]>
Commit | Line | Data |
---|---|---|
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 | ||
80 | static void print_counters(struct timespec *ts, int argc, const char **argv); | |
81 | ||
82 | /* Default events used for perf stat -T */ | |
83 | static 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. */ | |
96 | static const char * transaction_limited_attrs = { | |
97 | "task-clock," | |
98 | "{" | |
99 | "instructions," | |
100 | "cycles," | |
101 | "cpu/cycles-t/," | |
102 | "cpu/tx-start/" | |
103 | "}" | |
104 | }; | |
105 | ||
106 | static 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 | ||
115 | static struct perf_evlist *evsel_list; | |
116 | ||
117 | static struct target target = { | |
118 | .uid = UINT_MAX, | |
119 | }; | |
120 | ||
121 | typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu); | |
122 | ||
123 | static int run_count = 1; | |
124 | static bool no_inherit = false; | |
125 | static volatile pid_t child_pid = -1; | |
126 | static bool null_run = false; | |
127 | static int detailed_run = 0; | |
128 | static bool transaction_run; | |
129 | static bool topdown_run = false; | |
130 | static bool big_num = true; | |
131 | static int big_num_opt = -1; | |
132 | static const char *csv_sep = NULL; | |
133 | static bool csv_output = false; | |
134 | static bool group = false; | |
135 | static const char *pre_cmd = NULL; | |
136 | static const char *post_cmd = NULL; | |
137 | static bool sync_run = false; | |
138 | static unsigned int initial_delay = 0; | |
139 | static unsigned int unit_width = 4; /* strlen("unit") */ | |
140 | static bool forever = false; | |
141 | static bool metric_only = false; | |
142 | static bool force_metric_only = false; | |
143 | static bool no_merge = false; | |
144 | static struct timespec ref_time; | |
145 | static struct cpu_map *aggr_map; | |
146 | static aggr_get_id_t aggr_get_id; | |
147 | static bool append_file; | |
148 | static const char *output_name; | |
149 | static int output_fd; | |
150 | static int print_free_counters_hint; | |
151 | ||
152 | struct perf_stat { | |
153 | bool record; | |
154 | struct perf_data_file file; | |
155 | struct perf_session *session; | |
156 | u64 bytes_written; | |
157 | struct perf_tool tool; | |
158 | bool maps_allocated; | |
159 | struct cpu_map *cpus; | |
160 | struct thread_map *threads; | |
161 | enum aggr_mode aggr_mode; | |
162 | }; | |
163 | ||
164 | static struct perf_stat perf_stat; | |
165 | #define STAT_RECORD perf_stat.record | |
166 | ||
167 | static volatile int done = 0; | |
168 | ||
169 | static struct perf_stat_config stat_config = { | |
170 | .aggr_mode = AGGR_GLOBAL, | |
171 | .scale = true, | |
172 | }; | |
173 | ||
174 | static inline void diff_timespec(struct timespec *r, struct timespec *a, | |
175 | struct timespec *b) | |
176 | { | |
177 | r->tv_sec = a->tv_sec - b->tv_sec; | |
178 | if (a->tv_nsec < b->tv_nsec) { | |
179 | r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec; | |
180 | r->tv_sec--; | |
181 | } else { | |
182 | r->tv_nsec = a->tv_nsec - b->tv_nsec ; | |
183 | } | |
184 | } | |
185 | ||
186 | static void perf_stat__reset_stats(void) | |
187 | { | |
188 | perf_evlist__reset_stats(evsel_list); | |
189 | perf_stat__reset_shadow_stats(); | |
190 | } | |
191 | ||
192 | static int create_perf_stat_counter(struct perf_evsel *evsel) | |
193 | { | |
194 | struct perf_event_attr *attr = &evsel->attr; | |
195 | ||
196 | if (stat_config.scale) | |
197 | attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | | |
198 | PERF_FORMAT_TOTAL_TIME_RUNNING; | |
199 | ||
200 | attr->inherit = !no_inherit; | |
201 | ||
202 | /* | |
203 | * Some events get initialized with sample_(period/type) set, | |
204 | * like tracepoints. Clear it up for counting. | |
205 | */ | |
206 | attr->sample_period = 0; | |
207 | ||
208 | /* | |
209 | * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless | |
210 | * while avoiding that older tools show confusing messages. | |
211 | * | |
212 | * However for pipe sessions we need to keep it zero, | |
213 | * because script's perf_evsel__check_attr is triggered | |
214 | * by attr->sample_type != 0, and we can't run it on | |
215 | * stat sessions. | |
216 | */ | |
217 | if (!(STAT_RECORD && perf_stat.file.is_pipe)) | |
218 | attr->sample_type = PERF_SAMPLE_IDENTIFIER; | |
219 | ||
220 | /* | |
221 | * Disabling all counters initially, they will be enabled | |
222 | * either manually by us or by kernel via enable_on_exec | |
223 | * set later. | |
224 | */ | |
225 | if (perf_evsel__is_group_leader(evsel)) { | |
226 | attr->disabled = 1; | |
227 | ||
228 | /* | |
229 | * In case of initial_delay we enable tracee | |
230 | * events manually. | |
231 | */ | |
232 | if (target__none(&target) && !initial_delay) | |
233 | attr->enable_on_exec = 1; | |
234 | } | |
235 | ||
236 | if (target__has_cpu(&target)) | |
237 | return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel)); | |
238 | ||
239 | return perf_evsel__open_per_thread(evsel, evsel_list->threads); | |
240 | } | |
241 | ||
242 | /* | |
243 | * Does the counter have nsecs as a unit? | |
244 | */ | |
245 | static inline int nsec_counter(struct perf_evsel *evsel) | |
246 | { | |
247 | if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || | |
248 | perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) | |
249 | return 1; | |
250 | ||
251 | return 0; | |
252 | } | |
253 | ||
254 | static int process_synthesized_event(struct perf_tool *tool __maybe_unused, | |
255 | union perf_event *event, | |
256 | struct perf_sample *sample __maybe_unused, | |
257 | struct machine *machine __maybe_unused) | |
258 | { | |
259 | if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) { | |
260 | pr_err("failed to write perf data, error: %m\n"); | |
261 | return -1; | |
262 | } | |
263 | ||
264 | perf_stat.bytes_written += event->header.size; | |
265 | return 0; | |
266 | } | |
267 | ||
268 | static int write_stat_round_event(u64 tm, u64 type) | |
269 | { | |
270 | return perf_event__synthesize_stat_round(NULL, tm, type, | |
271 | process_synthesized_event, | |
272 | NULL); | |
273 | } | |
274 | ||
275 | #define WRITE_STAT_ROUND_EVENT(time, interval) \ | |
276 | write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval) | |
277 | ||
278 | #define SID(e, x, y) xyarray__entry(e->sample_id, x, y) | |
279 | ||
280 | static int | |
281 | perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread, | |
282 | struct perf_counts_values *count) | |
283 | { | |
284 | struct perf_sample_id *sid = SID(counter, cpu, thread); | |
285 | ||
286 | return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count, | |
287 | process_synthesized_event, NULL); | |
288 | } | |
289 | ||
290 | /* | |
291 | * Read out the results of a single counter: | |
292 | * do not aggregate counts across CPUs in system-wide mode | |
293 | */ | |
294 | static int read_counter(struct perf_evsel *counter) | |
295 | { | |
296 | int nthreads = thread_map__nr(evsel_list->threads); | |
297 | int ncpus, cpu, thread; | |
298 | ||
299 | if (target__has_cpu(&target)) | |
300 | ncpus = perf_evsel__nr_cpus(counter); | |
301 | else | |
302 | ncpus = 1; | |
303 | ||
304 | if (!counter->supported) | |
305 | return -ENOENT; | |
306 | ||
307 | if (counter->system_wide) | |
308 | nthreads = 1; | |
309 | ||
310 | for (thread = 0; thread < nthreads; thread++) { | |
311 | for (cpu = 0; cpu < ncpus; cpu++) { | |
312 | struct perf_counts_values *count; | |
313 | ||
314 | count = perf_counts(counter->counts, cpu, thread); | |
315 | if (perf_evsel__read(counter, cpu, thread, count)) | |
316 | return -1; | |
317 | ||
318 | if (STAT_RECORD) { | |
319 | if (perf_evsel__write_stat_event(counter, cpu, thread, count)) { | |
320 | pr_err("failed to write stat event\n"); | |
321 | return -1; | |
322 | } | |
323 | } | |
324 | ||
325 | if (verbose > 1) { | |
326 | fprintf(stat_config.output, | |
327 | "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", | |
328 | perf_evsel__name(counter), | |
329 | cpu, | |
330 | count->val, count->ena, count->run); | |
331 | } | |
332 | } | |
333 | } | |
334 | ||
335 | return 0; | |
336 | } | |
337 | ||
338 | static void read_counters(void) | |
339 | { | |
340 | struct perf_evsel *counter; | |
341 | ||
342 | evlist__for_each_entry(evsel_list, counter) { | |
343 | if (read_counter(counter)) | |
344 | pr_debug("failed to read counter %s\n", counter->name); | |
345 | ||
346 | if (perf_stat_process_counter(&stat_config, counter)) | |
347 | pr_warning("failed to process counter %s\n", counter->name); | |
348 | } | |
349 | } | |
350 | ||
351 | static void process_interval(void) | |
352 | { | |
353 | struct timespec ts, rs; | |
354 | ||
355 | read_counters(); | |
356 | ||
357 | clock_gettime(CLOCK_MONOTONIC, &ts); | |
358 | diff_timespec(&rs, &ts, &ref_time); | |
359 | ||
360 | if (STAT_RECORD) { | |
361 | if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL)) | |
362 | pr_err("failed to write stat round event\n"); | |
363 | } | |
364 | ||
365 | print_counters(&rs, 0, NULL); | |
366 | } | |
367 | ||
368 | static void enable_counters(void) | |
369 | { | |
370 | if (initial_delay) | |
371 | usleep(initial_delay * USEC_PER_MSEC); | |
372 | ||
373 | /* | |
374 | * We need to enable counters only if: | |
375 | * - we don't have tracee (attaching to task or cpu) | |
376 | * - we have initial delay configured | |
377 | */ | |
378 | if (!target__none(&target) || initial_delay) | |
379 | perf_evlist__enable(evsel_list); | |
380 | } | |
381 | ||
382 | static void disable_counters(void) | |
383 | { | |
384 | /* | |
385 | * If we don't have tracee (attaching to task or cpu), counters may | |
386 | * still be running. To get accurate group ratios, we must stop groups | |
387 | * from counting before reading their constituent counters. | |
388 | */ | |
389 | if (!target__none(&target)) | |
390 | perf_evlist__disable(evsel_list); | |
391 | } | |
392 | ||
393 | static volatile int workload_exec_errno; | |
394 | ||
395 | /* | |
396 | * perf_evlist__prepare_workload will send a SIGUSR1 | |
397 | * if the fork fails, since we asked by setting its | |
398 | * want_signal to true. | |
399 | */ | |
400 | static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, | |
401 | void *ucontext __maybe_unused) | |
402 | { | |
403 | workload_exec_errno = info->si_value.sival_int; | |
404 | } | |
405 | ||
406 | static bool has_unit(struct perf_evsel *counter) | |
407 | { | |
408 | return counter->unit && *counter->unit; | |
409 | } | |
410 | ||
411 | static bool has_scale(struct perf_evsel *counter) | |
412 | { | |
413 | return counter->scale != 1; | |
414 | } | |
415 | ||
416 | static int perf_stat_synthesize_config(bool is_pipe) | |
417 | { | |
418 | struct perf_evsel *counter; | |
419 | int err; | |
420 | ||
421 | if (is_pipe) { | |
422 | err = perf_event__synthesize_attrs(NULL, perf_stat.session, | |
423 | process_synthesized_event); | |
424 | if (err < 0) { | |
425 | pr_err("Couldn't synthesize attrs.\n"); | |
426 | return err; | |
427 | } | |
428 | } | |
429 | ||
430 | /* | |
431 | * Synthesize other events stuff not carried within | |
432 | * attr event - unit, scale, name | |
433 | */ | |
434 | evlist__for_each_entry(evsel_list, counter) { | |
435 | if (!counter->supported) | |
436 | continue; | |
437 | ||
438 | /* | |
439 | * Synthesize unit and scale only if it's defined. | |
440 | */ | |
441 | if (has_unit(counter)) { | |
442 | err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event); | |
443 | if (err < 0) { | |
444 | pr_err("Couldn't synthesize evsel unit.\n"); | |
445 | return err; | |
446 | } | |
447 | } | |
448 | ||
449 | if (has_scale(counter)) { | |
450 | err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event); | |
451 | if (err < 0) { | |
452 | pr_err("Couldn't synthesize evsel scale.\n"); | |
453 | return err; | |
454 | } | |
455 | } | |
456 | ||
457 | if (counter->own_cpus) { | |
458 | err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event); | |
459 | if (err < 0) { | |
460 | pr_err("Couldn't synthesize evsel scale.\n"); | |
461 | return err; | |
462 | } | |
463 | } | |
464 | ||
465 | /* | |
466 | * Name is needed only for pipe output, | |
467 | * perf.data carries event names. | |
468 | */ | |
469 | if (is_pipe) { | |
470 | err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event); | |
471 | if (err < 0) { | |
472 | pr_err("Couldn't synthesize evsel name.\n"); | |
473 | return err; | |
474 | } | |
475 | } | |
476 | } | |
477 | ||
478 | err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads, | |
479 | process_synthesized_event, | |
480 | NULL); | |
481 | if (err < 0) { | |
482 | pr_err("Couldn't synthesize thread map.\n"); | |
483 | return err; | |
484 | } | |
485 | ||
486 | err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus, | |
487 | process_synthesized_event, NULL); | |
488 | if (err < 0) { | |
489 | pr_err("Couldn't synthesize thread map.\n"); | |
490 | return err; | |
491 | } | |
492 | ||
493 | err = perf_event__synthesize_stat_config(NULL, &stat_config, | |
494 | process_synthesized_event, NULL); | |
495 | if (err < 0) { | |
496 | pr_err("Couldn't synthesize config.\n"); | |
497 | return err; | |
498 | } | |
499 | ||
500 | return 0; | |
501 | } | |
502 | ||
503 | #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y)) | |
504 | ||
505 | static int __store_counter_ids(struct perf_evsel *counter, | |
506 | struct cpu_map *cpus, | |
507 | struct thread_map *threads) | |
508 | { | |
509 | int cpu, thread; | |
510 | ||
511 | for (cpu = 0; cpu < cpus->nr; cpu++) { | |
512 | for (thread = 0; thread < threads->nr; thread++) { | |
513 | int fd = FD(counter, cpu, thread); | |
514 | ||
515 | if (perf_evlist__id_add_fd(evsel_list, counter, | |
516 | cpu, thread, fd) < 0) | |
517 | return -1; | |
518 | } | |
519 | } | |
520 | ||
521 | return 0; | |
522 | } | |
523 | ||
524 | static int store_counter_ids(struct perf_evsel *counter) | |
525 | { | |
526 | struct cpu_map *cpus = counter->cpus; | |
527 | struct thread_map *threads = counter->threads; | |
528 | ||
529 | if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr)) | |
530 | return -ENOMEM; | |
531 | ||
532 | return __store_counter_ids(counter, cpus, threads); | |
533 | } | |
534 | ||
535 | static int __run_perf_stat(int argc, const char **argv) | |
536 | { | |
537 | int interval = stat_config.interval; | |
538 | char msg[BUFSIZ]; | |
539 | unsigned long long t0, t1; | |
540 | struct perf_evsel *counter; | |
541 | struct timespec ts; | |
542 | size_t l; | |
543 | int status = 0; | |
544 | const bool forks = (argc > 0); | |
545 | bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false; | |
546 | struct perf_evsel_config_term *err_term; | |
547 | ||
548 | if (interval) { | |
549 | ts.tv_sec = interval / USEC_PER_MSEC; | |
550 | ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC; | |
551 | } else { | |
552 | ts.tv_sec = 1; | |
553 | ts.tv_nsec = 0; | |
554 | } | |
555 | ||
556 | if (forks) { | |
557 | if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe, | |
558 | workload_exec_failed_signal) < 0) { | |
559 | perror("failed to prepare workload"); | |
560 | return -1; | |
561 | } | |
562 | child_pid = evsel_list->workload.pid; | |
563 | } | |
564 | ||
565 | if (group) | |
566 | perf_evlist__set_leader(evsel_list); | |
567 | ||
568 | evlist__for_each_entry(evsel_list, counter) { | |
569 | try_again: | |
570 | if (create_perf_stat_counter(counter) < 0) { | |
571 | /* | |
572 | * PPC returns ENXIO for HW counters until 2.6.37 | |
573 | * (behavior changed with commit b0a873e). | |
574 | */ | |
575 | if (errno == EINVAL || errno == ENOSYS || | |
576 | errno == ENOENT || errno == EOPNOTSUPP || | |
577 | errno == ENXIO) { | |
578 | if (verbose > 0) | |
579 | ui__warning("%s event is not supported by the kernel.\n", | |
580 | perf_evsel__name(counter)); | |
581 | counter->supported = false; | |
582 | ||
583 | if ((counter->leader != counter) || | |
584 | !(counter->leader->nr_members > 1)) | |
585 | continue; | |
586 | } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) { | |
587 | if (verbose > 0) | |
588 | ui__warning("%s\n", msg); | |
589 | goto try_again; | |
590 | } | |
591 | ||
592 | perf_evsel__open_strerror(counter, &target, | |
593 | errno, msg, sizeof(msg)); | |
594 | ui__error("%s\n", msg); | |
595 | ||
596 | if (child_pid != -1) | |
597 | kill(child_pid, SIGTERM); | |
598 | ||
599 | return -1; | |
600 | } | |
601 | counter->supported = true; | |
602 | ||
603 | l = strlen(counter->unit); | |
604 | if (l > unit_width) | |
605 | unit_width = l; | |
606 | ||
607 | if (STAT_RECORD && store_counter_ids(counter)) | |
608 | return -1; | |
609 | } | |
610 | ||
611 | if (perf_evlist__apply_filters(evsel_list, &counter)) { | |
612 | error("failed to set filter \"%s\" on event %s with %d (%s)\n", | |
613 | counter->filter, perf_evsel__name(counter), errno, | |
614 | str_error_r(errno, msg, sizeof(msg))); | |
615 | return -1; | |
616 | } | |
617 | ||
618 | if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) { | |
619 | error("failed to set config \"%s\" on event %s with %d (%s)\n", | |
620 | err_term->val.drv_cfg, perf_evsel__name(counter), errno, | |
621 | str_error_r(errno, msg, sizeof(msg))); | |
622 | return -1; | |
623 | } | |
624 | ||
625 | if (STAT_RECORD) { | |
626 | int err, fd = perf_data_file__fd(&perf_stat.file); | |
627 | ||
628 | if (is_pipe) { | |
629 | err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file)); | |
630 | } else { | |
631 | err = perf_session__write_header(perf_stat.session, evsel_list, | |
632 | fd, false); | |
633 | } | |
634 | ||
635 | if (err < 0) | |
636 | return err; | |
637 | ||
638 | err = perf_stat_synthesize_config(is_pipe); | |
639 | if (err < 0) | |
640 | return err; | |
641 | } | |
642 | ||
643 | /* | |
644 | * Enable counters and exec the command: | |
645 | */ | |
646 | t0 = rdclock(); | |
647 | clock_gettime(CLOCK_MONOTONIC, &ref_time); | |
648 | ||
649 | if (forks) { | |
650 | perf_evlist__start_workload(evsel_list); | |
651 | enable_counters(); | |
652 | ||
653 | if (interval) { | |
654 | while (!waitpid(child_pid, &status, WNOHANG)) { | |
655 | nanosleep(&ts, NULL); | |
656 | process_interval(); | |
657 | } | |
658 | } | |
659 | wait(&status); | |
660 | ||
661 | if (workload_exec_errno) { | |
662 | const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg)); | |
663 | pr_err("Workload failed: %s\n", emsg); | |
664 | return -1; | |
665 | } | |
666 | ||
667 | if (WIFSIGNALED(status)) | |
668 | psignal(WTERMSIG(status), argv[0]); | |
669 | } else { | |
670 | enable_counters(); | |
671 | while (!done) { | |
672 | nanosleep(&ts, NULL); | |
673 | if (interval) | |
674 | process_interval(); | |
675 | } | |
676 | } | |
677 | ||
678 | disable_counters(); | |
679 | ||
680 | t1 = rdclock(); | |
681 | ||
682 | update_stats(&walltime_nsecs_stats, t1 - t0); | |
683 | ||
684 | /* | |
685 | * Closing a group leader splits the group, and as we only disable | |
686 | * group leaders, results in remaining events becoming enabled. To | |
687 | * avoid arbitrary skew, we must read all counters before closing any | |
688 | * group leaders. | |
689 | */ | |
690 | read_counters(); | |
691 | perf_evlist__close(evsel_list); | |
692 | ||
693 | return WEXITSTATUS(status); | |
694 | } | |
695 | ||
696 | static int run_perf_stat(int argc, const char **argv) | |
697 | { | |
698 | int ret; | |
699 | ||
700 | if (pre_cmd) { | |
701 | ret = system(pre_cmd); | |
702 | if (ret) | |
703 | return ret; | |
704 | } | |
705 | ||
706 | if (sync_run) | |
707 | sync(); | |
708 | ||
709 | ret = __run_perf_stat(argc, argv); | |
710 | if (ret) | |
711 | return ret; | |
712 | ||
713 | if (post_cmd) { | |
714 | ret = system(post_cmd); | |
715 | if (ret) | |
716 | return ret; | |
717 | } | |
718 | ||
719 | return ret; | |
720 | } | |
721 | ||
722 | static void print_running(u64 run, u64 ena) | |
723 | { | |
724 | if (csv_output) { | |
725 | fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f", | |
726 | csv_sep, | |
727 | run, | |
728 | csv_sep, | |
729 | ena ? 100.0 * run / ena : 100.0); | |
730 | } else if (run != ena) { | |
731 | fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena); | |
732 | } | |
733 | } | |
734 | ||
735 | static void print_noise_pct(double total, double avg) | |
736 | { | |
737 | double pct = rel_stddev_stats(total, avg); | |
738 | ||
739 | if (csv_output) | |
740 | fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct); | |
741 | else if (pct) | |
742 | fprintf(stat_config.output, " ( +-%6.2f%% )", pct); | |
743 | } | |
744 | ||
745 | static void print_noise(struct perf_evsel *evsel, double avg) | |
746 | { | |
747 | struct perf_stat_evsel *ps; | |
748 | ||
749 | if (run_count == 1) | |
750 | return; | |
751 | ||
752 | ps = evsel->priv; | |
753 | print_noise_pct(stddev_stats(&ps->res_stats[0]), avg); | |
754 | } | |
755 | ||
756 | static void aggr_printout(struct perf_evsel *evsel, int id, int nr) | |
757 | { | |
758 | switch (stat_config.aggr_mode) { | |
759 | case AGGR_CORE: | |
760 | fprintf(stat_config.output, "S%d-C%*d%s%*d%s", | |
761 | cpu_map__id_to_socket(id), | |
762 | csv_output ? 0 : -8, | |
763 | cpu_map__id_to_cpu(id), | |
764 | csv_sep, | |
765 | csv_output ? 0 : 4, | |
766 | nr, | |
767 | csv_sep); | |
768 | break; | |
769 | case AGGR_SOCKET: | |
770 | fprintf(stat_config.output, "S%*d%s%*d%s", | |
771 | csv_output ? 0 : -5, | |
772 | id, | |
773 | csv_sep, | |
774 | csv_output ? 0 : 4, | |
775 | nr, | |
776 | csv_sep); | |
777 | break; | |
778 | case AGGR_NONE: | |
779 | fprintf(stat_config.output, "CPU%*d%s", | |
780 | csv_output ? 0 : -4, | |
781 | perf_evsel__cpus(evsel)->map[id], csv_sep); | |
782 | break; | |
783 | case AGGR_THREAD: | |
784 | fprintf(stat_config.output, "%*s-%*d%s", | |
785 | csv_output ? 0 : 16, | |
786 | thread_map__comm(evsel->threads, id), | |
787 | csv_output ? 0 : -8, | |
788 | thread_map__pid(evsel->threads, id), | |
789 | csv_sep); | |
790 | break; | |
791 | case AGGR_GLOBAL: | |
792 | case AGGR_UNSET: | |
793 | default: | |
794 | break; | |
795 | } | |
796 | } | |
797 | ||
798 | struct outstate { | |
799 | FILE *fh; | |
800 | bool newline; | |
801 | const char *prefix; | |
802 | int nfields; | |
803 | int id, nr; | |
804 | struct perf_evsel *evsel; | |
805 | }; | |
806 | ||
807 | #define METRIC_LEN 35 | |
808 | ||
809 | static void new_line_std(void *ctx) | |
810 | { | |
811 | struct outstate *os = ctx; | |
812 | ||
813 | os->newline = true; | |
814 | } | |
815 | ||
816 | static void do_new_line_std(struct outstate *os) | |
817 | { | |
818 | fputc('\n', os->fh); | |
819 | fputs(os->prefix, os->fh); | |
820 | aggr_printout(os->evsel, os->id, os->nr); | |
821 | if (stat_config.aggr_mode == AGGR_NONE) | |
822 | fprintf(os->fh, " "); | |
823 | fprintf(os->fh, " "); | |
824 | } | |
825 | ||
826 | static void print_metric_std(void *ctx, const char *color, const char *fmt, | |
827 | const char *unit, double val) | |
828 | { | |
829 | struct outstate *os = ctx; | |
830 | FILE *out = os->fh; | |
831 | int n; | |
832 | bool newline = os->newline; | |
833 | ||
834 | os->newline = false; | |
835 | ||
836 | if (unit == NULL || fmt == NULL) { | |
837 | fprintf(out, "%-*s", METRIC_LEN, ""); | |
838 | return; | |
839 | } | |
840 | ||
841 | if (newline) | |
842 | do_new_line_std(os); | |
843 | ||
844 | n = fprintf(out, " # "); | |
845 | if (color) | |
846 | n += color_fprintf(out, color, fmt, val); | |
847 | else | |
848 | n += fprintf(out, fmt, val); | |
849 | fprintf(out, " %-*s", METRIC_LEN - n - 1, unit); | |
850 | } | |
851 | ||
852 | static void new_line_csv(void *ctx) | |
853 | { | |
854 | struct outstate *os = ctx; | |
855 | int i; | |
856 | ||
857 | fputc('\n', os->fh); | |
858 | if (os->prefix) | |
859 | fprintf(os->fh, "%s%s", os->prefix, csv_sep); | |
860 | aggr_printout(os->evsel, os->id, os->nr); | |
861 | for (i = 0; i < os->nfields; i++) | |
862 | fputs(csv_sep, os->fh); | |
863 | } | |
864 | ||
865 | static void print_metric_csv(void *ctx, | |
866 | const char *color __maybe_unused, | |
867 | const char *fmt, const char *unit, double val) | |
868 | { | |
869 | struct outstate *os = ctx; | |
870 | FILE *out = os->fh; | |
871 | char buf[64], *vals, *ends; | |
872 | ||
873 | if (unit == NULL || fmt == NULL) { | |
874 | fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep); | |
875 | return; | |
876 | } | |
877 | snprintf(buf, sizeof(buf), fmt, val); | |
878 | vals = buf; | |
879 | while (isspace(*vals)) | |
880 | vals++; | |
881 | ends = vals; | |
882 | while (isdigit(*ends) || *ends == '.') | |
883 | ends++; | |
884 | *ends = 0; | |
885 | while (isspace(*unit)) | |
886 | unit++; | |
887 | fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit); | |
888 | } | |
889 | ||
890 | #define METRIC_ONLY_LEN 20 | |
891 | ||
892 | /* Filter out some columns that don't work well in metrics only mode */ | |
893 | ||
894 | static bool valid_only_metric(const char *unit) | |
895 | { | |
896 | if (!unit) | |
897 | return false; | |
898 | if (strstr(unit, "/sec") || | |
899 | strstr(unit, "hz") || | |
900 | strstr(unit, "Hz") || | |
901 | strstr(unit, "CPUs utilized")) | |
902 | return false; | |
903 | return true; | |
904 | } | |
905 | ||
906 | static const char *fixunit(char *buf, struct perf_evsel *evsel, | |
907 | const char *unit) | |
908 | { | |
909 | if (!strncmp(unit, "of all", 6)) { | |
910 | snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel), | |
911 | unit); | |
912 | return buf; | |
913 | } | |
914 | return unit; | |
915 | } | |
916 | ||
917 | static void print_metric_only(void *ctx, const char *color, const char *fmt, | |
918 | const char *unit, double val) | |
919 | { | |
920 | struct outstate *os = ctx; | |
921 | FILE *out = os->fh; | |
922 | int n; | |
923 | char buf[1024]; | |
924 | unsigned mlen = METRIC_ONLY_LEN; | |
925 | ||
926 | if (!valid_only_metric(unit)) | |
927 | return; | |
928 | unit = fixunit(buf, os->evsel, unit); | |
929 | if (color) | |
930 | n = color_fprintf(out, color, fmt, val); | |
931 | else | |
932 | n = fprintf(out, fmt, val); | |
933 | if (n > METRIC_ONLY_LEN) | |
934 | n = METRIC_ONLY_LEN; | |
935 | if (mlen < strlen(unit)) | |
936 | mlen = strlen(unit) + 1; | |
937 | fprintf(out, "%*s", mlen - n, ""); | |
938 | } | |
939 | ||
940 | static void print_metric_only_csv(void *ctx, const char *color __maybe_unused, | |
941 | const char *fmt, | |
942 | const char *unit, double val) | |
943 | { | |
944 | struct outstate *os = ctx; | |
945 | FILE *out = os->fh; | |
946 | char buf[64], *vals, *ends; | |
947 | char tbuf[1024]; | |
948 | ||
949 | if (!valid_only_metric(unit)) | |
950 | return; | |
951 | unit = fixunit(tbuf, os->evsel, unit); | |
952 | snprintf(buf, sizeof buf, fmt, val); | |
953 | vals = buf; | |
954 | while (isspace(*vals)) | |
955 | vals++; | |
956 | ends = vals; | |
957 | while (isdigit(*ends) || *ends == '.') | |
958 | ends++; | |
959 | *ends = 0; | |
960 | fprintf(out, "%s%s", vals, csv_sep); | |
961 | } | |
962 | ||
963 | static void new_line_metric(void *ctx __maybe_unused) | |
964 | { | |
965 | } | |
966 | ||
967 | static void print_metric_header(void *ctx, const char *color __maybe_unused, | |
968 | const char *fmt __maybe_unused, | |
969 | const char *unit, double val __maybe_unused) | |
970 | { | |
971 | struct outstate *os = ctx; | |
972 | char tbuf[1024]; | |
973 | ||
974 | if (!valid_only_metric(unit)) | |
975 | return; | |
976 | unit = fixunit(tbuf, os->evsel, unit); | |
977 | if (csv_output) | |
978 | fprintf(os->fh, "%s%s", unit, csv_sep); | |
979 | else | |
980 | fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit); | |
981 | } | |
982 | ||
983 | static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg) | |
984 | { | |
985 | FILE *output = stat_config.output; | |
986 | double msecs = avg / NSEC_PER_MSEC; | |
987 | const char *fmt_v, *fmt_n; | |
988 | char name[25]; | |
989 | ||
990 | fmt_v = csv_output ? "%.6f%s" : "%18.6f%s"; | |
991 | fmt_n = csv_output ? "%s" : "%-25s"; | |
992 | ||
993 | aggr_printout(evsel, id, nr); | |
994 | ||
995 | scnprintf(name, sizeof(name), "%s%s", | |
996 | perf_evsel__name(evsel), csv_output ? "" : " (msec)"); | |
997 | ||
998 | fprintf(output, fmt_v, msecs, csv_sep); | |
999 | ||
1000 | if (csv_output) | |
1001 | fprintf(output, "%s%s", evsel->unit, csv_sep); | |
1002 | else | |
1003 | fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep); | |
1004 | ||
1005 | fprintf(output, fmt_n, name); | |
1006 | ||
1007 | if (evsel->cgrp) | |
1008 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); | |
1009 | } | |
1010 | ||
1011 | static int first_shadow_cpu(struct perf_evsel *evsel, int id) | |
1012 | { | |
1013 | int i; | |
1014 | ||
1015 | if (!aggr_get_id) | |
1016 | return 0; | |
1017 | ||
1018 | if (stat_config.aggr_mode == AGGR_NONE) | |
1019 | return id; | |
1020 | ||
1021 | if (stat_config.aggr_mode == AGGR_GLOBAL) | |
1022 | return 0; | |
1023 | ||
1024 | for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) { | |
1025 | int cpu2 = perf_evsel__cpus(evsel)->map[i]; | |
1026 | ||
1027 | if (aggr_get_id(evsel_list->cpus, cpu2) == id) | |
1028 | return cpu2; | |
1029 | } | |
1030 | return 0; | |
1031 | } | |
1032 | ||
1033 | static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg) | |
1034 | { | |
1035 | FILE *output = stat_config.output; | |
1036 | double sc = evsel->scale; | |
1037 | const char *fmt; | |
1038 | ||
1039 | if (csv_output) { | |
1040 | fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s"; | |
1041 | } else { | |
1042 | if (big_num) | |
1043 | fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s"; | |
1044 | else | |
1045 | fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s"; | |
1046 | } | |
1047 | ||
1048 | aggr_printout(evsel, id, nr); | |
1049 | ||
1050 | fprintf(output, fmt, avg, csv_sep); | |
1051 | ||
1052 | if (evsel->unit) | |
1053 | fprintf(output, "%-*s%s", | |
1054 | csv_output ? 0 : unit_width, | |
1055 | evsel->unit, csv_sep); | |
1056 | ||
1057 | fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel)); | |
1058 | ||
1059 | if (evsel->cgrp) | |
1060 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); | |
1061 | } | |
1062 | ||
1063 | static void printout(int id, int nr, struct perf_evsel *counter, double uval, | |
1064 | char *prefix, u64 run, u64 ena, double noise) | |
1065 | { | |
1066 | struct perf_stat_output_ctx out; | |
1067 | struct outstate os = { | |
1068 | .fh = stat_config.output, | |
1069 | .prefix = prefix ? prefix : "", | |
1070 | .id = id, | |
1071 | .nr = nr, | |
1072 | .evsel = counter, | |
1073 | }; | |
1074 | print_metric_t pm = print_metric_std; | |
1075 | void (*nl)(void *); | |
1076 | ||
1077 | if (metric_only) { | |
1078 | nl = new_line_metric; | |
1079 | if (csv_output) | |
1080 | pm = print_metric_only_csv; | |
1081 | else | |
1082 | pm = print_metric_only; | |
1083 | } else | |
1084 | nl = new_line_std; | |
1085 | ||
1086 | if (csv_output && !metric_only) { | |
1087 | static int aggr_fields[] = { | |
1088 | [AGGR_GLOBAL] = 0, | |
1089 | [AGGR_THREAD] = 1, | |
1090 | [AGGR_NONE] = 1, | |
1091 | [AGGR_SOCKET] = 2, | |
1092 | [AGGR_CORE] = 2, | |
1093 | }; | |
1094 | ||
1095 | pm = print_metric_csv; | |
1096 | nl = new_line_csv; | |
1097 | os.nfields = 3; | |
1098 | os.nfields += aggr_fields[stat_config.aggr_mode]; | |
1099 | if (counter->cgrp) | |
1100 | os.nfields++; | |
1101 | } | |
1102 | if (run == 0 || ena == 0 || counter->counts->scaled == -1) { | |
1103 | if (metric_only) { | |
1104 | pm(&os, NULL, "", "", 0); | |
1105 | return; | |
1106 | } | |
1107 | aggr_printout(counter, id, nr); | |
1108 | ||
1109 | fprintf(stat_config.output, "%*s%s", | |
1110 | csv_output ? 0 : 18, | |
1111 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, | |
1112 | csv_sep); | |
1113 | ||
1114 | if (counter->supported) | |
1115 | print_free_counters_hint = 1; | |
1116 | ||
1117 | fprintf(stat_config.output, "%-*s%s", | |
1118 | csv_output ? 0 : unit_width, | |
1119 | counter->unit, csv_sep); | |
1120 | ||
1121 | fprintf(stat_config.output, "%*s", | |
1122 | csv_output ? 0 : -25, | |
1123 | perf_evsel__name(counter)); | |
1124 | ||
1125 | if (counter->cgrp) | |
1126 | fprintf(stat_config.output, "%s%s", | |
1127 | csv_sep, counter->cgrp->name); | |
1128 | ||
1129 | if (!csv_output) | |
1130 | pm(&os, NULL, NULL, "", 0); | |
1131 | print_noise(counter, noise); | |
1132 | print_running(run, ena); | |
1133 | if (csv_output) | |
1134 | pm(&os, NULL, NULL, "", 0); | |
1135 | return; | |
1136 | } | |
1137 | ||
1138 | if (metric_only) | |
1139 | /* nothing */; | |
1140 | else if (nsec_counter(counter)) | |
1141 | nsec_printout(id, nr, counter, uval); | |
1142 | else | |
1143 | abs_printout(id, nr, counter, uval); | |
1144 | ||
1145 | out.print_metric = pm; | |
1146 | out.new_line = nl; | |
1147 | out.ctx = &os; | |
1148 | ||
1149 | if (csv_output && !metric_only) { | |
1150 | print_noise(counter, noise); | |
1151 | print_running(run, ena); | |
1152 | } | |
1153 | ||
1154 | perf_stat__print_shadow_stats(counter, uval, | |
1155 | first_shadow_cpu(counter, id), | |
1156 | &out); | |
1157 | if (!csv_output && !metric_only) { | |
1158 | print_noise(counter, noise); | |
1159 | print_running(run, ena); | |
1160 | } | |
1161 | } | |
1162 | ||
1163 | static void aggr_update_shadow(void) | |
1164 | { | |
1165 | int cpu, s2, id, s; | |
1166 | u64 val; | |
1167 | struct perf_evsel *counter; | |
1168 | ||
1169 | for (s = 0; s < aggr_map->nr; s++) { | |
1170 | id = aggr_map->map[s]; | |
1171 | evlist__for_each_entry(evsel_list, counter) { | |
1172 | val = 0; | |
1173 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { | |
1174 | s2 = aggr_get_id(evsel_list->cpus, cpu); | |
1175 | if (s2 != id) | |
1176 | continue; | |
1177 | val += perf_counts(counter->counts, cpu, 0)->val; | |
1178 | } | |
1179 | val = val * counter->scale; | |
1180 | perf_stat__update_shadow_stats(counter, &val, | |
1181 | first_shadow_cpu(counter, id)); | |
1182 | } | |
1183 | } | |
1184 | } | |
1185 | ||
1186 | static void collect_all_aliases(struct perf_evsel *counter, | |
1187 | void (*cb)(struct perf_evsel *counter, void *data, | |
1188 | bool first), | |
1189 | void *data) | |
1190 | { | |
1191 | struct perf_evsel *alias; | |
1192 | ||
1193 | alias = list_prepare_entry(counter, &(evsel_list->entries), node); | |
1194 | list_for_each_entry_continue (alias, &evsel_list->entries, node) { | |
1195 | if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) || | |
1196 | alias->scale != counter->scale || | |
1197 | alias->cgrp != counter->cgrp || | |
1198 | strcmp(alias->unit, counter->unit) || | |
1199 | nsec_counter(alias) != nsec_counter(counter)) | |
1200 | break; | |
1201 | alias->merged_stat = true; | |
1202 | cb(alias, data, false); | |
1203 | } | |
1204 | } | |
1205 | ||
1206 | static bool collect_data(struct perf_evsel *counter, | |
1207 | void (*cb)(struct perf_evsel *counter, void *data, | |
1208 | bool first), | |
1209 | void *data) | |
1210 | { | |
1211 | if (counter->merged_stat) | |
1212 | return false; | |
1213 | cb(counter, data, true); | |
1214 | if (!no_merge) | |
1215 | collect_all_aliases(counter, cb, data); | |
1216 | return true; | |
1217 | } | |
1218 | ||
1219 | struct aggr_data { | |
1220 | u64 ena, run, val; | |
1221 | int id; | |
1222 | int nr; | |
1223 | int cpu; | |
1224 | }; | |
1225 | ||
1226 | static void aggr_cb(struct perf_evsel *counter, void *data, bool first) | |
1227 | { | |
1228 | struct aggr_data *ad = data; | |
1229 | int cpu, s2; | |
1230 | ||
1231 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { | |
1232 | struct perf_counts_values *counts; | |
1233 | ||
1234 | s2 = aggr_get_id(perf_evsel__cpus(counter), cpu); | |
1235 | if (s2 != ad->id) | |
1236 | continue; | |
1237 | if (first) | |
1238 | ad->nr++; | |
1239 | counts = perf_counts(counter->counts, cpu, 0); | |
1240 | ad->val += counts->val; | |
1241 | ad->ena += counts->ena; | |
1242 | ad->run += counts->run; | |
1243 | } | |
1244 | } | |
1245 | ||
1246 | static void print_aggr(char *prefix) | |
1247 | { | |
1248 | FILE *output = stat_config.output; | |
1249 | struct perf_evsel *counter; | |
1250 | int s, id, nr; | |
1251 | double uval; | |
1252 | u64 ena, run, val; | |
1253 | bool first; | |
1254 | ||
1255 | if (!(aggr_map || aggr_get_id)) | |
1256 | return; | |
1257 | ||
1258 | aggr_update_shadow(); | |
1259 | ||
1260 | /* | |
1261 | * With metric_only everything is on a single line. | |
1262 | * Without each counter has its own line. | |
1263 | */ | |
1264 | for (s = 0; s < aggr_map->nr; s++) { | |
1265 | struct aggr_data ad; | |
1266 | if (prefix && metric_only) | |
1267 | fprintf(output, "%s", prefix); | |
1268 | ||
1269 | ad.id = id = aggr_map->map[s]; | |
1270 | first = true; | |
1271 | evlist__for_each_entry(evsel_list, counter) { | |
1272 | ad.val = ad.ena = ad.run = 0; | |
1273 | ad.nr = 0; | |
1274 | if (!collect_data(counter, aggr_cb, &ad)) | |
1275 | continue; | |
1276 | nr = ad.nr; | |
1277 | ena = ad.ena; | |
1278 | run = ad.run; | |
1279 | val = ad.val; | |
1280 | if (first && metric_only) { | |
1281 | first = false; | |
1282 | aggr_printout(counter, id, nr); | |
1283 | } | |
1284 | if (prefix && !metric_only) | |
1285 | fprintf(output, "%s", prefix); | |
1286 | ||
1287 | uval = val * counter->scale; | |
1288 | printout(id, nr, counter, uval, prefix, run, ena, 1.0); | |
1289 | if (!metric_only) | |
1290 | fputc('\n', output); | |
1291 | } | |
1292 | if (metric_only) | |
1293 | fputc('\n', output); | |
1294 | } | |
1295 | } | |
1296 | ||
1297 | static void print_aggr_thread(struct perf_evsel *counter, char *prefix) | |
1298 | { | |
1299 | FILE *output = stat_config.output; | |
1300 | int nthreads = thread_map__nr(counter->threads); | |
1301 | int ncpus = cpu_map__nr(counter->cpus); | |
1302 | int cpu, thread; | |
1303 | double uval; | |
1304 | ||
1305 | for (thread = 0; thread < nthreads; thread++) { | |
1306 | u64 ena = 0, run = 0, val = 0; | |
1307 | ||
1308 | for (cpu = 0; cpu < ncpus; cpu++) { | |
1309 | val += perf_counts(counter->counts, cpu, thread)->val; | |
1310 | ena += perf_counts(counter->counts, cpu, thread)->ena; | |
1311 | run += perf_counts(counter->counts, cpu, thread)->run; | |
1312 | } | |
1313 | ||
1314 | if (prefix) | |
1315 | fprintf(output, "%s", prefix); | |
1316 | ||
1317 | uval = val * counter->scale; | |
1318 | printout(thread, 0, counter, uval, prefix, run, ena, 1.0); | |
1319 | fputc('\n', output); | |
1320 | } | |
1321 | } | |
1322 | ||
1323 | struct caggr_data { | |
1324 | double avg, avg_enabled, avg_running; | |
1325 | }; | |
1326 | ||
1327 | static void counter_aggr_cb(struct perf_evsel *counter, void *data, | |
1328 | bool first __maybe_unused) | |
1329 | { | |
1330 | struct caggr_data *cd = data; | |
1331 | struct perf_stat_evsel *ps = counter->priv; | |
1332 | ||
1333 | cd->avg += avg_stats(&ps->res_stats[0]); | |
1334 | cd->avg_enabled += avg_stats(&ps->res_stats[1]); | |
1335 | cd->avg_running += avg_stats(&ps->res_stats[2]); | |
1336 | } | |
1337 | ||
1338 | /* | |
1339 | * Print out the results of a single counter: | |
1340 | * aggregated counts in system-wide mode | |
1341 | */ | |
1342 | static void print_counter_aggr(struct perf_evsel *counter, char *prefix) | |
1343 | { | |
1344 | FILE *output = stat_config.output; | |
1345 | double uval; | |
1346 | struct caggr_data cd = { .avg = 0.0 }; | |
1347 | ||
1348 | if (!collect_data(counter, counter_aggr_cb, &cd)) | |
1349 | return; | |
1350 | ||
1351 | if (prefix && !metric_only) | |
1352 | fprintf(output, "%s", prefix); | |
1353 | ||
1354 | uval = cd.avg * counter->scale; | |
1355 | printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled, cd.avg); | |
1356 | if (!metric_only) | |
1357 | fprintf(output, "\n"); | |
1358 | } | |
1359 | ||
1360 | static void counter_cb(struct perf_evsel *counter, void *data, | |
1361 | bool first __maybe_unused) | |
1362 | { | |
1363 | struct aggr_data *ad = data; | |
1364 | ||
1365 | ad->val += perf_counts(counter->counts, ad->cpu, 0)->val; | |
1366 | ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena; | |
1367 | ad->run += perf_counts(counter->counts, ad->cpu, 0)->run; | |
1368 | } | |
1369 | ||
1370 | /* | |
1371 | * Print out the results of a single counter: | |
1372 | * does not use aggregated count in system-wide | |
1373 | */ | |
1374 | static void print_counter(struct perf_evsel *counter, char *prefix) | |
1375 | { | |
1376 | FILE *output = stat_config.output; | |
1377 | u64 ena, run, val; | |
1378 | double uval; | |
1379 | int cpu; | |
1380 | ||
1381 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { | |
1382 | struct aggr_data ad = { .cpu = cpu }; | |
1383 | ||
1384 | if (!collect_data(counter, counter_cb, &ad)) | |
1385 | return; | |
1386 | val = ad.val; | |
1387 | ena = ad.ena; | |
1388 | run = ad.run; | |
1389 | ||
1390 | if (prefix) | |
1391 | fprintf(output, "%s", prefix); | |
1392 | ||
1393 | uval = val * counter->scale; | |
1394 | printout(cpu, 0, counter, uval, prefix, run, ena, 1.0); | |
1395 | ||
1396 | fputc('\n', output); | |
1397 | } | |
1398 | } | |
1399 | ||
1400 | static void print_no_aggr_metric(char *prefix) | |
1401 | { | |
1402 | int cpu; | |
1403 | int nrcpus = 0; | |
1404 | struct perf_evsel *counter; | |
1405 | u64 ena, run, val; | |
1406 | double uval; | |
1407 | ||
1408 | nrcpus = evsel_list->cpus->nr; | |
1409 | for (cpu = 0; cpu < nrcpus; cpu++) { | |
1410 | bool first = true; | |
1411 | ||
1412 | if (prefix) | |
1413 | fputs(prefix, stat_config.output); | |
1414 | evlist__for_each_entry(evsel_list, counter) { | |
1415 | if (first) { | |
1416 | aggr_printout(counter, cpu, 0); | |
1417 | first = false; | |
1418 | } | |
1419 | val = perf_counts(counter->counts, cpu, 0)->val; | |
1420 | ena = perf_counts(counter->counts, cpu, 0)->ena; | |
1421 | run = perf_counts(counter->counts, cpu, 0)->run; | |
1422 | ||
1423 | uval = val * counter->scale; | |
1424 | printout(cpu, 0, counter, uval, prefix, run, ena, 1.0); | |
1425 | } | |
1426 | fputc('\n', stat_config.output); | |
1427 | } | |
1428 | } | |
1429 | ||
1430 | static int aggr_header_lens[] = { | |
1431 | [AGGR_CORE] = 18, | |
1432 | [AGGR_SOCKET] = 12, | |
1433 | [AGGR_NONE] = 6, | |
1434 | [AGGR_THREAD] = 24, | |
1435 | [AGGR_GLOBAL] = 0, | |
1436 | }; | |
1437 | ||
1438 | static const char *aggr_header_csv[] = { | |
1439 | [AGGR_CORE] = "core,cpus,", | |
1440 | [AGGR_SOCKET] = "socket,cpus", | |
1441 | [AGGR_NONE] = "cpu,", | |
1442 | [AGGR_THREAD] = "comm-pid,", | |
1443 | [AGGR_GLOBAL] = "" | |
1444 | }; | |
1445 | ||
1446 | static void print_metric_headers(const char *prefix, bool no_indent) | |
1447 | { | |
1448 | struct perf_stat_output_ctx out; | |
1449 | struct perf_evsel *counter; | |
1450 | struct outstate os = { | |
1451 | .fh = stat_config.output | |
1452 | }; | |
1453 | ||
1454 | if (prefix) | |
1455 | fprintf(stat_config.output, "%s", prefix); | |
1456 | ||
1457 | if (!csv_output && !no_indent) | |
1458 | fprintf(stat_config.output, "%*s", | |
1459 | aggr_header_lens[stat_config.aggr_mode], ""); | |
1460 | if (csv_output) { | |
1461 | if (stat_config.interval) | |
1462 | fputs("time,", stat_config.output); | |
1463 | fputs(aggr_header_csv[stat_config.aggr_mode], | |
1464 | stat_config.output); | |
1465 | } | |
1466 | ||
1467 | /* Print metrics headers only */ | |
1468 | evlist__for_each_entry(evsel_list, counter) { | |
1469 | os.evsel = counter; | |
1470 | out.ctx = &os; | |
1471 | out.print_metric = print_metric_header; | |
1472 | out.new_line = new_line_metric; | |
1473 | os.evsel = counter; | |
1474 | perf_stat__print_shadow_stats(counter, 0, | |
1475 | 0, | |
1476 | &out); | |
1477 | } | |
1478 | fputc('\n', stat_config.output); | |
1479 | } | |
1480 | ||
1481 | static void print_interval(char *prefix, struct timespec *ts) | |
1482 | { | |
1483 | FILE *output = stat_config.output; | |
1484 | static int num_print_interval; | |
1485 | ||
1486 | sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep); | |
1487 | ||
1488 | if (num_print_interval == 0 && !csv_output) { | |
1489 | switch (stat_config.aggr_mode) { | |
1490 | case AGGR_SOCKET: | |
1491 | fprintf(output, "# time socket cpus"); | |
1492 | if (!metric_only) | |
1493 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
1494 | break; | |
1495 | case AGGR_CORE: | |
1496 | fprintf(output, "# time core cpus"); | |
1497 | if (!metric_only) | |
1498 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
1499 | break; | |
1500 | case AGGR_NONE: | |
1501 | fprintf(output, "# time CPU"); | |
1502 | if (!metric_only) | |
1503 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
1504 | break; | |
1505 | case AGGR_THREAD: | |
1506 | fprintf(output, "# time comm-pid"); | |
1507 | if (!metric_only) | |
1508 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
1509 | break; | |
1510 | case AGGR_GLOBAL: | |
1511 | default: | |
1512 | fprintf(output, "# time"); | |
1513 | if (!metric_only) | |
1514 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
1515 | case AGGR_UNSET: | |
1516 | break; | |
1517 | } | |
1518 | } | |
1519 | ||
1520 | if (num_print_interval == 0 && metric_only) | |
1521 | print_metric_headers(" ", true); | |
1522 | if (++num_print_interval == 25) | |
1523 | num_print_interval = 0; | |
1524 | } | |
1525 | ||
1526 | static void print_header(int argc, const char **argv) | |
1527 | { | |
1528 | FILE *output = stat_config.output; | |
1529 | int i; | |
1530 | ||
1531 | fflush(stdout); | |
1532 | ||
1533 | if (!csv_output) { | |
1534 | fprintf(output, "\n"); | |
1535 | fprintf(output, " Performance counter stats for "); | |
1536 | if (target.system_wide) | |
1537 | fprintf(output, "\'system wide"); | |
1538 | else if (target.cpu_list) | |
1539 | fprintf(output, "\'CPU(s) %s", target.cpu_list); | |
1540 | else if (!target__has_task(&target)) { | |
1541 | fprintf(output, "\'%s", argv ? argv[0] : "pipe"); | |
1542 | for (i = 1; argv && (i < argc); i++) | |
1543 | fprintf(output, " %s", argv[i]); | |
1544 | } else if (target.pid) | |
1545 | fprintf(output, "process id \'%s", target.pid); | |
1546 | else | |
1547 | fprintf(output, "thread id \'%s", target.tid); | |
1548 | ||
1549 | fprintf(output, "\'"); | |
1550 | if (run_count > 1) | |
1551 | fprintf(output, " (%d runs)", run_count); | |
1552 | fprintf(output, ":\n\n"); | |
1553 | } | |
1554 | } | |
1555 | ||
1556 | static void print_footer(void) | |
1557 | { | |
1558 | FILE *output = stat_config.output; | |
1559 | ||
1560 | if (!null_run) | |
1561 | fprintf(output, "\n"); | |
1562 | fprintf(output, " %17.9f seconds time elapsed", | |
1563 | avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC); | |
1564 | if (run_count > 1) { | |
1565 | fprintf(output, " "); | |
1566 | print_noise_pct(stddev_stats(&walltime_nsecs_stats), | |
1567 | avg_stats(&walltime_nsecs_stats)); | |
1568 | } | |
1569 | fprintf(output, "\n\n"); | |
1570 | ||
1571 | if (print_free_counters_hint) | |
1572 | fprintf(output, | |
1573 | "Some events weren't counted. Try disabling the NMI watchdog:\n" | |
1574 | " echo 0 > /proc/sys/kernel/nmi_watchdog\n" | |
1575 | " perf stat ...\n" | |
1576 | " echo 1 > /proc/sys/kernel/nmi_watchdog\n"); | |
1577 | } | |
1578 | ||
1579 | static void print_counters(struct timespec *ts, int argc, const char **argv) | |
1580 | { | |
1581 | int interval = stat_config.interval; | |
1582 | struct perf_evsel *counter; | |
1583 | char buf[64], *prefix = NULL; | |
1584 | ||
1585 | /* Do not print anything if we record to the pipe. */ | |
1586 | if (STAT_RECORD && perf_stat.file.is_pipe) | |
1587 | return; | |
1588 | ||
1589 | if (interval) | |
1590 | print_interval(prefix = buf, ts); | |
1591 | else | |
1592 | print_header(argc, argv); | |
1593 | ||
1594 | if (metric_only) { | |
1595 | static int num_print_iv; | |
1596 | ||
1597 | if (num_print_iv == 0 && !interval) | |
1598 | print_metric_headers(prefix, false); | |
1599 | if (num_print_iv++ == 25) | |
1600 | num_print_iv = 0; | |
1601 | if (stat_config.aggr_mode == AGGR_GLOBAL && prefix) | |
1602 | fprintf(stat_config.output, "%s", prefix); | |
1603 | } | |
1604 | ||
1605 | switch (stat_config.aggr_mode) { | |
1606 | case AGGR_CORE: | |
1607 | case AGGR_SOCKET: | |
1608 | print_aggr(prefix); | |
1609 | break; | |
1610 | case AGGR_THREAD: | |
1611 | evlist__for_each_entry(evsel_list, counter) | |
1612 | print_aggr_thread(counter, prefix); | |
1613 | break; | |
1614 | case AGGR_GLOBAL: | |
1615 | evlist__for_each_entry(evsel_list, counter) | |
1616 | print_counter_aggr(counter, prefix); | |
1617 | if (metric_only) | |
1618 | fputc('\n', stat_config.output); | |
1619 | break; | |
1620 | case AGGR_NONE: | |
1621 | if (metric_only) | |
1622 | print_no_aggr_metric(prefix); | |
1623 | else { | |
1624 | evlist__for_each_entry(evsel_list, counter) | |
1625 | print_counter(counter, prefix); | |
1626 | } | |
1627 | break; | |
1628 | case AGGR_UNSET: | |
1629 | default: | |
1630 | break; | |
1631 | } | |
1632 | ||
1633 | if (!interval && !csv_output) | |
1634 | print_footer(); | |
1635 | ||
1636 | fflush(stat_config.output); | |
1637 | } | |
1638 | ||
1639 | static volatile int signr = -1; | |
1640 | ||
1641 | static void skip_signal(int signo) | |
1642 | { | |
1643 | if ((child_pid == -1) || stat_config.interval) | |
1644 | done = 1; | |
1645 | ||
1646 | signr = signo; | |
1647 | /* | |
1648 | * render child_pid harmless | |
1649 | * won't send SIGTERM to a random | |
1650 | * process in case of race condition | |
1651 | * and fast PID recycling | |
1652 | */ | |
1653 | child_pid = -1; | |
1654 | } | |
1655 | ||
1656 | static void sig_atexit(void) | |
1657 | { | |
1658 | sigset_t set, oset; | |
1659 | ||
1660 | /* | |
1661 | * avoid race condition with SIGCHLD handler | |
1662 | * in skip_signal() which is modifying child_pid | |
1663 | * goal is to avoid send SIGTERM to a random | |
1664 | * process | |
1665 | */ | |
1666 | sigemptyset(&set); | |
1667 | sigaddset(&set, SIGCHLD); | |
1668 | sigprocmask(SIG_BLOCK, &set, &oset); | |
1669 | ||
1670 | if (child_pid != -1) | |
1671 | kill(child_pid, SIGTERM); | |
1672 | ||
1673 | sigprocmask(SIG_SETMASK, &oset, NULL); | |
1674 | ||
1675 | if (signr == -1) | |
1676 | return; | |
1677 | ||
1678 | signal(signr, SIG_DFL); | |
1679 | kill(getpid(), signr); | |
1680 | } | |
1681 | ||
1682 | static int stat__set_big_num(const struct option *opt __maybe_unused, | |
1683 | const char *s __maybe_unused, int unset) | |
1684 | { | |
1685 | big_num_opt = unset ? 0 : 1; | |
1686 | return 0; | |
1687 | } | |
1688 | ||
1689 | static int enable_metric_only(const struct option *opt __maybe_unused, | |
1690 | const char *s __maybe_unused, int unset) | |
1691 | { | |
1692 | force_metric_only = true; | |
1693 | metric_only = !unset; | |
1694 | return 0; | |
1695 | } | |
1696 | ||
1697 | static const struct option stat_options[] = { | |
1698 | OPT_BOOLEAN('T', "transaction", &transaction_run, | |
1699 | "hardware transaction statistics"), | |
1700 | OPT_CALLBACK('e', "event", &evsel_list, "event", | |
1701 | "event selector. use 'perf list' to list available events", | |
1702 | parse_events_option), | |
1703 | OPT_CALLBACK(0, "filter", &evsel_list, "filter", | |
1704 | "event filter", parse_filter), | |
1705 | OPT_BOOLEAN('i', "no-inherit", &no_inherit, | |
1706 | "child tasks do not inherit counters"), | |
1707 | OPT_STRING('p', "pid", &target.pid, "pid", | |
1708 | "stat events on existing process id"), | |
1709 | OPT_STRING('t', "tid", &target.tid, "tid", | |
1710 | "stat events on existing thread id"), | |
1711 | OPT_BOOLEAN('a', "all-cpus", &target.system_wide, | |
1712 | "system-wide collection from all CPUs"), | |
1713 | OPT_BOOLEAN('g', "group", &group, | |
1714 | "put the counters into a counter group"), | |
1715 | OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"), | |
1716 | OPT_INCR('v', "verbose", &verbose, | |
1717 | "be more verbose (show counter open errors, etc)"), | |
1718 | OPT_INTEGER('r', "repeat", &run_count, | |
1719 | "repeat command and print average + stddev (max: 100, forever: 0)"), | |
1720 | OPT_BOOLEAN('n', "null", &null_run, | |
1721 | "null run - dont start any counters"), | |
1722 | OPT_INCR('d', "detailed", &detailed_run, | |
1723 | "detailed run - start a lot of events"), | |
1724 | OPT_BOOLEAN('S', "sync", &sync_run, | |
1725 | "call sync() before starting a run"), | |
1726 | OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, | |
1727 | "print large numbers with thousands\' separators", | |
1728 | stat__set_big_num), | |
1729 | OPT_STRING('C', "cpu", &target.cpu_list, "cpu", | |
1730 | "list of cpus to monitor in system-wide"), | |
1731 | OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode, | |
1732 | "disable CPU count aggregation", AGGR_NONE), | |
1733 | OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"), | |
1734 | OPT_STRING('x', "field-separator", &csv_sep, "separator", | |
1735 | "print counts with custom separator"), | |
1736 | OPT_CALLBACK('G', "cgroup", &evsel_list, "name", | |
1737 | "monitor event in cgroup name only", parse_cgroups), | |
1738 | OPT_STRING('o', "output", &output_name, "file", "output file name"), | |
1739 | OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), | |
1740 | OPT_INTEGER(0, "log-fd", &output_fd, | |
1741 | "log output to fd, instead of stderr"), | |
1742 | OPT_STRING(0, "pre", &pre_cmd, "command", | |
1743 | "command to run prior to the measured command"), | |
1744 | OPT_STRING(0, "post", &post_cmd, "command", | |
1745 | "command to run after to the measured command"), | |
1746 | OPT_UINTEGER('I', "interval-print", &stat_config.interval, | |
1747 | "print counts at regular interval in ms (>= 10)"), | |
1748 | OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode, | |
1749 | "aggregate counts per processor socket", AGGR_SOCKET), | |
1750 | OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode, | |
1751 | "aggregate counts per physical processor core", AGGR_CORE), | |
1752 | OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode, | |
1753 | "aggregate counts per thread", AGGR_THREAD), | |
1754 | OPT_UINTEGER('D', "delay", &initial_delay, | |
1755 | "ms to wait before starting measurement after program start"), | |
1756 | OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL, | |
1757 | "Only print computed metrics. No raw values", enable_metric_only), | |
1758 | OPT_BOOLEAN(0, "topdown", &topdown_run, | |
1759 | "measure topdown level 1 statistics"), | |
1760 | OPT_END() | |
1761 | }; | |
1762 | ||
1763 | static int perf_stat__get_socket(struct cpu_map *map, int cpu) | |
1764 | { | |
1765 | return cpu_map__get_socket(map, cpu, NULL); | |
1766 | } | |
1767 | ||
1768 | static int perf_stat__get_core(struct cpu_map *map, int cpu) | |
1769 | { | |
1770 | return cpu_map__get_core(map, cpu, NULL); | |
1771 | } | |
1772 | ||
1773 | static int cpu_map__get_max(struct cpu_map *map) | |
1774 | { | |
1775 | int i, max = -1; | |
1776 | ||
1777 | for (i = 0; i < map->nr; i++) { | |
1778 | if (map->map[i] > max) | |
1779 | max = map->map[i]; | |
1780 | } | |
1781 | ||
1782 | return max; | |
1783 | } | |
1784 | ||
1785 | static struct cpu_map *cpus_aggr_map; | |
1786 | ||
1787 | static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx) | |
1788 | { | |
1789 | int cpu; | |
1790 | ||
1791 | if (idx >= map->nr) | |
1792 | return -1; | |
1793 | ||
1794 | cpu = map->map[idx]; | |
1795 | ||
1796 | if (cpus_aggr_map->map[cpu] == -1) | |
1797 | cpus_aggr_map->map[cpu] = get_id(map, idx); | |
1798 | ||
1799 | return cpus_aggr_map->map[cpu]; | |
1800 | } | |
1801 | ||
1802 | static int perf_stat__get_socket_cached(struct cpu_map *map, int idx) | |
1803 | { | |
1804 | return perf_stat__get_aggr(perf_stat__get_socket, map, idx); | |
1805 | } | |
1806 | ||
1807 | static int perf_stat__get_core_cached(struct cpu_map *map, int idx) | |
1808 | { | |
1809 | return perf_stat__get_aggr(perf_stat__get_core, map, idx); | |
1810 | } | |
1811 | ||
1812 | static int perf_stat_init_aggr_mode(void) | |
1813 | { | |
1814 | int nr; | |
1815 | ||
1816 | switch (stat_config.aggr_mode) { | |
1817 | case AGGR_SOCKET: | |
1818 | if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) { | |
1819 | perror("cannot build socket map"); | |
1820 | return -1; | |
1821 | } | |
1822 | aggr_get_id = perf_stat__get_socket_cached; | |
1823 | break; | |
1824 | case AGGR_CORE: | |
1825 | if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) { | |
1826 | perror("cannot build core map"); | |
1827 | return -1; | |
1828 | } | |
1829 | aggr_get_id = perf_stat__get_core_cached; | |
1830 | break; | |
1831 | case AGGR_NONE: | |
1832 | case AGGR_GLOBAL: | |
1833 | case AGGR_THREAD: | |
1834 | case AGGR_UNSET: | |
1835 | default: | |
1836 | break; | |
1837 | } | |
1838 | ||
1839 | /* | |
1840 | * The evsel_list->cpus is the base we operate on, | |
1841 | * taking the highest cpu number to be the size of | |
1842 | * the aggregation translate cpumap. | |
1843 | */ | |
1844 | nr = cpu_map__get_max(evsel_list->cpus); | |
1845 | cpus_aggr_map = cpu_map__empty_new(nr + 1); | |
1846 | return cpus_aggr_map ? 0 : -ENOMEM; | |
1847 | } | |
1848 | ||
1849 | static void perf_stat__exit_aggr_mode(void) | |
1850 | { | |
1851 | cpu_map__put(aggr_map); | |
1852 | cpu_map__put(cpus_aggr_map); | |
1853 | aggr_map = NULL; | |
1854 | cpus_aggr_map = NULL; | |
1855 | } | |
1856 | ||
1857 | static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx) | |
1858 | { | |
1859 | int cpu; | |
1860 | ||
1861 | if (idx > map->nr) | |
1862 | return -1; | |
1863 | ||
1864 | cpu = map->map[idx]; | |
1865 | ||
1866 | if (cpu >= env->nr_cpus_avail) | |
1867 | return -1; | |
1868 | ||
1869 | return cpu; | |
1870 | } | |
1871 | ||
1872 | static int perf_env__get_socket(struct cpu_map *map, int idx, void *data) | |
1873 | { | |
1874 | struct perf_env *env = data; | |
1875 | int cpu = perf_env__get_cpu(env, map, idx); | |
1876 | ||
1877 | return cpu == -1 ? -1 : env->cpu[cpu].socket_id; | |
1878 | } | |
1879 | ||
1880 | static int perf_env__get_core(struct cpu_map *map, int idx, void *data) | |
1881 | { | |
1882 | struct perf_env *env = data; | |
1883 | int core = -1, cpu = perf_env__get_cpu(env, map, idx); | |
1884 | ||
1885 | if (cpu != -1) { | |
1886 | int socket_id = env->cpu[cpu].socket_id; | |
1887 | ||
1888 | /* | |
1889 | * Encode socket in upper 16 bits | |
1890 | * core_id is relative to socket, and | |
1891 | * we need a global id. So we combine | |
1892 | * socket + core id. | |
1893 | */ | |
1894 | core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff); | |
1895 | } | |
1896 | ||
1897 | return core; | |
1898 | } | |
1899 | ||
1900 | static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus, | |
1901 | struct cpu_map **sockp) | |
1902 | { | |
1903 | return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env); | |
1904 | } | |
1905 | ||
1906 | static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus, | |
1907 | struct cpu_map **corep) | |
1908 | { | |
1909 | return cpu_map__build_map(cpus, corep, perf_env__get_core, env); | |
1910 | } | |
1911 | ||
1912 | static int perf_stat__get_socket_file(struct cpu_map *map, int idx) | |
1913 | { | |
1914 | return perf_env__get_socket(map, idx, &perf_stat.session->header.env); | |
1915 | } | |
1916 | ||
1917 | static int perf_stat__get_core_file(struct cpu_map *map, int idx) | |
1918 | { | |
1919 | return perf_env__get_core(map, idx, &perf_stat.session->header.env); | |
1920 | } | |
1921 | ||
1922 | static int perf_stat_init_aggr_mode_file(struct perf_stat *st) | |
1923 | { | |
1924 | struct perf_env *env = &st->session->header.env; | |
1925 | ||
1926 | switch (stat_config.aggr_mode) { | |
1927 | case AGGR_SOCKET: | |
1928 | if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) { | |
1929 | perror("cannot build socket map"); | |
1930 | return -1; | |
1931 | } | |
1932 | aggr_get_id = perf_stat__get_socket_file; | |
1933 | break; | |
1934 | case AGGR_CORE: | |
1935 | if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) { | |
1936 | perror("cannot build core map"); | |
1937 | return -1; | |
1938 | } | |
1939 | aggr_get_id = perf_stat__get_core_file; | |
1940 | break; | |
1941 | case AGGR_NONE: | |
1942 | case AGGR_GLOBAL: | |
1943 | case AGGR_THREAD: | |
1944 | case AGGR_UNSET: | |
1945 | default: | |
1946 | break; | |
1947 | } | |
1948 | ||
1949 | return 0; | |
1950 | } | |
1951 | ||
1952 | static int topdown_filter_events(const char **attr, char **str, bool use_group) | |
1953 | { | |
1954 | int off = 0; | |
1955 | int i; | |
1956 | int len = 0; | |
1957 | char *s; | |
1958 | ||
1959 | for (i = 0; attr[i]; i++) { | |
1960 | if (pmu_have_event("cpu", attr[i])) { | |
1961 | len += strlen(attr[i]) + 1; | |
1962 | attr[i - off] = attr[i]; | |
1963 | } else | |
1964 | off++; | |
1965 | } | |
1966 | attr[i - off] = NULL; | |
1967 | ||
1968 | *str = malloc(len + 1 + 2); | |
1969 | if (!*str) | |
1970 | return -1; | |
1971 | s = *str; | |
1972 | if (i - off == 0) { | |
1973 | *s = 0; | |
1974 | return 0; | |
1975 | } | |
1976 | if (use_group) | |
1977 | *s++ = '{'; | |
1978 | for (i = 0; attr[i]; i++) { | |
1979 | strcpy(s, attr[i]); | |
1980 | s += strlen(s); | |
1981 | *s++ = ','; | |
1982 | } | |
1983 | if (use_group) { | |
1984 | s[-1] = '}'; | |
1985 | *s = 0; | |
1986 | } else | |
1987 | s[-1] = 0; | |
1988 | return 0; | |
1989 | } | |
1990 | ||
1991 | __weak bool arch_topdown_check_group(bool *warn) | |
1992 | { | |
1993 | *warn = false; | |
1994 | return false; | |
1995 | } | |
1996 | ||
1997 | __weak void arch_topdown_group_warn(void) | |
1998 | { | |
1999 | } | |
2000 | ||
2001 | /* | |
2002 | * Add default attributes, if there were no attributes specified or | |
2003 | * if -d/--detailed, -d -d or -d -d -d is used: | |
2004 | */ | |
2005 | static int add_default_attributes(void) | |
2006 | { | |
2007 | int err; | |
2008 | struct perf_event_attr default_attrs0[] = { | |
2009 | ||
2010 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, | |
2011 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, | |
2012 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, | |
2013 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, | |
2014 | ||
2015 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, | |
2016 | }; | |
2017 | struct perf_event_attr frontend_attrs[] = { | |
2018 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, | |
2019 | }; | |
2020 | struct perf_event_attr backend_attrs[] = { | |
2021 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, | |
2022 | }; | |
2023 | struct perf_event_attr default_attrs1[] = { | |
2024 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, | |
2025 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, | |
2026 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, | |
2027 | ||
2028 | }; | |
2029 | ||
2030 | /* | |
2031 | * Detailed stats (-d), covering the L1 and last level data caches: | |
2032 | */ | |
2033 | struct perf_event_attr detailed_attrs[] = { | |
2034 | ||
2035 | { .type = PERF_TYPE_HW_CACHE, | |
2036 | .config = | |
2037 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
2038 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2039 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2040 | ||
2041 | { .type = PERF_TYPE_HW_CACHE, | |
2042 | .config = | |
2043 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
2044 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2045 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2046 | ||
2047 | { .type = PERF_TYPE_HW_CACHE, | |
2048 | .config = | |
2049 | PERF_COUNT_HW_CACHE_LL << 0 | | |
2050 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2051 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2052 | ||
2053 | { .type = PERF_TYPE_HW_CACHE, | |
2054 | .config = | |
2055 | PERF_COUNT_HW_CACHE_LL << 0 | | |
2056 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2057 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2058 | }; | |
2059 | ||
2060 | /* | |
2061 | * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: | |
2062 | */ | |
2063 | struct perf_event_attr very_detailed_attrs[] = { | |
2064 | ||
2065 | { .type = PERF_TYPE_HW_CACHE, | |
2066 | .config = | |
2067 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
2068 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2069 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2070 | ||
2071 | { .type = PERF_TYPE_HW_CACHE, | |
2072 | .config = | |
2073 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
2074 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2075 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2076 | ||
2077 | { .type = PERF_TYPE_HW_CACHE, | |
2078 | .config = | |
2079 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
2080 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2081 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2082 | ||
2083 | { .type = PERF_TYPE_HW_CACHE, | |
2084 | .config = | |
2085 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
2086 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2087 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2088 | ||
2089 | { .type = PERF_TYPE_HW_CACHE, | |
2090 | .config = | |
2091 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
2092 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2093 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2094 | ||
2095 | { .type = PERF_TYPE_HW_CACHE, | |
2096 | .config = | |
2097 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
2098 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2099 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2100 | ||
2101 | }; | |
2102 | ||
2103 | /* | |
2104 | * Very, very detailed stats (-d -d -d), adding prefetch events: | |
2105 | */ | |
2106 | struct perf_event_attr very_very_detailed_attrs[] = { | |
2107 | ||
2108 | { .type = PERF_TYPE_HW_CACHE, | |
2109 | .config = | |
2110 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
2111 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
2112 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2113 | ||
2114 | { .type = PERF_TYPE_HW_CACHE, | |
2115 | .config = | |
2116 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
2117 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
2118 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2119 | }; | |
2120 | ||
2121 | /* Set attrs if no event is selected and !null_run: */ | |
2122 | if (null_run) | |
2123 | return 0; | |
2124 | ||
2125 | if (transaction_run) { | |
2126 | if (pmu_have_event("cpu", "cycles-ct") && | |
2127 | pmu_have_event("cpu", "el-start")) | |
2128 | err = parse_events(evsel_list, transaction_attrs, NULL); | |
2129 | else | |
2130 | err = parse_events(evsel_list, transaction_limited_attrs, NULL); | |
2131 | if (err) { | |
2132 | fprintf(stderr, "Cannot set up transaction events\n"); | |
2133 | return -1; | |
2134 | } | |
2135 | return 0; | |
2136 | } | |
2137 | ||
2138 | if (topdown_run) { | |
2139 | char *str = NULL; | |
2140 | bool warn = false; | |
2141 | ||
2142 | if (stat_config.aggr_mode != AGGR_GLOBAL && | |
2143 | stat_config.aggr_mode != AGGR_CORE) { | |
2144 | pr_err("top down event configuration requires --per-core mode\n"); | |
2145 | return -1; | |
2146 | } | |
2147 | stat_config.aggr_mode = AGGR_CORE; | |
2148 | if (nr_cgroups || !target__has_cpu(&target)) { | |
2149 | pr_err("top down event configuration requires system-wide mode (-a)\n"); | |
2150 | return -1; | |
2151 | } | |
2152 | ||
2153 | if (!force_metric_only) | |
2154 | metric_only = true; | |
2155 | if (topdown_filter_events(topdown_attrs, &str, | |
2156 | arch_topdown_check_group(&warn)) < 0) { | |
2157 | pr_err("Out of memory\n"); | |
2158 | return -1; | |
2159 | } | |
2160 | if (topdown_attrs[0] && str) { | |
2161 | if (warn) | |
2162 | arch_topdown_group_warn(); | |
2163 | err = parse_events(evsel_list, str, NULL); | |
2164 | if (err) { | |
2165 | fprintf(stderr, | |
2166 | "Cannot set up top down events %s: %d\n", | |
2167 | str, err); | |
2168 | free(str); | |
2169 | return -1; | |
2170 | } | |
2171 | } else { | |
2172 | fprintf(stderr, "System does not support topdown\n"); | |
2173 | return -1; | |
2174 | } | |
2175 | free(str); | |
2176 | } | |
2177 | ||
2178 | if (!evsel_list->nr_entries) { | |
2179 | if (target__has_cpu(&target)) | |
2180 | default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK; | |
2181 | ||
2182 | if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0) | |
2183 | return -1; | |
2184 | if (pmu_have_event("cpu", "stalled-cycles-frontend")) { | |
2185 | if (perf_evlist__add_default_attrs(evsel_list, | |
2186 | frontend_attrs) < 0) | |
2187 | return -1; | |
2188 | } | |
2189 | if (pmu_have_event("cpu", "stalled-cycles-backend")) { | |
2190 | if (perf_evlist__add_default_attrs(evsel_list, | |
2191 | backend_attrs) < 0) | |
2192 | return -1; | |
2193 | } | |
2194 | if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0) | |
2195 | return -1; | |
2196 | } | |
2197 | ||
2198 | /* Detailed events get appended to the event list: */ | |
2199 | ||
2200 | if (detailed_run < 1) | |
2201 | return 0; | |
2202 | ||
2203 | /* Append detailed run extra attributes: */ | |
2204 | if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) | |
2205 | return -1; | |
2206 | ||
2207 | if (detailed_run < 2) | |
2208 | return 0; | |
2209 | ||
2210 | /* Append very detailed run extra attributes: */ | |
2211 | if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) | |
2212 | return -1; | |
2213 | ||
2214 | if (detailed_run < 3) | |
2215 | return 0; | |
2216 | ||
2217 | /* Append very, very detailed run extra attributes: */ | |
2218 | return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); | |
2219 | } | |
2220 | ||
2221 | static const char * const stat_record_usage[] = { | |
2222 | "perf stat record [<options>]", | |
2223 | NULL, | |
2224 | }; | |
2225 | ||
2226 | static void init_features(struct perf_session *session) | |
2227 | { | |
2228 | int feat; | |
2229 | ||
2230 | for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++) | |
2231 | perf_header__set_feat(&session->header, feat); | |
2232 | ||
2233 | perf_header__clear_feat(&session->header, HEADER_BUILD_ID); | |
2234 | perf_header__clear_feat(&session->header, HEADER_TRACING_DATA); | |
2235 | perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK); | |
2236 | perf_header__clear_feat(&session->header, HEADER_AUXTRACE); | |
2237 | } | |
2238 | ||
2239 | static int __cmd_record(int argc, const char **argv) | |
2240 | { | |
2241 | struct perf_session *session; | |
2242 | struct perf_data_file *file = &perf_stat.file; | |
2243 | ||
2244 | argc = parse_options(argc, argv, stat_options, stat_record_usage, | |
2245 | PARSE_OPT_STOP_AT_NON_OPTION); | |
2246 | ||
2247 | if (output_name) | |
2248 | file->path = output_name; | |
2249 | ||
2250 | if (run_count != 1 || forever) { | |
2251 | pr_err("Cannot use -r option with perf stat record.\n"); | |
2252 | return -1; | |
2253 | } | |
2254 | ||
2255 | session = perf_session__new(file, false, NULL); | |
2256 | if (session == NULL) { | |
2257 | pr_err("Perf session creation failed.\n"); | |
2258 | return -1; | |
2259 | } | |
2260 | ||
2261 | init_features(session); | |
2262 | ||
2263 | session->evlist = evsel_list; | |
2264 | perf_stat.session = session; | |
2265 | perf_stat.record = true; | |
2266 | return argc; | |
2267 | } | |
2268 | ||
2269 | static int process_stat_round_event(struct perf_tool *tool __maybe_unused, | |
2270 | union perf_event *event, | |
2271 | struct perf_session *session) | |
2272 | { | |
2273 | struct stat_round_event *stat_round = &event->stat_round; | |
2274 | struct perf_evsel *counter; | |
2275 | struct timespec tsh, *ts = NULL; | |
2276 | const char **argv = session->header.env.cmdline_argv; | |
2277 | int argc = session->header.env.nr_cmdline; | |
2278 | ||
2279 | evlist__for_each_entry(evsel_list, counter) | |
2280 | perf_stat_process_counter(&stat_config, counter); | |
2281 | ||
2282 | if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL) | |
2283 | update_stats(&walltime_nsecs_stats, stat_round->time); | |
2284 | ||
2285 | if (stat_config.interval && stat_round->time) { | |
2286 | tsh.tv_sec = stat_round->time / NSEC_PER_SEC; | |
2287 | tsh.tv_nsec = stat_round->time % NSEC_PER_SEC; | |
2288 | ts = &tsh; | |
2289 | } | |
2290 | ||
2291 | print_counters(ts, argc, argv); | |
2292 | return 0; | |
2293 | } | |
2294 | ||
2295 | static | |
2296 | int process_stat_config_event(struct perf_tool *tool, | |
2297 | union perf_event *event, | |
2298 | struct perf_session *session __maybe_unused) | |
2299 | { | |
2300 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); | |
2301 | ||
2302 | perf_event__read_stat_config(&stat_config, &event->stat_config); | |
2303 | ||
2304 | if (cpu_map__empty(st->cpus)) { | |
2305 | if (st->aggr_mode != AGGR_UNSET) | |
2306 | pr_warning("warning: processing task data, aggregation mode not set\n"); | |
2307 | return 0; | |
2308 | } | |
2309 | ||
2310 | if (st->aggr_mode != AGGR_UNSET) | |
2311 | stat_config.aggr_mode = st->aggr_mode; | |
2312 | ||
2313 | if (perf_stat.file.is_pipe) | |
2314 | perf_stat_init_aggr_mode(); | |
2315 | else | |
2316 | perf_stat_init_aggr_mode_file(st); | |
2317 | ||
2318 | return 0; | |
2319 | } | |
2320 | ||
2321 | static int set_maps(struct perf_stat *st) | |
2322 | { | |
2323 | if (!st->cpus || !st->threads) | |
2324 | return 0; | |
2325 | ||
2326 | if (WARN_ONCE(st->maps_allocated, "stats double allocation\n")) | |
2327 | return -EINVAL; | |
2328 | ||
2329 | perf_evlist__set_maps(evsel_list, st->cpus, st->threads); | |
2330 | ||
2331 | if (perf_evlist__alloc_stats(evsel_list, true)) | |
2332 | return -ENOMEM; | |
2333 | ||
2334 | st->maps_allocated = true; | |
2335 | return 0; | |
2336 | } | |
2337 | ||
2338 | static | |
2339 | int process_thread_map_event(struct perf_tool *tool, | |
2340 | union perf_event *event, | |
2341 | struct perf_session *session __maybe_unused) | |
2342 | { | |
2343 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); | |
2344 | ||
2345 | if (st->threads) { | |
2346 | pr_warning("Extra thread map event, ignoring.\n"); | |
2347 | return 0; | |
2348 | } | |
2349 | ||
2350 | st->threads = thread_map__new_event(&event->thread_map); | |
2351 | if (!st->threads) | |
2352 | return -ENOMEM; | |
2353 | ||
2354 | return set_maps(st); | |
2355 | } | |
2356 | ||
2357 | static | |
2358 | int process_cpu_map_event(struct perf_tool *tool, | |
2359 | union perf_event *event, | |
2360 | struct perf_session *session __maybe_unused) | |
2361 | { | |
2362 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); | |
2363 | struct cpu_map *cpus; | |
2364 | ||
2365 | if (st->cpus) { | |
2366 | pr_warning("Extra cpu map event, ignoring.\n"); | |
2367 | return 0; | |
2368 | } | |
2369 | ||
2370 | cpus = cpu_map__new_data(&event->cpu_map.data); | |
2371 | if (!cpus) | |
2372 | return -ENOMEM; | |
2373 | ||
2374 | st->cpus = cpus; | |
2375 | return set_maps(st); | |
2376 | } | |
2377 | ||
2378 | static const char * const stat_report_usage[] = { | |
2379 | "perf stat report [<options>]", | |
2380 | NULL, | |
2381 | }; | |
2382 | ||
2383 | static struct perf_stat perf_stat = { | |
2384 | .tool = { | |
2385 | .attr = perf_event__process_attr, | |
2386 | .event_update = perf_event__process_event_update, | |
2387 | .thread_map = process_thread_map_event, | |
2388 | .cpu_map = process_cpu_map_event, | |
2389 | .stat_config = process_stat_config_event, | |
2390 | .stat = perf_event__process_stat_event, | |
2391 | .stat_round = process_stat_round_event, | |
2392 | }, | |
2393 | .aggr_mode = AGGR_UNSET, | |
2394 | }; | |
2395 | ||
2396 | static int __cmd_report(int argc, const char **argv) | |
2397 | { | |
2398 | struct perf_session *session; | |
2399 | const struct option options[] = { | |
2400 | OPT_STRING('i', "input", &input_name, "file", "input file name"), | |
2401 | OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode, | |
2402 | "aggregate counts per processor socket", AGGR_SOCKET), | |
2403 | OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode, | |
2404 | "aggregate counts per physical processor core", AGGR_CORE), | |
2405 | OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode, | |
2406 | "disable CPU count aggregation", AGGR_NONE), | |
2407 | OPT_END() | |
2408 | }; | |
2409 | struct stat st; | |
2410 | int ret; | |
2411 | ||
2412 | argc = parse_options(argc, argv, options, stat_report_usage, 0); | |
2413 | ||
2414 | if (!input_name || !strlen(input_name)) { | |
2415 | if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) | |
2416 | input_name = "-"; | |
2417 | else | |
2418 | input_name = "perf.data"; | |
2419 | } | |
2420 | ||
2421 | perf_stat.file.path = input_name; | |
2422 | perf_stat.file.mode = PERF_DATA_MODE_READ; | |
2423 | ||
2424 | session = perf_session__new(&perf_stat.file, false, &perf_stat.tool); | |
2425 | if (session == NULL) | |
2426 | return -1; | |
2427 | ||
2428 | perf_stat.session = session; | |
2429 | stat_config.output = stderr; | |
2430 | evsel_list = session->evlist; | |
2431 | ||
2432 | ret = perf_session__process_events(session); | |
2433 | if (ret) | |
2434 | return ret; | |
2435 | ||
2436 | perf_session__delete(session); | |
2437 | return 0; | |
2438 | } | |
2439 | ||
2440 | static void setup_system_wide(int forks) | |
2441 | { | |
2442 | /* | |
2443 | * Make system wide (-a) the default target if | |
2444 | * no target was specified and one of following | |
2445 | * conditions is met: | |
2446 | * | |
2447 | * - there's no workload specified | |
2448 | * - there is workload specified but all requested | |
2449 | * events are system wide events | |
2450 | */ | |
2451 | if (!target__none(&target)) | |
2452 | return; | |
2453 | ||
2454 | if (!forks) | |
2455 | target.system_wide = true; | |
2456 | else { | |
2457 | struct perf_evsel *counter; | |
2458 | ||
2459 | evlist__for_each_entry(evsel_list, counter) { | |
2460 | if (!counter->system_wide) | |
2461 | return; | |
2462 | } | |
2463 | ||
2464 | if (evsel_list->nr_entries) | |
2465 | target.system_wide = true; | |
2466 | } | |
2467 | } | |
2468 | ||
2469 | int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused) | |
2470 | { | |
2471 | const char * const stat_usage[] = { | |
2472 | "perf stat [<options>] [<command>]", | |
2473 | NULL | |
2474 | }; | |
2475 | int status = -EINVAL, run_idx; | |
2476 | const char *mode; | |
2477 | FILE *output = stderr; | |
2478 | unsigned int interval; | |
2479 | const char * const stat_subcommands[] = { "record", "report" }; | |
2480 | ||
2481 | setlocale(LC_ALL, ""); | |
2482 | ||
2483 | evsel_list = perf_evlist__new(); | |
2484 | if (evsel_list == NULL) | |
2485 | return -ENOMEM; | |
2486 | ||
2487 | parse_events__shrink_config_terms(); | |
2488 | argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands, | |
2489 | (const char **) stat_usage, | |
2490 | PARSE_OPT_STOP_AT_NON_OPTION); | |
2491 | perf_stat__init_shadow_stats(); | |
2492 | ||
2493 | if (csv_sep) { | |
2494 | csv_output = true; | |
2495 | if (!strcmp(csv_sep, "\\t")) | |
2496 | csv_sep = "\t"; | |
2497 | } else | |
2498 | csv_sep = DEFAULT_SEPARATOR; | |
2499 | ||
2500 | if (argc && !strncmp(argv[0], "rec", 3)) { | |
2501 | argc = __cmd_record(argc, argv); | |
2502 | if (argc < 0) | |
2503 | return -1; | |
2504 | } else if (argc && !strncmp(argv[0], "rep", 3)) | |
2505 | return __cmd_report(argc, argv); | |
2506 | ||
2507 | interval = stat_config.interval; | |
2508 | ||
2509 | /* | |
2510 | * For record command the -o is already taken care of. | |
2511 | */ | |
2512 | if (!STAT_RECORD && output_name && strcmp(output_name, "-")) | |
2513 | output = NULL; | |
2514 | ||
2515 | if (output_name && output_fd) { | |
2516 | fprintf(stderr, "cannot use both --output and --log-fd\n"); | |
2517 | parse_options_usage(stat_usage, stat_options, "o", 1); | |
2518 | parse_options_usage(NULL, stat_options, "log-fd", 0); | |
2519 | goto out; | |
2520 | } | |
2521 | ||
2522 | if (metric_only && stat_config.aggr_mode == AGGR_THREAD) { | |
2523 | fprintf(stderr, "--metric-only is not supported with --per-thread\n"); | |
2524 | goto out; | |
2525 | } | |
2526 | ||
2527 | if (metric_only && run_count > 1) { | |
2528 | fprintf(stderr, "--metric-only is not supported with -r\n"); | |
2529 | goto out; | |
2530 | } | |
2531 | ||
2532 | if (output_fd < 0) { | |
2533 | fprintf(stderr, "argument to --log-fd must be a > 0\n"); | |
2534 | parse_options_usage(stat_usage, stat_options, "log-fd", 0); | |
2535 | goto out; | |
2536 | } | |
2537 | ||
2538 | if (!output) { | |
2539 | struct timespec tm; | |
2540 | mode = append_file ? "a" : "w"; | |
2541 | ||
2542 | output = fopen(output_name, mode); | |
2543 | if (!output) { | |
2544 | perror("failed to create output file"); | |
2545 | return -1; | |
2546 | } | |
2547 | clock_gettime(CLOCK_REALTIME, &tm); | |
2548 | fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); | |
2549 | } else if (output_fd > 0) { | |
2550 | mode = append_file ? "a" : "w"; | |
2551 | output = fdopen(output_fd, mode); | |
2552 | if (!output) { | |
2553 | perror("Failed opening logfd"); | |
2554 | return -errno; | |
2555 | } | |
2556 | } | |
2557 | ||
2558 | stat_config.output = output; | |
2559 | ||
2560 | /* | |
2561 | * let the spreadsheet do the pretty-printing | |
2562 | */ | |
2563 | if (csv_output) { | |
2564 | /* User explicitly passed -B? */ | |
2565 | if (big_num_opt == 1) { | |
2566 | fprintf(stderr, "-B option not supported with -x\n"); | |
2567 | parse_options_usage(stat_usage, stat_options, "B", 1); | |
2568 | parse_options_usage(NULL, stat_options, "x", 1); | |
2569 | goto out; | |
2570 | } else /* Nope, so disable big number formatting */ | |
2571 | big_num = false; | |
2572 | } else if (big_num_opt == 0) /* User passed --no-big-num */ | |
2573 | big_num = false; | |
2574 | ||
2575 | setup_system_wide(argc); | |
2576 | ||
2577 | if (run_count < 0) { | |
2578 | pr_err("Run count must be a positive number\n"); | |
2579 | parse_options_usage(stat_usage, stat_options, "r", 1); | |
2580 | goto out; | |
2581 | } else if (run_count == 0) { | |
2582 | forever = true; | |
2583 | run_count = 1; | |
2584 | } | |
2585 | ||
2586 | if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) { | |
2587 | fprintf(stderr, "The --per-thread option is only available " | |
2588 | "when monitoring via -p -t options.\n"); | |
2589 | parse_options_usage(NULL, stat_options, "p", 1); | |
2590 | parse_options_usage(NULL, stat_options, "t", 1); | |
2591 | goto out; | |
2592 | } | |
2593 | ||
2594 | /* | |
2595 | * no_aggr, cgroup are for system-wide only | |
2596 | * --per-thread is aggregated per thread, we dont mix it with cpu mode | |
2597 | */ | |
2598 | if (((stat_config.aggr_mode != AGGR_GLOBAL && | |
2599 | stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) && | |
2600 | !target__has_cpu(&target)) { | |
2601 | fprintf(stderr, "both cgroup and no-aggregation " | |
2602 | "modes only available in system-wide mode\n"); | |
2603 | ||
2604 | parse_options_usage(stat_usage, stat_options, "G", 1); | |
2605 | parse_options_usage(NULL, stat_options, "A", 1); | |
2606 | parse_options_usage(NULL, stat_options, "a", 1); | |
2607 | goto out; | |
2608 | } | |
2609 | ||
2610 | if (add_default_attributes()) | |
2611 | goto out; | |
2612 | ||
2613 | target__validate(&target); | |
2614 | ||
2615 | if (perf_evlist__create_maps(evsel_list, &target) < 0) { | |
2616 | if (target__has_task(&target)) { | |
2617 | pr_err("Problems finding threads of monitor\n"); | |
2618 | parse_options_usage(stat_usage, stat_options, "p", 1); | |
2619 | parse_options_usage(NULL, stat_options, "t", 1); | |
2620 | } else if (target__has_cpu(&target)) { | |
2621 | perror("failed to parse CPUs map"); | |
2622 | parse_options_usage(stat_usage, stat_options, "C", 1); | |
2623 | parse_options_usage(NULL, stat_options, "a", 1); | |
2624 | } | |
2625 | goto out; | |
2626 | } | |
2627 | ||
2628 | /* | |
2629 | * Initialize thread_map with comm names, | |
2630 | * so we could print it out on output. | |
2631 | */ | |
2632 | if (stat_config.aggr_mode == AGGR_THREAD) | |
2633 | thread_map__read_comms(evsel_list->threads); | |
2634 | ||
2635 | if (interval && interval < 100) { | |
2636 | if (interval < 10) { | |
2637 | pr_err("print interval must be >= 10ms\n"); | |
2638 | parse_options_usage(stat_usage, stat_options, "I", 1); | |
2639 | goto out; | |
2640 | } else | |
2641 | pr_warning("print interval < 100ms. " | |
2642 | "The overhead percentage could be high in some cases. " | |
2643 | "Please proceed with caution.\n"); | |
2644 | } | |
2645 | ||
2646 | if (perf_evlist__alloc_stats(evsel_list, interval)) | |
2647 | goto out; | |
2648 | ||
2649 | if (perf_stat_init_aggr_mode()) | |
2650 | goto out; | |
2651 | ||
2652 | /* | |
2653 | * We dont want to block the signals - that would cause | |
2654 | * child tasks to inherit that and Ctrl-C would not work. | |
2655 | * What we want is for Ctrl-C to work in the exec()-ed | |
2656 | * task, but being ignored by perf stat itself: | |
2657 | */ | |
2658 | atexit(sig_atexit); | |
2659 | if (!forever) | |
2660 | signal(SIGINT, skip_signal); | |
2661 | signal(SIGCHLD, skip_signal); | |
2662 | signal(SIGALRM, skip_signal); | |
2663 | signal(SIGABRT, skip_signal); | |
2664 | ||
2665 | status = 0; | |
2666 | for (run_idx = 0; forever || run_idx < run_count; run_idx++) { | |
2667 | if (run_count != 1 && verbose > 0) | |
2668 | fprintf(output, "[ perf stat: executing run #%d ... ]\n", | |
2669 | run_idx + 1); | |
2670 | ||
2671 | status = run_perf_stat(argc, argv); | |
2672 | if (forever && status != -1) { | |
2673 | print_counters(NULL, argc, argv); | |
2674 | perf_stat__reset_stats(); | |
2675 | } | |
2676 | } | |
2677 | ||
2678 | if (!forever && status != -1 && !interval) | |
2679 | print_counters(NULL, argc, argv); | |
2680 | ||
2681 | if (STAT_RECORD) { | |
2682 | /* | |
2683 | * We synthesize the kernel mmap record just so that older tools | |
2684 | * don't emit warnings about not being able to resolve symbols | |
2685 | * due to /proc/sys/kernel/kptr_restrict settings and instear provide | |
2686 | * a saner message about no samples being in the perf.data file. | |
2687 | * | |
2688 | * This also serves to suppress a warning about f_header.data.size == 0 | |
2689 | * in header.c at the moment 'perf stat record' gets introduced, which | |
2690 | * is not really needed once we start adding the stat specific PERF_RECORD_ | |
2691 | * records, but the need to suppress the kptr_restrict messages in older | |
2692 | * tools remain -acme | |
2693 | */ | |
2694 | int fd = perf_data_file__fd(&perf_stat.file); | |
2695 | int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat, | |
2696 | process_synthesized_event, | |
2697 | &perf_stat.session->machines.host); | |
2698 | if (err) { | |
2699 | pr_warning("Couldn't synthesize the kernel mmap record, harmless, " | |
2700 | "older tools may produce warnings about this file\n."); | |
2701 | } | |
2702 | ||
2703 | if (!interval) { | |
2704 | if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL)) | |
2705 | pr_err("failed to write stat round event\n"); | |
2706 | } | |
2707 | ||
2708 | if (!perf_stat.file.is_pipe) { | |
2709 | perf_stat.session->header.data_size += perf_stat.bytes_written; | |
2710 | perf_session__write_header(perf_stat.session, evsel_list, fd, true); | |
2711 | } | |
2712 | ||
2713 | perf_session__delete(perf_stat.session); | |
2714 | } | |
2715 | ||
2716 | perf_stat__exit_aggr_mode(); | |
2717 | perf_evlist__free_stats(evsel_list); | |
2718 | out: | |
2719 | perf_evlist__delete(evsel_list); | |
2720 | return status; | |
2721 | } |