Commit | Line | Data |
---|---|---|
ddcacfa0 | 1 | /* |
bf9e1876 IM |
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: | |
ddcacfa0 | 8 | |
2cba3ffb | 9 | $ perf stat ./hackbench 10 |
ddcacfa0 | 10 | |
2cba3ffb | 11 | Time: 0.118 |
ddcacfa0 | 12 | |
2cba3ffb | 13 | Performance counter stats for './hackbench 10': |
ddcacfa0 | 14 | |
2cba3ffb IM |
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 | |
ddcacfa0 | 28 | |
5242519b | 29 | * |
2cba3ffb | 30 | * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com> |
5242519b IM |
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> | |
6e750a8f | 39 | * Jaswinder Singh Rajput <jaswinder@kernel.org> |
5242519b IM |
40 | * |
41 | * Released under the GPL v2. (and only v2, not any later version) | |
ddcacfa0 IM |
42 | */ |
43 | ||
1a482f38 | 44 | #include "perf.h" |
16f762a2 | 45 | #include "builtin.h" |
f14d5707 | 46 | #include "util/cgroup.h" |
148be2c1 | 47 | #include "util/util.h" |
4b6ab94e | 48 | #include <subcmd/parse-options.h> |
5242519b | 49 | #include "util/parse-events.h" |
4cabc3d1 | 50 | #include "util/pmu.h" |
8f28827a | 51 | #include "util/event.h" |
361c99a6 | 52 | #include "util/evlist.h" |
69aad6f1 | 53 | #include "util/evsel.h" |
8f28827a | 54 | #include "util/debug.h" |
a5d243d0 | 55 | #include "util/color.h" |
0007ecea | 56 | #include "util/stat.h" |
60666c63 | 57 | #include "util/header.h" |
a12b51c4 | 58 | #include "util/cpumap.h" |
d6d901c2 | 59 | #include "util/thread.h" |
fd78260b | 60 | #include "util/thread_map.h" |
d809560b | 61 | #include "util/counts.h" |
44b1e60a | 62 | #include "util/group.h" |
4979d0c7 | 63 | #include "util/session.h" |
ba6039b6 | 64 | #include "util/tool.h" |
44b1e60a | 65 | #include "util/group.h" |
ba6039b6 | 66 | #include "asm/bug.h" |
ddcacfa0 | 67 | |
44b1e60a | 68 | #include <api/fs/fs.h> |
1f16c575 | 69 | #include <stdlib.h> |
ddcacfa0 | 70 | #include <sys/prctl.h> |
5af52b51 | 71 | #include <locale.h> |
e3b03b6c | 72 | #include <math.h> |
16c8a109 | 73 | |
d7470b6a | 74 | #define DEFAULT_SEPARATOR " " |
2cee77c4 DA |
75 | #define CNTR_NOT_SUPPORTED "<not supported>" |
76 | #define CNTR_NOT_COUNTED "<not counted>" | |
d7470b6a | 77 | |
d4f63a47 | 78 | static void print_counters(struct timespec *ts, int argc, const char **argv); |
13370a9b | 79 | |
4cabc3d1 | 80 | /* Default events used for perf stat -T */ |
a454742c JO |
81 | static const char *transaction_attrs = { |
82 | "task-clock," | |
4cabc3d1 AK |
83 | "{" |
84 | "instructions," | |
85 | "cycles," | |
86 | "cpu/cycles-t/," | |
87 | "cpu/tx-start/," | |
88 | "cpu/el-start/," | |
89 | "cpu/cycles-ct/" | |
90 | "}" | |
91 | }; | |
92 | ||
93 | /* More limited version when the CPU does not have all events. */ | |
a454742c JO |
94 | static const char * transaction_limited_attrs = { |
95 | "task-clock," | |
4cabc3d1 AK |
96 | "{" |
97 | "instructions," | |
98 | "cycles," | |
99 | "cpu/cycles-t/," | |
100 | "cpu/tx-start/" | |
101 | "}" | |
102 | }; | |
103 | ||
44b1e60a AK |
104 | static const char * topdown_attrs[] = { |
105 | "topdown-total-slots", | |
106 | "topdown-slots-retired", | |
107 | "topdown-recovery-bubbles", | |
108 | "topdown-fetch-bubbles", | |
109 | "topdown-slots-issued", | |
110 | NULL, | |
111 | }; | |
112 | ||
666e6d48 | 113 | static struct perf_evlist *evsel_list; |
361c99a6 | 114 | |
602ad878 | 115 | static struct target target = { |
77a6f014 NK |
116 | .uid = UINT_MAX, |
117 | }; | |
ddcacfa0 | 118 | |
1e5a2931 JO |
119 | typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu); |
120 | ||
3d632595 | 121 | static int run_count = 1; |
2e6cdf99 | 122 | static bool no_inherit = false; |
d07f0b12 | 123 | static volatile pid_t child_pid = -1; |
c0555642 | 124 | static bool null_run = false; |
2cba3ffb | 125 | static int detailed_run = 0; |
4cabc3d1 | 126 | static bool transaction_run; |
44b1e60a | 127 | static bool topdown_run = false; |
201e0b06 | 128 | static bool big_num = true; |
d7470b6a | 129 | static int big_num_opt = -1; |
d7470b6a SE |
130 | static const char *csv_sep = NULL; |
131 | static bool csv_output = false; | |
43bece79 | 132 | static bool group = false; |
1f16c575 PZ |
133 | static const char *pre_cmd = NULL; |
134 | static const char *post_cmd = NULL; | |
135 | static bool sync_run = false; | |
41191688 | 136 | static unsigned int initial_delay = 0; |
410136f5 | 137 | static unsigned int unit_width = 4; /* strlen("unit") */ |
a7e191c3 | 138 | static bool forever = false; |
54b50916 | 139 | static bool metric_only = false; |
44b1e60a | 140 | static bool force_metric_only = false; |
13370a9b | 141 | static struct timespec ref_time; |
86ee6e18 | 142 | static struct cpu_map *aggr_map; |
1e5a2931 | 143 | static aggr_get_id_t aggr_get_id; |
e0547311 JO |
144 | static bool append_file; |
145 | static const char *output_name; | |
146 | static int output_fd; | |
5af52b51 | 147 | |
4979d0c7 JO |
148 | struct perf_stat { |
149 | bool record; | |
150 | struct perf_data_file file; | |
151 | struct perf_session *session; | |
152 | u64 bytes_written; | |
ba6039b6 | 153 | struct perf_tool tool; |
1975d36e JO |
154 | bool maps_allocated; |
155 | struct cpu_map *cpus; | |
156 | struct thread_map *threads; | |
89af4e05 | 157 | enum aggr_mode aggr_mode; |
4979d0c7 JO |
158 | }; |
159 | ||
160 | static struct perf_stat perf_stat; | |
161 | #define STAT_RECORD perf_stat.record | |
162 | ||
60666c63 LW |
163 | static volatile int done = 0; |
164 | ||
421a50f3 JO |
165 | static struct perf_stat_config stat_config = { |
166 | .aggr_mode = AGGR_GLOBAL, | |
711a572e | 167 | .scale = true, |
421a50f3 JO |
168 | }; |
169 | ||
13370a9b SE |
170 | static inline void diff_timespec(struct timespec *r, struct timespec *a, |
171 | struct timespec *b) | |
172 | { | |
173 | r->tv_sec = a->tv_sec - b->tv_sec; | |
174 | if (a->tv_nsec < b->tv_nsec) { | |
175 | r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec; | |
176 | r->tv_sec--; | |
177 | } else { | |
178 | r->tv_nsec = a->tv_nsec - b->tv_nsec ; | |
179 | } | |
180 | } | |
181 | ||
254ecbc7 JO |
182 | static void perf_stat__reset_stats(void) |
183 | { | |
184 | perf_evlist__reset_stats(evsel_list); | |
f87027b9 | 185 | perf_stat__reset_shadow_stats(); |
1eda3b21 JO |
186 | } |
187 | ||
cac21425 | 188 | static int create_perf_stat_counter(struct perf_evsel *evsel) |
ddcacfa0 | 189 | { |
69aad6f1 | 190 | struct perf_event_attr *attr = &evsel->attr; |
727ab04e | 191 | |
711a572e | 192 | if (stat_config.scale) |
a21ca2ca IM |
193 | attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | |
194 | PERF_FORMAT_TOTAL_TIME_RUNNING; | |
ddcacfa0 | 195 | |
5d2cd909 ACM |
196 | attr->inherit = !no_inherit; |
197 | ||
6acd8e92 JO |
198 | /* |
199 | * Some events get initialized with sample_(period/type) set, | |
200 | * like tracepoints. Clear it up for counting. | |
201 | */ | |
202 | attr->sample_period = 0; | |
6db1a5c1 | 203 | |
4979d0c7 JO |
204 | /* |
205 | * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless | |
206 | * while avoiding that older tools show confusing messages. | |
6db1a5c1 JO |
207 | * |
208 | * However for pipe sessions we need to keep it zero, | |
209 | * because script's perf_evsel__check_attr is triggered | |
210 | * by attr->sample_type != 0, and we can't run it on | |
211 | * stat sessions. | |
4979d0c7 | 212 | */ |
6db1a5c1 JO |
213 | if (!(STAT_RECORD && perf_stat.file.is_pipe)) |
214 | attr->sample_type = PERF_SAMPLE_IDENTIFIER; | |
6acd8e92 | 215 | |
67ccdecd JO |
216 | /* |
217 | * Disabling all counters initially, they will be enabled | |
218 | * either manually by us or by kernel via enable_on_exec | |
219 | * set later. | |
220 | */ | |
c8280cec | 221 | if (perf_evsel__is_group_leader(evsel)) { |
67ccdecd JO |
222 | attr->disabled = 1; |
223 | ||
c8280cec JO |
224 | /* |
225 | * In case of initial_delay we enable tracee | |
226 | * events manually. | |
227 | */ | |
228 | if (target__none(&target) && !initial_delay) | |
41191688 | 229 | attr->enable_on_exec = 1; |
ddcacfa0 | 230 | } |
084ab9f8 | 231 | |
c8280cec JO |
232 | if (target__has_cpu(&target)) |
233 | return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel)); | |
234 | ||
594ac61a | 235 | return perf_evsel__open_per_thread(evsel, evsel_list->threads); |
ddcacfa0 IM |
236 | } |
237 | ||
c04f5e5d IM |
238 | /* |
239 | * Does the counter have nsecs as a unit? | |
240 | */ | |
daec78a0 | 241 | static inline int nsec_counter(struct perf_evsel *evsel) |
c04f5e5d | 242 | { |
daec78a0 ACM |
243 | if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || |
244 | perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) | |
c04f5e5d IM |
245 | return 1; |
246 | ||
247 | return 0; | |
248 | } | |
249 | ||
8b99b1a4 JO |
250 | static int process_synthesized_event(struct perf_tool *tool __maybe_unused, |
251 | union perf_event *event, | |
252 | struct perf_sample *sample __maybe_unused, | |
253 | struct machine *machine __maybe_unused) | |
4979d0c7 | 254 | { |
8b99b1a4 | 255 | if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) { |
4979d0c7 JO |
256 | pr_err("failed to write perf data, error: %m\n"); |
257 | return -1; | |
258 | } | |
259 | ||
8b99b1a4 | 260 | perf_stat.bytes_written += event->header.size; |
4979d0c7 JO |
261 | return 0; |
262 | } | |
263 | ||
1975d36e | 264 | static int write_stat_round_event(u64 tm, u64 type) |
7aad0c32 | 265 | { |
1975d36e | 266 | return perf_event__synthesize_stat_round(NULL, tm, type, |
7aad0c32 JO |
267 | process_synthesized_event, |
268 | NULL); | |
269 | } | |
270 | ||
271 | #define WRITE_STAT_ROUND_EVENT(time, interval) \ | |
272 | write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval) | |
273 | ||
5a6ea81b JO |
274 | #define SID(e, x, y) xyarray__entry(e->sample_id, x, y) |
275 | ||
276 | static int | |
277 | perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread, | |
278 | struct perf_counts_values *count) | |
279 | { | |
280 | struct perf_sample_id *sid = SID(counter, cpu, thread); | |
281 | ||
282 | return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count, | |
283 | process_synthesized_event, NULL); | |
284 | } | |
285 | ||
f5b4a9c3 SE |
286 | /* |
287 | * Read out the results of a single counter: | |
288 | * do not aggregate counts across CPUs in system-wide mode | |
289 | */ | |
c52b12ed | 290 | static int read_counter(struct perf_evsel *counter) |
f5b4a9c3 | 291 | { |
9bf1a529 | 292 | int nthreads = thread_map__nr(evsel_list->threads); |
00e727bb MR |
293 | int ncpus, cpu, thread; |
294 | ||
295 | if (target__has_cpu(&target)) | |
296 | ncpus = perf_evsel__nr_cpus(counter); | |
297 | else | |
298 | ncpus = 1; | |
f5b4a9c3 | 299 | |
3b4331d9 SP |
300 | if (!counter->supported) |
301 | return -ENOENT; | |
302 | ||
9bf1a529 JO |
303 | if (counter->system_wide) |
304 | nthreads = 1; | |
305 | ||
306 | for (thread = 0; thread < nthreads; thread++) { | |
307 | for (cpu = 0; cpu < ncpus; cpu++) { | |
3b3eb044 JO |
308 | struct perf_counts_values *count; |
309 | ||
310 | count = perf_counts(counter->counts, cpu, thread); | |
311 | if (perf_evsel__read(counter, cpu, thread, count)) | |
9bf1a529 | 312 | return -1; |
5a6ea81b JO |
313 | |
314 | if (STAT_RECORD) { | |
315 | if (perf_evsel__write_stat_event(counter, cpu, thread, count)) { | |
316 | pr_err("failed to write stat event\n"); | |
317 | return -1; | |
318 | } | |
319 | } | |
0b1abbf4 AK |
320 | |
321 | if (verbose > 1) { | |
322 | fprintf(stat_config.output, | |
323 | "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", | |
324 | perf_evsel__name(counter), | |
325 | cpu, | |
326 | count->val, count->ena, count->run); | |
327 | } | |
9bf1a529 | 328 | } |
f5b4a9c3 | 329 | } |
c52b12ed ACM |
330 | |
331 | return 0; | |
2996f5dd IM |
332 | } |
333 | ||
3df33eff | 334 | static void read_counters(void) |
13370a9b | 335 | { |
13370a9b | 336 | struct perf_evsel *counter; |
13370a9b | 337 | |
e5cadb93 | 338 | evlist__for_each_entry(evsel_list, counter) { |
3b3eb044 | 339 | if (read_counter(counter)) |
245bad8e | 340 | pr_debug("failed to read counter %s\n", counter->name); |
3b3eb044 | 341 | |
f80010eb | 342 | if (perf_stat_process_counter(&stat_config, counter)) |
3b3eb044 | 343 | pr_warning("failed to process counter %s\n", counter->name); |
13370a9b | 344 | } |
106a94a0 JO |
345 | } |
346 | ||
ba411a95 | 347 | static void process_interval(void) |
106a94a0 | 348 | { |
106a94a0 | 349 | struct timespec ts, rs; |
106a94a0 | 350 | |
3df33eff | 351 | read_counters(); |
86ee6e18 | 352 | |
13370a9b SE |
353 | clock_gettime(CLOCK_MONOTONIC, &ts); |
354 | diff_timespec(&rs, &ts, &ref_time); | |
13370a9b | 355 | |
7aad0c32 JO |
356 | if (STAT_RECORD) { |
357 | if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSECS_PER_SEC + rs.tv_nsec, INTERVAL)) | |
358 | pr_err("failed to write stat round event\n"); | |
359 | } | |
360 | ||
d4f63a47 | 361 | print_counters(&rs, 0, NULL); |
13370a9b SE |
362 | } |
363 | ||
67ccdecd | 364 | static void enable_counters(void) |
41191688 | 365 | { |
67ccdecd | 366 | if (initial_delay) |
41191688 | 367 | usleep(initial_delay * 1000); |
67ccdecd JO |
368 | |
369 | /* | |
370 | * We need to enable counters only if: | |
371 | * - we don't have tracee (attaching to task or cpu) | |
372 | * - we have initial delay configured | |
373 | */ | |
374 | if (!target__none(&target) || initial_delay) | |
ab46db0a | 375 | perf_evlist__enable(evsel_list); |
41191688 AK |
376 | } |
377 | ||
3df33eff MR |
378 | static void disable_counters(void) |
379 | { | |
380 | /* | |
381 | * If we don't have tracee (attaching to task or cpu), counters may | |
382 | * still be running. To get accurate group ratios, we must stop groups | |
383 | * from counting before reading their constituent counters. | |
384 | */ | |
385 | if (!target__none(&target)) | |
386 | perf_evlist__disable(evsel_list); | |
387 | } | |
388 | ||
f33cbe72 | 389 | static volatile int workload_exec_errno; |
6af206fd ACM |
390 | |
391 | /* | |
392 | * perf_evlist__prepare_workload will send a SIGUSR1 | |
393 | * if the fork fails, since we asked by setting its | |
394 | * want_signal to true. | |
395 | */ | |
f33cbe72 ACM |
396 | static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, |
397 | void *ucontext __maybe_unused) | |
6af206fd | 398 | { |
f33cbe72 | 399 | workload_exec_errno = info->si_value.sival_int; |
6af206fd ACM |
400 | } |
401 | ||
7b60a7e3 JO |
402 | static bool has_unit(struct perf_evsel *counter) |
403 | { | |
404 | return counter->unit && *counter->unit; | |
405 | } | |
406 | ||
407 | static bool has_scale(struct perf_evsel *counter) | |
408 | { | |
409 | return counter->scale != 1; | |
410 | } | |
411 | ||
664c98d4 | 412 | static int perf_stat_synthesize_config(bool is_pipe) |
8b99b1a4 | 413 | { |
7b60a7e3 | 414 | struct perf_evsel *counter; |
8b99b1a4 JO |
415 | int err; |
416 | ||
664c98d4 JO |
417 | if (is_pipe) { |
418 | err = perf_event__synthesize_attrs(NULL, perf_stat.session, | |
419 | process_synthesized_event); | |
420 | if (err < 0) { | |
421 | pr_err("Couldn't synthesize attrs.\n"); | |
422 | return err; | |
423 | } | |
424 | } | |
425 | ||
7b60a7e3 JO |
426 | /* |
427 | * Synthesize other events stuff not carried within | |
428 | * attr event - unit, scale, name | |
429 | */ | |
e5cadb93 | 430 | evlist__for_each_entry(evsel_list, counter) { |
7b60a7e3 JO |
431 | if (!counter->supported) |
432 | continue; | |
433 | ||
434 | /* | |
435 | * Synthesize unit and scale only if it's defined. | |
436 | */ | |
437 | if (has_unit(counter)) { | |
438 | err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event); | |
439 | if (err < 0) { | |
440 | pr_err("Couldn't synthesize evsel unit.\n"); | |
441 | return err; | |
442 | } | |
443 | } | |
444 | ||
445 | if (has_scale(counter)) { | |
446 | err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event); | |
447 | if (err < 0) { | |
448 | pr_err("Couldn't synthesize evsel scale.\n"); | |
449 | return err; | |
450 | } | |
451 | } | |
452 | ||
453 | if (counter->own_cpus) { | |
454 | err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event); | |
455 | if (err < 0) { | |
456 | pr_err("Couldn't synthesize evsel scale.\n"); | |
457 | return err; | |
458 | } | |
459 | } | |
460 | ||
461 | /* | |
462 | * Name is needed only for pipe output, | |
463 | * perf.data carries event names. | |
464 | */ | |
465 | if (is_pipe) { | |
466 | err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event); | |
467 | if (err < 0) { | |
468 | pr_err("Couldn't synthesize evsel name.\n"); | |
469 | return err; | |
470 | } | |
471 | } | |
472 | } | |
473 | ||
8b99b1a4 JO |
474 | err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads, |
475 | process_synthesized_event, | |
476 | NULL); | |
477 | if (err < 0) { | |
478 | pr_err("Couldn't synthesize thread map.\n"); | |
479 | return err; | |
480 | } | |
481 | ||
482 | err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus, | |
483 | process_synthesized_event, NULL); | |
484 | if (err < 0) { | |
485 | pr_err("Couldn't synthesize thread map.\n"); | |
486 | return err; | |
487 | } | |
488 | ||
489 | err = perf_event__synthesize_stat_config(NULL, &stat_config, | |
490 | process_synthesized_event, NULL); | |
491 | if (err < 0) { | |
492 | pr_err("Couldn't synthesize config.\n"); | |
493 | return err; | |
494 | } | |
495 | ||
496 | return 0; | |
497 | } | |
498 | ||
2af4646d JO |
499 | #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y)) |
500 | ||
501 | static int __store_counter_ids(struct perf_evsel *counter, | |
502 | struct cpu_map *cpus, | |
503 | struct thread_map *threads) | |
504 | { | |
505 | int cpu, thread; | |
506 | ||
507 | for (cpu = 0; cpu < cpus->nr; cpu++) { | |
508 | for (thread = 0; thread < threads->nr; thread++) { | |
509 | int fd = FD(counter, cpu, thread); | |
510 | ||
511 | if (perf_evlist__id_add_fd(evsel_list, counter, | |
512 | cpu, thread, fd) < 0) | |
513 | return -1; | |
514 | } | |
515 | } | |
516 | ||
517 | return 0; | |
518 | } | |
519 | ||
520 | static int store_counter_ids(struct perf_evsel *counter) | |
521 | { | |
522 | struct cpu_map *cpus = counter->cpus; | |
523 | struct thread_map *threads = counter->threads; | |
524 | ||
525 | if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr)) | |
526 | return -ENOMEM; | |
527 | ||
528 | return __store_counter_ids(counter, cpus, threads); | |
529 | } | |
530 | ||
acf28922 | 531 | static int __run_perf_stat(int argc, const char **argv) |
42202dd5 | 532 | { |
ec0d3d1f | 533 | int interval = stat_config.interval; |
56e52e85 | 534 | char msg[512]; |
42202dd5 | 535 | unsigned long long t0, t1; |
cac21425 | 536 | struct perf_evsel *counter; |
13370a9b | 537 | struct timespec ts; |
410136f5 | 538 | size_t l; |
42202dd5 | 539 | int status = 0; |
6be2850e | 540 | const bool forks = (argc > 0); |
664c98d4 | 541 | bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false; |
42202dd5 | 542 | |
13370a9b SE |
543 | if (interval) { |
544 | ts.tv_sec = interval / 1000; | |
545 | ts.tv_nsec = (interval % 1000) * 1000000; | |
546 | } else { | |
547 | ts.tv_sec = 1; | |
548 | ts.tv_nsec = 0; | |
549 | } | |
550 | ||
60666c63 | 551 | if (forks) { |
664c98d4 | 552 | if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe, |
735f7e0b | 553 | workload_exec_failed_signal) < 0) { |
acf28922 NK |
554 | perror("failed to prepare workload"); |
555 | return -1; | |
60666c63 | 556 | } |
d20a47e7 | 557 | child_pid = evsel_list->workload.pid; |
051ae7f7 PM |
558 | } |
559 | ||
6a4bb04c | 560 | if (group) |
63dab225 | 561 | perf_evlist__set_leader(evsel_list); |
6a4bb04c | 562 | |
e5cadb93 | 563 | evlist__for_each_entry(evsel_list, counter) { |
42ef8a78 | 564 | try_again: |
cac21425 | 565 | if (create_perf_stat_counter(counter) < 0) { |
979987a5 DA |
566 | /* |
567 | * PPC returns ENXIO for HW counters until 2.6.37 | |
568 | * (behavior changed with commit b0a873e). | |
569 | */ | |
38f6ae1e | 570 | if (errno == EINVAL || errno == ENOSYS || |
979987a5 DA |
571 | errno == ENOENT || errno == EOPNOTSUPP || |
572 | errno == ENXIO) { | |
c63ca0c0 DA |
573 | if (verbose) |
574 | ui__warning("%s event is not supported by the kernel.\n", | |
7289f83c | 575 | perf_evsel__name(counter)); |
2cee77c4 | 576 | counter->supported = false; |
cb5ef600 KL |
577 | |
578 | if ((counter->leader != counter) || | |
579 | !(counter->leader->nr_members > 1)) | |
580 | continue; | |
42ef8a78 ACM |
581 | } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) { |
582 | if (verbose) | |
583 | ui__warning("%s\n", msg); | |
584 | goto try_again; | |
585 | } | |
ede70290 | 586 | |
56e52e85 ACM |
587 | perf_evsel__open_strerror(counter, &target, |
588 | errno, msg, sizeof(msg)); | |
589 | ui__error("%s\n", msg); | |
590 | ||
48290609 ACM |
591 | if (child_pid != -1) |
592 | kill(child_pid, SIGTERM); | |
fceda7fe | 593 | |
48290609 ACM |
594 | return -1; |
595 | } | |
2cee77c4 | 596 | counter->supported = true; |
410136f5 SE |
597 | |
598 | l = strlen(counter->unit); | |
599 | if (l > unit_width) | |
600 | unit_width = l; | |
2af4646d JO |
601 | |
602 | if (STAT_RECORD && store_counter_ids(counter)) | |
603 | return -1; | |
084ab9f8 | 604 | } |
42202dd5 | 605 | |
23d4aad4 ACM |
606 | if (perf_evlist__apply_filters(evsel_list, &counter)) { |
607 | error("failed to set filter \"%s\" on event %s with %d (%s)\n", | |
608 | counter->filter, perf_evsel__name(counter), errno, | |
c8b5f2c9 | 609 | str_error_r(errno, msg, sizeof(msg))); |
cfd748ae FW |
610 | return -1; |
611 | } | |
612 | ||
4979d0c7 JO |
613 | if (STAT_RECORD) { |
614 | int err, fd = perf_data_file__fd(&perf_stat.file); | |
615 | ||
664c98d4 JO |
616 | if (is_pipe) { |
617 | err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file)); | |
618 | } else { | |
619 | err = perf_session__write_header(perf_stat.session, evsel_list, | |
620 | fd, false); | |
621 | } | |
622 | ||
4979d0c7 JO |
623 | if (err < 0) |
624 | return err; | |
8b99b1a4 | 625 | |
664c98d4 | 626 | err = perf_stat_synthesize_config(is_pipe); |
8b99b1a4 JO |
627 | if (err < 0) |
628 | return err; | |
4979d0c7 JO |
629 | } |
630 | ||
42202dd5 IM |
631 | /* |
632 | * Enable counters and exec the command: | |
633 | */ | |
634 | t0 = rdclock(); | |
13370a9b | 635 | clock_gettime(CLOCK_MONOTONIC, &ref_time); |
42202dd5 | 636 | |
60666c63 | 637 | if (forks) { |
acf28922 | 638 | perf_evlist__start_workload(evsel_list); |
67ccdecd | 639 | enable_counters(); |
acf28922 | 640 | |
13370a9b SE |
641 | if (interval) { |
642 | while (!waitpid(child_pid, &status, WNOHANG)) { | |
643 | nanosleep(&ts, NULL); | |
ba411a95 | 644 | process_interval(); |
13370a9b SE |
645 | } |
646 | } | |
60666c63 | 647 | wait(&status); |
6af206fd | 648 | |
f33cbe72 | 649 | if (workload_exec_errno) { |
c8b5f2c9 | 650 | const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg)); |
f33cbe72 | 651 | pr_err("Workload failed: %s\n", emsg); |
6af206fd | 652 | return -1; |
f33cbe72 | 653 | } |
6af206fd | 654 | |
33e49ea7 AK |
655 | if (WIFSIGNALED(status)) |
656 | psignal(WTERMSIG(status), argv[0]); | |
60666c63 | 657 | } else { |
67ccdecd | 658 | enable_counters(); |
13370a9b SE |
659 | while (!done) { |
660 | nanosleep(&ts, NULL); | |
661 | if (interval) | |
ba411a95 | 662 | process_interval(); |
13370a9b | 663 | } |
60666c63 | 664 | } |
42202dd5 | 665 | |
3df33eff MR |
666 | disable_counters(); |
667 | ||
42202dd5 IM |
668 | t1 = rdclock(); |
669 | ||
9e9772c4 | 670 | update_stats(&walltime_nsecs_stats, t1 - t0); |
42202dd5 | 671 | |
3df33eff MR |
672 | /* |
673 | * Closing a group leader splits the group, and as we only disable | |
674 | * group leaders, results in remaining events becoming enabled. To | |
675 | * avoid arbitrary skew, we must read all counters before closing any | |
676 | * group leaders. | |
677 | */ | |
678 | read_counters(); | |
679 | perf_evlist__close(evsel_list); | |
c52b12ed | 680 | |
42202dd5 IM |
681 | return WEXITSTATUS(status); |
682 | } | |
683 | ||
41cde476 | 684 | static int run_perf_stat(int argc, const char **argv) |
1f16c575 PZ |
685 | { |
686 | int ret; | |
687 | ||
688 | if (pre_cmd) { | |
689 | ret = system(pre_cmd); | |
690 | if (ret) | |
691 | return ret; | |
692 | } | |
693 | ||
694 | if (sync_run) | |
695 | sync(); | |
696 | ||
697 | ret = __run_perf_stat(argc, argv); | |
698 | if (ret) | |
699 | return ret; | |
700 | ||
701 | if (post_cmd) { | |
702 | ret = system(post_cmd); | |
703 | if (ret) | |
704 | return ret; | |
705 | } | |
706 | ||
707 | return ret; | |
708 | } | |
709 | ||
d73515c0 AK |
710 | static void print_running(u64 run, u64 ena) |
711 | { | |
712 | if (csv_output) { | |
5821522e | 713 | fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f", |
d73515c0 AK |
714 | csv_sep, |
715 | run, | |
716 | csv_sep, | |
717 | ena ? 100.0 * run / ena : 100.0); | |
718 | } else if (run != ena) { | |
5821522e | 719 | fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena); |
d73515c0 AK |
720 | } |
721 | } | |
722 | ||
f99844cb IM |
723 | static void print_noise_pct(double total, double avg) |
724 | { | |
0007ecea | 725 | double pct = rel_stddev_stats(total, avg); |
f99844cb | 726 | |
3ae9a34d | 727 | if (csv_output) |
5821522e | 728 | fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct); |
a1bca6cc | 729 | else if (pct) |
5821522e | 730 | fprintf(stat_config.output, " ( +-%6.2f%% )", pct); |
f99844cb IM |
731 | } |
732 | ||
69aad6f1 | 733 | static void print_noise(struct perf_evsel *evsel, double avg) |
42202dd5 | 734 | { |
581cc8a2 | 735 | struct perf_stat_evsel *ps; |
69aad6f1 | 736 | |
849abde9 PZ |
737 | if (run_count == 1) |
738 | return; | |
739 | ||
69aad6f1 | 740 | ps = evsel->priv; |
f99844cb | 741 | print_noise_pct(stddev_stats(&ps->res_stats[0]), avg); |
42202dd5 IM |
742 | } |
743 | ||
12c08a9f | 744 | static void aggr_printout(struct perf_evsel *evsel, int id, int nr) |
44175b6f | 745 | { |
421a50f3 | 746 | switch (stat_config.aggr_mode) { |
12c08a9f | 747 | case AGGR_CORE: |
5821522e | 748 | fprintf(stat_config.output, "S%d-C%*d%s%*d%s", |
12c08a9f SE |
749 | cpu_map__id_to_socket(id), |
750 | csv_output ? 0 : -8, | |
751 | cpu_map__id_to_cpu(id), | |
752 | csv_sep, | |
753 | csv_output ? 0 : 4, | |
754 | nr, | |
755 | csv_sep); | |
756 | break; | |
86ee6e18 | 757 | case AGGR_SOCKET: |
5821522e | 758 | fprintf(stat_config.output, "S%*d%s%*d%s", |
d7e7a451 | 759 | csv_output ? 0 : -5, |
12c08a9f | 760 | id, |
d7e7a451 SE |
761 | csv_sep, |
762 | csv_output ? 0 : 4, | |
763 | nr, | |
764 | csv_sep); | |
86ee6e18 SE |
765 | break; |
766 | case AGGR_NONE: | |
5821522e | 767 | fprintf(stat_config.output, "CPU%*d%s", |
d7470b6a | 768 | csv_output ? 0 : -4, |
12c08a9f | 769 | perf_evsel__cpus(evsel)->map[id], csv_sep); |
86ee6e18 | 770 | break; |
32b8af82 | 771 | case AGGR_THREAD: |
5821522e | 772 | fprintf(stat_config.output, "%*s-%*d%s", |
32b8af82 JO |
773 | csv_output ? 0 : 16, |
774 | thread_map__comm(evsel->threads, id), | |
775 | csv_output ? 0 : -8, | |
776 | thread_map__pid(evsel->threads, id), | |
777 | csv_sep); | |
778 | break; | |
86ee6e18 | 779 | case AGGR_GLOBAL: |
208df99e | 780 | case AGGR_UNSET: |
86ee6e18 SE |
781 | default: |
782 | break; | |
783 | } | |
784 | } | |
785 | ||
140aeadc AK |
786 | struct outstate { |
787 | FILE *fh; | |
788 | bool newline; | |
f9483392 | 789 | const char *prefix; |
92a61f64 | 790 | int nfields; |
44d49a60 AK |
791 | int id, nr; |
792 | struct perf_evsel *evsel; | |
140aeadc AK |
793 | }; |
794 | ||
795 | #define METRIC_LEN 35 | |
796 | ||
797 | static void new_line_std(void *ctx) | |
798 | { | |
799 | struct outstate *os = ctx; | |
800 | ||
801 | os->newline = true; | |
802 | } | |
803 | ||
804 | static void do_new_line_std(struct outstate *os) | |
805 | { | |
806 | fputc('\n', os->fh); | |
f9483392 | 807 | fputs(os->prefix, os->fh); |
44d49a60 | 808 | aggr_printout(os->evsel, os->id, os->nr); |
140aeadc AK |
809 | if (stat_config.aggr_mode == AGGR_NONE) |
810 | fprintf(os->fh, " "); | |
140aeadc AK |
811 | fprintf(os->fh, " "); |
812 | } | |
813 | ||
814 | static void print_metric_std(void *ctx, const char *color, const char *fmt, | |
815 | const char *unit, double val) | |
816 | { | |
817 | struct outstate *os = ctx; | |
818 | FILE *out = os->fh; | |
819 | int n; | |
820 | bool newline = os->newline; | |
821 | ||
822 | os->newline = false; | |
823 | ||
824 | if (unit == NULL || fmt == NULL) { | |
825 | fprintf(out, "%-*s", METRIC_LEN, ""); | |
826 | return; | |
827 | } | |
828 | ||
829 | if (newline) | |
830 | do_new_line_std(os); | |
831 | ||
832 | n = fprintf(out, " # "); | |
833 | if (color) | |
834 | n += color_fprintf(out, color, fmt, val); | |
835 | else | |
836 | n += fprintf(out, fmt, val); | |
837 | fprintf(out, " %-*s", METRIC_LEN - n - 1, unit); | |
838 | } | |
839 | ||
92a61f64 AK |
840 | static void new_line_csv(void *ctx) |
841 | { | |
842 | struct outstate *os = ctx; | |
843 | int i; | |
844 | ||
845 | fputc('\n', os->fh); | |
846 | if (os->prefix) | |
847 | fprintf(os->fh, "%s%s", os->prefix, csv_sep); | |
44d49a60 | 848 | aggr_printout(os->evsel, os->id, os->nr); |
92a61f64 AK |
849 | for (i = 0; i < os->nfields; i++) |
850 | fputs(csv_sep, os->fh); | |
851 | } | |
852 | ||
853 | static void print_metric_csv(void *ctx, | |
854 | const char *color __maybe_unused, | |
855 | const char *fmt, const char *unit, double val) | |
856 | { | |
857 | struct outstate *os = ctx; | |
858 | FILE *out = os->fh; | |
859 | char buf[64], *vals, *ends; | |
860 | ||
861 | if (unit == NULL || fmt == NULL) { | |
862 | fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep); | |
863 | return; | |
864 | } | |
865 | snprintf(buf, sizeof(buf), fmt, val); | |
866 | vals = buf; | |
867 | while (isspace(*vals)) | |
868 | vals++; | |
869 | ends = vals; | |
870 | while (isdigit(*ends) || *ends == '.') | |
871 | ends++; | |
872 | *ends = 0; | |
873 | while (isspace(*unit)) | |
874 | unit++; | |
875 | fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit); | |
876 | } | |
877 | ||
54b50916 AK |
878 | #define METRIC_ONLY_LEN 20 |
879 | ||
880 | /* Filter out some columns that don't work well in metrics only mode */ | |
881 | ||
882 | static bool valid_only_metric(const char *unit) | |
883 | { | |
884 | if (!unit) | |
885 | return false; | |
886 | if (strstr(unit, "/sec") || | |
887 | strstr(unit, "hz") || | |
888 | strstr(unit, "Hz") || | |
889 | strstr(unit, "CPUs utilized")) | |
890 | return false; | |
891 | return true; | |
892 | } | |
893 | ||
894 | static const char *fixunit(char *buf, struct perf_evsel *evsel, | |
895 | const char *unit) | |
896 | { | |
897 | if (!strncmp(unit, "of all", 6)) { | |
898 | snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel), | |
899 | unit); | |
900 | return buf; | |
901 | } | |
902 | return unit; | |
903 | } | |
904 | ||
905 | static void print_metric_only(void *ctx, const char *color, const char *fmt, | |
906 | const char *unit, double val) | |
907 | { | |
908 | struct outstate *os = ctx; | |
909 | FILE *out = os->fh; | |
910 | int n; | |
911 | char buf[1024]; | |
912 | unsigned mlen = METRIC_ONLY_LEN; | |
913 | ||
914 | if (!valid_only_metric(unit)) | |
915 | return; | |
916 | unit = fixunit(buf, os->evsel, unit); | |
917 | if (color) | |
918 | n = color_fprintf(out, color, fmt, val); | |
919 | else | |
920 | n = fprintf(out, fmt, val); | |
921 | if (n > METRIC_ONLY_LEN) | |
922 | n = METRIC_ONLY_LEN; | |
923 | if (mlen < strlen(unit)) | |
924 | mlen = strlen(unit) + 1; | |
925 | fprintf(out, "%*s", mlen - n, ""); | |
926 | } | |
927 | ||
928 | static void print_metric_only_csv(void *ctx, const char *color __maybe_unused, | |
929 | const char *fmt, | |
930 | const char *unit, double val) | |
931 | { | |
932 | struct outstate *os = ctx; | |
933 | FILE *out = os->fh; | |
934 | char buf[64], *vals, *ends; | |
935 | char tbuf[1024]; | |
936 | ||
937 | if (!valid_only_metric(unit)) | |
938 | return; | |
939 | unit = fixunit(tbuf, os->evsel, unit); | |
940 | snprintf(buf, sizeof buf, fmt, val); | |
941 | vals = buf; | |
942 | while (isspace(*vals)) | |
943 | vals++; | |
944 | ends = vals; | |
945 | while (isdigit(*ends) || *ends == '.') | |
946 | ends++; | |
947 | *ends = 0; | |
948 | fprintf(out, "%s%s", vals, csv_sep); | |
949 | } | |
950 | ||
951 | static void new_line_metric(void *ctx __maybe_unused) | |
952 | { | |
953 | } | |
954 | ||
955 | static void print_metric_header(void *ctx, const char *color __maybe_unused, | |
956 | const char *fmt __maybe_unused, | |
957 | const char *unit, double val __maybe_unused) | |
958 | { | |
959 | struct outstate *os = ctx; | |
960 | char tbuf[1024]; | |
961 | ||
962 | if (!valid_only_metric(unit)) | |
963 | return; | |
964 | unit = fixunit(tbuf, os->evsel, unit); | |
965 | if (csv_output) | |
966 | fprintf(os->fh, "%s%s", unit, csv_sep); | |
967 | else | |
968 | fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit); | |
969 | } | |
970 | ||
da88c7f7 | 971 | static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg) |
86ee6e18 | 972 | { |
5821522e | 973 | FILE *output = stat_config.output; |
86ee6e18 | 974 | double msecs = avg / 1e6; |
410136f5 | 975 | const char *fmt_v, *fmt_n; |
4bbe5a61 | 976 | char name[25]; |
86ee6e18 | 977 | |
410136f5 SE |
978 | fmt_v = csv_output ? "%.6f%s" : "%18.6f%s"; |
979 | fmt_n = csv_output ? "%s" : "%-25s"; | |
980 | ||
da88c7f7 | 981 | aggr_printout(evsel, id, nr); |
d7470b6a | 982 | |
4bbe5a61 DA |
983 | scnprintf(name, sizeof(name), "%s%s", |
984 | perf_evsel__name(evsel), csv_output ? "" : " (msec)"); | |
410136f5 SE |
985 | |
986 | fprintf(output, fmt_v, msecs, csv_sep); | |
987 | ||
988 | if (csv_output) | |
989 | fprintf(output, "%s%s", evsel->unit, csv_sep); | |
990 | else | |
991 | fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep); | |
992 | ||
993 | fprintf(output, fmt_n, name); | |
d7470b6a | 994 | |
023695d9 | 995 | if (evsel->cgrp) |
4aa9015f | 996 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); |
44175b6f IM |
997 | } |
998 | ||
44d49a60 AK |
999 | static int first_shadow_cpu(struct perf_evsel *evsel, int id) |
1000 | { | |
1001 | int i; | |
1002 | ||
1003 | if (!aggr_get_id) | |
1004 | return 0; | |
1005 | ||
1006 | if (stat_config.aggr_mode == AGGR_NONE) | |
1007 | return id; | |
1008 | ||
1009 | if (stat_config.aggr_mode == AGGR_GLOBAL) | |
1010 | return 0; | |
1011 | ||
1012 | for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) { | |
1013 | int cpu2 = perf_evsel__cpus(evsel)->map[i]; | |
1014 | ||
1015 | if (aggr_get_id(evsel_list->cpus, cpu2) == id) | |
1016 | return cpu2; | |
1017 | } | |
1018 | return 0; | |
1019 | } | |
1020 | ||
556b1fb7 JO |
1021 | static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg) |
1022 | { | |
5821522e | 1023 | FILE *output = stat_config.output; |
556b1fb7 JO |
1024 | double sc = evsel->scale; |
1025 | const char *fmt; | |
556b1fb7 JO |
1026 | |
1027 | if (csv_output) { | |
e3b03b6c | 1028 | fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s"; |
556b1fb7 JO |
1029 | } else { |
1030 | if (big_num) | |
e3b03b6c | 1031 | fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s"; |
556b1fb7 | 1032 | else |
e3b03b6c | 1033 | fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s"; |
556b1fb7 JO |
1034 | } |
1035 | ||
1036 | aggr_printout(evsel, id, nr); | |
1037 | ||
556b1fb7 JO |
1038 | fprintf(output, fmt, avg, csv_sep); |
1039 | ||
1040 | if (evsel->unit) | |
1041 | fprintf(output, "%-*s%s", | |
1042 | csv_output ? 0 : unit_width, | |
1043 | evsel->unit, csv_sep); | |
1044 | ||
1045 | fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel)); | |
1046 | ||
1047 | if (evsel->cgrp) | |
1048 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); | |
eedfcb4b | 1049 | } |
556b1fb7 | 1050 | |
f9483392 | 1051 | static void printout(int id, int nr, struct perf_evsel *counter, double uval, |
cb110f47 | 1052 | char *prefix, u64 run, u64 ena, double noise) |
eedfcb4b | 1053 | { |
140aeadc | 1054 | struct perf_stat_output_ctx out; |
f9483392 AK |
1055 | struct outstate os = { |
1056 | .fh = stat_config.output, | |
44d49a60 AK |
1057 | .prefix = prefix ? prefix : "", |
1058 | .id = id, | |
1059 | .nr = nr, | |
1060 | .evsel = counter, | |
f9483392 | 1061 | }; |
140aeadc AK |
1062 | print_metric_t pm = print_metric_std; |
1063 | void (*nl)(void *); | |
eedfcb4b | 1064 | |
54b50916 AK |
1065 | if (metric_only) { |
1066 | nl = new_line_metric; | |
1067 | if (csv_output) | |
1068 | pm = print_metric_only_csv; | |
1069 | else | |
1070 | pm = print_metric_only; | |
1071 | } else | |
1072 | nl = new_line_std; | |
eedfcb4b | 1073 | |
54b50916 | 1074 | if (csv_output && !metric_only) { |
92a61f64 AK |
1075 | static int aggr_fields[] = { |
1076 | [AGGR_GLOBAL] = 0, | |
1077 | [AGGR_THREAD] = 1, | |
1078 | [AGGR_NONE] = 1, | |
1079 | [AGGR_SOCKET] = 2, | |
1080 | [AGGR_CORE] = 2, | |
1081 | }; | |
1082 | ||
1083 | pm = print_metric_csv; | |
1084 | nl = new_line_csv; | |
1085 | os.nfields = 3; | |
1086 | os.nfields += aggr_fields[stat_config.aggr_mode]; | |
1087 | if (counter->cgrp) | |
1088 | os.nfields++; | |
1089 | } | |
b002f3bb | 1090 | if (run == 0 || ena == 0 || counter->counts->scaled == -1) { |
54b50916 AK |
1091 | if (metric_only) { |
1092 | pm(&os, NULL, "", "", 0); | |
1093 | return; | |
1094 | } | |
cb110f47 AK |
1095 | aggr_printout(counter, id, nr); |
1096 | ||
1097 | fprintf(stat_config.output, "%*s%s", | |
1098 | csv_output ? 0 : 18, | |
1099 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, | |
1100 | csv_sep); | |
1101 | ||
1102 | fprintf(stat_config.output, "%-*s%s", | |
1103 | csv_output ? 0 : unit_width, | |
1104 | counter->unit, csv_sep); | |
1105 | ||
1106 | fprintf(stat_config.output, "%*s", | |
1107 | csv_output ? 0 : -25, | |
1108 | perf_evsel__name(counter)); | |
1109 | ||
1110 | if (counter->cgrp) | |
1111 | fprintf(stat_config.output, "%s%s", | |
1112 | csv_sep, counter->cgrp->name); | |
1113 | ||
92a61f64 AK |
1114 | if (!csv_output) |
1115 | pm(&os, NULL, NULL, "", 0); | |
1116 | print_noise(counter, noise); | |
cb110f47 | 1117 | print_running(run, ena); |
92a61f64 AK |
1118 | if (csv_output) |
1119 | pm(&os, NULL, NULL, "", 0); | |
cb110f47 AK |
1120 | return; |
1121 | } | |
1122 | ||
54b50916 AK |
1123 | if (metric_only) |
1124 | /* nothing */; | |
1125 | else if (nsec_counter(counter)) | |
eedfcb4b AK |
1126 | nsec_printout(id, nr, counter, uval); |
1127 | else | |
1128 | abs_printout(id, nr, counter, uval); | |
556b1fb7 | 1129 | |
140aeadc AK |
1130 | out.print_metric = pm; |
1131 | out.new_line = nl; | |
1132 | out.ctx = &os; | |
1133 | ||
54b50916 | 1134 | if (csv_output && !metric_only) { |
92a61f64 AK |
1135 | print_noise(counter, noise); |
1136 | print_running(run, ena); | |
1137 | } | |
1138 | ||
1139 | perf_stat__print_shadow_stats(counter, uval, | |
44d49a60 | 1140 | first_shadow_cpu(counter, id), |
140aeadc | 1141 | &out); |
54b50916 | 1142 | if (!csv_output && !metric_only) { |
92a61f64 AK |
1143 | print_noise(counter, noise); |
1144 | print_running(run, ena); | |
1145 | } | |
556b1fb7 JO |
1146 | } |
1147 | ||
44d49a60 AK |
1148 | static void aggr_update_shadow(void) |
1149 | { | |
1150 | int cpu, s2, id, s; | |
1151 | u64 val; | |
1152 | struct perf_evsel *counter; | |
1153 | ||
1154 | for (s = 0; s < aggr_map->nr; s++) { | |
1155 | id = aggr_map->map[s]; | |
e5cadb93 | 1156 | evlist__for_each_entry(evsel_list, counter) { |
44d49a60 AK |
1157 | val = 0; |
1158 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { | |
1159 | s2 = aggr_get_id(evsel_list->cpus, cpu); | |
1160 | if (s2 != id) | |
1161 | continue; | |
1162 | val += perf_counts(counter->counts, cpu, 0)->val; | |
1163 | } | |
1164 | val = val * counter->scale; | |
1165 | perf_stat__update_shadow_stats(counter, &val, | |
1166 | first_shadow_cpu(counter, id)); | |
1167 | } | |
1168 | } | |
1169 | } | |
1170 | ||
86ee6e18 | 1171 | static void print_aggr(char *prefix) |
d7e7a451 | 1172 | { |
5821522e | 1173 | FILE *output = stat_config.output; |
d7e7a451 | 1174 | struct perf_evsel *counter; |
601083cf | 1175 | int cpu, s, s2, id, nr; |
410136f5 | 1176 | double uval; |
d7e7a451 | 1177 | u64 ena, run, val; |
54b50916 | 1178 | bool first; |
d7e7a451 | 1179 | |
86ee6e18 | 1180 | if (!(aggr_map || aggr_get_id)) |
d7e7a451 SE |
1181 | return; |
1182 | ||
44d49a60 AK |
1183 | aggr_update_shadow(); |
1184 | ||
54b50916 AK |
1185 | /* |
1186 | * With metric_only everything is on a single line. | |
1187 | * Without each counter has its own line. | |
1188 | */ | |
86ee6e18 | 1189 | for (s = 0; s < aggr_map->nr; s++) { |
54b50916 AK |
1190 | if (prefix && metric_only) |
1191 | fprintf(output, "%s", prefix); | |
1192 | ||
86ee6e18 | 1193 | id = aggr_map->map[s]; |
54b50916 | 1194 | first = true; |
e5cadb93 | 1195 | evlist__for_each_entry(evsel_list, counter) { |
d7e7a451 SE |
1196 | val = ena = run = 0; |
1197 | nr = 0; | |
1198 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { | |
601083cf | 1199 | s2 = aggr_get_id(perf_evsel__cpus(counter), cpu); |
86ee6e18 | 1200 | if (s2 != id) |
d7e7a451 | 1201 | continue; |
a6fa0038 JO |
1202 | val += perf_counts(counter->counts, cpu, 0)->val; |
1203 | ena += perf_counts(counter->counts, cpu, 0)->ena; | |
1204 | run += perf_counts(counter->counts, cpu, 0)->run; | |
d7e7a451 SE |
1205 | nr++; |
1206 | } | |
54b50916 AK |
1207 | if (first && metric_only) { |
1208 | first = false; | |
1209 | aggr_printout(counter, id, nr); | |
1210 | } | |
1211 | if (prefix && !metric_only) | |
d7e7a451 SE |
1212 | fprintf(output, "%s", prefix); |
1213 | ||
410136f5 | 1214 | uval = val * counter->scale; |
cb110f47 | 1215 | printout(id, nr, counter, uval, prefix, run, ena, 1.0); |
54b50916 AK |
1216 | if (!metric_only) |
1217 | fputc('\n', output); | |
d7e7a451 | 1218 | } |
54b50916 AK |
1219 | if (metric_only) |
1220 | fputc('\n', output); | |
d7e7a451 SE |
1221 | } |
1222 | } | |
1223 | ||
32b8af82 JO |
1224 | static void print_aggr_thread(struct perf_evsel *counter, char *prefix) |
1225 | { | |
5821522e | 1226 | FILE *output = stat_config.output; |
32b8af82 JO |
1227 | int nthreads = thread_map__nr(counter->threads); |
1228 | int ncpus = cpu_map__nr(counter->cpus); | |
1229 | int cpu, thread; | |
1230 | double uval; | |
1231 | ||
1232 | for (thread = 0; thread < nthreads; thread++) { | |
1233 | u64 ena = 0, run = 0, val = 0; | |
1234 | ||
1235 | for (cpu = 0; cpu < ncpus; cpu++) { | |
1236 | val += perf_counts(counter->counts, cpu, thread)->val; | |
1237 | ena += perf_counts(counter->counts, cpu, thread)->ena; | |
1238 | run += perf_counts(counter->counts, cpu, thread)->run; | |
1239 | } | |
1240 | ||
1241 | if (prefix) | |
1242 | fprintf(output, "%s", prefix); | |
1243 | ||
1244 | uval = val * counter->scale; | |
cb110f47 | 1245 | printout(thread, 0, counter, uval, prefix, run, ena, 1.0); |
32b8af82 JO |
1246 | fputc('\n', output); |
1247 | } | |
1248 | } | |
1249 | ||
2996f5dd IM |
1250 | /* |
1251 | * Print out the results of a single counter: | |
f5b4a9c3 | 1252 | * aggregated counts in system-wide mode |
2996f5dd | 1253 | */ |
13370a9b | 1254 | static void print_counter_aggr(struct perf_evsel *counter, char *prefix) |
2996f5dd | 1255 | { |
5821522e | 1256 | FILE *output = stat_config.output; |
581cc8a2 | 1257 | struct perf_stat_evsel *ps = counter->priv; |
69aad6f1 | 1258 | double avg = avg_stats(&ps->res_stats[0]); |
410136f5 | 1259 | double uval; |
d73515c0 AK |
1260 | double avg_enabled, avg_running; |
1261 | ||
1262 | avg_enabled = avg_stats(&ps->res_stats[1]); | |
1263 | avg_running = avg_stats(&ps->res_stats[2]); | |
2996f5dd | 1264 | |
54b50916 | 1265 | if (prefix && !metric_only) |
13370a9b SE |
1266 | fprintf(output, "%s", prefix); |
1267 | ||
410136f5 | 1268 | uval = avg * counter->scale; |
cb110f47 | 1269 | printout(-1, 0, counter, uval, prefix, avg_running, avg_enabled, avg); |
54b50916 AK |
1270 | if (!metric_only) |
1271 | fprintf(output, "\n"); | |
c04f5e5d IM |
1272 | } |
1273 | ||
f5b4a9c3 SE |
1274 | /* |
1275 | * Print out the results of a single counter: | |
1276 | * does not use aggregated count in system-wide | |
1277 | */ | |
13370a9b | 1278 | static void print_counter(struct perf_evsel *counter, char *prefix) |
f5b4a9c3 | 1279 | { |
5821522e | 1280 | FILE *output = stat_config.output; |
f5b4a9c3 | 1281 | u64 ena, run, val; |
410136f5 | 1282 | double uval; |
f5b4a9c3 SE |
1283 | int cpu; |
1284 | ||
7ae92e74 | 1285 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { |
a6fa0038 JO |
1286 | val = perf_counts(counter->counts, cpu, 0)->val; |
1287 | ena = perf_counts(counter->counts, cpu, 0)->ena; | |
1288 | run = perf_counts(counter->counts, cpu, 0)->run; | |
13370a9b SE |
1289 | |
1290 | if (prefix) | |
1291 | fprintf(output, "%s", prefix); | |
1292 | ||
410136f5 | 1293 | uval = val * counter->scale; |
cb110f47 | 1294 | printout(cpu, 0, counter, uval, prefix, run, ena, 1.0); |
f5b4a9c3 | 1295 | |
4aa9015f | 1296 | fputc('\n', output); |
f5b4a9c3 SE |
1297 | } |
1298 | } | |
1299 | ||
206cab65 AK |
1300 | static void print_no_aggr_metric(char *prefix) |
1301 | { | |
1302 | int cpu; | |
1303 | int nrcpus = 0; | |
1304 | struct perf_evsel *counter; | |
1305 | u64 ena, run, val; | |
1306 | double uval; | |
1307 | ||
1308 | nrcpus = evsel_list->cpus->nr; | |
1309 | for (cpu = 0; cpu < nrcpus; cpu++) { | |
1310 | bool first = true; | |
1311 | ||
1312 | if (prefix) | |
1313 | fputs(prefix, stat_config.output); | |
e5cadb93 | 1314 | evlist__for_each_entry(evsel_list, counter) { |
206cab65 AK |
1315 | if (first) { |
1316 | aggr_printout(counter, cpu, 0); | |
1317 | first = false; | |
1318 | } | |
1319 | val = perf_counts(counter->counts, cpu, 0)->val; | |
1320 | ena = perf_counts(counter->counts, cpu, 0)->ena; | |
1321 | run = perf_counts(counter->counts, cpu, 0)->run; | |
1322 | ||
1323 | uval = val * counter->scale; | |
1324 | printout(cpu, 0, counter, uval, prefix, run, ena, 1.0); | |
1325 | } | |
1326 | fputc('\n', stat_config.output); | |
1327 | } | |
1328 | } | |
1329 | ||
54b50916 AK |
1330 | static int aggr_header_lens[] = { |
1331 | [AGGR_CORE] = 18, | |
1332 | [AGGR_SOCKET] = 12, | |
206cab65 | 1333 | [AGGR_NONE] = 6, |
54b50916 AK |
1334 | [AGGR_THREAD] = 24, |
1335 | [AGGR_GLOBAL] = 0, | |
1336 | }; | |
1337 | ||
c51fd639 AK |
1338 | static const char *aggr_header_csv[] = { |
1339 | [AGGR_CORE] = "core,cpus,", | |
1340 | [AGGR_SOCKET] = "socket,cpus", | |
1341 | [AGGR_NONE] = "cpu,", | |
1342 | [AGGR_THREAD] = "comm-pid,", | |
1343 | [AGGR_GLOBAL] = "" | |
1344 | }; | |
1345 | ||
41c8ca2a | 1346 | static void print_metric_headers(const char *prefix, bool no_indent) |
54b50916 AK |
1347 | { |
1348 | struct perf_stat_output_ctx out; | |
1349 | struct perf_evsel *counter; | |
1350 | struct outstate os = { | |
1351 | .fh = stat_config.output | |
1352 | }; | |
1353 | ||
1354 | if (prefix) | |
1355 | fprintf(stat_config.output, "%s", prefix); | |
1356 | ||
41c8ca2a | 1357 | if (!csv_output && !no_indent) |
54b50916 AK |
1358 | fprintf(stat_config.output, "%*s", |
1359 | aggr_header_lens[stat_config.aggr_mode], ""); | |
c51fd639 AK |
1360 | if (csv_output) { |
1361 | if (stat_config.interval) | |
1362 | fputs("time,", stat_config.output); | |
1363 | fputs(aggr_header_csv[stat_config.aggr_mode], | |
1364 | stat_config.output); | |
1365 | } | |
54b50916 AK |
1366 | |
1367 | /* Print metrics headers only */ | |
e5cadb93 | 1368 | evlist__for_each_entry(evsel_list, counter) { |
54b50916 AK |
1369 | os.evsel = counter; |
1370 | out.ctx = &os; | |
1371 | out.print_metric = print_metric_header; | |
1372 | out.new_line = new_line_metric; | |
1373 | os.evsel = counter; | |
1374 | perf_stat__print_shadow_stats(counter, 0, | |
1375 | 0, | |
1376 | &out); | |
1377 | } | |
1378 | fputc('\n', stat_config.output); | |
1379 | } | |
1380 | ||
d4f63a47 JO |
1381 | static void print_interval(char *prefix, struct timespec *ts) |
1382 | { | |
5821522e | 1383 | FILE *output = stat_config.output; |
d4f63a47 JO |
1384 | static int num_print_interval; |
1385 | ||
1386 | sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep); | |
1387 | ||
41c8ca2a | 1388 | if (num_print_interval == 0 && !csv_output) { |
421a50f3 | 1389 | switch (stat_config.aggr_mode) { |
d4f63a47 | 1390 | case AGGR_SOCKET: |
41c8ca2a AK |
1391 | fprintf(output, "# time socket cpus"); |
1392 | if (!metric_only) | |
1393 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
d4f63a47 JO |
1394 | break; |
1395 | case AGGR_CORE: | |
41c8ca2a AK |
1396 | fprintf(output, "# time core cpus"); |
1397 | if (!metric_only) | |
1398 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
d4f63a47 JO |
1399 | break; |
1400 | case AGGR_NONE: | |
41c8ca2a AK |
1401 | fprintf(output, "# time CPU"); |
1402 | if (!metric_only) | |
1403 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
d4f63a47 | 1404 | break; |
32b8af82 | 1405 | case AGGR_THREAD: |
41c8ca2a AK |
1406 | fprintf(output, "# time comm-pid"); |
1407 | if (!metric_only) | |
1408 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
32b8af82 | 1409 | break; |
d4f63a47 JO |
1410 | case AGGR_GLOBAL: |
1411 | default: | |
41c8ca2a AK |
1412 | fprintf(output, "# time"); |
1413 | if (!metric_only) | |
1414 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
208df99e JO |
1415 | case AGGR_UNSET: |
1416 | break; | |
d4f63a47 JO |
1417 | } |
1418 | } | |
1419 | ||
41c8ca2a AK |
1420 | if (num_print_interval == 0 && metric_only) |
1421 | print_metric_headers(" ", true); | |
d4f63a47 JO |
1422 | if (++num_print_interval == 25) |
1423 | num_print_interval = 0; | |
1424 | } | |
1425 | ||
1426 | static void print_header(int argc, const char **argv) | |
42202dd5 | 1427 | { |
5821522e | 1428 | FILE *output = stat_config.output; |
69aad6f1 | 1429 | int i; |
42202dd5 | 1430 | |
ddcacfa0 IM |
1431 | fflush(stdout); |
1432 | ||
d7470b6a | 1433 | if (!csv_output) { |
4aa9015f SE |
1434 | fprintf(output, "\n"); |
1435 | fprintf(output, " Performance counter stats for "); | |
62d3b617 DA |
1436 | if (target.system_wide) |
1437 | fprintf(output, "\'system wide"); | |
1438 | else if (target.cpu_list) | |
1439 | fprintf(output, "\'CPU(s) %s", target.cpu_list); | |
602ad878 | 1440 | else if (!target__has_task(&target)) { |
ba6039b6 JO |
1441 | fprintf(output, "\'%s", argv ? argv[0] : "pipe"); |
1442 | for (i = 1; argv && (i < argc); i++) | |
4aa9015f | 1443 | fprintf(output, " %s", argv[i]); |
20f946b4 NK |
1444 | } else if (target.pid) |
1445 | fprintf(output, "process id \'%s", target.pid); | |
d7470b6a | 1446 | else |
20f946b4 | 1447 | fprintf(output, "thread id \'%s", target.tid); |
44db76c8 | 1448 | |
4aa9015f | 1449 | fprintf(output, "\'"); |
d7470b6a | 1450 | if (run_count > 1) |
4aa9015f SE |
1451 | fprintf(output, " (%d runs)", run_count); |
1452 | fprintf(output, ":\n\n"); | |
d7470b6a | 1453 | } |
d4f63a47 JO |
1454 | } |
1455 | ||
1456 | static void print_footer(void) | |
1457 | { | |
5821522e JO |
1458 | FILE *output = stat_config.output; |
1459 | ||
d4f63a47 JO |
1460 | if (!null_run) |
1461 | fprintf(output, "\n"); | |
1462 | fprintf(output, " %17.9f seconds time elapsed", | |
1463 | avg_stats(&walltime_nsecs_stats)/1e9); | |
1464 | if (run_count > 1) { | |
1465 | fprintf(output, " "); | |
1466 | print_noise_pct(stddev_stats(&walltime_nsecs_stats), | |
1467 | avg_stats(&walltime_nsecs_stats)); | |
1468 | } | |
1469 | fprintf(output, "\n\n"); | |
1470 | } | |
1471 | ||
1472 | static void print_counters(struct timespec *ts, int argc, const char **argv) | |
1473 | { | |
ec0d3d1f | 1474 | int interval = stat_config.interval; |
d4f63a47 JO |
1475 | struct perf_evsel *counter; |
1476 | char buf[64], *prefix = NULL; | |
1477 | ||
664c98d4 JO |
1478 | /* Do not print anything if we record to the pipe. */ |
1479 | if (STAT_RECORD && perf_stat.file.is_pipe) | |
1480 | return; | |
1481 | ||
d4f63a47 JO |
1482 | if (interval) |
1483 | print_interval(prefix = buf, ts); | |
1484 | else | |
1485 | print_header(argc, argv); | |
2996f5dd | 1486 | |
54b50916 AK |
1487 | if (metric_only) { |
1488 | static int num_print_iv; | |
1489 | ||
41c8ca2a AK |
1490 | if (num_print_iv == 0 && !interval) |
1491 | print_metric_headers(prefix, false); | |
54b50916 AK |
1492 | if (num_print_iv++ == 25) |
1493 | num_print_iv = 0; | |
1494 | if (stat_config.aggr_mode == AGGR_GLOBAL && prefix) | |
1495 | fprintf(stat_config.output, "%s", prefix); | |
1496 | } | |
1497 | ||
421a50f3 | 1498 | switch (stat_config.aggr_mode) { |
12c08a9f | 1499 | case AGGR_CORE: |
86ee6e18 | 1500 | case AGGR_SOCKET: |
d4f63a47 | 1501 | print_aggr(prefix); |
86ee6e18 | 1502 | break; |
32b8af82 | 1503 | case AGGR_THREAD: |
e5cadb93 | 1504 | evlist__for_each_entry(evsel_list, counter) |
32b8af82 JO |
1505 | print_aggr_thread(counter, prefix); |
1506 | break; | |
86ee6e18 | 1507 | case AGGR_GLOBAL: |
e5cadb93 | 1508 | evlist__for_each_entry(evsel_list, counter) |
d4f63a47 | 1509 | print_counter_aggr(counter, prefix); |
54b50916 AK |
1510 | if (metric_only) |
1511 | fputc('\n', stat_config.output); | |
86ee6e18 SE |
1512 | break; |
1513 | case AGGR_NONE: | |
206cab65 AK |
1514 | if (metric_only) |
1515 | print_no_aggr_metric(prefix); | |
1516 | else { | |
e5cadb93 | 1517 | evlist__for_each_entry(evsel_list, counter) |
206cab65 AK |
1518 | print_counter(counter, prefix); |
1519 | } | |
86ee6e18 | 1520 | break; |
208df99e | 1521 | case AGGR_UNSET: |
86ee6e18 SE |
1522 | default: |
1523 | break; | |
f5b4a9c3 | 1524 | } |
ddcacfa0 | 1525 | |
d4f63a47 JO |
1526 | if (!interval && !csv_output) |
1527 | print_footer(); | |
1528 | ||
5821522e | 1529 | fflush(stat_config.output); |
ddcacfa0 IM |
1530 | } |
1531 | ||
f7b7c26e PZ |
1532 | static volatile int signr = -1; |
1533 | ||
5242519b | 1534 | static void skip_signal(int signo) |
ddcacfa0 | 1535 | { |
ec0d3d1f | 1536 | if ((child_pid == -1) || stat_config.interval) |
60666c63 LW |
1537 | done = 1; |
1538 | ||
f7b7c26e | 1539 | signr = signo; |
d07f0b12 SE |
1540 | /* |
1541 | * render child_pid harmless | |
1542 | * won't send SIGTERM to a random | |
1543 | * process in case of race condition | |
1544 | * and fast PID recycling | |
1545 | */ | |
1546 | child_pid = -1; | |
f7b7c26e PZ |
1547 | } |
1548 | ||
1549 | static void sig_atexit(void) | |
1550 | { | |
d07f0b12 SE |
1551 | sigset_t set, oset; |
1552 | ||
1553 | /* | |
1554 | * avoid race condition with SIGCHLD handler | |
1555 | * in skip_signal() which is modifying child_pid | |
1556 | * goal is to avoid send SIGTERM to a random | |
1557 | * process | |
1558 | */ | |
1559 | sigemptyset(&set); | |
1560 | sigaddset(&set, SIGCHLD); | |
1561 | sigprocmask(SIG_BLOCK, &set, &oset); | |
1562 | ||
933da83a CW |
1563 | if (child_pid != -1) |
1564 | kill(child_pid, SIGTERM); | |
1565 | ||
d07f0b12 SE |
1566 | sigprocmask(SIG_SETMASK, &oset, NULL); |
1567 | ||
f7b7c26e PZ |
1568 | if (signr == -1) |
1569 | return; | |
1570 | ||
1571 | signal(signr, SIG_DFL); | |
1572 | kill(getpid(), signr); | |
5242519b IM |
1573 | } |
1574 | ||
1d037ca1 IT |
1575 | static int stat__set_big_num(const struct option *opt __maybe_unused, |
1576 | const char *s __maybe_unused, int unset) | |
d7470b6a SE |
1577 | { |
1578 | big_num_opt = unset ? 0 : 1; | |
1579 | return 0; | |
1580 | } | |
1581 | ||
44b1e60a AK |
1582 | static int enable_metric_only(const struct option *opt __maybe_unused, |
1583 | const char *s __maybe_unused, int unset) | |
1584 | { | |
1585 | force_metric_only = true; | |
1586 | metric_only = !unset; | |
1587 | return 0; | |
1588 | } | |
1589 | ||
e0547311 JO |
1590 | static const struct option stat_options[] = { |
1591 | OPT_BOOLEAN('T', "transaction", &transaction_run, | |
1592 | "hardware transaction statistics"), | |
1593 | OPT_CALLBACK('e', "event", &evsel_list, "event", | |
1594 | "event selector. use 'perf list' to list available events", | |
1595 | parse_events_option), | |
1596 | OPT_CALLBACK(0, "filter", &evsel_list, "filter", | |
1597 | "event filter", parse_filter), | |
1598 | OPT_BOOLEAN('i', "no-inherit", &no_inherit, | |
1599 | "child tasks do not inherit counters"), | |
1600 | OPT_STRING('p', "pid", &target.pid, "pid", | |
1601 | "stat events on existing process id"), | |
1602 | OPT_STRING('t', "tid", &target.tid, "tid", | |
1603 | "stat events on existing thread id"), | |
1604 | OPT_BOOLEAN('a', "all-cpus", &target.system_wide, | |
1605 | "system-wide collection from all CPUs"), | |
1606 | OPT_BOOLEAN('g', "group", &group, | |
1607 | "put the counters into a counter group"), | |
1608 | OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"), | |
1609 | OPT_INCR('v', "verbose", &verbose, | |
1610 | "be more verbose (show counter open errors, etc)"), | |
1611 | OPT_INTEGER('r', "repeat", &run_count, | |
1612 | "repeat command and print average + stddev (max: 100, forever: 0)"), | |
1613 | OPT_BOOLEAN('n', "null", &null_run, | |
1614 | "null run - dont start any counters"), | |
1615 | OPT_INCR('d', "detailed", &detailed_run, | |
1616 | "detailed run - start a lot of events"), | |
1617 | OPT_BOOLEAN('S', "sync", &sync_run, | |
1618 | "call sync() before starting a run"), | |
1619 | OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, | |
1620 | "print large numbers with thousands\' separators", | |
1621 | stat__set_big_num), | |
1622 | OPT_STRING('C', "cpu", &target.cpu_list, "cpu", | |
1623 | "list of cpus to monitor in system-wide"), | |
1624 | OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode, | |
1625 | "disable CPU count aggregation", AGGR_NONE), | |
1626 | OPT_STRING('x', "field-separator", &csv_sep, "separator", | |
1627 | "print counts with custom separator"), | |
1628 | OPT_CALLBACK('G', "cgroup", &evsel_list, "name", | |
1629 | "monitor event in cgroup name only", parse_cgroups), | |
1630 | OPT_STRING('o', "output", &output_name, "file", "output file name"), | |
1631 | OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), | |
1632 | OPT_INTEGER(0, "log-fd", &output_fd, | |
1633 | "log output to fd, instead of stderr"), | |
1634 | OPT_STRING(0, "pre", &pre_cmd, "command", | |
1635 | "command to run prior to the measured command"), | |
1636 | OPT_STRING(0, "post", &post_cmd, "command", | |
1637 | "command to run after to the measured command"), | |
1638 | OPT_UINTEGER('I', "interval-print", &stat_config.interval, | |
1639 | "print counts at regular interval in ms (>= 10)"), | |
1640 | OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode, | |
1641 | "aggregate counts per processor socket", AGGR_SOCKET), | |
1642 | OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode, | |
1643 | "aggregate counts per physical processor core", AGGR_CORE), | |
1644 | OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode, | |
1645 | "aggregate counts per thread", AGGR_THREAD), | |
1646 | OPT_UINTEGER('D', "delay", &initial_delay, | |
1647 | "ms to wait before starting measurement after program start"), | |
44b1e60a AK |
1648 | OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL, |
1649 | "Only print computed metrics. No raw values", enable_metric_only), | |
1650 | OPT_BOOLEAN(0, "topdown", &topdown_run, | |
1651 | "measure topdown level 1 statistics"), | |
e0547311 JO |
1652 | OPT_END() |
1653 | }; | |
1654 | ||
1fe7a300 JO |
1655 | static int perf_stat__get_socket(struct cpu_map *map, int cpu) |
1656 | { | |
1657 | return cpu_map__get_socket(map, cpu, NULL); | |
1658 | } | |
1659 | ||
1660 | static int perf_stat__get_core(struct cpu_map *map, int cpu) | |
1661 | { | |
1662 | return cpu_map__get_core(map, cpu, NULL); | |
1663 | } | |
1664 | ||
1e5a2931 JO |
1665 | static int cpu_map__get_max(struct cpu_map *map) |
1666 | { | |
1667 | int i, max = -1; | |
1668 | ||
1669 | for (i = 0; i < map->nr; i++) { | |
1670 | if (map->map[i] > max) | |
1671 | max = map->map[i]; | |
1672 | } | |
1673 | ||
1674 | return max; | |
1675 | } | |
1676 | ||
1677 | static struct cpu_map *cpus_aggr_map; | |
1678 | ||
1679 | static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx) | |
1680 | { | |
1681 | int cpu; | |
1682 | ||
1683 | if (idx >= map->nr) | |
1684 | return -1; | |
1685 | ||
1686 | cpu = map->map[idx]; | |
1687 | ||
1688 | if (cpus_aggr_map->map[cpu] == -1) | |
1689 | cpus_aggr_map->map[cpu] = get_id(map, idx); | |
1690 | ||
1691 | return cpus_aggr_map->map[cpu]; | |
1692 | } | |
1693 | ||
1694 | static int perf_stat__get_socket_cached(struct cpu_map *map, int idx) | |
1695 | { | |
1696 | return perf_stat__get_aggr(perf_stat__get_socket, map, idx); | |
1697 | } | |
1698 | ||
1699 | static int perf_stat__get_core_cached(struct cpu_map *map, int idx) | |
1700 | { | |
1701 | return perf_stat__get_aggr(perf_stat__get_core, map, idx); | |
1702 | } | |
1703 | ||
86ee6e18 SE |
1704 | static int perf_stat_init_aggr_mode(void) |
1705 | { | |
1e5a2931 JO |
1706 | int nr; |
1707 | ||
421a50f3 | 1708 | switch (stat_config.aggr_mode) { |
86ee6e18 SE |
1709 | case AGGR_SOCKET: |
1710 | if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) { | |
1711 | perror("cannot build socket map"); | |
1712 | return -1; | |
1713 | } | |
1e5a2931 | 1714 | aggr_get_id = perf_stat__get_socket_cached; |
86ee6e18 | 1715 | break; |
12c08a9f SE |
1716 | case AGGR_CORE: |
1717 | if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) { | |
1718 | perror("cannot build core map"); | |
1719 | return -1; | |
1720 | } | |
1e5a2931 | 1721 | aggr_get_id = perf_stat__get_core_cached; |
12c08a9f | 1722 | break; |
86ee6e18 SE |
1723 | case AGGR_NONE: |
1724 | case AGGR_GLOBAL: | |
32b8af82 | 1725 | case AGGR_THREAD: |
208df99e | 1726 | case AGGR_UNSET: |
86ee6e18 SE |
1727 | default: |
1728 | break; | |
1729 | } | |
1e5a2931 JO |
1730 | |
1731 | /* | |
1732 | * The evsel_list->cpus is the base we operate on, | |
1733 | * taking the highest cpu number to be the size of | |
1734 | * the aggregation translate cpumap. | |
1735 | */ | |
1736 | nr = cpu_map__get_max(evsel_list->cpus); | |
1737 | cpus_aggr_map = cpu_map__empty_new(nr + 1); | |
1738 | return cpus_aggr_map ? 0 : -ENOMEM; | |
86ee6e18 SE |
1739 | } |
1740 | ||
544c2ae7 MH |
1741 | static void perf_stat__exit_aggr_mode(void) |
1742 | { | |
1743 | cpu_map__put(aggr_map); | |
1744 | cpu_map__put(cpus_aggr_map); | |
1745 | aggr_map = NULL; | |
1746 | cpus_aggr_map = NULL; | |
1747 | } | |
1748 | ||
68d702f7 JO |
1749 | static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx) |
1750 | { | |
1751 | int cpu; | |
1752 | ||
1753 | if (idx > map->nr) | |
1754 | return -1; | |
1755 | ||
1756 | cpu = map->map[idx]; | |
1757 | ||
1758 | if (cpu >= env->nr_cpus_online) | |
1759 | return -1; | |
1760 | ||
1761 | return cpu; | |
1762 | } | |
1763 | ||
1764 | static int perf_env__get_socket(struct cpu_map *map, int idx, void *data) | |
1765 | { | |
1766 | struct perf_env *env = data; | |
1767 | int cpu = perf_env__get_cpu(env, map, idx); | |
1768 | ||
1769 | return cpu == -1 ? -1 : env->cpu[cpu].socket_id; | |
1770 | } | |
1771 | ||
1772 | static int perf_env__get_core(struct cpu_map *map, int idx, void *data) | |
1773 | { | |
1774 | struct perf_env *env = data; | |
1775 | int core = -1, cpu = perf_env__get_cpu(env, map, idx); | |
1776 | ||
1777 | if (cpu != -1) { | |
1778 | int socket_id = env->cpu[cpu].socket_id; | |
1779 | ||
1780 | /* | |
1781 | * Encode socket in upper 16 bits | |
1782 | * core_id is relative to socket, and | |
1783 | * we need a global id. So we combine | |
1784 | * socket + core id. | |
1785 | */ | |
1786 | core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff); | |
1787 | } | |
1788 | ||
1789 | return core; | |
1790 | } | |
1791 | ||
1792 | static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus, | |
1793 | struct cpu_map **sockp) | |
1794 | { | |
1795 | return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env); | |
1796 | } | |
1797 | ||
1798 | static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus, | |
1799 | struct cpu_map **corep) | |
1800 | { | |
1801 | return cpu_map__build_map(cpus, corep, perf_env__get_core, env); | |
1802 | } | |
1803 | ||
1804 | static int perf_stat__get_socket_file(struct cpu_map *map, int idx) | |
1805 | { | |
1806 | return perf_env__get_socket(map, idx, &perf_stat.session->header.env); | |
1807 | } | |
1808 | ||
1809 | static int perf_stat__get_core_file(struct cpu_map *map, int idx) | |
1810 | { | |
1811 | return perf_env__get_core(map, idx, &perf_stat.session->header.env); | |
1812 | } | |
1813 | ||
1814 | static int perf_stat_init_aggr_mode_file(struct perf_stat *st) | |
1815 | { | |
1816 | struct perf_env *env = &st->session->header.env; | |
1817 | ||
1818 | switch (stat_config.aggr_mode) { | |
1819 | case AGGR_SOCKET: | |
1820 | if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) { | |
1821 | perror("cannot build socket map"); | |
1822 | return -1; | |
1823 | } | |
1824 | aggr_get_id = perf_stat__get_socket_file; | |
1825 | break; | |
1826 | case AGGR_CORE: | |
1827 | if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) { | |
1828 | perror("cannot build core map"); | |
1829 | return -1; | |
1830 | } | |
1831 | aggr_get_id = perf_stat__get_core_file; | |
1832 | break; | |
1833 | case AGGR_NONE: | |
1834 | case AGGR_GLOBAL: | |
1835 | case AGGR_THREAD: | |
1836 | case AGGR_UNSET: | |
1837 | default: | |
1838 | break; | |
1839 | } | |
1840 | ||
1841 | return 0; | |
1842 | } | |
1843 | ||
44b1e60a AK |
1844 | static int topdown_filter_events(const char **attr, char **str, bool use_group) |
1845 | { | |
1846 | int off = 0; | |
1847 | int i; | |
1848 | int len = 0; | |
1849 | char *s; | |
1850 | ||
1851 | for (i = 0; attr[i]; i++) { | |
1852 | if (pmu_have_event("cpu", attr[i])) { | |
1853 | len += strlen(attr[i]) + 1; | |
1854 | attr[i - off] = attr[i]; | |
1855 | } else | |
1856 | off++; | |
1857 | } | |
1858 | attr[i - off] = NULL; | |
1859 | ||
1860 | *str = malloc(len + 1 + 2); | |
1861 | if (!*str) | |
1862 | return -1; | |
1863 | s = *str; | |
1864 | if (i - off == 0) { | |
1865 | *s = 0; | |
1866 | return 0; | |
1867 | } | |
1868 | if (use_group) | |
1869 | *s++ = '{'; | |
1870 | for (i = 0; attr[i]; i++) { | |
1871 | strcpy(s, attr[i]); | |
1872 | s += strlen(s); | |
1873 | *s++ = ','; | |
1874 | } | |
1875 | if (use_group) { | |
1876 | s[-1] = '}'; | |
1877 | *s = 0; | |
1878 | } else | |
1879 | s[-1] = 0; | |
1880 | return 0; | |
1881 | } | |
1882 | ||
1883 | __weak bool arch_topdown_check_group(bool *warn) | |
1884 | { | |
1885 | *warn = false; | |
1886 | return false; | |
1887 | } | |
1888 | ||
1889 | __weak void arch_topdown_group_warn(void) | |
1890 | { | |
1891 | } | |
1892 | ||
2cba3ffb IM |
1893 | /* |
1894 | * Add default attributes, if there were no attributes specified or | |
1895 | * if -d/--detailed, -d -d or -d -d -d is used: | |
1896 | */ | |
1897 | static int add_default_attributes(void) | |
1898 | { | |
44b1e60a | 1899 | int err; |
9dec4473 | 1900 | struct perf_event_attr default_attrs0[] = { |
b070a547 ACM |
1901 | |
1902 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, | |
1903 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, | |
1904 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, | |
1905 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, | |
1906 | ||
1907 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, | |
9dec4473 AK |
1908 | }; |
1909 | struct perf_event_attr frontend_attrs[] = { | |
b070a547 | 1910 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, |
9dec4473 AK |
1911 | }; |
1912 | struct perf_event_attr backend_attrs[] = { | |
b070a547 | 1913 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, |
9dec4473 AK |
1914 | }; |
1915 | struct perf_event_attr default_attrs1[] = { | |
b070a547 ACM |
1916 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, |
1917 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, | |
1918 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, | |
1919 | ||
1920 | }; | |
1921 | ||
1922 | /* | |
1923 | * Detailed stats (-d), covering the L1 and last level data caches: | |
1924 | */ | |
1925 | struct perf_event_attr detailed_attrs[] = { | |
1926 | ||
1927 | { .type = PERF_TYPE_HW_CACHE, | |
1928 | .config = | |
1929 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1930 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1931 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1932 | ||
1933 | { .type = PERF_TYPE_HW_CACHE, | |
1934 | .config = | |
1935 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1936 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1937 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1938 | ||
1939 | { .type = PERF_TYPE_HW_CACHE, | |
1940 | .config = | |
1941 | PERF_COUNT_HW_CACHE_LL << 0 | | |
1942 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1943 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1944 | ||
1945 | { .type = PERF_TYPE_HW_CACHE, | |
1946 | .config = | |
1947 | PERF_COUNT_HW_CACHE_LL << 0 | | |
1948 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1949 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1950 | }; | |
1951 | ||
1952 | /* | |
1953 | * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: | |
1954 | */ | |
1955 | struct perf_event_attr very_detailed_attrs[] = { | |
1956 | ||
1957 | { .type = PERF_TYPE_HW_CACHE, | |
1958 | .config = | |
1959 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
1960 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1961 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1962 | ||
1963 | { .type = PERF_TYPE_HW_CACHE, | |
1964 | .config = | |
1965 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
1966 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1967 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1968 | ||
1969 | { .type = PERF_TYPE_HW_CACHE, | |
1970 | .config = | |
1971 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1972 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1973 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1974 | ||
1975 | { .type = PERF_TYPE_HW_CACHE, | |
1976 | .config = | |
1977 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1978 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1979 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1980 | ||
1981 | { .type = PERF_TYPE_HW_CACHE, | |
1982 | .config = | |
1983 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
1984 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1985 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1986 | ||
1987 | { .type = PERF_TYPE_HW_CACHE, | |
1988 | .config = | |
1989 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
1990 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1991 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1992 | ||
1993 | }; | |
1994 | ||
1995 | /* | |
1996 | * Very, very detailed stats (-d -d -d), adding prefetch events: | |
1997 | */ | |
1998 | struct perf_event_attr very_very_detailed_attrs[] = { | |
1999 | ||
2000 | { .type = PERF_TYPE_HW_CACHE, | |
2001 | .config = | |
2002 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
2003 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
2004 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2005 | ||
2006 | { .type = PERF_TYPE_HW_CACHE, | |
2007 | .config = | |
2008 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
2009 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
2010 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2011 | }; | |
2012 | ||
2cba3ffb IM |
2013 | /* Set attrs if no event is selected and !null_run: */ |
2014 | if (null_run) | |
2015 | return 0; | |
2016 | ||
4cabc3d1 | 2017 | if (transaction_run) { |
4cabc3d1 AK |
2018 | if (pmu_have_event("cpu", "cycles-ct") && |
2019 | pmu_have_event("cpu", "el-start")) | |
a454742c | 2020 | err = parse_events(evsel_list, transaction_attrs, NULL); |
4cabc3d1 | 2021 | else |
a454742c JO |
2022 | err = parse_events(evsel_list, transaction_limited_attrs, NULL); |
2023 | if (err) { | |
4cabc3d1 AK |
2024 | fprintf(stderr, "Cannot set up transaction events\n"); |
2025 | return -1; | |
2026 | } | |
2027 | return 0; | |
2028 | } | |
2029 | ||
44b1e60a AK |
2030 | if (topdown_run) { |
2031 | char *str = NULL; | |
2032 | bool warn = false; | |
2033 | ||
2034 | if (stat_config.aggr_mode != AGGR_GLOBAL && | |
2035 | stat_config.aggr_mode != AGGR_CORE) { | |
2036 | pr_err("top down event configuration requires --per-core mode\n"); | |
2037 | return -1; | |
2038 | } | |
2039 | stat_config.aggr_mode = AGGR_CORE; | |
2040 | if (nr_cgroups || !target__has_cpu(&target)) { | |
2041 | pr_err("top down event configuration requires system-wide mode (-a)\n"); | |
2042 | return -1; | |
2043 | } | |
2044 | ||
2045 | if (!force_metric_only) | |
2046 | metric_only = true; | |
2047 | if (topdown_filter_events(topdown_attrs, &str, | |
2048 | arch_topdown_check_group(&warn)) < 0) { | |
2049 | pr_err("Out of memory\n"); | |
2050 | return -1; | |
2051 | } | |
2052 | if (topdown_attrs[0] && str) { | |
2053 | if (warn) | |
2054 | arch_topdown_group_warn(); | |
2055 | err = parse_events(evsel_list, str, NULL); | |
2056 | if (err) { | |
2057 | fprintf(stderr, | |
2058 | "Cannot set up top down events %s: %d\n", | |
2059 | str, err); | |
2060 | free(str); | |
2061 | return -1; | |
2062 | } | |
2063 | } else { | |
2064 | fprintf(stderr, "System does not support topdown\n"); | |
2065 | return -1; | |
2066 | } | |
2067 | free(str); | |
2068 | } | |
2069 | ||
2cba3ffb | 2070 | if (!evsel_list->nr_entries) { |
a1f3d567 NK |
2071 | if (target__has_cpu(&target)) |
2072 | default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK; | |
2073 | ||
9dec4473 AK |
2074 | if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0) |
2075 | return -1; | |
2076 | if (pmu_have_event("cpu", "stalled-cycles-frontend")) { | |
2077 | if (perf_evlist__add_default_attrs(evsel_list, | |
2078 | frontend_attrs) < 0) | |
2079 | return -1; | |
2080 | } | |
2081 | if (pmu_have_event("cpu", "stalled-cycles-backend")) { | |
2082 | if (perf_evlist__add_default_attrs(evsel_list, | |
2083 | backend_attrs) < 0) | |
2084 | return -1; | |
2085 | } | |
2086 | if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0) | |
50d08e47 | 2087 | return -1; |
2cba3ffb IM |
2088 | } |
2089 | ||
2090 | /* Detailed events get appended to the event list: */ | |
2091 | ||
2092 | if (detailed_run < 1) | |
2093 | return 0; | |
2094 | ||
2095 | /* Append detailed run extra attributes: */ | |
79695e1b | 2096 | if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) |
50d08e47 | 2097 | return -1; |
2cba3ffb IM |
2098 | |
2099 | if (detailed_run < 2) | |
2100 | return 0; | |
2101 | ||
2102 | /* Append very detailed run extra attributes: */ | |
79695e1b | 2103 | if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) |
50d08e47 | 2104 | return -1; |
2cba3ffb IM |
2105 | |
2106 | if (detailed_run < 3) | |
2107 | return 0; | |
2108 | ||
2109 | /* Append very, very detailed run extra attributes: */ | |
79695e1b | 2110 | return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); |
2cba3ffb IM |
2111 | } |
2112 | ||
8a59f3cc | 2113 | static const char * const stat_record_usage[] = { |
4979d0c7 JO |
2114 | "perf stat record [<options>]", |
2115 | NULL, | |
2116 | }; | |
2117 | ||
3ba78bd0 JO |
2118 | static void init_features(struct perf_session *session) |
2119 | { | |
2120 | int feat; | |
2121 | ||
2122 | for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++) | |
2123 | perf_header__set_feat(&session->header, feat); | |
2124 | ||
2125 | perf_header__clear_feat(&session->header, HEADER_BUILD_ID); | |
2126 | perf_header__clear_feat(&session->header, HEADER_TRACING_DATA); | |
2127 | perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK); | |
2128 | perf_header__clear_feat(&session->header, HEADER_AUXTRACE); | |
2129 | } | |
2130 | ||
4979d0c7 JO |
2131 | static int __cmd_record(int argc, const char **argv) |
2132 | { | |
2133 | struct perf_session *session; | |
2134 | struct perf_data_file *file = &perf_stat.file; | |
2135 | ||
8a59f3cc | 2136 | argc = parse_options(argc, argv, stat_options, stat_record_usage, |
4979d0c7 JO |
2137 | PARSE_OPT_STOP_AT_NON_OPTION); |
2138 | ||
2139 | if (output_name) | |
2140 | file->path = output_name; | |
2141 | ||
e9d6db8e JO |
2142 | if (run_count != 1 || forever) { |
2143 | pr_err("Cannot use -r option with perf stat record.\n"); | |
2144 | return -1; | |
2145 | } | |
2146 | ||
4979d0c7 JO |
2147 | session = perf_session__new(file, false, NULL); |
2148 | if (session == NULL) { | |
2149 | pr_err("Perf session creation failed.\n"); | |
2150 | return -1; | |
2151 | } | |
2152 | ||
3ba78bd0 JO |
2153 | init_features(session); |
2154 | ||
4979d0c7 JO |
2155 | session->evlist = evsel_list; |
2156 | perf_stat.session = session; | |
2157 | perf_stat.record = true; | |
2158 | return argc; | |
2159 | } | |
2160 | ||
a56f9390 JO |
2161 | static int process_stat_round_event(struct perf_tool *tool __maybe_unused, |
2162 | union perf_event *event, | |
2163 | struct perf_session *session) | |
2164 | { | |
e3b03b6c | 2165 | struct stat_round_event *stat_round = &event->stat_round; |
a56f9390 JO |
2166 | struct perf_evsel *counter; |
2167 | struct timespec tsh, *ts = NULL; | |
2168 | const char **argv = session->header.env.cmdline_argv; | |
2169 | int argc = session->header.env.nr_cmdline; | |
2170 | ||
e5cadb93 | 2171 | evlist__for_each_entry(evsel_list, counter) |
a56f9390 JO |
2172 | perf_stat_process_counter(&stat_config, counter); |
2173 | ||
e3b03b6c AK |
2174 | if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL) |
2175 | update_stats(&walltime_nsecs_stats, stat_round->time); | |
a56f9390 | 2176 | |
e3b03b6c AK |
2177 | if (stat_config.interval && stat_round->time) { |
2178 | tsh.tv_sec = stat_round->time / NSECS_PER_SEC; | |
2179 | tsh.tv_nsec = stat_round->time % NSECS_PER_SEC; | |
a56f9390 JO |
2180 | ts = &tsh; |
2181 | } | |
2182 | ||
2183 | print_counters(ts, argc, argv); | |
2184 | return 0; | |
2185 | } | |
2186 | ||
62ba18ba JO |
2187 | static |
2188 | int process_stat_config_event(struct perf_tool *tool __maybe_unused, | |
2189 | union perf_event *event, | |
2190 | struct perf_session *session __maybe_unused) | |
2191 | { | |
68d702f7 JO |
2192 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); |
2193 | ||
62ba18ba | 2194 | perf_event__read_stat_config(&stat_config, &event->stat_config); |
68d702f7 | 2195 | |
89af4e05 JO |
2196 | if (cpu_map__empty(st->cpus)) { |
2197 | if (st->aggr_mode != AGGR_UNSET) | |
2198 | pr_warning("warning: processing task data, aggregation mode not set\n"); | |
2199 | return 0; | |
2200 | } | |
2201 | ||
2202 | if (st->aggr_mode != AGGR_UNSET) | |
2203 | stat_config.aggr_mode = st->aggr_mode; | |
2204 | ||
68d702f7 JO |
2205 | if (perf_stat.file.is_pipe) |
2206 | perf_stat_init_aggr_mode(); | |
2207 | else | |
2208 | perf_stat_init_aggr_mode_file(st); | |
2209 | ||
62ba18ba JO |
2210 | return 0; |
2211 | } | |
2212 | ||
1975d36e JO |
2213 | static int set_maps(struct perf_stat *st) |
2214 | { | |
2215 | if (!st->cpus || !st->threads) | |
2216 | return 0; | |
2217 | ||
2218 | if (WARN_ONCE(st->maps_allocated, "stats double allocation\n")) | |
2219 | return -EINVAL; | |
2220 | ||
2221 | perf_evlist__set_maps(evsel_list, st->cpus, st->threads); | |
2222 | ||
2223 | if (perf_evlist__alloc_stats(evsel_list, true)) | |
2224 | return -ENOMEM; | |
2225 | ||
2226 | st->maps_allocated = true; | |
2227 | return 0; | |
2228 | } | |
2229 | ||
2230 | static | |
2231 | int process_thread_map_event(struct perf_tool *tool __maybe_unused, | |
2232 | union perf_event *event, | |
2233 | struct perf_session *session __maybe_unused) | |
2234 | { | |
2235 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); | |
2236 | ||
2237 | if (st->threads) { | |
2238 | pr_warning("Extra thread map event, ignoring.\n"); | |
2239 | return 0; | |
2240 | } | |
2241 | ||
2242 | st->threads = thread_map__new_event(&event->thread_map); | |
2243 | if (!st->threads) | |
2244 | return -ENOMEM; | |
2245 | ||
2246 | return set_maps(st); | |
2247 | } | |
2248 | ||
2249 | static | |
2250 | int process_cpu_map_event(struct perf_tool *tool __maybe_unused, | |
2251 | union perf_event *event, | |
2252 | struct perf_session *session __maybe_unused) | |
2253 | { | |
2254 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); | |
2255 | struct cpu_map *cpus; | |
2256 | ||
2257 | if (st->cpus) { | |
2258 | pr_warning("Extra cpu map event, ignoring.\n"); | |
2259 | return 0; | |
2260 | } | |
2261 | ||
2262 | cpus = cpu_map__new_data(&event->cpu_map.data); | |
2263 | if (!cpus) | |
2264 | return -ENOMEM; | |
2265 | ||
2266 | st->cpus = cpus; | |
2267 | return set_maps(st); | |
2268 | } | |
2269 | ||
8a59f3cc | 2270 | static const char * const stat_report_usage[] = { |
ba6039b6 JO |
2271 | "perf stat report [<options>]", |
2272 | NULL, | |
2273 | }; | |
2274 | ||
2275 | static struct perf_stat perf_stat = { | |
2276 | .tool = { | |
2277 | .attr = perf_event__process_attr, | |
fa6ea781 | 2278 | .event_update = perf_event__process_event_update, |
1975d36e JO |
2279 | .thread_map = process_thread_map_event, |
2280 | .cpu_map = process_cpu_map_event, | |
62ba18ba | 2281 | .stat_config = process_stat_config_event, |
a56f9390 JO |
2282 | .stat = perf_event__process_stat_event, |
2283 | .stat_round = process_stat_round_event, | |
ba6039b6 | 2284 | }, |
89af4e05 | 2285 | .aggr_mode = AGGR_UNSET, |
ba6039b6 JO |
2286 | }; |
2287 | ||
2288 | static int __cmd_report(int argc, const char **argv) | |
2289 | { | |
2290 | struct perf_session *session; | |
2291 | const struct option options[] = { | |
2292 | OPT_STRING('i', "input", &input_name, "file", "input file name"), | |
89af4e05 JO |
2293 | OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode, |
2294 | "aggregate counts per processor socket", AGGR_SOCKET), | |
2295 | OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode, | |
2296 | "aggregate counts per physical processor core", AGGR_CORE), | |
2297 | OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode, | |
2298 | "disable CPU count aggregation", AGGR_NONE), | |
ba6039b6 JO |
2299 | OPT_END() |
2300 | }; | |
2301 | struct stat st; | |
2302 | int ret; | |
2303 | ||
8a59f3cc | 2304 | argc = parse_options(argc, argv, options, stat_report_usage, 0); |
ba6039b6 JO |
2305 | |
2306 | if (!input_name || !strlen(input_name)) { | |
2307 | if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) | |
2308 | input_name = "-"; | |
2309 | else | |
2310 | input_name = "perf.data"; | |
2311 | } | |
2312 | ||
2313 | perf_stat.file.path = input_name; | |
2314 | perf_stat.file.mode = PERF_DATA_MODE_READ; | |
2315 | ||
2316 | session = perf_session__new(&perf_stat.file, false, &perf_stat.tool); | |
2317 | if (session == NULL) | |
2318 | return -1; | |
2319 | ||
2320 | perf_stat.session = session; | |
2321 | stat_config.output = stderr; | |
2322 | evsel_list = session->evlist; | |
2323 | ||
2324 | ret = perf_session__process_events(session); | |
2325 | if (ret) | |
2326 | return ret; | |
2327 | ||
2328 | perf_session__delete(session); | |
2329 | return 0; | |
2330 | } | |
2331 | ||
1d037ca1 | 2332 | int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused) |
5242519b | 2333 | { |
b070a547 ACM |
2334 | const char * const stat_usage[] = { |
2335 | "perf stat [<options>] [<command>]", | |
2336 | NULL | |
2337 | }; | |
cc03c542 | 2338 | int status = -EINVAL, run_idx; |
4aa9015f | 2339 | const char *mode; |
5821522e | 2340 | FILE *output = stderr; |
ec0d3d1f | 2341 | unsigned int interval; |
ba6039b6 | 2342 | const char * const stat_subcommands[] = { "record", "report" }; |
42202dd5 | 2343 | |
5af52b51 SE |
2344 | setlocale(LC_ALL, ""); |
2345 | ||
334fe7a3 | 2346 | evsel_list = perf_evlist__new(); |
361c99a6 ACM |
2347 | if (evsel_list == NULL) |
2348 | return -ENOMEM; | |
2349 | ||
1669e509 | 2350 | parse_events__shrink_config_terms(); |
4979d0c7 JO |
2351 | argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands, |
2352 | (const char **) stat_usage, | |
2353 | PARSE_OPT_STOP_AT_NON_OPTION); | |
fb4605ba | 2354 | perf_stat__init_shadow_stats(); |
4979d0c7 | 2355 | |
6edb78a2 JO |
2356 | if (csv_sep) { |
2357 | csv_output = true; | |
2358 | if (!strcmp(csv_sep, "\\t")) | |
2359 | csv_sep = "\t"; | |
2360 | } else | |
2361 | csv_sep = DEFAULT_SEPARATOR; | |
2362 | ||
4979d0c7 JO |
2363 | if (argc && !strncmp(argv[0], "rec", 3)) { |
2364 | argc = __cmd_record(argc, argv); | |
2365 | if (argc < 0) | |
2366 | return -1; | |
ba6039b6 JO |
2367 | } else if (argc && !strncmp(argv[0], "rep", 3)) |
2368 | return __cmd_report(argc, argv); | |
d7470b6a | 2369 | |
ec0d3d1f JO |
2370 | interval = stat_config.interval; |
2371 | ||
4979d0c7 JO |
2372 | /* |
2373 | * For record command the -o is already taken care of. | |
2374 | */ | |
2375 | if (!STAT_RECORD && output_name && strcmp(output_name, "-")) | |
4aa9015f SE |
2376 | output = NULL; |
2377 | ||
56f3bae7 JC |
2378 | if (output_name && output_fd) { |
2379 | fprintf(stderr, "cannot use both --output and --log-fd\n"); | |
e0547311 JO |
2380 | parse_options_usage(stat_usage, stat_options, "o", 1); |
2381 | parse_options_usage(NULL, stat_options, "log-fd", 0); | |
cc03c542 | 2382 | goto out; |
56f3bae7 | 2383 | } |
fc3e4d07 | 2384 | |
54b50916 AK |
2385 | if (metric_only && stat_config.aggr_mode == AGGR_THREAD) { |
2386 | fprintf(stderr, "--metric-only is not supported with --per-thread\n"); | |
2387 | goto out; | |
2388 | } | |
2389 | ||
54b50916 AK |
2390 | if (metric_only && run_count > 1) { |
2391 | fprintf(stderr, "--metric-only is not supported with -r\n"); | |
2392 | goto out; | |
2393 | } | |
2394 | ||
fc3e4d07 SE |
2395 | if (output_fd < 0) { |
2396 | fprintf(stderr, "argument to --log-fd must be a > 0\n"); | |
e0547311 | 2397 | parse_options_usage(stat_usage, stat_options, "log-fd", 0); |
cc03c542 | 2398 | goto out; |
fc3e4d07 SE |
2399 | } |
2400 | ||
4aa9015f SE |
2401 | if (!output) { |
2402 | struct timespec tm; | |
2403 | mode = append_file ? "a" : "w"; | |
2404 | ||
2405 | output = fopen(output_name, mode); | |
2406 | if (!output) { | |
2407 | perror("failed to create output file"); | |
fceda7fe | 2408 | return -1; |
4aa9015f SE |
2409 | } |
2410 | clock_gettime(CLOCK_REALTIME, &tm); | |
2411 | fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); | |
fc3e4d07 | 2412 | } else if (output_fd > 0) { |
56f3bae7 JC |
2413 | mode = append_file ? "a" : "w"; |
2414 | output = fdopen(output_fd, mode); | |
2415 | if (!output) { | |
2416 | perror("Failed opening logfd"); | |
2417 | return -errno; | |
2418 | } | |
4aa9015f SE |
2419 | } |
2420 | ||
5821522e JO |
2421 | stat_config.output = output; |
2422 | ||
d7470b6a SE |
2423 | /* |
2424 | * let the spreadsheet do the pretty-printing | |
2425 | */ | |
2426 | if (csv_output) { | |
61a9f324 | 2427 | /* User explicitly passed -B? */ |
d7470b6a SE |
2428 | if (big_num_opt == 1) { |
2429 | fprintf(stderr, "-B option not supported with -x\n"); | |
e0547311 JO |
2430 | parse_options_usage(stat_usage, stat_options, "B", 1); |
2431 | parse_options_usage(NULL, stat_options, "x", 1); | |
cc03c542 | 2432 | goto out; |
d7470b6a SE |
2433 | } else /* Nope, so disable big number formatting */ |
2434 | big_num = false; | |
2435 | } else if (big_num_opt == 0) /* User passed --no-big-num */ | |
2436 | big_num = false; | |
2437 | ||
602ad878 | 2438 | if (!argc && target__none(&target)) |
e0547311 | 2439 | usage_with_options(stat_usage, stat_options); |
ac3063bd | 2440 | |
a7e191c3 | 2441 | if (run_count < 0) { |
cc03c542 | 2442 | pr_err("Run count must be a positive number\n"); |
e0547311 | 2443 | parse_options_usage(stat_usage, stat_options, "r", 1); |
cc03c542 | 2444 | goto out; |
a7e191c3 FD |
2445 | } else if (run_count == 0) { |
2446 | forever = true; | |
2447 | run_count = 1; | |
2448 | } | |
ddcacfa0 | 2449 | |
421a50f3 | 2450 | if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) { |
32b8af82 JO |
2451 | fprintf(stderr, "The --per-thread option is only available " |
2452 | "when monitoring via -p -t options.\n"); | |
e0547311 JO |
2453 | parse_options_usage(NULL, stat_options, "p", 1); |
2454 | parse_options_usage(NULL, stat_options, "t", 1); | |
32b8af82 JO |
2455 | goto out; |
2456 | } | |
2457 | ||
2458 | /* | |
2459 | * no_aggr, cgroup are for system-wide only | |
2460 | * --per-thread is aggregated per thread, we dont mix it with cpu mode | |
2461 | */ | |
421a50f3 JO |
2462 | if (((stat_config.aggr_mode != AGGR_GLOBAL && |
2463 | stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) && | |
602ad878 | 2464 | !target__has_cpu(&target)) { |
023695d9 SE |
2465 | fprintf(stderr, "both cgroup and no-aggregation " |
2466 | "modes only available in system-wide mode\n"); | |
2467 | ||
e0547311 JO |
2468 | parse_options_usage(stat_usage, stat_options, "G", 1); |
2469 | parse_options_usage(NULL, stat_options, "A", 1); | |
2470 | parse_options_usage(NULL, stat_options, "a", 1); | |
cc03c542 | 2471 | goto out; |
d7e7a451 SE |
2472 | } |
2473 | ||
2cba3ffb IM |
2474 | if (add_default_attributes()) |
2475 | goto out; | |
ddcacfa0 | 2476 | |
602ad878 | 2477 | target__validate(&target); |
5c98d466 | 2478 | |
77a6f014 | 2479 | if (perf_evlist__create_maps(evsel_list, &target) < 0) { |
602ad878 | 2480 | if (target__has_task(&target)) { |
77a6f014 | 2481 | pr_err("Problems finding threads of monitor\n"); |
e0547311 JO |
2482 | parse_options_usage(stat_usage, stat_options, "p", 1); |
2483 | parse_options_usage(NULL, stat_options, "t", 1); | |
602ad878 | 2484 | } else if (target__has_cpu(&target)) { |
77a6f014 | 2485 | perror("failed to parse CPUs map"); |
e0547311 JO |
2486 | parse_options_usage(stat_usage, stat_options, "C", 1); |
2487 | parse_options_usage(NULL, stat_options, "a", 1); | |
cc03c542 NK |
2488 | } |
2489 | goto out; | |
60d567e2 | 2490 | } |
32b8af82 JO |
2491 | |
2492 | /* | |
2493 | * Initialize thread_map with comm names, | |
2494 | * so we could print it out on output. | |
2495 | */ | |
421a50f3 | 2496 | if (stat_config.aggr_mode == AGGR_THREAD) |
32b8af82 JO |
2497 | thread_map__read_comms(evsel_list->threads); |
2498 | ||
13370a9b | 2499 | if (interval && interval < 100) { |
19afd104 KL |
2500 | if (interval < 10) { |
2501 | pr_err("print interval must be >= 10ms\n"); | |
e0547311 | 2502 | parse_options_usage(stat_usage, stat_options, "I", 1); |
19afd104 KL |
2503 | goto out; |
2504 | } else | |
2505 | pr_warning("print interval < 100ms. " | |
2506 | "The overhead percentage could be high in some cases. " | |
2507 | "Please proceed with caution.\n"); | |
13370a9b | 2508 | } |
c45c6ea2 | 2509 | |
d134ffb9 | 2510 | if (perf_evlist__alloc_stats(evsel_list, interval)) |
03ad9747 | 2511 | goto out; |
d6d901c2 | 2512 | |
86ee6e18 | 2513 | if (perf_stat_init_aggr_mode()) |
03ad9747 | 2514 | goto out; |
86ee6e18 | 2515 | |
58d7e993 IM |
2516 | /* |
2517 | * We dont want to block the signals - that would cause | |
2518 | * child tasks to inherit that and Ctrl-C would not work. | |
2519 | * What we want is for Ctrl-C to work in the exec()-ed | |
2520 | * task, but being ignored by perf stat itself: | |
2521 | */ | |
f7b7c26e | 2522 | atexit(sig_atexit); |
a7e191c3 FD |
2523 | if (!forever) |
2524 | signal(SIGINT, skip_signal); | |
13370a9b | 2525 | signal(SIGCHLD, skip_signal); |
58d7e993 IM |
2526 | signal(SIGALRM, skip_signal); |
2527 | signal(SIGABRT, skip_signal); | |
2528 | ||
42202dd5 | 2529 | status = 0; |
a7e191c3 | 2530 | for (run_idx = 0; forever || run_idx < run_count; run_idx++) { |
42202dd5 | 2531 | if (run_count != 1 && verbose) |
4aa9015f SE |
2532 | fprintf(output, "[ perf stat: executing run #%d ... ]\n", |
2533 | run_idx + 1); | |
f9cef0a9 | 2534 | |
42202dd5 | 2535 | status = run_perf_stat(argc, argv); |
a7e191c3 | 2536 | if (forever && status != -1) { |
d4f63a47 | 2537 | print_counters(NULL, argc, argv); |
254ecbc7 | 2538 | perf_stat__reset_stats(); |
a7e191c3 | 2539 | } |
42202dd5 IM |
2540 | } |
2541 | ||
a7e191c3 | 2542 | if (!forever && status != -1 && !interval) |
d4f63a47 | 2543 | print_counters(NULL, argc, argv); |
d134ffb9 | 2544 | |
4979d0c7 JO |
2545 | if (STAT_RECORD) { |
2546 | /* | |
2547 | * We synthesize the kernel mmap record just so that older tools | |
2548 | * don't emit warnings about not being able to resolve symbols | |
2549 | * due to /proc/sys/kernel/kptr_restrict settings and instear provide | |
2550 | * a saner message about no samples being in the perf.data file. | |
2551 | * | |
2552 | * This also serves to suppress a warning about f_header.data.size == 0 | |
8b99b1a4 JO |
2553 | * in header.c at the moment 'perf stat record' gets introduced, which |
2554 | * is not really needed once we start adding the stat specific PERF_RECORD_ | |
2555 | * records, but the need to suppress the kptr_restrict messages in older | |
2556 | * tools remain -acme | |
4979d0c7 JO |
2557 | */ |
2558 | int fd = perf_data_file__fd(&perf_stat.file); | |
2559 | int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat, | |
2560 | process_synthesized_event, | |
2561 | &perf_stat.session->machines.host); | |
2562 | if (err) { | |
2563 | pr_warning("Couldn't synthesize the kernel mmap record, harmless, " | |
2564 | "older tools may produce warnings about this file\n."); | |
2565 | } | |
2566 | ||
7aad0c32 JO |
2567 | if (!interval) { |
2568 | if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL)) | |
2569 | pr_err("failed to write stat round event\n"); | |
2570 | } | |
2571 | ||
664c98d4 JO |
2572 | if (!perf_stat.file.is_pipe) { |
2573 | perf_stat.session->header.data_size += perf_stat.bytes_written; | |
2574 | perf_session__write_header(perf_stat.session, evsel_list, fd, true); | |
2575 | } | |
4979d0c7 JO |
2576 | |
2577 | perf_session__delete(perf_stat.session); | |
2578 | } | |
2579 | ||
544c2ae7 | 2580 | perf_stat__exit_aggr_mode(); |
d134ffb9 | 2581 | perf_evlist__free_stats(evsel_list); |
0015e2e1 ACM |
2582 | out: |
2583 | perf_evlist__delete(evsel_list); | |
42202dd5 | 2584 | return status; |
ddcacfa0 | 2585 | } |