) {
1167 val = 0;
1168 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1169 s2 = aggr_get_id(evsel_list->cpus, cpu);
1170 if (s2 != id)
1171 continue;
1172 val += perf_counts(counter->counts, cpu, 0)->val;
1173 }
1174 val = val * counter->scale;
1175 perf_stat__update_shadow_stats(counter, &val,
1176 first_shadow_cpu(counter, id));
1177 }
1178 }
1179}
1180
1181static void print_aggr(char *prefix)
1182{
1183 FILE *output = stat_config.output;
1184 struct perf_evsel *counter;
1185 int cpu, s, s2, id, nr;
1186 double uval;
1187 u64 ena, run, val;
1188 bool first;
1189
1190 if (!(aggr_map || aggr_get_id))
1191 return;
1192
1193 aggr_update_shadow();
1194
1195 /*
1196 * With metric_only everything is on a single line.
1197 * Without each counter has its own line.
1198 */
1199 for (s = 0; s < aggr_map->nr; s++) {
1200 if (prefix && metric_only)
1201 fprintf(output, "%s", prefix);
1202
1203 id = aggr_map->map[s];
1204 first = true;
1205 evlist__for_each_entry(evsel_list, counter) {
1206 val = ena = run = 0;
1207 nr = 0;
1208 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1209 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1210 if (s2 != id)
1211 continue;
1212 val += perf_counts(counter->counts, cpu, 0)->val;
1213 ena += perf_counts(counter->counts, cpu, 0)->ena;
1214 run += perf_counts(counter->counts, cpu, 0)->run;
1215 nr++;
1216 }
1217 if (first && metric_only) {
1218 first = false;
1219 aggr_printout(counter, id, nr);
1220 }
1221 if (prefix && !metric_only)
1222 fprintf(output, "%s", prefix);
1223
1224 uval = val * counter->scale;
1225 printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1226 if (!metric_only)
1227 fputc('\n', output);
1228 }
1229 if (metric_only)
1230 fputc('\n', output);
1231 }
1232}
1233
1234static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1235{
1236 FILE *output = stat_config.output;
1237 int nthreads = thread_map__nr(counter->threads);
1238 int ncpus = cpu_map__nr(counter->cpus);
1239 int cpu, thread;
1240 double uval;
1241
1242 for (thread = 0; thread < nthreads; thread++) {
1243 u64 ena = 0, run = 0, val = 0;
1244
1245 for (cpu = 0; cpu < ncpus; cpu++) {
1246 val += perf_counts(counter->counts, cpu, thread)->val;
1247 ena += perf_counts(counter->counts, cpu, thread)->ena;
1248 run += perf_counts(counter->counts, cpu, thread)->run;
1249 }
1250
1251 if (prefix)
1252 fprintf(output, "%s", prefix);
1253
1254 uval = val * counter->scale;
1255 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1256 fputc('\n', output);
1257 }
1258}
1259
1260/*
1261 * Print out the results of a single counter:
1262 * aggregated counts in system-wide mode
1263 */
1264static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1265{
1266 FILE *output = stat_config.output;
1267 struct perf_stat_evsel *ps = counter->priv;
1268 double avg = avg_stats(&ps->res_stats[0]);
1269 double uval;
1270 double avg_enabled, avg_running;
1271
1272 avg_enabled = avg_stats(&ps->res_stats[1]);
1273 avg_running = avg_stats(&ps->res_stats[2]);
1274
1275 if (prefix && !metric_only)
1276 fprintf(output, "%s", prefix);
1277
1278 uval = avg * counter->scale;
1279 printout(-1, 0, counter, uval, prefix, avg_running, avg_enabled, avg);
1280 if (!metric_only)
1281 fprintf(output, "\n");
1282}
1283
1284/*
1285 * Print out the results of a single counter:
1286 * does not use aggregated count in system-wide
1287 */
1288static void print_counter(struct perf_evsel *counter, char *prefix)
1289{
1290 FILE *output = stat_config.output;
1291 u64 ena, run, val;
1292 double uval;
1293 int cpu;
1294
1295 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1296 val = perf_counts(counter->counts, cpu, 0)->val;
1297 ena = perf_counts(counter->counts, cpu, 0)->ena;
1298 run = perf_counts(counter->counts, cpu, 0)->run;
1299
1300 if (prefix)
1301 fprintf(output, "%s", prefix);
1302
1303 uval = val * counter->scale;
1304 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1305
1306 fputc('\n', output);
1307 }
1308}
1309
1310static void print_no_aggr_metric(char *prefix)
1311{
1312 int cpu;
1313 int nrcpus = 0;
1314 struct perf_evsel *counter;
1315 u64 ena, run, val;
1316 double uval;
1317
1318 nrcpus = evsel_list->cpus->nr;
1319 for (cpu = 0; cpu < nrcpus; cpu++) {
1320 bool first = true;
1321
1322 if (prefix)
1323 fputs(prefix, stat_config.output);
1324 evlist__for_each_entry(evsel_list, counter) {
1325 if (first) {
1326 aggr_printout(counter, cpu, 0);
1327 first = false;
1328 }
1329 val = perf_counts(counter->counts, cpu, 0)->val;
1330 ena = perf_counts(counter->counts, cpu, 0)->ena;
1331 run = perf_counts(counter->counts, cpu, 0)->run;
1332
1333 uval = val * counter->scale;
1334 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1335 }
1336 fputc('\n', stat_config.output);
1337 }
1338}
1339
1340static int aggr_header_lens[] = {
1341 [AGGR_CORE] = 18,
1342 [AGGR_SOCKET] = 12,
1343 [AGGR_NONE] = 6,
1344 [AGGR_THREAD] = 24,
1345 [AGGR_GLOBAL] = 0,
1346};
1347
1348static const char *aggr_header_csv[] = {
1349 [AGGR_CORE] = "core,cpus,",
1350 [AGGR_SOCKET] = "socket,cpus",
1351 [AGGR_NONE] = "cpu,",
1352 [AGGR_THREAD] = "comm-pid,",
1353 [AGGR_GLOBAL] = ""
1354};
1355
1356static void print_metric_headers(const char *prefix, bool no_indent)
1357{
1358 struct perf_stat_output_ctx out;
1359 struct perf_evsel *counter;
1360 struct outstate os = {
1361 .fh = stat_config.output
1362 };
1363
1364 if (prefix)
1365 fprintf(stat_config.output, "%s", prefix);
1366
1367 if (!csv_output && !no_indent)
1368 fprintf(stat_config.output, "%*s",
1369 aggr_header_lens[stat_config.aggr_mode], "");
1370 if (csv_output) {
1371 if (stat_config.interval)
1372 fputs("time,", stat_config.output);
1373 fputs(aggr_header_csv[stat_config.aggr_mode],
1374 stat_config.output);
1375 }
1376
1377 /* Print metrics headers only */
1378 evlist__for_each_entry(evsel_list, counter) {
1379 os.evsel = counter;
1380 out.ctx = &os;
1381 out.print_metric = print_metric_header;
1382 out.new_line = new_line_metric;
1383 os.evsel = counter;
1384 perf_stat__print_shadow_stats(counter, 0,
1385 0,
1386 &out);
1387 }
1388 fputc('\n', stat_config.output);
1389}
1390
1391static void print_interval(char *prefix, struct timespec *ts)
1392{
1393 FILE *output = stat_config.output;
1394 static int num_print_interval;
1395
1396 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1397
1398 if (num_print_interval == 0 && !csv_output) {
1399 switch (stat_config.aggr_mode) {
1400 case AGGR_SOCKET:
1401 fprintf(output, "# time socket cpus");
1402 if (!metric_only)
1403 fprintf(output, " counts %*s events\n", unit_width, "unit");
1404 break;
1405 case AGGR_CORE:
1406 fprintf(output, "# time core cpus");
1407 if (!metric_only)
1408 fprintf(output, " counts %*s events\n", unit_width, "unit");
1409 break;
1410 case AGGR_NONE:
1411 fprintf(output, "# time CPU");
1412 if (!metric_only)
1413 fprintf(output, " counts %*s events\n", unit_width, "unit");
1414 break;
1415 case AGGR_THREAD:
1416 fprintf(output, "# time comm-pid");
1417 if (!metric_only)
1418 fprintf(output, " counts %*s events\n", unit_width, "unit");
1419 break;
1420 case AGGR_GLOBAL:
1421 default:
1422 fprintf(output, "# time");
1423 if (!metric_only)
1424 fprintf(output, " counts %*s events\n", unit_width, "unit");
1425 case AGGR_UNSET:
1426 break;
1427 }
1428 }
1429
1430 if (num_print_interval == 0 && metric_only)
1431 print_metric_headers(" ", true);
1432 if (++num_print_interval == 25)
1433 num_print_interval = 0;
1434}
1435
1436static void print_header(int argc, const char **argv)
1437{
1438 FILE *output = stat_config.output;
1439 int i;
1440
1441 fflush(stdout);
1442
1443 if (!csv_output) {
1444 fprintf(output, "\n");
1445 fprintf(output, " Performance counter stats for ");
1446 if (target.system_wide)
1447 fprintf(output, "\'system wide");
1448 else if (target.cpu_list)
1449 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1450 else if (!target__has_task(&target)) {
1451 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1452 for (i = 1; argv && (i < argc); i++)
1453 fprintf(output, " %s", argv[i]);
1454 } else if (target.pid)
1455 fprintf(output, "process id \'%s", target.pid);
1456 else
1457 fprintf(output, "thread id \'%s", target.tid);
1458
1459 fprintf(output, "\'");
1460 if (run_count > 1)
1461 fprintf(output, " (%d runs)", run_count);
1462 fprintf(output, ":\n\n");
1463 }
1464}
1465
1466static void print_footer(void)
1467{
1468 FILE *output = stat_config.output;
1469
1470 if (!null_run)
1471 fprintf(output, "\n");
1472 fprintf(output, " %17.9f seconds time elapsed",
1473 avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC);
1474 if (run_count > 1) {
1475 fprintf(output, " ");
1476 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1477 avg_stats(&walltime_nsecs_stats));
1478 }
1479 fprintf(output, "\n\n");
1480}
1481
1482static void print_counters(struct timespec *ts, int argc, const char **argv)
1483{
1484 int interval = stat_config.interval;
1485 struct perf_evsel *counter;
1486 char buf[64], *prefix = NULL;
1487
1488 /* Do not print anything if we record to the pipe. */
1489 if (STAT_RECORD && perf_stat.file.is_pipe)
1490 return;
1491
1492 if (interval)
1493 print_interval(prefix = buf, ts);
1494 else
1495 print_header(argc, argv);
1496
1497 if (metric_only) {
1498 static int num_print_iv;
1499
1500 if (num_print_iv == 0 && !interval)
1501 print_metric_headers(prefix, false);
1502 if (num_print_iv++ == 25)
1503 num_print_iv = 0;
1504 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1505 fprintf(stat_config.output, "%s", prefix);
1506 }
1507
1508 switch (stat_config.aggr_mode) {
1509 case AGGR_CORE:
1510 case AGGR_SOCKET:
1511 print_aggr(prefix);
1512 break;
1513 case AGGR_THREAD:
1514 evlist__for_each_entry(evsel_list, counter)
1515 print_aggr_thread(counter, prefix);
1516 break;
1517 case AGGR_GLOBAL:
1518 evlist__for_each_entry(evsel_list, counter)
1519 print_counter_aggr(counter, prefix);
1520 if (metric_only)
1521 fputc('\n', stat_config.output);
1522 break;
1523 case AGGR_NONE:
1524 if (metric_only)
1525 print_no_aggr_metric(prefix);
1526 else {
1527 evlist__for_each_entry(evsel_list, counter)
1528 print_counter(counter, prefix);
1529 }
1530 break;
1531 case AGGR_UNSET:
1532 default:
1533 break;
1534 }
1535
1536 if (!interval && !csv_output)
1537 print_footer();
1538
1539 fflush(stat_config.output);
1540}
1541
1542static volatile int signr = -1;
1543
1544static void skip_signal(int signo)
1545{
1546 if ((child_pid == -1) || stat_config.interval)
1547 done = 1;
1548
1549 signr = signo;
1550 /*
1551 * render child_pid harmless
1552 * won't send SIGTERM to a random
1553 * process in case of race condition
1554 * and fast PID recycling
1555 */
1556 child_pid = -1;
1557}
1558
1559static void sig_atexit(void)
1560{
1561 sigset_t set, oset;
1562
1563 /*
1564 * avoid race condition with SIGCHLD handler
1565 * in skip_signal() which is modifying child_pid
1566 * goal is to avoid send SIGTERM to a random
1567 * process
1568 */
1569 sigemptyset(&set);
1570 sigaddset(&set, SIGCHLD);
1571 sigprocmask(SIG_BLOCK, &set, &oset);
1572
1573 if (child_pid != -1)
1574 kill(child_pid, SIGTERM);
1575
1576 sigprocmask(SIG_SETMASK, &oset, NULL);
1577
1578 if (signr == -1)
1579 return;
1580
1581 signal(signr, SIG_DFL);
1582 kill(getpid(), signr);
1583}
1584
1585static int stat__set_big_num(const struct option *opt __maybe_unused,
1586 const char *s __maybe_unused, int unset)
1587{
1588 big_num_opt = unset ? 0 : 1;
1589 return 0;
1590}
1591
1592static int enable_metric_only(const struct option *opt __maybe_unused,
1593 const char *s __maybe_unused, int unset)
1594{
1595 force_metric_only = true;
1596 metric_only = !unset;
1597 return 0;
1598}
1599
1600static const struct option stat_options[] = {
1601 OPT_BOOLEAN('T', "transaction", &transaction_run,
1602 "hardware transaction statistics"),
1603 OPT_CALLBACK('e', "event", &evsel_list, "event",
1604 "event selector. use 'perf list' to list available events",
1605 parse_events_option),
1606 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1607 "event filter", parse_filter),
1608 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1609 "child tasks do not inherit counters"),
1610 OPT_STRING('p', "pid", &target.pid, "pid",
1611 "stat events on existing process id"),
1612 OPT_STRING('t', "tid", &target.tid, "tid",
1613 "stat events on existing thread id"),
1614 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1615 "system-wide collection from all CPUs"),
1616 OPT_BOOLEAN('g', "group", &group,
1617 "put the counters into a counter group"),
1618 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1619 OPT_INCR('v', "verbose", &verbose,
1620 "be more verbose (show counter open errors, etc)"),
1621 OPT_INTEGER('r', "repeat", &run_count,
1622 "repeat command and print average + stddev (max: 100, forever: 0)"),
1623 OPT_BOOLEAN('n', "null", &null_run,
1624 "null run - dont start any counters"),
1625 OPT_INCR('d', "detailed", &detailed_run,
1626 "detailed run - start a lot of events"),
1627 OPT_BOOLEAN('S', "sync", &sync_run,
1628 "call sync() before starting a run"),
1629 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1630 "print large numbers with thousands\' separators",
1631 stat__set_big_num),
1632 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1633 "list of cpus to monitor in system-wide"),
1634 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1635 "disable CPU count aggregation", AGGR_NONE),
1636 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1637 "print counts with custom separator"),
1638 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1639 "monitor event in cgroup name only", parse_cgroups),
1640 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1641 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1642 OPT_INTEGER(0, "log-fd", &output_fd,
1643 "log output to fd, instead of stderr"),
1644 OPT_STRING(0, "pre", &pre_cmd, "command",
1645 "command to run prior to the measured command"),
1646 OPT_STRING(0, "post", &post_cmd, "command",
1647 "command to run after to the measured command"),
1648 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1649 "print counts at regular interval in ms (>= 10)"),
1650 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1651 "aggregate counts per processor socket", AGGR_SOCKET),
1652 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1653 "aggregate counts per physical processor core", AGGR_CORE),
1654 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1655 "aggregate counts per thread", AGGR_THREAD),
1656 OPT_UINTEGER('D', "delay", &initial_delay,
1657 "ms to wait before starting measurement after program start"),
1658 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1659 "Only print computed metrics. No raw values", enable_metric_only),
1660 OPT_BOOLEAN(0, "topdown", &topdown_run,
1661 "measure topdown level 1 statistics"),
1662 OPT_END()
1663};
1664
1665static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1666{
1667 return cpu_map__get_socket(map, cpu, NULL);
1668}
1669
1670static int perf_stat__get_core(struct cpu_map *map, int cpu)
1671{
1672 return cpu_map__get_core(map, cpu, NULL);
1673}
1674
1675static int cpu_map__get_max(struct cpu_map *map)
1676{
1677 int i, max = -1;
1678
1679 for (i = 0; i < map->nr; i++) {
1680 if (map->map[i] > max)
1681 max = map->map[i];
1682 }
1683
1684 return max;
1685}
1686
1687static struct cpu_map *cpus_aggr_map;
1688
1689static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1690{
1691 int cpu;
1692
1693 if (idx >= map->nr)
1694 return -1;
1695
1696 cpu = map->map[idx];
1697
1698 if (cpus_aggr_map->map[cpu] == -1)
1699 cpus_aggr_map->map[cpu] = get_id(map, idx);
1700
1701 return cpus_aggr_map->map[cpu];
1702}
1703
1704static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1705{
1706 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1707}
1708
1709static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1710{
1711 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1712}
1713
1714static int perf_stat_init_aggr_mode(void)
1715{
1716 int nr;
1717
1718 switch (stat_config.aggr_mode) {
1719 case AGGR_SOCKET:
1720 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1721 perror("cannot build socket map");
1722 return -1;
1723 }
1724 aggr_get_id = perf_stat__get_socket_cached;
1725 break;
1726 case AGGR_CORE:
1727 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1728 perror("cannot build core map");
1729 return -1;
1730 }
1731 aggr_get_id = perf_stat__get_core_cached;
1732 break;
1733 case AGGR_NONE:
1734 case AGGR_GLOBAL:
1735 case AGGR_THREAD:
1736 case AGGR_UNSET:
1737 default:
1738 break;
1739 }
1740
1741 /*
1742 * The evsel_list->cpus is the base we operate on,
1743 * taking the highest cpu number to be the size of
1744 * the aggregation translate cpumap.
1745 */
1746 nr = cpu_map__get_max(evsel_list->cpus);
1747 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1748 return cpus_aggr_map ? 0 : -ENOMEM;
1749}
1750
1751static void perf_stat__exit_aggr_mode(void)
1752{
1753 cpu_map__put(aggr_map);
1754 cpu_map__put(cpus_aggr_map);
1755 aggr_map = NULL;
1756 cpus_aggr_map = NULL;
1757}
1758
1759static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1760{
1761 int cpu;
1762
1763 if (idx > map->nr)
1764 return -1;
1765
1766 cpu = map->map[idx];
1767
1768 if (cpu >= env->nr_cpus_online)
1769 return -1;
1770
1771 return cpu;
1772}
1773
1774static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1775{
1776 struct perf_env *env = data;
1777 int cpu = perf_env__get_cpu(env, map, idx);
1778
1779 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1780}
1781
1782static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1783{
1784 struct perf_env *env = data;
1785 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1786
1787 if (cpu != -1) {
1788 int socket_id = env->cpu[cpu].socket_id;
1789
1790 /*
1791 * Encode socket in upper 16 bits
1792 * core_id is relative to socket, and
1793 * we need a global id. So we combine
1794 * socket + core id.
1795 */
1796 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1797 }
1798
1799 return core;
1800}
1801
1802static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1803 struct cpu_map **sockp)
1804{
1805 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1806}
1807
1808static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1809 struct cpu_map **corep)
1810{
1811 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1812}
1813
1814static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1815{
1816 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1817}
1818
1819static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1820{
1821 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1822}
1823
1824static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1825{
1826 struct perf_env *env = &st->session->header.env;
1827
1828 switch (stat_config.aggr_mode) {
1829 case AGGR_SOCKET:
1830 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1831 perror("cannot build socket map");
1832 return -1;
1833 }
1834 aggr_get_id = perf_stat__get_socket_file;
1835 break;
1836 case AGGR_CORE:
1837 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1838 perror("cannot build core map");
1839 return -1;
1840 }
1841 aggr_get_id = perf_stat__get_core_file;
1842 break;
1843 case AGGR_NONE:
1844 case AGGR_GLOBAL:
1845 case AGGR_THREAD:
1846 case AGGR_UNSET:
1847 default:
1848 break;
1849 }
1850
1851 return 0;
1852}
1853
1854static int topdown_filter_events(const char **attr, char **str, bool use_group)
1855{
1856 int off = 0;
1857 int i;
1858 int len = 0;
1859 char *s;
1860
1861 for (i = 0; attr[i]; i++) {
1862 if (pmu_have_event("cpu", attr[i])) {
1863 len += strlen(attr[i]) + 1;
1864 attr[i - off] = attr[i];
1865 } else
1866 off++;
1867 }
1868 attr[i - off] = NULL;
1869
1870 *str = malloc(len + 1 + 2);
1871 if (!*str)
1872 return -1;
1873 s = *str;
1874 if (i - off == 0) {
1875 *s = 0;
1876 return 0;
1877 }
1878 if (use_group)
1879 *s++ = '{';
1880 for (i = 0; attr[i]; i++) {
1881 strcpy(s, attr[i]);
1882 s += strlen(s);
1883 *s++ = ',';
1884 }
1885 if (use_group) {
1886 s[-1] = '}';
1887 *s = 0;
1888 } else
1889 s[-1] = 0;
1890 return 0;
1891}
1892
1893__weak bool arch_topdown_check_group(bool *warn)
1894{
1895 *warn = false;
1896 return false;
1897}
1898
1899__weak void arch_topdown_group_warn(void)
1900{
1901}
1902
1903/*
1904 * Add default attributes, if there were no attributes specified or
1905 * if -d/--detailed, -d -d or -d -d -d is used:
1906 */
1907static int add_default_attributes(void)
1908{
1909 int err;
1910 struct perf_event_attr default_attrs0[] = {
1911
1912 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1913 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1914 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1915 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1916
1917 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1918};
1919 struct perf_event_attr frontend_attrs[] = {
1920 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1921};
1922 struct perf_event_attr backend_attrs[] = {
1923 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1924};
1925 struct perf_event_attr default_attrs1[] = {
1926 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1927 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1928 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1929
1930};
1931
1932/*
1933 * Detailed stats (-d), covering the L1 and last level data caches:
1934 */
1935 struct perf_event_attr detailed_attrs[] = {
1936
1937 { .type = PERF_TYPE_HW_CACHE,
1938 .config =
1939 PERF_COUNT_HW_CACHE_L1D << 0 |
1940 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1941 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1942
1943 { .type = PERF_TYPE_HW_CACHE,
1944 .config =
1945 PERF_COUNT_HW_CACHE_L1D << 0 |
1946 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1947 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1948
1949 { .type = PERF_TYPE_HW_CACHE,
1950 .config =
1951 PERF_COUNT_HW_CACHE_LL << 0 |
1952 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1953 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1954
1955 { .type = PERF_TYPE_HW_CACHE,
1956 .config =
1957 PERF_COUNT_HW_CACHE_LL << 0 |
1958 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1959 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1960};
1961
1962/*
1963 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1964 */
1965 struct perf_event_attr very_detailed_attrs[] = {
1966
1967 { .type = PERF_TYPE_HW_CACHE,
1968 .config =
1969 PERF_COUNT_HW_CACHE_L1I << 0 |
1970 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1971 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1972
1973 { .type = PERF_TYPE_HW_CACHE,
1974 .config =
1975 PERF_COUNT_HW_CACHE_L1I << 0 |
1976 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1977 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1978
1979 { .type = PERF_TYPE_HW_CACHE,
1980 .config =
1981 PERF_COUNT_HW_CACHE_DTLB << 0 |
1982 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1983 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1984
1985 { .type = PERF_TYPE_HW_CACHE,
1986 .config =
1987 PERF_COUNT_HW_CACHE_DTLB << 0 |
1988 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1989 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1990
1991 { .type = PERF_TYPE_HW_CACHE,
1992 .config =
1993 PERF_COUNT_HW_CACHE_ITLB << 0 |
1994 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1995 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1996
1997 { .type = PERF_TYPE_HW_CACHE,
1998 .config =
1999 PERF_COUNT_HW_CACHE_ITLB << 0 |
2000 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2001 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2002
2003};
2004
2005/*
2006 * Very, very detailed stats (-d -d -d), adding prefetch events:
2007 */
2008 struct perf_event_attr very_very_detailed_attrs[] = {
2009
2010 { .type = PERF_TYPE_HW_CACHE,
2011 .config =
2012 PERF_COUNT_HW_CACHE_L1D << 0 |
2013 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2014 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2015
2016 { .type = PERF_TYPE_HW_CACHE,
2017 .config =
2018 PERF_COUNT_HW_CACHE_L1D << 0 |
2019 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2020 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2021};
2022
2023 /* Set attrs if no event is selected and !null_run: */
2024 if (null_run)
2025 return 0;
2026
2027 if (transaction_run) {
2028 if (pmu_have_event("cpu", "cycles-ct") &&
2029 pmu_have_event("cpu", "el-start"))
2030 err = parse_events(evsel_list, transaction_attrs, NULL);
2031 else
2032 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
2033 if (err) {
2034 fprintf(stderr, "Cannot set up transaction events\n");
2035 return -1;
2036 }
2037 return 0;
2038 }
2039
2040 if (topdown_run) {
2041 char *str = NULL;
2042 bool warn = false;
2043
2044 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2045 stat_config.aggr_mode != AGGR_CORE) {
2046 pr_err("top down event configuration requires --per-core mode\n");
2047 return -1;
2048 }
2049 stat_config.aggr_mode = AGGR_CORE;
2050 if (nr_cgroups || !target__has_cpu(&target)) {
2051 pr_err("top down event configuration requires system-wide mode (-a)\n");
2052 return -1;
2053 }
2054
2055 if (!force_metric_only)
2056 metric_only = true;
2057 if (topdown_filter_events(topdown_attrs, &str,
2058 arch_topdown_check_group(&warn)) < 0) {
2059 pr_err("Out of memory\n");
2060 return -1;
2061 }
2062 if (topdown_attrs[0] && str) {
2063 if (warn)
2064 arch_topdown_group_warn();
2065 err = parse_events(evsel_list, str, NULL);
2066 if (err) {
2067 fprintf(stderr,
2068 "Cannot set up top down events %s: %d\n",
2069 str, err);
2070 free(str);
2071 return -1;
2072 }
2073 } else {
2074 fprintf(stderr, "System does not support topdown\n");
2075 return -1;
2076 }
2077 free(str);
2078 }
2079
2080 if (!evsel_list->nr_entries) {
2081 if (target__has_cpu(&target))
2082 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2083
2084 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2085 return -1;
2086 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2087 if (perf_evlist__add_default_attrs(evsel_list,
2088 frontend_attrs) < 0)
2089 return -1;
2090 }
2091 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2092 if (perf_evlist__add_default_attrs(evsel_list,
2093 backend_attrs) < 0)
2094 return -1;
2095 }
2096 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2097 return -1;
2098 }
2099
2100 /* Detailed events get appended to the event list: */
2101
2102 if (detailed_run < 1)
2103 return 0;
2104
2105 /* Append detailed run extra attributes: */
2106 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2107 return -1;
2108
2109 if (detailed_run < 2)
2110 return 0;
2111
2112 /* Append very detailed run extra attributes: */
2113 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2114 return -1;
2115
2116 if (detailed_run < 3)
2117 return 0;
2118
2119 /* Append very, very detailed run extra attributes: */
2120 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2121}
2122
2123static const char * const stat_record_usage[] = {
2124 "perf stat record [<options>]",
2125 NULL,
2126};
2127
2128static void init_features(struct perf_session *session)
2129{
2130 int feat;
2131
2132 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2133 perf_header__set_feat(&session->header, feat);
2134
2135 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2136 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2137 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2138 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2139}
2140
2141static int __cmd_record(int argc, const char **argv)
2142{
2143 struct perf_session *session;
2144 struct perf_data_file *file = &perf_stat.file;
2145
2146 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2147 PARSE_OPT_STOP_AT_NON_OPTION);
2148
2149 if (output_name)
2150 file->path = output_name;
2151
2152 if (run_count != 1 || forever) {
2153 pr_err("Cannot use -r option with perf stat record.\n");
2154 return -1;
2155 }
2156
2157 session = perf_session__new(file, false, NULL);
2158 if (session == NULL) {
2159 pr_err("Perf session creation failed.\n");
2160 return -1;
2161 }
2162
2163 init_features(session);
2164
2165 session->evlist = evsel_list;
2166 perf_stat.session = session;
2167 perf_stat.record = true;
2168 return argc;
2169}
2170
2171static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2172 union perf_event *event,
2173 struct perf_session *session)
2174{
2175 struct stat_round_event *stat_round = &event->stat_round;
2176 struct perf_evsel *counter;
2177 struct timespec tsh, *ts = NULL;
2178 const char **argv = session->header.env.cmdline_argv;
2179 int argc = session->header.env.nr_cmdline;
2180
2181 evlist__for_each_entry(evsel_list, counter)
2182 perf_stat_process_counter(&stat_config, counter);
2183
2184 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2185 update_stats(&walltime_nsecs_stats, stat_round->time);
2186
2187 if (stat_config.interval && stat_round->time) {
2188 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2189 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2190 ts = &tsh;
2191 }
2192
2193 print_counters(ts, argc, argv);
2194 return 0;
2195}
2196
2197static
2198int process_stat_config_event(struct perf_tool *tool,
2199 union perf_event *event,
2200 struct perf_session *session __maybe_unused)
2201{
2202 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2203
2204 perf_event__read_stat_config(&stat_config, &event->stat_config);
2205
2206 if (cpu_map__empty(st->cpus)) {
2207 if (st->aggr_mode != AGGR_UNSET)
2208 pr_warning("warning: processing task data, aggregation mode not set\n");
2209 return 0;
2210 }
2211
2212 if (st->aggr_mode != AGGR_UNSET)
2213 stat_config.aggr_mode = st->aggr_mode;
2214
2215 if (perf_stat.file.is_pipe)
2216 perf_stat_init_aggr_mode();
2217 else
2218 perf_stat_init_aggr_mode_file(st);
2219
2220 return 0;
2221}
2222
2223static int set_maps(struct perf_stat *st)
2224{
2225 if (!st->cpus || !st->threads)
2226 return 0;
2227
2228 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2229 return -EINVAL;
2230
2231 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2232
2233 if (perf_evlist__alloc_stats(evsel_list, true))
2234 return -ENOMEM;
2235
2236 st->maps_allocated = true;
2237 return 0;
2238}
2239
2240static
2241int process_thread_map_event(struct perf_tool *tool,
2242 union perf_event *event,
2243 struct perf_session *session __maybe_unused)
2244{
2245 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2246
2247 if (st->threads) {
2248 pr_warning("Extra thread map event, ignoring.\n");
2249 return 0;
2250 }
2251
2252 st->threads = thread_map__new_event(&event->thread_map);
2253 if (!st->threads)
2254 return -ENOMEM;
2255
2256 return set_maps(st);
2257}
2258
2259static
2260int process_cpu_map_event(struct perf_tool *tool,
2261 union perf_event *event,
2262 struct perf_session *session __maybe_unused)
2263{
2264 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2265 struct cpu_map *cpus;
2266
2267 if (st->cpus) {
2268 pr_warning("Extra cpu map event, ignoring.\n");
2269 return 0;
2270 }
2271
2272 cpus = cpu_map__new_data(&event->cpu_map.data);
2273 if (!cpus)
2274 return -ENOMEM;
2275
2276 st->cpus = cpus;
2277 return set_maps(st);
2278}
2279
2280static const char * const stat_report_usage[] = {
2281 "perf stat report [<options>]",
2282 NULL,
2283};
2284
2285static struct perf_stat perf_stat = {
2286 .tool = {
2287 .attr = perf_event__process_attr,
2288 .event_update = perf_event__process_event_update,
2289 .thread_map = process_thread_map_event,
2290 .cpu_map = process_cpu_map_event,
2291 .stat_config = process_stat_config_event,
2292 .stat = perf_event__process_stat_event,
2293 .stat_round = process_stat_round_event,
2294 },
2295 .aggr_mode = AGGR_UNSET,
2296};
2297
2298static int __cmd_report(int argc, const char **argv)
2299{
2300 struct perf_session *session;
2301 const struct option options[] = {
2302 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2303 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2304 "aggregate counts per processor socket", AGGR_SOCKET),
2305 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2306 "aggregate counts per physical processor core", AGGR_CORE),
2307 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2308 "disable CPU count aggregation", AGGR_NONE),
2309 OPT_END()
2310 };
2311 struct stat st;
2312 int ret;
2313
2314 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2315
2316 if (!input_name || !strlen(input_name)) {
2317 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2318 input_name = "-";
2319 else
2320 input_name = "perf.data";
2321 }
2322
2323 perf_stat.file.path = input_name;
2324 perf_stat.file.mode = PERF_DATA_MODE_READ;
2325
2326 session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
2327 if (session == NULL)
2328 return -1;
2329
2330 perf_stat.session = session;
2331 stat_config.output = stderr;
2332 evsel_list = session->evlist;
2333
2334 ret = perf_session__process_events(session);
2335 if (ret)
2336 return ret;
2337
2338 perf_session__delete(session);
2339 return 0;
2340}
2341
2342int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
2343{
2344 const char * const stat_usage[] = {
2345 "perf stat [<options>] [<command>]",
2346 NULL
2347 };
2348 int status = -EINVAL, run_idx;
2349 const char *mode;
2350 FILE *output = stderr;
2351 unsigned int interval;
2352 const char * const stat_subcommands[] = { "record", "report" };
2353
2354 setlocale(LC_ALL, "");
2355
2356 evsel_list = perf_evlist__new();
2357 if (evsel_list == NULL)
2358 return -ENOMEM;
2359
2360 parse_events__shrink_config_terms();
2361 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2362 (const char **) stat_usage,
2363 PARSE_OPT_STOP_AT_NON_OPTION);
2364 perf_stat__init_shadow_stats();
2365
2366 if (csv_sep) {
2367 csv_output = true;
2368 if (!strcmp(csv_sep, "\\t"))
2369 csv_sep = "\t";
2370 } else
2371 csv_sep = DEFAULT_SEPARATOR;
2372
2373 if (argc && !strncmp(argv[0], "rec", 3)) {
2374 argc = __cmd_record(argc, argv);
2375 if (argc < 0)
2376 return -1;
2377 } else if (argc && !strncmp(argv[0], "rep", 3))
2378 return __cmd_report(argc, argv);
2379
2380 interval = stat_config.interval;
2381
2382 /*
2383 * For record command the -o is already taken care of.
2384 */
2385 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2386 output = NULL;
2387
2388 if (output_name && output_fd) {
2389 fprintf(stderr, "cannot use both --output and --log-fd\n");
2390 parse_options_usage(stat_usage, stat_options, "o", 1);
2391 parse_options_usage(NULL, stat_options, "log-fd", 0);
2392 goto out;
2393 }
2394
2395 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2396 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2397 goto out;
2398 }
2399
2400 if (metric_only && run_count > 1) {
2401 fprintf(stderr, "--metric-only is not supported with -r\n");
2402 goto out;
2403 }
2404
2405 if (output_fd < 0) {
2406 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2407 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2408 goto out;
2409 }
2410
2411 if (!output) {
2412 struct timespec tm;
2413 mode = append_file ? "a" : "w";
2414
2415 output = fopen(output_name, mode);
2416 if (!output) {
2417 perror("failed to create output file");
2418 return -1;
2419 }
2420 clock_gettime(CLOCK_REALTIME, &tm);
2421 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2422 } else if (output_fd > 0) {
2423 mode = append_file ? "a" : "w";
2424 output = fdopen(output_fd, mode);
2425 if (!output) {
2426 perror("Failed opening logfd");
2427 return -errno;
2428 }
2429 }
2430
2431 stat_config.output = output;
2432
2433 /*
2434 * let the spreadsheet do the pretty-printing
2435 */
2436 if (csv_output) {
2437 /* User explicitly passed -B? */
2438 if (big_num_opt == 1) {
2439 fprintf(stderr, "-B option not supported with -x\n");
2440 parse_options_usage(stat_usage, stat_options, "B", 1);
2441 parse_options_usage(NULL, stat_options, "x", 1);
2442 goto out;
2443 } else /* Nope, so disable big number formatting */
2444 big_num = false;
2445 } else if (big_num_opt == 0) /* User passed --no-big-num */
2446 big_num = false;
2447
2448 if (!argc && target__none(&target))
2449 usage_with_options(stat_usage, stat_options);
2450
2451 if (run_count < 0) {
2452 pr_err("Run count must be a positive number\n");
2453 parse_options_usage(stat_usage, stat_options, "r", 1);
2454 goto out;
2455 } else if (run_count == 0) {
2456 forever = true;
2457 run_count = 1;
2458 }
2459
2460 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2461 fprintf(stderr, "The --per-thread option is only available "
2462 "when monitoring via -p -t options.\n");
2463 parse_options_usage(NULL, stat_options, "p", 1);
2464 parse_options_usage(NULL, stat_options, "t", 1);
2465 goto out;
2466 }
2467
2468 /*
2469 * no_aggr, cgroup are for system-wide only
2470 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2471 */
2472 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2473 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2474 !target__has_cpu(&target)) {
2475 fprintf(stderr, "both cgroup and no-aggregation "
2476 "modes only available in system-wide mode\n");
2477
2478 parse_options_usage(stat_usage, stat_options, "G", 1);
2479 parse_options_usage(NULL, stat_options, "A", 1);
2480 parse_options_usage(NULL, stat_options, "a", 1);
2481 goto out;
2482 }
2483
2484 if (add_default_attributes())
2485 goto out;
2486
2487 target__validate(&target);
2488
2489 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2490 if (target__has_task(&target)) {
2491 pr_err("Problems finding threads of monitor\n");
2492 parse_options_usage(stat_usage, stat_options, "p", 1);
2493 parse_options_usage(NULL, stat_options, "t", 1);
2494 } else if (target__has_cpu(&target)) {
2495 perror("failed to parse CPUs map");
2496 parse_options_usage(stat_usage, stat_options, "C", 1);
2497 parse_options_usage(NULL, stat_options, "a", 1);
2498 }
2499 goto out;
2500 }
2501
2502 /*
2503 * Initialize thread_map with comm names,
2504 * so we could print it out on output.
2505 */
2506 if (stat_config.aggr_mode == AGGR_THREAD)
2507 thread_map__read_comms(evsel_list->threads);
2508
2509 if (interval && interval < 100) {
2510 if (interval < 10) {
2511 pr_err("print interval must be >= 10ms\n");
2512 parse_options_usage(stat_usage, stat_options, "I", 1);
2513 goto out;
2514 } else
2515 pr_warning("print interval < 100ms. "
2516 "The overhead percentage could be high in some cases. "
2517 "Please proceed with caution.\n");
2518 }
2519
2520 if (perf_evlist__alloc_stats(evsel_list, interval))
2521 goto out;
2522
2523 if (perf_stat_init_aggr_mode())
2524 goto out;
2525
2526 /*
2527 * We dont want to block the signals - that would cause
2528 * child tasks to inherit that and Ctrl-C would not work.
2529 * What we want is for Ctrl-C to work in the exec()-ed
2530 * task, but being ignored by perf stat itself:
2531 */
2532 atexit(sig_atexit);
2533 if (!forever)
2534 signal(SIGINT, skip_signal);
2535 signal(SIGCHLD, skip_signal);
2536 signal(SIGALRM, skip_signal);
2537 signal(SIGABRT, skip_signal);
2538
2539 status = 0;
2540 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2541 if (run_count != 1 && verbose)
2542 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2543 run_idx + 1);
2544
2545 status = run_perf_stat(argc, argv);
2546 if (forever && status != -1) {
2547 print_counters(NULL, argc, argv);
2548 perf_stat__reset_stats();
2549 }
2550 }
2551
2552 if (!forever && status != -1 && !interval)
2553 print_counters(NULL, argc, argv);
2554
2555 if (STAT_RECORD) {
2556 /*
2557 * We synthesize the kernel mmap record just so that older tools
2558 * don't emit warnings about not being able to resolve symbols
2559 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2560 * a saner message about no samples being in the perf.data file.
2561 *
2562 * This also serves to suppress a warning about f_header.data.size == 0
2563 * in header.c at the moment 'perf stat record' gets introduced, which
2564 * is not really needed once we start adding the stat specific PERF_RECORD_
2565 * records, but the need to suppress the kptr_restrict messages in older
2566 * tools remain -acme
2567 */
2568 int fd = perf_data_file__fd(&perf_stat.file);
2569 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2570 process_synthesized_event,
2571 &perf_stat.session->machines.host);
2572 if (err) {
2573 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2574 "older tools may produce warnings about this file\n.");
2575 }
2576
2577 if (!interval) {
2578 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2579 pr_err("failed to write stat round event\n");
2580 }
2581
2582 if (!perf_stat.file.is_pipe) {
2583 perf_stat.session->header.data_size += perf_stat.bytes_written;
2584 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2585 }
2586
2587 perf_session__delete(perf_stat.session);
2588 }
2589
2590 perf_stat__exit_aggr_mode();
2591 perf_evlist__free_stats(evsel_list);
2592out:
2593 perf_evlist__delete(evsel_list);
2594 return status;
2595}