]>
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" |
5d8bb1ec | 55 | #include "util/drv_configs.h" |
a5d243d0 | 56 | #include "util/color.h" |
0007ecea | 57 | #include "util/stat.h" |
60666c63 | 58 | #include "util/header.h" |
a12b51c4 | 59 | #include "util/cpumap.h" |
d6d901c2 | 60 | #include "util/thread.h" |
fd78260b | 61 | #include "util/thread_map.h" |
d809560b | 62 | #include "util/counts.h" |
44b1e60a | 63 | #include "util/group.h" |
4979d0c7 | 64 | #include "util/session.h" |
ba6039b6 | 65 | #include "util/tool.h" |
44b1e60a | 66 | #include "util/group.h" |
ba6039b6 | 67 | #include "asm/bug.h" |
ddcacfa0 | 68 | |
bd48c63e | 69 | #include <linux/time64.h> |
44b1e60a | 70 | #include <api/fs/fs.h> |
1f16c575 | 71 | #include <stdlib.h> |
ddcacfa0 | 72 | #include <sys/prctl.h> |
5af52b51 | 73 | #include <locale.h> |
e3b03b6c | 74 | #include <math.h> |
16c8a109 | 75 | |
d7470b6a | 76 | #define DEFAULT_SEPARATOR " " |
2cee77c4 DA |
77 | #define CNTR_NOT_SUPPORTED "<not supported>" |
78 | #define CNTR_NOT_COUNTED "<not counted>" | |
d7470b6a | 79 | |
d4f63a47 | 80 | static void print_counters(struct timespec *ts, int argc, const char **argv); |
13370a9b | 81 | |
4cabc3d1 | 82 | /* Default events used for perf stat -T */ |
a454742c JO |
83 | static const char *transaction_attrs = { |
84 | "task-clock," | |
4cabc3d1 AK |
85 | "{" |
86 | "instructions," | |
87 | "cycles," | |
88 | "cpu/cycles-t/," | |
89 | "cpu/tx-start/," | |
90 | "cpu/el-start/," | |
91 | "cpu/cycles-ct/" | |
92 | "}" | |
93 | }; | |
94 | ||
95 | /* More limited version when the CPU does not have all events. */ | |
a454742c JO |
96 | static const char * transaction_limited_attrs = { |
97 | "task-clock," | |
4cabc3d1 AK |
98 | "{" |
99 | "instructions," | |
100 | "cycles," | |
101 | "cpu/cycles-t/," | |
102 | "cpu/tx-start/" | |
103 | "}" | |
104 | }; | |
105 | ||
44b1e60a AK |
106 | static const char * topdown_attrs[] = { |
107 | "topdown-total-slots", | |
108 | "topdown-slots-retired", | |
109 | "topdown-recovery-bubbles", | |
110 | "topdown-fetch-bubbles", | |
111 | "topdown-slots-issued", | |
112 | NULL, | |
113 | }; | |
114 | ||
666e6d48 | 115 | static struct perf_evlist *evsel_list; |
361c99a6 | 116 | |
602ad878 | 117 | static struct target target = { |
77a6f014 NK |
118 | .uid = UINT_MAX, |
119 | }; | |
ddcacfa0 | 120 | |
1e5a2931 JO |
121 | typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu); |
122 | ||
3d632595 | 123 | static int run_count = 1; |
2e6cdf99 | 124 | static bool no_inherit = false; |
d07f0b12 | 125 | static volatile pid_t child_pid = -1; |
c0555642 | 126 | static bool null_run = false; |
2cba3ffb | 127 | static int detailed_run = 0; |
4cabc3d1 | 128 | static bool transaction_run; |
44b1e60a | 129 | static bool topdown_run = false; |
201e0b06 | 130 | static bool big_num = true; |
d7470b6a | 131 | static int big_num_opt = -1; |
d7470b6a SE |
132 | static const char *csv_sep = NULL; |
133 | static bool csv_output = false; | |
43bece79 | 134 | static bool group = false; |
1f16c575 PZ |
135 | static const char *pre_cmd = NULL; |
136 | static const char *post_cmd = NULL; | |
137 | static bool sync_run = false; | |
41191688 | 138 | static unsigned int initial_delay = 0; |
410136f5 | 139 | static unsigned int unit_width = 4; /* strlen("unit") */ |
a7e191c3 | 140 | static bool forever = false; |
54b50916 | 141 | static bool metric_only = false; |
44b1e60a | 142 | static bool force_metric_only = false; |
13370a9b | 143 | static struct timespec ref_time; |
86ee6e18 | 144 | static struct cpu_map *aggr_map; |
1e5a2931 | 145 | static aggr_get_id_t aggr_get_id; |
e0547311 JO |
146 | static bool append_file; |
147 | static const char *output_name; | |
148 | static int output_fd; | |
02d492e5 | 149 | static int print_free_counters_hint; |
5af52b51 | 150 | |
4979d0c7 JO |
151 | struct perf_stat { |
152 | bool record; | |
153 | struct perf_data_file file; | |
154 | struct perf_session *session; | |
155 | u64 bytes_written; | |
ba6039b6 | 156 | struct perf_tool tool; |
1975d36e JO |
157 | bool maps_allocated; |
158 | struct cpu_map *cpus; | |
159 | struct thread_map *threads; | |
89af4e05 | 160 | enum aggr_mode aggr_mode; |
4979d0c7 JO |
161 | }; |
162 | ||
163 | static struct perf_stat perf_stat; | |
164 | #define STAT_RECORD perf_stat.record | |
165 | ||
60666c63 LW |
166 | static volatile int done = 0; |
167 | ||
421a50f3 JO |
168 | static struct perf_stat_config stat_config = { |
169 | .aggr_mode = AGGR_GLOBAL, | |
711a572e | 170 | .scale = true, |
421a50f3 JO |
171 | }; |
172 | ||
13370a9b SE |
173 | static inline void diff_timespec(struct timespec *r, struct timespec *a, |
174 | struct timespec *b) | |
175 | { | |
176 | r->tv_sec = a->tv_sec - b->tv_sec; | |
177 | if (a->tv_nsec < b->tv_nsec) { | |
310ebb93 | 178 | r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec; |
13370a9b SE |
179 | r->tv_sec--; |
180 | } else { | |
181 | r->tv_nsec = a->tv_nsec - b->tv_nsec ; | |
182 | } | |
183 | } | |
184 | ||
254ecbc7 JO |
185 | static void perf_stat__reset_stats(void) |
186 | { | |
187 | perf_evlist__reset_stats(evsel_list); | |
f87027b9 | 188 | perf_stat__reset_shadow_stats(); |
1eda3b21 JO |
189 | } |
190 | ||
cac21425 | 191 | static int create_perf_stat_counter(struct perf_evsel *evsel) |
ddcacfa0 | 192 | { |
69aad6f1 | 193 | struct perf_event_attr *attr = &evsel->attr; |
727ab04e | 194 | |
711a572e | 195 | if (stat_config.scale) |
a21ca2ca IM |
196 | attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | |
197 | PERF_FORMAT_TOTAL_TIME_RUNNING; | |
ddcacfa0 | 198 | |
5d2cd909 ACM |
199 | attr->inherit = !no_inherit; |
200 | ||
6acd8e92 JO |
201 | /* |
202 | * Some events get initialized with sample_(period/type) set, | |
203 | * like tracepoints. Clear it up for counting. | |
204 | */ | |
205 | attr->sample_period = 0; | |
6db1a5c1 | 206 | |
4979d0c7 JO |
207 | /* |
208 | * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless | |
209 | * while avoiding that older tools show confusing messages. | |
6db1a5c1 JO |
210 | * |
211 | * However for pipe sessions we need to keep it zero, | |
212 | * because script's perf_evsel__check_attr is triggered | |
213 | * by attr->sample_type != 0, and we can't run it on | |
214 | * stat sessions. | |
4979d0c7 | 215 | */ |
6db1a5c1 JO |
216 | if (!(STAT_RECORD && perf_stat.file.is_pipe)) |
217 | attr->sample_type = PERF_SAMPLE_IDENTIFIER; | |
6acd8e92 | 218 | |
67ccdecd JO |
219 | /* |
220 | * Disabling all counters initially, they will be enabled | |
221 | * either manually by us or by kernel via enable_on_exec | |
222 | * set later. | |
223 | */ | |
c8280cec | 224 | if (perf_evsel__is_group_leader(evsel)) { |
67ccdecd JO |
225 | attr->disabled = 1; |
226 | ||
c8280cec JO |
227 | /* |
228 | * In case of initial_delay we enable tracee | |
229 | * events manually. | |
230 | */ | |
231 | if (target__none(&target) && !initial_delay) | |
41191688 | 232 | attr->enable_on_exec = 1; |
ddcacfa0 | 233 | } |
084ab9f8 | 234 | |
c8280cec JO |
235 | if (target__has_cpu(&target)) |
236 | return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel)); | |
237 | ||
594ac61a | 238 | return perf_evsel__open_per_thread(evsel, evsel_list->threads); |
ddcacfa0 IM |
239 | } |
240 | ||
c04f5e5d IM |
241 | /* |
242 | * Does the counter have nsecs as a unit? | |
243 | */ | |
daec78a0 | 244 | static inline int nsec_counter(struct perf_evsel *evsel) |
c04f5e5d | 245 | { |
daec78a0 ACM |
246 | if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || |
247 | perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) | |
c04f5e5d IM |
248 | return 1; |
249 | ||
250 | return 0; | |
251 | } | |
252 | ||
8b99b1a4 JO |
253 | static int process_synthesized_event(struct perf_tool *tool __maybe_unused, |
254 | union perf_event *event, | |
255 | struct perf_sample *sample __maybe_unused, | |
256 | struct machine *machine __maybe_unused) | |
4979d0c7 | 257 | { |
8b99b1a4 | 258 | if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) { |
4979d0c7 JO |
259 | pr_err("failed to write perf data, error: %m\n"); |
260 | return -1; | |
261 | } | |
262 | ||
8b99b1a4 | 263 | perf_stat.bytes_written += event->header.size; |
4979d0c7 JO |
264 | return 0; |
265 | } | |
266 | ||
1975d36e | 267 | static int write_stat_round_event(u64 tm, u64 type) |
7aad0c32 | 268 | { |
1975d36e | 269 | return perf_event__synthesize_stat_round(NULL, tm, type, |
7aad0c32 JO |
270 | process_synthesized_event, |
271 | NULL); | |
272 | } | |
273 | ||
274 | #define WRITE_STAT_ROUND_EVENT(time, interval) \ | |
275 | write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval) | |
276 | ||
5a6ea81b JO |
277 | #define SID(e, x, y) xyarray__entry(e->sample_id, x, y) |
278 | ||
279 | static int | |
280 | perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread, | |
281 | struct perf_counts_values *count) | |
282 | { | |
283 | struct perf_sample_id *sid = SID(counter, cpu, thread); | |
284 | ||
285 | return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count, | |
286 | process_synthesized_event, NULL); | |
287 | } | |
288 | ||
f5b4a9c3 SE |
289 | /* |
290 | * Read out the results of a single counter: | |
291 | * do not aggregate counts across CPUs in system-wide mode | |
292 | */ | |
c52b12ed | 293 | static int read_counter(struct perf_evsel *counter) |
f5b4a9c3 | 294 | { |
9bf1a529 | 295 | int nthreads = thread_map__nr(evsel_list->threads); |
00e727bb MR |
296 | int ncpus, cpu, thread; |
297 | ||
298 | if (target__has_cpu(&target)) | |
299 | ncpus = perf_evsel__nr_cpus(counter); | |
300 | else | |
301 | ncpus = 1; | |
f5b4a9c3 | 302 | |
3b4331d9 SP |
303 | if (!counter->supported) |
304 | return -ENOENT; | |
305 | ||
9bf1a529 JO |
306 | if (counter->system_wide) |
307 | nthreads = 1; | |
308 | ||
309 | for (thread = 0; thread < nthreads; thread++) { | |
310 | for (cpu = 0; cpu < ncpus; cpu++) { | |
3b3eb044 JO |
311 | struct perf_counts_values *count; |
312 | ||
313 | count = perf_counts(counter->counts, cpu, thread); | |
314 | if (perf_evsel__read(counter, cpu, thread, count)) | |
9bf1a529 | 315 | return -1; |
5a6ea81b JO |
316 | |
317 | if (STAT_RECORD) { | |
318 | if (perf_evsel__write_stat_event(counter, cpu, thread, count)) { | |
319 | pr_err("failed to write stat event\n"); | |
320 | return -1; | |
321 | } | |
322 | } | |
0b1abbf4 AK |
323 | |
324 | if (verbose > 1) { | |
325 | fprintf(stat_config.output, | |
326 | "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", | |
327 | perf_evsel__name(counter), | |
328 | cpu, | |
329 | count->val, count->ena, count->run); | |
330 | } | |
9bf1a529 | 331 | } |
f5b4a9c3 | 332 | } |
c52b12ed ACM |
333 | |
334 | return 0; | |
2996f5dd IM |
335 | } |
336 | ||
3df33eff | 337 | static void read_counters(void) |
13370a9b | 338 | { |
13370a9b | 339 | struct perf_evsel *counter; |
13370a9b | 340 | |
e5cadb93 | 341 | evlist__for_each_entry(evsel_list, counter) { |
3b3eb044 | 342 | if (read_counter(counter)) |
245bad8e | 343 | pr_debug("failed to read counter %s\n", counter->name); |
3b3eb044 | 344 | |
f80010eb | 345 | if (perf_stat_process_counter(&stat_config, counter)) |
3b3eb044 | 346 | pr_warning("failed to process counter %s\n", counter->name); |
13370a9b | 347 | } |
106a94a0 JO |
348 | } |
349 | ||
ba411a95 | 350 | static void process_interval(void) |
106a94a0 | 351 | { |
106a94a0 | 352 | struct timespec ts, rs; |
106a94a0 | 353 | |
3df33eff | 354 | read_counters(); |
86ee6e18 | 355 | |
13370a9b SE |
356 | clock_gettime(CLOCK_MONOTONIC, &ts); |
357 | diff_timespec(&rs, &ts, &ref_time); | |
13370a9b | 358 | |
7aad0c32 | 359 | if (STAT_RECORD) { |
bd48c63e | 360 | if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL)) |
7aad0c32 JO |
361 | pr_err("failed to write stat round event\n"); |
362 | } | |
363 | ||
d4f63a47 | 364 | print_counters(&rs, 0, NULL); |
13370a9b SE |
365 | } |
366 | ||
67ccdecd | 367 | static void enable_counters(void) |
41191688 | 368 | { |
67ccdecd | 369 | if (initial_delay) |
310ebb93 | 370 | usleep(initial_delay * USEC_PER_MSEC); |
67ccdecd JO |
371 | |
372 | /* | |
373 | * We need to enable counters only if: | |
374 | * - we don't have tracee (attaching to task or cpu) | |
375 | * - we have initial delay configured | |
376 | */ | |
377 | if (!target__none(&target) || initial_delay) | |
ab46db0a | 378 | perf_evlist__enable(evsel_list); |
41191688 AK |
379 | } |
380 | ||
3df33eff MR |
381 | static void disable_counters(void) |
382 | { | |
383 | /* | |
384 | * If we don't have tracee (attaching to task or cpu), counters may | |
385 | * still be running. To get accurate group ratios, we must stop groups | |
386 | * from counting before reading their constituent counters. | |
387 | */ | |
388 | if (!target__none(&target)) | |
389 | perf_evlist__disable(evsel_list); | |
390 | } | |
391 | ||
f33cbe72 | 392 | static volatile int workload_exec_errno; |
6af206fd ACM |
393 | |
394 | /* | |
395 | * perf_evlist__prepare_workload will send a SIGUSR1 | |
396 | * if the fork fails, since we asked by setting its | |
397 | * want_signal to true. | |
398 | */ | |
f33cbe72 ACM |
399 | static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, |
400 | void *ucontext __maybe_unused) | |
6af206fd | 401 | { |
f33cbe72 | 402 | workload_exec_errno = info->si_value.sival_int; |
6af206fd ACM |
403 | } |
404 | ||
7b60a7e3 JO |
405 | static bool has_unit(struct perf_evsel *counter) |
406 | { | |
407 | return counter->unit && *counter->unit; | |
408 | } | |
409 | ||
410 | static bool has_scale(struct perf_evsel *counter) | |
411 | { | |
412 | return counter->scale != 1; | |
413 | } | |
414 | ||
664c98d4 | 415 | static int perf_stat_synthesize_config(bool is_pipe) |
8b99b1a4 | 416 | { |
7b60a7e3 | 417 | struct perf_evsel *counter; |
8b99b1a4 JO |
418 | int err; |
419 | ||
664c98d4 JO |
420 | if (is_pipe) { |
421 | err = perf_event__synthesize_attrs(NULL, perf_stat.session, | |
422 | process_synthesized_event); | |
423 | if (err < 0) { | |
424 | pr_err("Couldn't synthesize attrs.\n"); | |
425 | return err; | |
426 | } | |
427 | } | |
428 | ||
7b60a7e3 JO |
429 | /* |
430 | * Synthesize other events stuff not carried within | |
431 | * attr event - unit, scale, name | |
432 | */ | |
e5cadb93 | 433 | evlist__for_each_entry(evsel_list, counter) { |
7b60a7e3 JO |
434 | if (!counter->supported) |
435 | continue; | |
436 | ||
437 | /* | |
438 | * Synthesize unit and scale only if it's defined. | |
439 | */ | |
440 | if (has_unit(counter)) { | |
441 | err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event); | |
442 | if (err < 0) { | |
443 | pr_err("Couldn't synthesize evsel unit.\n"); | |
444 | return err; | |
445 | } | |
446 | } | |
447 | ||
448 | if (has_scale(counter)) { | |
449 | err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event); | |
450 | if (err < 0) { | |
451 | pr_err("Couldn't synthesize evsel scale.\n"); | |
452 | return err; | |
453 | } | |
454 | } | |
455 | ||
456 | if (counter->own_cpus) { | |
457 | err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event); | |
458 | if (err < 0) { | |
459 | pr_err("Couldn't synthesize evsel scale.\n"); | |
460 | return err; | |
461 | } | |
462 | } | |
463 | ||
464 | /* | |
465 | * Name is needed only for pipe output, | |
466 | * perf.data carries event names. | |
467 | */ | |
468 | if (is_pipe) { | |
469 | err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event); | |
470 | if (err < 0) { | |
471 | pr_err("Couldn't synthesize evsel name.\n"); | |
472 | return err; | |
473 | } | |
474 | } | |
475 | } | |
476 | ||
8b99b1a4 JO |
477 | err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads, |
478 | process_synthesized_event, | |
479 | NULL); | |
480 | if (err < 0) { | |
481 | pr_err("Couldn't synthesize thread map.\n"); | |
482 | return err; | |
483 | } | |
484 | ||
485 | err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus, | |
486 | process_synthesized_event, NULL); | |
487 | if (err < 0) { | |
488 | pr_err("Couldn't synthesize thread map.\n"); | |
489 | return err; | |
490 | } | |
491 | ||
492 | err = perf_event__synthesize_stat_config(NULL, &stat_config, | |
493 | process_synthesized_event, NULL); | |
494 | if (err < 0) { | |
495 | pr_err("Couldn't synthesize config.\n"); | |
496 | return err; | |
497 | } | |
498 | ||
499 | return 0; | |
500 | } | |
501 | ||
2af4646d JO |
502 | #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y)) |
503 | ||
504 | static int __store_counter_ids(struct perf_evsel *counter, | |
505 | struct cpu_map *cpus, | |
506 | struct thread_map *threads) | |
507 | { | |
508 | int cpu, thread; | |
509 | ||
510 | for (cpu = 0; cpu < cpus->nr; cpu++) { | |
511 | for (thread = 0; thread < threads->nr; thread++) { | |
512 | int fd = FD(counter, cpu, thread); | |
513 | ||
514 | if (perf_evlist__id_add_fd(evsel_list, counter, | |
515 | cpu, thread, fd) < 0) | |
516 | return -1; | |
517 | } | |
518 | } | |
519 | ||
520 | return 0; | |
521 | } | |
522 | ||
523 | static int store_counter_ids(struct perf_evsel *counter) | |
524 | { | |
525 | struct cpu_map *cpus = counter->cpus; | |
526 | struct thread_map *threads = counter->threads; | |
527 | ||
528 | if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr)) | |
529 | return -ENOMEM; | |
530 | ||
531 | return __store_counter_ids(counter, cpus, threads); | |
532 | } | |
533 | ||
acf28922 | 534 | static int __run_perf_stat(int argc, const char **argv) |
42202dd5 | 535 | { |
ec0d3d1f | 536 | int interval = stat_config.interval; |
d6195a6a | 537 | char msg[BUFSIZ]; |
42202dd5 | 538 | unsigned long long t0, t1; |
cac21425 | 539 | struct perf_evsel *counter; |
13370a9b | 540 | struct timespec ts; |
410136f5 | 541 | size_t l; |
42202dd5 | 542 | int status = 0; |
6be2850e | 543 | const bool forks = (argc > 0); |
664c98d4 | 544 | bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false; |
5d8bb1ec | 545 | struct perf_evsel_config_term *err_term; |
42202dd5 | 546 | |
13370a9b | 547 | if (interval) { |
310ebb93 ACM |
548 | ts.tv_sec = interval / USEC_PER_MSEC; |
549 | ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC; | |
13370a9b SE |
550 | } else { |
551 | ts.tv_sec = 1; | |
552 | ts.tv_nsec = 0; | |
553 | } | |
554 | ||
60666c63 | 555 | if (forks) { |
664c98d4 | 556 | if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe, |
735f7e0b | 557 | workload_exec_failed_signal) < 0) { |
acf28922 NK |
558 | perror("failed to prepare workload"); |
559 | return -1; | |
60666c63 | 560 | } |
d20a47e7 | 561 | child_pid = evsel_list->workload.pid; |
051ae7f7 PM |
562 | } |
563 | ||
6a4bb04c | 564 | if (group) |
63dab225 | 565 | perf_evlist__set_leader(evsel_list); |
6a4bb04c | 566 | |
e5cadb93 | 567 | evlist__for_each_entry(evsel_list, counter) { |
42ef8a78 | 568 | try_again: |
cac21425 | 569 | if (create_perf_stat_counter(counter) < 0) { |
979987a5 DA |
570 | /* |
571 | * PPC returns ENXIO for HW counters until 2.6.37 | |
572 | * (behavior changed with commit b0a873e). | |
573 | */ | |
38f6ae1e | 574 | if (errno == EINVAL || errno == ENOSYS || |
979987a5 DA |
575 | errno == ENOENT || errno == EOPNOTSUPP || |
576 | errno == ENXIO) { | |
bb963e16 | 577 | if (verbose > 0) |
c63ca0c0 | 578 | ui__warning("%s event is not supported by the kernel.\n", |
7289f83c | 579 | perf_evsel__name(counter)); |
2cee77c4 | 580 | counter->supported = false; |
cb5ef600 KL |
581 | |
582 | if ((counter->leader != counter) || | |
583 | !(counter->leader->nr_members > 1)) | |
584 | continue; | |
42ef8a78 | 585 | } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) { |
bb963e16 | 586 | if (verbose > 0) |
42ef8a78 ACM |
587 | ui__warning("%s\n", msg); |
588 | goto try_again; | |
589 | } | |
ede70290 | 590 | |
56e52e85 ACM |
591 | perf_evsel__open_strerror(counter, &target, |
592 | errno, msg, sizeof(msg)); | |
593 | ui__error("%s\n", msg); | |
594 | ||
48290609 ACM |
595 | if (child_pid != -1) |
596 | kill(child_pid, SIGTERM); | |
fceda7fe | 597 | |
48290609 ACM |
598 | return -1; |
599 | } | |
2cee77c4 | 600 | counter->supported = true; |
410136f5 SE |
601 | |
602 | l = strlen(counter->unit); | |
603 | if (l > unit_width) | |
604 | unit_width = l; | |
2af4646d JO |
605 | |
606 | if (STAT_RECORD && store_counter_ids(counter)) | |
607 | return -1; | |
084ab9f8 | 608 | } |
42202dd5 | 609 | |
23d4aad4 ACM |
610 | if (perf_evlist__apply_filters(evsel_list, &counter)) { |
611 | error("failed to set filter \"%s\" on event %s with %d (%s)\n", | |
612 | counter->filter, perf_evsel__name(counter), errno, | |
c8b5f2c9 | 613 | str_error_r(errno, msg, sizeof(msg))); |
cfd748ae FW |
614 | return -1; |
615 | } | |
616 | ||
5d8bb1ec MP |
617 | if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) { |
618 | error("failed to set config \"%s\" on event %s with %d (%s)\n", | |
619 | err_term->val.drv_cfg, perf_evsel__name(counter), errno, | |
620 | str_error_r(errno, msg, sizeof(msg))); | |
621 | return -1; | |
622 | } | |
623 | ||
4979d0c7 JO |
624 | if (STAT_RECORD) { |
625 | int err, fd = perf_data_file__fd(&perf_stat.file); | |
626 | ||
664c98d4 JO |
627 | if (is_pipe) { |
628 | err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file)); | |
629 | } else { | |
630 | err = perf_session__write_header(perf_stat.session, evsel_list, | |
631 | fd, false); | |
632 | } | |
633 | ||
4979d0c7 JO |
634 | if (err < 0) |
635 | return err; | |
8b99b1a4 | 636 | |
664c98d4 | 637 | err = perf_stat_synthesize_config(is_pipe); |
8b99b1a4 JO |
638 | if (err < 0) |
639 | return err; | |
4979d0c7 JO |
640 | } |
641 | ||
42202dd5 IM |
642 | /* |
643 | * Enable counters and exec the command: | |
644 | */ | |
645 | t0 = rdclock(); | |
13370a9b | 646 | clock_gettime(CLOCK_MONOTONIC, &ref_time); |
42202dd5 | 647 | |
60666c63 | 648 | if (forks) { |
acf28922 | 649 | perf_evlist__start_workload(evsel_list); |
67ccdecd | 650 | enable_counters(); |
acf28922 | 651 | |
13370a9b SE |
652 | if (interval) { |
653 | while (!waitpid(child_pid, &status, WNOHANG)) { | |
654 | nanosleep(&ts, NULL); | |
ba411a95 | 655 | process_interval(); |
13370a9b SE |
656 | } |
657 | } | |
60666c63 | 658 | wait(&status); |
6af206fd | 659 | |
f33cbe72 | 660 | if (workload_exec_errno) { |
c8b5f2c9 | 661 | const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg)); |
f33cbe72 | 662 | pr_err("Workload failed: %s\n", emsg); |
6af206fd | 663 | return -1; |
f33cbe72 | 664 | } |
6af206fd | 665 | |
33e49ea7 AK |
666 | if (WIFSIGNALED(status)) |
667 | psignal(WTERMSIG(status), argv[0]); | |
60666c63 | 668 | } else { |
67ccdecd | 669 | enable_counters(); |
13370a9b SE |
670 | while (!done) { |
671 | nanosleep(&ts, NULL); | |
672 | if (interval) | |
ba411a95 | 673 | process_interval(); |
13370a9b | 674 | } |
60666c63 | 675 | } |
42202dd5 | 676 | |
3df33eff MR |
677 | disable_counters(); |
678 | ||
42202dd5 IM |
679 | t1 = rdclock(); |
680 | ||
9e9772c4 | 681 | update_stats(&walltime_nsecs_stats, t1 - t0); |
42202dd5 | 682 | |
3df33eff MR |
683 | /* |
684 | * Closing a group leader splits the group, and as we only disable | |
685 | * group leaders, results in remaining events becoming enabled. To | |
686 | * avoid arbitrary skew, we must read all counters before closing any | |
687 | * group leaders. | |
688 | */ | |
689 | read_counters(); | |
690 | perf_evlist__close(evsel_list); | |
c52b12ed | 691 | |
42202dd5 IM |
692 | return WEXITSTATUS(status); |
693 | } | |
694 | ||
41cde476 | 695 | static int run_perf_stat(int argc, const char **argv) |
1f16c575 PZ |
696 | { |
697 | int ret; | |
698 | ||
699 | if (pre_cmd) { | |
700 | ret = system(pre_cmd); | |
701 | if (ret) | |
702 | return ret; | |
703 | } | |
704 | ||
705 | if (sync_run) | |
706 | sync(); | |
707 | ||
708 | ret = __run_perf_stat(argc, argv); | |
709 | if (ret) | |
710 | return ret; | |
711 | ||
712 | if (post_cmd) { | |
713 | ret = system(post_cmd); | |
714 | if (ret) | |
715 | return ret; | |
716 | } | |
717 | ||
718 | return ret; | |
719 | } | |
720 | ||
d73515c0 AK |
721 | static void print_running(u64 run, u64 ena) |
722 | { | |
723 | if (csv_output) { | |
5821522e | 724 | fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f", |
d73515c0 AK |
725 | csv_sep, |
726 | run, | |
727 | csv_sep, | |
728 | ena ? 100.0 * run / ena : 100.0); | |
729 | } else if (run != ena) { | |
5821522e | 730 | fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena); |
d73515c0 AK |
731 | } |
732 | } | |
733 | ||
f99844cb IM |
734 | static void print_noise_pct(double total, double avg) |
735 | { | |
0007ecea | 736 | double pct = rel_stddev_stats(total, avg); |
f99844cb | 737 | |
3ae9a34d | 738 | if (csv_output) |
5821522e | 739 | fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct); |
a1bca6cc | 740 | else if (pct) |
5821522e | 741 | fprintf(stat_config.output, " ( +-%6.2f%% )", pct); |
f99844cb IM |
742 | } |
743 | ||
69aad6f1 | 744 | static void print_noise(struct perf_evsel *evsel, double avg) |
42202dd5 | 745 | { |
581cc8a2 | 746 | struct perf_stat_evsel *ps; |
69aad6f1 | 747 | |
849abde9 PZ |
748 | if (run_count == 1) |
749 | return; | |
750 | ||
69aad6f1 | 751 | ps = evsel->priv; |
f99844cb | 752 | print_noise_pct(stddev_stats(&ps->res_stats[0]), avg); |
42202dd5 IM |
753 | } |
754 | ||
12c08a9f | 755 | static void aggr_printout(struct perf_evsel *evsel, int id, int nr) |
44175b6f | 756 | { |
421a50f3 | 757 | switch (stat_config.aggr_mode) { |
12c08a9f | 758 | case AGGR_CORE: |
5821522e | 759 | fprintf(stat_config.output, "S%d-C%*d%s%*d%s", |
12c08a9f SE |
760 | cpu_map__id_to_socket(id), |
761 | csv_output ? 0 : -8, | |
762 | cpu_map__id_to_cpu(id), | |
763 | csv_sep, | |
764 | csv_output ? 0 : 4, | |
765 | nr, | |
766 | csv_sep); | |
767 | break; | |
86ee6e18 | 768 | case AGGR_SOCKET: |
5821522e | 769 | fprintf(stat_config.output, "S%*d%s%*d%s", |
d7e7a451 | 770 | csv_output ? 0 : -5, |
12c08a9f | 771 | id, |
d7e7a451 SE |
772 | csv_sep, |
773 | csv_output ? 0 : 4, | |
774 | nr, | |
775 | csv_sep); | |
86ee6e18 SE |
776 | break; |
777 | case AGGR_NONE: | |
5821522e | 778 | fprintf(stat_config.output, "CPU%*d%s", |
d7470b6a | 779 | csv_output ? 0 : -4, |
12c08a9f | 780 | perf_evsel__cpus(evsel)->map[id], csv_sep); |
86ee6e18 | 781 | break; |
32b8af82 | 782 | case AGGR_THREAD: |
5821522e | 783 | fprintf(stat_config.output, "%*s-%*d%s", |
32b8af82 JO |
784 | csv_output ? 0 : 16, |
785 | thread_map__comm(evsel->threads, id), | |
786 | csv_output ? 0 : -8, | |
787 | thread_map__pid(evsel->threads, id), | |
788 | csv_sep); | |
789 | break; | |
86ee6e18 | 790 | case AGGR_GLOBAL: |
208df99e | 791 | case AGGR_UNSET: |
86ee6e18 SE |
792 | default: |
793 | break; | |
794 | } | |
795 | } | |
796 | ||
140aeadc AK |
797 | struct outstate { |
798 | FILE *fh; | |
799 | bool newline; | |
f9483392 | 800 | const char *prefix; |
92a61f64 | 801 | int nfields; |
44d49a60 AK |
802 | int id, nr; |
803 | struct perf_evsel *evsel; | |
140aeadc AK |
804 | }; |
805 | ||
806 | #define METRIC_LEN 35 | |
807 | ||
808 | static void new_line_std(void *ctx) | |
809 | { | |
810 | struct outstate *os = ctx; | |
811 | ||
812 | os->newline = true; | |
813 | } | |
814 | ||
815 | static void do_new_line_std(struct outstate *os) | |
816 | { | |
817 | fputc('\n', os->fh); | |
f9483392 | 818 | fputs(os->prefix, os->fh); |
44d49a60 | 819 | aggr_printout(os->evsel, os->id, os->nr); |
140aeadc AK |
820 | if (stat_config.aggr_mode == AGGR_NONE) |
821 | fprintf(os->fh, " "); | |
140aeadc AK |
822 | fprintf(os->fh, " "); |
823 | } | |
824 | ||
825 | static void print_metric_std(void *ctx, const char *color, const char *fmt, | |
826 | const char *unit, double val) | |
827 | { | |
828 | struct outstate *os = ctx; | |
829 | FILE *out = os->fh; | |
830 | int n; | |
831 | bool newline = os->newline; | |
832 | ||
833 | os->newline = false; | |
834 | ||
835 | if (unit == NULL || fmt == NULL) { | |
836 | fprintf(out, "%-*s", METRIC_LEN, ""); | |
837 | return; | |
838 | } | |
839 | ||
840 | if (newline) | |
841 | do_new_line_std(os); | |
842 | ||
843 | n = fprintf(out, " # "); | |
844 | if (color) | |
845 | n += color_fprintf(out, color, fmt, val); | |
846 | else | |
847 | n += fprintf(out, fmt, val); | |
848 | fprintf(out, " %-*s", METRIC_LEN - n - 1, unit); | |
849 | } | |
850 | ||
92a61f64 AK |
851 | static void new_line_csv(void *ctx) |
852 | { | |
853 | struct outstate *os = ctx; | |
854 | int i; | |
855 | ||
856 | fputc('\n', os->fh); | |
857 | if (os->prefix) | |
858 | fprintf(os->fh, "%s%s", os->prefix, csv_sep); | |
44d49a60 | 859 | aggr_printout(os->evsel, os->id, os->nr); |
92a61f64 AK |
860 | for (i = 0; i < os->nfields; i++) |
861 | fputs(csv_sep, os->fh); | |
862 | } | |
863 | ||
864 | static void print_metric_csv(void *ctx, | |
865 | const char *color __maybe_unused, | |
866 | const char *fmt, const char *unit, double val) | |
867 | { | |
868 | struct outstate *os = ctx; | |
869 | FILE *out = os->fh; | |
870 | char buf[64], *vals, *ends; | |
871 | ||
872 | if (unit == NULL || fmt == NULL) { | |
873 | fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep); | |
874 | return; | |
875 | } | |
876 | snprintf(buf, sizeof(buf), fmt, val); | |
877 | vals = buf; | |
878 | while (isspace(*vals)) | |
879 | vals++; | |
880 | ends = vals; | |
881 | while (isdigit(*ends) || *ends == '.') | |
882 | ends++; | |
883 | *ends = 0; | |
884 | while (isspace(*unit)) | |
885 | unit++; | |
886 | fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit); | |
887 | } | |
888 | ||
54b50916 AK |
889 | #define METRIC_ONLY_LEN 20 |
890 | ||
891 | /* Filter out some columns that don't work well in metrics only mode */ | |
892 | ||
893 | static bool valid_only_metric(const char *unit) | |
894 | { | |
895 | if (!unit) | |
896 | return false; | |
897 | if (strstr(unit, "/sec") || | |
898 | strstr(unit, "hz") || | |
899 | strstr(unit, "Hz") || | |
900 | strstr(unit, "CPUs utilized")) | |
901 | return false; | |
902 | return true; | |
903 | } | |
904 | ||
905 | static const char *fixunit(char *buf, struct perf_evsel *evsel, | |
906 | const char *unit) | |
907 | { | |
908 | if (!strncmp(unit, "of all", 6)) { | |
909 | snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel), | |
910 | unit); | |
911 | return buf; | |
912 | } | |
913 | return unit; | |
914 | } | |
915 | ||
916 | static void print_metric_only(void *ctx, const char *color, const char *fmt, | |
917 | const char *unit, double val) | |
918 | { | |
919 | struct outstate *os = ctx; | |
920 | FILE *out = os->fh; | |
921 | int n; | |
922 | char buf[1024]; | |
923 | unsigned mlen = METRIC_ONLY_LEN; | |
924 | ||
925 | if (!valid_only_metric(unit)) | |
926 | return; | |
927 | unit = fixunit(buf, os->evsel, unit); | |
928 | if (color) | |
929 | n = color_fprintf(out, color, fmt, val); | |
930 | else | |
931 | n = fprintf(out, fmt, val); | |
932 | if (n > METRIC_ONLY_LEN) | |
933 | n = METRIC_ONLY_LEN; | |
934 | if (mlen < strlen(unit)) | |
935 | mlen = strlen(unit) + 1; | |
936 | fprintf(out, "%*s", mlen - n, ""); | |
937 | } | |
938 | ||
939 | static void print_metric_only_csv(void *ctx, const char *color __maybe_unused, | |
940 | const char *fmt, | |
941 | const char *unit, double val) | |
942 | { | |
943 | struct outstate *os = ctx; | |
944 | FILE *out = os->fh; | |
945 | char buf[64], *vals, *ends; | |
946 | char tbuf[1024]; | |
947 | ||
948 | if (!valid_only_metric(unit)) | |
949 | return; | |
950 | unit = fixunit(tbuf, os->evsel, unit); | |
951 | snprintf(buf, sizeof buf, fmt, val); | |
952 | vals = buf; | |
953 | while (isspace(*vals)) | |
954 | vals++; | |
955 | ends = vals; | |
956 | while (isdigit(*ends) || *ends == '.') | |
957 | ends++; | |
958 | *ends = 0; | |
959 | fprintf(out, "%s%s", vals, csv_sep); | |
960 | } | |
961 | ||
962 | static void new_line_metric(void *ctx __maybe_unused) | |
963 | { | |
964 | } | |
965 | ||
966 | static void print_metric_header(void *ctx, const char *color __maybe_unused, | |
967 | const char *fmt __maybe_unused, | |
968 | const char *unit, double val __maybe_unused) | |
969 | { | |
970 | struct outstate *os = ctx; | |
971 | char tbuf[1024]; | |
972 | ||
973 | if (!valid_only_metric(unit)) | |
974 | return; | |
975 | unit = fixunit(tbuf, os->evsel, unit); | |
976 | if (csv_output) | |
977 | fprintf(os->fh, "%s%s", unit, csv_sep); | |
978 | else | |
979 | fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit); | |
980 | } | |
981 | ||
da88c7f7 | 982 | static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg) |
86ee6e18 | 983 | { |
5821522e | 984 | FILE *output = stat_config.output; |
310ebb93 | 985 | double msecs = avg / NSEC_PER_MSEC; |
410136f5 | 986 | const char *fmt_v, *fmt_n; |
4bbe5a61 | 987 | char name[25]; |
86ee6e18 | 988 | |
410136f5 SE |
989 | fmt_v = csv_output ? "%.6f%s" : "%18.6f%s"; |
990 | fmt_n = csv_output ? "%s" : "%-25s"; | |
991 | ||
da88c7f7 | 992 | aggr_printout(evsel, id, nr); |
d7470b6a | 993 | |
4bbe5a61 DA |
994 | scnprintf(name, sizeof(name), "%s%s", |
995 | perf_evsel__name(evsel), csv_output ? "" : " (msec)"); | |
410136f5 SE |
996 | |
997 | fprintf(output, fmt_v, msecs, csv_sep); | |
998 | ||
999 | if (csv_output) | |
1000 | fprintf(output, "%s%s", evsel->unit, csv_sep); | |
1001 | else | |
1002 | fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep); | |
1003 | ||
1004 | fprintf(output, fmt_n, name); | |
d7470b6a | 1005 | |
023695d9 | 1006 | if (evsel->cgrp) |
4aa9015f | 1007 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); |
44175b6f IM |
1008 | } |
1009 | ||
44d49a60 AK |
1010 | static int first_shadow_cpu(struct perf_evsel *evsel, int id) |
1011 | { | |
1012 | int i; | |
1013 | ||
1014 | if (!aggr_get_id) | |
1015 | return 0; | |
1016 | ||
1017 | if (stat_config.aggr_mode == AGGR_NONE) | |
1018 | return id; | |
1019 | ||
1020 | if (stat_config.aggr_mode == AGGR_GLOBAL) | |
1021 | return 0; | |
1022 | ||
1023 | for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) { | |
1024 | int cpu2 = perf_evsel__cpus(evsel)->map[i]; | |
1025 | ||
1026 | if (aggr_get_id(evsel_list->cpus, cpu2) == id) | |
1027 | return cpu2; | |
1028 | } | |
1029 | return 0; | |
1030 | } | |
1031 | ||
556b1fb7 JO |
1032 | static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg) |
1033 | { | |
5821522e | 1034 | FILE *output = stat_config.output; |
556b1fb7 JO |
1035 | double sc = evsel->scale; |
1036 | const char *fmt; | |
556b1fb7 JO |
1037 | |
1038 | if (csv_output) { | |
e3b03b6c | 1039 | fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s"; |
556b1fb7 JO |
1040 | } else { |
1041 | if (big_num) | |
e3b03b6c | 1042 | fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s"; |
556b1fb7 | 1043 | else |
e3b03b6c | 1044 | fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s"; |
556b1fb7 JO |
1045 | } |
1046 | ||
1047 | aggr_printout(evsel, id, nr); | |
1048 | ||
556b1fb7 JO |
1049 | fprintf(output, fmt, avg, csv_sep); |
1050 | ||
1051 | if (evsel->unit) | |
1052 | fprintf(output, "%-*s%s", | |
1053 | csv_output ? 0 : unit_width, | |
1054 | evsel->unit, csv_sep); | |
1055 | ||
1056 | fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel)); | |
1057 | ||
1058 | if (evsel->cgrp) | |
1059 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); | |
eedfcb4b | 1060 | } |
556b1fb7 | 1061 | |
f9483392 | 1062 | static void printout(int id, int nr, struct perf_evsel *counter, double uval, |
cb110f47 | 1063 | char *prefix, u64 run, u64 ena, double noise) |
eedfcb4b | 1064 | { |
140aeadc | 1065 | struct perf_stat_output_ctx out; |
f9483392 AK |
1066 | struct outstate os = { |
1067 | .fh = stat_config.output, | |
44d49a60 AK |
1068 | .prefix = prefix ? prefix : "", |
1069 | .id = id, | |
1070 | .nr = nr, | |
1071 | .evsel = counter, | |
f9483392 | 1072 | }; |
140aeadc AK |
1073 | print_metric_t pm = print_metric_std; |
1074 | void (*nl)(void *); | |
eedfcb4b | 1075 | |
54b50916 AK |
1076 | if (metric_only) { |
1077 | nl = new_line_metric; | |
1078 | if (csv_output) | |
1079 | pm = print_metric_only_csv; | |
1080 | else | |
1081 | pm = print_metric_only; | |
1082 | } else | |
1083 | nl = new_line_std; | |
eedfcb4b | 1084 | |
54b50916 | 1085 | if (csv_output && !metric_only) { |
92a61f64 AK |
1086 | static int aggr_fields[] = { |
1087 | [AGGR_GLOBAL] = 0, | |
1088 | [AGGR_THREAD] = 1, | |
1089 | [AGGR_NONE] = 1, | |
1090 | [AGGR_SOCKET] = 2, | |
1091 | [AGGR_CORE] = 2, | |
1092 | }; | |
1093 | ||
1094 | pm = print_metric_csv; | |
1095 | nl = new_line_csv; | |
1096 | os.nfields = 3; | |
1097 | os.nfields += aggr_fields[stat_config.aggr_mode]; | |
1098 | if (counter->cgrp) | |
1099 | os.nfields++; | |
1100 | } | |
b002f3bb | 1101 | if (run == 0 || ena == 0 || counter->counts->scaled == -1) { |
54b50916 AK |
1102 | if (metric_only) { |
1103 | pm(&os, NULL, "", "", 0); | |
1104 | return; | |
1105 | } | |
cb110f47 AK |
1106 | aggr_printout(counter, id, nr); |
1107 | ||
1108 | fprintf(stat_config.output, "%*s%s", | |
1109 | csv_output ? 0 : 18, | |
1110 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, | |
1111 | csv_sep); | |
1112 | ||
02d492e5 BP |
1113 | if (counter->supported) |
1114 | print_free_counters_hint = 1; | |
1115 | ||
cb110f47 AK |
1116 | fprintf(stat_config.output, "%-*s%s", |
1117 | csv_output ? 0 : unit_width, | |
1118 | counter->unit, csv_sep); | |
1119 | ||
1120 | fprintf(stat_config.output, "%*s", | |
1121 | csv_output ? 0 : -25, | |
1122 | perf_evsel__name(counter)); | |
1123 | ||
1124 | if (counter->cgrp) | |
1125 | fprintf(stat_config.output, "%s%s", | |
1126 | csv_sep, counter->cgrp->name); | |
1127 | ||
92a61f64 AK |
1128 | if (!csv_output) |
1129 | pm(&os, NULL, NULL, "", 0); | |
1130 | print_noise(counter, noise); | |
cb110f47 | 1131 | print_running(run, ena); |
92a61f64 AK |
1132 | if (csv_output) |
1133 | pm(&os, NULL, NULL, "", 0); | |
cb110f47 AK |
1134 | return; |
1135 | } | |
1136 | ||
54b50916 AK |
1137 | if (metric_only) |
1138 | /* nothing */; | |
1139 | else if (nsec_counter(counter)) | |
eedfcb4b AK |
1140 | nsec_printout(id, nr, counter, uval); |
1141 | else | |
1142 | abs_printout(id, nr, counter, uval); | |
556b1fb7 | 1143 | |
140aeadc AK |
1144 | out.print_metric = pm; |
1145 | out.new_line = nl; | |
1146 | out.ctx = &os; | |
1147 | ||
54b50916 | 1148 | if (csv_output && !metric_only) { |
92a61f64 AK |
1149 | print_noise(counter, noise); |
1150 | print_running(run, ena); | |
1151 | } | |
1152 | ||
1153 | perf_stat__print_shadow_stats(counter, uval, | |
44d49a60 | 1154 | first_shadow_cpu(counter, id), |
140aeadc | 1155 | &out); |
54b50916 | 1156 | if (!csv_output && !metric_only) { |
92a61f64 AK |
1157 | print_noise(counter, noise); |
1158 | print_running(run, ena); | |
1159 | } | |
556b1fb7 JO |
1160 | } |
1161 | ||
44d49a60 AK |
1162 | static void aggr_update_shadow(void) |
1163 | { | |
1164 | int cpu, s2, id, s; | |
1165 | u64 val; | |
1166 | struct perf_evsel *counter; | |
1167 | ||
1168 | for (s = 0; s < aggr_map->nr; s++) { | |
1169 | id = aggr_map->map[s]; | |
e5cadb93 | 1170 | evlist__for_each_entry(evsel_list, counter) { |
44d49a60 AK |
1171 | val = 0; |
1172 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { | |
1173 | s2 = aggr_get_id(evsel_list->cpus, cpu); | |
1174 | if (s2 != id) | |
1175 | continue; | |
1176 | val += perf_counts(counter->counts, cpu, 0)->val; | |
1177 | } | |
1178 | val = val * counter->scale; | |
1179 | perf_stat__update_shadow_stats(counter, &val, | |
1180 | first_shadow_cpu(counter, id)); | |
1181 | } | |
1182 | } | |
1183 | } | |
1184 | ||
86ee6e18 | 1185 | static void print_aggr(char *prefix) |
d7e7a451 | 1186 | { |
5821522e | 1187 | FILE *output = stat_config.output; |
d7e7a451 | 1188 | struct perf_evsel *counter; |
601083cf | 1189 | int cpu, s, s2, id, nr; |
410136f5 | 1190 | double uval; |
d7e7a451 | 1191 | u64 ena, run, val; |
54b50916 | 1192 | bool first; |
d7e7a451 | 1193 | |
86ee6e18 | 1194 | if (!(aggr_map || aggr_get_id)) |
d7e7a451 SE |
1195 | return; |
1196 | ||
44d49a60 AK |
1197 | aggr_update_shadow(); |
1198 | ||
54b50916 AK |
1199 | /* |
1200 | * With metric_only everything is on a single line. | |
1201 | * Without each counter has its own line. | |
1202 | */ | |
86ee6e18 | 1203 | for (s = 0; s < aggr_map->nr; s++) { |
54b50916 AK |
1204 | if (prefix && metric_only) |
1205 | fprintf(output, "%s", prefix); | |
1206 | ||
86ee6e18 | 1207 | id = aggr_map->map[s]; |
54b50916 | 1208 | first = true; |
e5cadb93 | 1209 | evlist__for_each_entry(evsel_list, counter) { |
d7e7a451 SE |
1210 | val = ena = run = 0; |
1211 | nr = 0; | |
1212 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { | |
601083cf | 1213 | s2 = aggr_get_id(perf_evsel__cpus(counter), cpu); |
86ee6e18 | 1214 | if (s2 != id) |
d7e7a451 | 1215 | continue; |
a6fa0038 JO |
1216 | val += perf_counts(counter->counts, cpu, 0)->val; |
1217 | ena += perf_counts(counter->counts, cpu, 0)->ena; | |
1218 | run += perf_counts(counter->counts, cpu, 0)->run; | |
d7e7a451 SE |
1219 | nr++; |
1220 | } | |
54b50916 AK |
1221 | if (first && metric_only) { |
1222 | first = false; | |
1223 | aggr_printout(counter, id, nr); | |
1224 | } | |
1225 | if (prefix && !metric_only) | |
d7e7a451 SE |
1226 | fprintf(output, "%s", prefix); |
1227 | ||
410136f5 | 1228 | uval = val * counter->scale; |
cb110f47 | 1229 | printout(id, nr, counter, uval, prefix, run, ena, 1.0); |
54b50916 AK |
1230 | if (!metric_only) |
1231 | fputc('\n', output); | |
d7e7a451 | 1232 | } |
54b50916 AK |
1233 | if (metric_only) |
1234 | fputc('\n', output); | |
d7e7a451 SE |
1235 | } |
1236 | } | |
1237 | ||
32b8af82 JO |
1238 | static void print_aggr_thread(struct perf_evsel *counter, char *prefix) |
1239 | { | |
5821522e | 1240 | FILE *output = stat_config.output; |
32b8af82 JO |
1241 | int nthreads = thread_map__nr(counter->threads); |
1242 | int ncpus = cpu_map__nr(counter->cpus); | |
1243 | int cpu, thread; | |
1244 | double uval; | |
1245 | ||
1246 | for (thread = 0; thread < nthreads; thread++) { | |
1247 | u64 ena = 0, run = 0, val = 0; | |
1248 | ||
1249 | for (cpu = 0; cpu < ncpus; cpu++) { | |
1250 | val += perf_counts(counter->counts, cpu, thread)->val; | |
1251 | ena += perf_counts(counter->counts, cpu, thread)->ena; | |
1252 | run += perf_counts(counter->counts, cpu, thread)->run; | |
1253 | } | |
1254 | ||
1255 | if (prefix) | |
1256 | fprintf(output, "%s", prefix); | |
1257 | ||
1258 | uval = val * counter->scale; | |
cb110f47 | 1259 | printout(thread, 0, counter, uval, prefix, run, ena, 1.0); |
32b8af82 JO |
1260 | fputc('\n', output); |
1261 | } | |
1262 | } | |
1263 | ||
2996f5dd IM |
1264 | /* |
1265 | * Print out the results of a single counter: | |
f5b4a9c3 | 1266 | * aggregated counts in system-wide mode |
2996f5dd | 1267 | */ |
13370a9b | 1268 | static void print_counter_aggr(struct perf_evsel *counter, char *prefix) |
2996f5dd | 1269 | { |
5821522e | 1270 | FILE *output = stat_config.output; |
581cc8a2 | 1271 | struct perf_stat_evsel *ps = counter->priv; |
69aad6f1 | 1272 | double avg = avg_stats(&ps->res_stats[0]); |
410136f5 | 1273 | double uval; |
d73515c0 AK |
1274 | double avg_enabled, avg_running; |
1275 | ||
1276 | avg_enabled = avg_stats(&ps->res_stats[1]); | |
1277 | avg_running = avg_stats(&ps->res_stats[2]); | |
2996f5dd | 1278 | |
54b50916 | 1279 | if (prefix && !metric_only) |
13370a9b SE |
1280 | fprintf(output, "%s", prefix); |
1281 | ||
410136f5 | 1282 | uval = avg * counter->scale; |
cb110f47 | 1283 | printout(-1, 0, counter, uval, prefix, avg_running, avg_enabled, avg); |
54b50916 AK |
1284 | if (!metric_only) |
1285 | fprintf(output, "\n"); | |
c04f5e5d IM |
1286 | } |
1287 | ||
f5b4a9c3 SE |
1288 | /* |
1289 | * Print out the results of a single counter: | |
1290 | * does not use aggregated count in system-wide | |
1291 | */ | |
13370a9b | 1292 | static void print_counter(struct perf_evsel *counter, char *prefix) |
f5b4a9c3 | 1293 | { |
5821522e | 1294 | FILE *output = stat_config.output; |
f5b4a9c3 | 1295 | u64 ena, run, val; |
410136f5 | 1296 | double uval; |
f5b4a9c3 SE |
1297 | int cpu; |
1298 | ||
7ae92e74 | 1299 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { |
a6fa0038 JO |
1300 | val = perf_counts(counter->counts, cpu, 0)->val; |
1301 | ena = perf_counts(counter->counts, cpu, 0)->ena; | |
1302 | run = perf_counts(counter->counts, cpu, 0)->run; | |
13370a9b SE |
1303 | |
1304 | if (prefix) | |
1305 | fprintf(output, "%s", prefix); | |
1306 | ||
410136f5 | 1307 | uval = val * counter->scale; |
cb110f47 | 1308 | printout(cpu, 0, counter, uval, prefix, run, ena, 1.0); |
f5b4a9c3 | 1309 | |
4aa9015f | 1310 | fputc('\n', output); |
f5b4a9c3 SE |
1311 | } |
1312 | } | |
1313 | ||
206cab65 AK |
1314 | static void print_no_aggr_metric(char *prefix) |
1315 | { | |
1316 | int cpu; | |
1317 | int nrcpus = 0; | |
1318 | struct perf_evsel *counter; | |
1319 | u64 ena, run, val; | |
1320 | double uval; | |
1321 | ||
1322 | nrcpus = evsel_list->cpus->nr; | |
1323 | for (cpu = 0; cpu < nrcpus; cpu++) { | |
1324 | bool first = true; | |
1325 | ||
1326 | if (prefix) | |
1327 | fputs(prefix, stat_config.output); | |
e5cadb93 | 1328 | evlist__for_each_entry(evsel_list, counter) { |
206cab65 AK |
1329 | if (first) { |
1330 | aggr_printout(counter, cpu, 0); | |
1331 | first = false; | |
1332 | } | |
1333 | val = perf_counts(counter->counts, cpu, 0)->val; | |
1334 | ena = perf_counts(counter->counts, cpu, 0)->ena; | |
1335 | run = perf_counts(counter->counts, cpu, 0)->run; | |
1336 | ||
1337 | uval = val * counter->scale; | |
1338 | printout(cpu, 0, counter, uval, prefix, run, ena, 1.0); | |
1339 | } | |
1340 | fputc('\n', stat_config.output); | |
1341 | } | |
1342 | } | |
1343 | ||
54b50916 AK |
1344 | static int aggr_header_lens[] = { |
1345 | [AGGR_CORE] = 18, | |
1346 | [AGGR_SOCKET] = 12, | |
206cab65 | 1347 | [AGGR_NONE] = 6, |
54b50916 AK |
1348 | [AGGR_THREAD] = 24, |
1349 | [AGGR_GLOBAL] = 0, | |
1350 | }; | |
1351 | ||
c51fd639 AK |
1352 | static const char *aggr_header_csv[] = { |
1353 | [AGGR_CORE] = "core,cpus,", | |
1354 | [AGGR_SOCKET] = "socket,cpus", | |
1355 | [AGGR_NONE] = "cpu,", | |
1356 | [AGGR_THREAD] = "comm-pid,", | |
1357 | [AGGR_GLOBAL] = "" | |
1358 | }; | |
1359 | ||
41c8ca2a | 1360 | static void print_metric_headers(const char *prefix, bool no_indent) |
54b50916 AK |
1361 | { |
1362 | struct perf_stat_output_ctx out; | |
1363 | struct perf_evsel *counter; | |
1364 | struct outstate os = { | |
1365 | .fh = stat_config.output | |
1366 | }; | |
1367 | ||
1368 | if (prefix) | |
1369 | fprintf(stat_config.output, "%s", prefix); | |
1370 | ||
41c8ca2a | 1371 | if (!csv_output && !no_indent) |
54b50916 AK |
1372 | fprintf(stat_config.output, "%*s", |
1373 | aggr_header_lens[stat_config.aggr_mode], ""); | |
c51fd639 AK |
1374 | if (csv_output) { |
1375 | if (stat_config.interval) | |
1376 | fputs("time,", stat_config.output); | |
1377 | fputs(aggr_header_csv[stat_config.aggr_mode], | |
1378 | stat_config.output); | |
1379 | } | |
54b50916 AK |
1380 | |
1381 | /* Print metrics headers only */ | |
e5cadb93 | 1382 | evlist__for_each_entry(evsel_list, counter) { |
54b50916 AK |
1383 | os.evsel = counter; |
1384 | out.ctx = &os; | |
1385 | out.print_metric = print_metric_header; | |
1386 | out.new_line = new_line_metric; | |
1387 | os.evsel = counter; | |
1388 | perf_stat__print_shadow_stats(counter, 0, | |
1389 | 0, | |
1390 | &out); | |
1391 | } | |
1392 | fputc('\n', stat_config.output); | |
1393 | } | |
1394 | ||
d4f63a47 JO |
1395 | static void print_interval(char *prefix, struct timespec *ts) |
1396 | { | |
5821522e | 1397 | FILE *output = stat_config.output; |
d4f63a47 JO |
1398 | static int num_print_interval; |
1399 | ||
1400 | sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep); | |
1401 | ||
41c8ca2a | 1402 | if (num_print_interval == 0 && !csv_output) { |
421a50f3 | 1403 | switch (stat_config.aggr_mode) { |
d4f63a47 | 1404 | case AGGR_SOCKET: |
41c8ca2a AK |
1405 | fprintf(output, "# time socket cpus"); |
1406 | if (!metric_only) | |
1407 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
d4f63a47 JO |
1408 | break; |
1409 | case AGGR_CORE: | |
41c8ca2a AK |
1410 | fprintf(output, "# time core cpus"); |
1411 | if (!metric_only) | |
1412 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
d4f63a47 JO |
1413 | break; |
1414 | case AGGR_NONE: | |
41c8ca2a AK |
1415 | fprintf(output, "# time CPU"); |
1416 | if (!metric_only) | |
1417 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
d4f63a47 | 1418 | break; |
32b8af82 | 1419 | case AGGR_THREAD: |
41c8ca2a AK |
1420 | fprintf(output, "# time comm-pid"); |
1421 | if (!metric_only) | |
1422 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
32b8af82 | 1423 | break; |
d4f63a47 JO |
1424 | case AGGR_GLOBAL: |
1425 | default: | |
41c8ca2a AK |
1426 | fprintf(output, "# time"); |
1427 | if (!metric_only) | |
1428 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
208df99e JO |
1429 | case AGGR_UNSET: |
1430 | break; | |
d4f63a47 JO |
1431 | } |
1432 | } | |
1433 | ||
41c8ca2a AK |
1434 | if (num_print_interval == 0 && metric_only) |
1435 | print_metric_headers(" ", true); | |
d4f63a47 JO |
1436 | if (++num_print_interval == 25) |
1437 | num_print_interval = 0; | |
1438 | } | |
1439 | ||
1440 | static void print_header(int argc, const char **argv) | |
42202dd5 | 1441 | { |
5821522e | 1442 | FILE *output = stat_config.output; |
69aad6f1 | 1443 | int i; |
42202dd5 | 1444 | |
ddcacfa0 IM |
1445 | fflush(stdout); |
1446 | ||
d7470b6a | 1447 | if (!csv_output) { |
4aa9015f SE |
1448 | fprintf(output, "\n"); |
1449 | fprintf(output, " Performance counter stats for "); | |
62d3b617 DA |
1450 | if (target.system_wide) |
1451 | fprintf(output, "\'system wide"); | |
1452 | else if (target.cpu_list) | |
1453 | fprintf(output, "\'CPU(s) %s", target.cpu_list); | |
602ad878 | 1454 | else if (!target__has_task(&target)) { |
ba6039b6 JO |
1455 | fprintf(output, "\'%s", argv ? argv[0] : "pipe"); |
1456 | for (i = 1; argv && (i < argc); i++) | |
4aa9015f | 1457 | fprintf(output, " %s", argv[i]); |
20f946b4 NK |
1458 | } else if (target.pid) |
1459 | fprintf(output, "process id \'%s", target.pid); | |
d7470b6a | 1460 | else |
20f946b4 | 1461 | fprintf(output, "thread id \'%s", target.tid); |
44db76c8 | 1462 | |
4aa9015f | 1463 | fprintf(output, "\'"); |
d7470b6a | 1464 | if (run_count > 1) |
4aa9015f SE |
1465 | fprintf(output, " (%d runs)", run_count); |
1466 | fprintf(output, ":\n\n"); | |
d7470b6a | 1467 | } |
d4f63a47 JO |
1468 | } |
1469 | ||
1470 | static void print_footer(void) | |
1471 | { | |
5821522e JO |
1472 | FILE *output = stat_config.output; |
1473 | ||
d4f63a47 JO |
1474 | if (!null_run) |
1475 | fprintf(output, "\n"); | |
1476 | fprintf(output, " %17.9f seconds time elapsed", | |
310ebb93 | 1477 | avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC); |
d4f63a47 JO |
1478 | if (run_count > 1) { |
1479 | fprintf(output, " "); | |
1480 | print_noise_pct(stddev_stats(&walltime_nsecs_stats), | |
1481 | avg_stats(&walltime_nsecs_stats)); | |
1482 | } | |
1483 | fprintf(output, "\n\n"); | |
02d492e5 BP |
1484 | |
1485 | if (print_free_counters_hint) | |
1486 | fprintf(output, | |
1487 | "Some events weren't counted. Try disabling the NMI watchdog:\n" | |
1488 | " echo 0 > /proc/sys/kernel/nmi_watchdog\n" | |
1489 | " perf stat ...\n" | |
1490 | " echo 1 > /proc/sys/kernel/nmi_watchdog\n"); | |
d4f63a47 JO |
1491 | } |
1492 | ||
1493 | static void print_counters(struct timespec *ts, int argc, const char **argv) | |
1494 | { | |
ec0d3d1f | 1495 | int interval = stat_config.interval; |
d4f63a47 JO |
1496 | struct perf_evsel *counter; |
1497 | char buf[64], *prefix = NULL; | |
1498 | ||
664c98d4 JO |
1499 | /* Do not print anything if we record to the pipe. */ |
1500 | if (STAT_RECORD && perf_stat.file.is_pipe) | |
1501 | return; | |
1502 | ||
d4f63a47 JO |
1503 | if (interval) |
1504 | print_interval(prefix = buf, ts); | |
1505 | else | |
1506 | print_header(argc, argv); | |
2996f5dd | 1507 | |
54b50916 AK |
1508 | if (metric_only) { |
1509 | static int num_print_iv; | |
1510 | ||
41c8ca2a AK |
1511 | if (num_print_iv == 0 && !interval) |
1512 | print_metric_headers(prefix, false); | |
54b50916 AK |
1513 | if (num_print_iv++ == 25) |
1514 | num_print_iv = 0; | |
1515 | if (stat_config.aggr_mode == AGGR_GLOBAL && prefix) | |
1516 | fprintf(stat_config.output, "%s", prefix); | |
1517 | } | |
1518 | ||
421a50f3 | 1519 | switch (stat_config.aggr_mode) { |
12c08a9f | 1520 | case AGGR_CORE: |
86ee6e18 | 1521 | case AGGR_SOCKET: |
d4f63a47 | 1522 | print_aggr(prefix); |
86ee6e18 | 1523 | break; |
32b8af82 | 1524 | case AGGR_THREAD: |
e5cadb93 | 1525 | evlist__for_each_entry(evsel_list, counter) |
32b8af82 JO |
1526 | print_aggr_thread(counter, prefix); |
1527 | break; | |
86ee6e18 | 1528 | case AGGR_GLOBAL: |
e5cadb93 | 1529 | evlist__for_each_entry(evsel_list, counter) |
d4f63a47 | 1530 | print_counter_aggr(counter, prefix); |
54b50916 AK |
1531 | if (metric_only) |
1532 | fputc('\n', stat_config.output); | |
86ee6e18 SE |
1533 | break; |
1534 | case AGGR_NONE: | |
206cab65 AK |
1535 | if (metric_only) |
1536 | print_no_aggr_metric(prefix); | |
1537 | else { | |
e5cadb93 | 1538 | evlist__for_each_entry(evsel_list, counter) |
206cab65 AK |
1539 | print_counter(counter, prefix); |
1540 | } | |
86ee6e18 | 1541 | break; |
208df99e | 1542 | case AGGR_UNSET: |
86ee6e18 SE |
1543 | default: |
1544 | break; | |
f5b4a9c3 | 1545 | } |
ddcacfa0 | 1546 | |
d4f63a47 JO |
1547 | if (!interval && !csv_output) |
1548 | print_footer(); | |
1549 | ||
5821522e | 1550 | fflush(stat_config.output); |
ddcacfa0 IM |
1551 | } |
1552 | ||
f7b7c26e PZ |
1553 | static volatile int signr = -1; |
1554 | ||
5242519b | 1555 | static void skip_signal(int signo) |
ddcacfa0 | 1556 | { |
ec0d3d1f | 1557 | if ((child_pid == -1) || stat_config.interval) |
60666c63 LW |
1558 | done = 1; |
1559 | ||
f7b7c26e | 1560 | signr = signo; |
d07f0b12 SE |
1561 | /* |
1562 | * render child_pid harmless | |
1563 | * won't send SIGTERM to a random | |
1564 | * process in case of race condition | |
1565 | * and fast PID recycling | |
1566 | */ | |
1567 | child_pid = -1; | |
f7b7c26e PZ |
1568 | } |
1569 | ||
1570 | static void sig_atexit(void) | |
1571 | { | |
d07f0b12 SE |
1572 | sigset_t set, oset; |
1573 | ||
1574 | /* | |
1575 | * avoid race condition with SIGCHLD handler | |
1576 | * in skip_signal() which is modifying child_pid | |
1577 | * goal is to avoid send SIGTERM to a random | |
1578 | * process | |
1579 | */ | |
1580 | sigemptyset(&set); | |
1581 | sigaddset(&set, SIGCHLD); | |
1582 | sigprocmask(SIG_BLOCK, &set, &oset); | |
1583 | ||
933da83a CW |
1584 | if (child_pid != -1) |
1585 | kill(child_pid, SIGTERM); | |
1586 | ||
d07f0b12 SE |
1587 | sigprocmask(SIG_SETMASK, &oset, NULL); |
1588 | ||
f7b7c26e PZ |
1589 | if (signr == -1) |
1590 | return; | |
1591 | ||
1592 | signal(signr, SIG_DFL); | |
1593 | kill(getpid(), signr); | |
5242519b IM |
1594 | } |
1595 | ||
1d037ca1 IT |
1596 | static int stat__set_big_num(const struct option *opt __maybe_unused, |
1597 | const char *s __maybe_unused, int unset) | |
d7470b6a SE |
1598 | { |
1599 | big_num_opt = unset ? 0 : 1; | |
1600 | return 0; | |
1601 | } | |
1602 | ||
44b1e60a AK |
1603 | static int enable_metric_only(const struct option *opt __maybe_unused, |
1604 | const char *s __maybe_unused, int unset) | |
1605 | { | |
1606 | force_metric_only = true; | |
1607 | metric_only = !unset; | |
1608 | return 0; | |
1609 | } | |
1610 | ||
e0547311 JO |
1611 | static const struct option stat_options[] = { |
1612 | OPT_BOOLEAN('T', "transaction", &transaction_run, | |
1613 | "hardware transaction statistics"), | |
1614 | OPT_CALLBACK('e', "event", &evsel_list, "event", | |
1615 | "event selector. use 'perf list' to list available events", | |
1616 | parse_events_option), | |
1617 | OPT_CALLBACK(0, "filter", &evsel_list, "filter", | |
1618 | "event filter", parse_filter), | |
1619 | OPT_BOOLEAN('i', "no-inherit", &no_inherit, | |
1620 | "child tasks do not inherit counters"), | |
1621 | OPT_STRING('p', "pid", &target.pid, "pid", | |
1622 | "stat events on existing process id"), | |
1623 | OPT_STRING('t', "tid", &target.tid, "tid", | |
1624 | "stat events on existing thread id"), | |
1625 | OPT_BOOLEAN('a', "all-cpus", &target.system_wide, | |
1626 | "system-wide collection from all CPUs"), | |
1627 | OPT_BOOLEAN('g', "group", &group, | |
1628 | "put the counters into a counter group"), | |
1629 | OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"), | |
1630 | OPT_INCR('v', "verbose", &verbose, | |
1631 | "be more verbose (show counter open errors, etc)"), | |
1632 | OPT_INTEGER('r', "repeat", &run_count, | |
1633 | "repeat command and print average + stddev (max: 100, forever: 0)"), | |
1634 | OPT_BOOLEAN('n', "null", &null_run, | |
1635 | "null run - dont start any counters"), | |
1636 | OPT_INCR('d', "detailed", &detailed_run, | |
1637 | "detailed run - start a lot of events"), | |
1638 | OPT_BOOLEAN('S', "sync", &sync_run, | |
1639 | "call sync() before starting a run"), | |
1640 | OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, | |
1641 | "print large numbers with thousands\' separators", | |
1642 | stat__set_big_num), | |
1643 | OPT_STRING('C', "cpu", &target.cpu_list, "cpu", | |
1644 | "list of cpus to monitor in system-wide"), | |
1645 | OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode, | |
1646 | "disable CPU count aggregation", AGGR_NONE), | |
1647 | OPT_STRING('x', "field-separator", &csv_sep, "separator", | |
1648 | "print counts with custom separator"), | |
1649 | OPT_CALLBACK('G', "cgroup", &evsel_list, "name", | |
1650 | "monitor event in cgroup name only", parse_cgroups), | |
1651 | OPT_STRING('o', "output", &output_name, "file", "output file name"), | |
1652 | OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), | |
1653 | OPT_INTEGER(0, "log-fd", &output_fd, | |
1654 | "log output to fd, instead of stderr"), | |
1655 | OPT_STRING(0, "pre", &pre_cmd, "command", | |
1656 | "command to run prior to the measured command"), | |
1657 | OPT_STRING(0, "post", &post_cmd, "command", | |
1658 | "command to run after to the measured command"), | |
1659 | OPT_UINTEGER('I', "interval-print", &stat_config.interval, | |
1660 | "print counts at regular interval in ms (>= 10)"), | |
1661 | OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode, | |
1662 | "aggregate counts per processor socket", AGGR_SOCKET), | |
1663 | OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode, | |
1664 | "aggregate counts per physical processor core", AGGR_CORE), | |
1665 | OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode, | |
1666 | "aggregate counts per thread", AGGR_THREAD), | |
1667 | OPT_UINTEGER('D', "delay", &initial_delay, | |
1668 | "ms to wait before starting measurement after program start"), | |
44b1e60a AK |
1669 | OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL, |
1670 | "Only print computed metrics. No raw values", enable_metric_only), | |
1671 | OPT_BOOLEAN(0, "topdown", &topdown_run, | |
1672 | "measure topdown level 1 statistics"), | |
e0547311 JO |
1673 | OPT_END() |
1674 | }; | |
1675 | ||
1fe7a300 JO |
1676 | static int perf_stat__get_socket(struct cpu_map *map, int cpu) |
1677 | { | |
1678 | return cpu_map__get_socket(map, cpu, NULL); | |
1679 | } | |
1680 | ||
1681 | static int perf_stat__get_core(struct cpu_map *map, int cpu) | |
1682 | { | |
1683 | return cpu_map__get_core(map, cpu, NULL); | |
1684 | } | |
1685 | ||
1e5a2931 JO |
1686 | static int cpu_map__get_max(struct cpu_map *map) |
1687 | { | |
1688 | int i, max = -1; | |
1689 | ||
1690 | for (i = 0; i < map->nr; i++) { | |
1691 | if (map->map[i] > max) | |
1692 | max = map->map[i]; | |
1693 | } | |
1694 | ||
1695 | return max; | |
1696 | } | |
1697 | ||
1698 | static struct cpu_map *cpus_aggr_map; | |
1699 | ||
1700 | static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx) | |
1701 | { | |
1702 | int cpu; | |
1703 | ||
1704 | if (idx >= map->nr) | |
1705 | return -1; | |
1706 | ||
1707 | cpu = map->map[idx]; | |
1708 | ||
1709 | if (cpus_aggr_map->map[cpu] == -1) | |
1710 | cpus_aggr_map->map[cpu] = get_id(map, idx); | |
1711 | ||
1712 | return cpus_aggr_map->map[cpu]; | |
1713 | } | |
1714 | ||
1715 | static int perf_stat__get_socket_cached(struct cpu_map *map, int idx) | |
1716 | { | |
1717 | return perf_stat__get_aggr(perf_stat__get_socket, map, idx); | |
1718 | } | |
1719 | ||
1720 | static int perf_stat__get_core_cached(struct cpu_map *map, int idx) | |
1721 | { | |
1722 | return perf_stat__get_aggr(perf_stat__get_core, map, idx); | |
1723 | } | |
1724 | ||
86ee6e18 SE |
1725 | static int perf_stat_init_aggr_mode(void) |
1726 | { | |
1e5a2931 JO |
1727 | int nr; |
1728 | ||
421a50f3 | 1729 | switch (stat_config.aggr_mode) { |
86ee6e18 SE |
1730 | case AGGR_SOCKET: |
1731 | if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) { | |
1732 | perror("cannot build socket map"); | |
1733 | return -1; | |
1734 | } | |
1e5a2931 | 1735 | aggr_get_id = perf_stat__get_socket_cached; |
86ee6e18 | 1736 | break; |
12c08a9f SE |
1737 | case AGGR_CORE: |
1738 | if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) { | |
1739 | perror("cannot build core map"); | |
1740 | return -1; | |
1741 | } | |
1e5a2931 | 1742 | aggr_get_id = perf_stat__get_core_cached; |
12c08a9f | 1743 | break; |
86ee6e18 SE |
1744 | case AGGR_NONE: |
1745 | case AGGR_GLOBAL: | |
32b8af82 | 1746 | case AGGR_THREAD: |
208df99e | 1747 | case AGGR_UNSET: |
86ee6e18 SE |
1748 | default: |
1749 | break; | |
1750 | } | |
1e5a2931 JO |
1751 | |
1752 | /* | |
1753 | * The evsel_list->cpus is the base we operate on, | |
1754 | * taking the highest cpu number to be the size of | |
1755 | * the aggregation translate cpumap. | |
1756 | */ | |
1757 | nr = cpu_map__get_max(evsel_list->cpus); | |
1758 | cpus_aggr_map = cpu_map__empty_new(nr + 1); | |
1759 | return cpus_aggr_map ? 0 : -ENOMEM; | |
86ee6e18 SE |
1760 | } |
1761 | ||
544c2ae7 MH |
1762 | static void perf_stat__exit_aggr_mode(void) |
1763 | { | |
1764 | cpu_map__put(aggr_map); | |
1765 | cpu_map__put(cpus_aggr_map); | |
1766 | aggr_map = NULL; | |
1767 | cpus_aggr_map = NULL; | |
1768 | } | |
1769 | ||
68d702f7 JO |
1770 | static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx) |
1771 | { | |
1772 | int cpu; | |
1773 | ||
1774 | if (idx > map->nr) | |
1775 | return -1; | |
1776 | ||
1777 | cpu = map->map[idx]; | |
1778 | ||
da8a58b5 | 1779 | if (cpu >= env->nr_cpus_avail) |
68d702f7 JO |
1780 | return -1; |
1781 | ||
1782 | return cpu; | |
1783 | } | |
1784 | ||
1785 | static int perf_env__get_socket(struct cpu_map *map, int idx, void *data) | |
1786 | { | |
1787 | struct perf_env *env = data; | |
1788 | int cpu = perf_env__get_cpu(env, map, idx); | |
1789 | ||
1790 | return cpu == -1 ? -1 : env->cpu[cpu].socket_id; | |
1791 | } | |
1792 | ||
1793 | static int perf_env__get_core(struct cpu_map *map, int idx, void *data) | |
1794 | { | |
1795 | struct perf_env *env = data; | |
1796 | int core = -1, cpu = perf_env__get_cpu(env, map, idx); | |
1797 | ||
1798 | if (cpu != -1) { | |
1799 | int socket_id = env->cpu[cpu].socket_id; | |
1800 | ||
1801 | /* | |
1802 | * Encode socket in upper 16 bits | |
1803 | * core_id is relative to socket, and | |
1804 | * we need a global id. So we combine | |
1805 | * socket + core id. | |
1806 | */ | |
1807 | core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff); | |
1808 | } | |
1809 | ||
1810 | return core; | |
1811 | } | |
1812 | ||
1813 | static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus, | |
1814 | struct cpu_map **sockp) | |
1815 | { | |
1816 | return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env); | |
1817 | } | |
1818 | ||
1819 | static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus, | |
1820 | struct cpu_map **corep) | |
1821 | { | |
1822 | return cpu_map__build_map(cpus, corep, perf_env__get_core, env); | |
1823 | } | |
1824 | ||
1825 | static int perf_stat__get_socket_file(struct cpu_map *map, int idx) | |
1826 | { | |
1827 | return perf_env__get_socket(map, idx, &perf_stat.session->header.env); | |
1828 | } | |
1829 | ||
1830 | static int perf_stat__get_core_file(struct cpu_map *map, int idx) | |
1831 | { | |
1832 | return perf_env__get_core(map, idx, &perf_stat.session->header.env); | |
1833 | } | |
1834 | ||
1835 | static int perf_stat_init_aggr_mode_file(struct perf_stat *st) | |
1836 | { | |
1837 | struct perf_env *env = &st->session->header.env; | |
1838 | ||
1839 | switch (stat_config.aggr_mode) { | |
1840 | case AGGR_SOCKET: | |
1841 | if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) { | |
1842 | perror("cannot build socket map"); | |
1843 | return -1; | |
1844 | } | |
1845 | aggr_get_id = perf_stat__get_socket_file; | |
1846 | break; | |
1847 | case AGGR_CORE: | |
1848 | if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) { | |
1849 | perror("cannot build core map"); | |
1850 | return -1; | |
1851 | } | |
1852 | aggr_get_id = perf_stat__get_core_file; | |
1853 | break; | |
1854 | case AGGR_NONE: | |
1855 | case AGGR_GLOBAL: | |
1856 | case AGGR_THREAD: | |
1857 | case AGGR_UNSET: | |
1858 | default: | |
1859 | break; | |
1860 | } | |
1861 | ||
1862 | return 0; | |
1863 | } | |
1864 | ||
44b1e60a AK |
1865 | static int topdown_filter_events(const char **attr, char **str, bool use_group) |
1866 | { | |
1867 | int off = 0; | |
1868 | int i; | |
1869 | int len = 0; | |
1870 | char *s; | |
1871 | ||
1872 | for (i = 0; attr[i]; i++) { | |
1873 | if (pmu_have_event("cpu", attr[i])) { | |
1874 | len += strlen(attr[i]) + 1; | |
1875 | attr[i - off] = attr[i]; | |
1876 | } else | |
1877 | off++; | |
1878 | } | |
1879 | attr[i - off] = NULL; | |
1880 | ||
1881 | *str = malloc(len + 1 + 2); | |
1882 | if (!*str) | |
1883 | return -1; | |
1884 | s = *str; | |
1885 | if (i - off == 0) { | |
1886 | *s = 0; | |
1887 | return 0; | |
1888 | } | |
1889 | if (use_group) | |
1890 | *s++ = '{'; | |
1891 | for (i = 0; attr[i]; i++) { | |
1892 | strcpy(s, attr[i]); | |
1893 | s += strlen(s); | |
1894 | *s++ = ','; | |
1895 | } | |
1896 | if (use_group) { | |
1897 | s[-1] = '}'; | |
1898 | *s = 0; | |
1899 | } else | |
1900 | s[-1] = 0; | |
1901 | return 0; | |
1902 | } | |
1903 | ||
1904 | __weak bool arch_topdown_check_group(bool *warn) | |
1905 | { | |
1906 | *warn = false; | |
1907 | return false; | |
1908 | } | |
1909 | ||
1910 | __weak void arch_topdown_group_warn(void) | |
1911 | { | |
1912 | } | |
1913 | ||
2cba3ffb IM |
1914 | /* |
1915 | * Add default attributes, if there were no attributes specified or | |
1916 | * if -d/--detailed, -d -d or -d -d -d is used: | |
1917 | */ | |
1918 | static int add_default_attributes(void) | |
1919 | { | |
44b1e60a | 1920 | int err; |
9dec4473 | 1921 | struct perf_event_attr default_attrs0[] = { |
b070a547 ACM |
1922 | |
1923 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, | |
1924 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, | |
1925 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, | |
1926 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, | |
1927 | ||
1928 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, | |
9dec4473 AK |
1929 | }; |
1930 | struct perf_event_attr frontend_attrs[] = { | |
b070a547 | 1931 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, |
9dec4473 AK |
1932 | }; |
1933 | struct perf_event_attr backend_attrs[] = { | |
b070a547 | 1934 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, |
9dec4473 AK |
1935 | }; |
1936 | struct perf_event_attr default_attrs1[] = { | |
b070a547 ACM |
1937 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, |
1938 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, | |
1939 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, | |
1940 | ||
1941 | }; | |
1942 | ||
1943 | /* | |
1944 | * Detailed stats (-d), covering the L1 and last level data caches: | |
1945 | */ | |
1946 | struct perf_event_attr detailed_attrs[] = { | |
1947 | ||
1948 | { .type = PERF_TYPE_HW_CACHE, | |
1949 | .config = | |
1950 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1951 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1952 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1953 | ||
1954 | { .type = PERF_TYPE_HW_CACHE, | |
1955 | .config = | |
1956 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1957 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1958 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1959 | ||
1960 | { .type = PERF_TYPE_HW_CACHE, | |
1961 | .config = | |
1962 | PERF_COUNT_HW_CACHE_LL << 0 | | |
1963 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1964 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1965 | ||
1966 | { .type = PERF_TYPE_HW_CACHE, | |
1967 | .config = | |
1968 | PERF_COUNT_HW_CACHE_LL << 0 | | |
1969 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1970 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1971 | }; | |
1972 | ||
1973 | /* | |
1974 | * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: | |
1975 | */ | |
1976 | struct perf_event_attr very_detailed_attrs[] = { | |
1977 | ||
1978 | { .type = PERF_TYPE_HW_CACHE, | |
1979 | .config = | |
1980 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
1981 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1982 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1983 | ||
1984 | { .type = PERF_TYPE_HW_CACHE, | |
1985 | .config = | |
1986 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
1987 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1988 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1989 | ||
1990 | { .type = PERF_TYPE_HW_CACHE, | |
1991 | .config = | |
1992 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1993 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1994 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1995 | ||
1996 | { .type = PERF_TYPE_HW_CACHE, | |
1997 | .config = | |
1998 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1999 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2000 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2001 | ||
2002 | { .type = PERF_TYPE_HW_CACHE, | |
2003 | .config = | |
2004 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
2005 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2006 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2007 | ||
2008 | { .type = PERF_TYPE_HW_CACHE, | |
2009 | .config = | |
2010 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
2011 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2012 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2013 | ||
2014 | }; | |
2015 | ||
2016 | /* | |
2017 | * Very, very detailed stats (-d -d -d), adding prefetch events: | |
2018 | */ | |
2019 | struct perf_event_attr very_very_detailed_attrs[] = { | |
2020 | ||
2021 | { .type = PERF_TYPE_HW_CACHE, | |
2022 | .config = | |
2023 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
2024 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
2025 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2026 | ||
2027 | { .type = PERF_TYPE_HW_CACHE, | |
2028 | .config = | |
2029 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
2030 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
2031 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2032 | }; | |
2033 | ||
2cba3ffb IM |
2034 | /* Set attrs if no event is selected and !null_run: */ |
2035 | if (null_run) | |
2036 | return 0; | |
2037 | ||
4cabc3d1 | 2038 | if (transaction_run) { |
4cabc3d1 AK |
2039 | if (pmu_have_event("cpu", "cycles-ct") && |
2040 | pmu_have_event("cpu", "el-start")) | |
a454742c | 2041 | err = parse_events(evsel_list, transaction_attrs, NULL); |
4cabc3d1 | 2042 | else |
a454742c JO |
2043 | err = parse_events(evsel_list, transaction_limited_attrs, NULL); |
2044 | if (err) { | |
4cabc3d1 AK |
2045 | fprintf(stderr, "Cannot set up transaction events\n"); |
2046 | return -1; | |
2047 | } | |
2048 | return 0; | |
2049 | } | |
2050 | ||
44b1e60a AK |
2051 | if (topdown_run) { |
2052 | char *str = NULL; | |
2053 | bool warn = false; | |
2054 | ||
2055 | if (stat_config.aggr_mode != AGGR_GLOBAL && | |
2056 | stat_config.aggr_mode != AGGR_CORE) { | |
2057 | pr_err("top down event configuration requires --per-core mode\n"); | |
2058 | return -1; | |
2059 | } | |
2060 | stat_config.aggr_mode = AGGR_CORE; | |
2061 | if (nr_cgroups || !target__has_cpu(&target)) { | |
2062 | pr_err("top down event configuration requires system-wide mode (-a)\n"); | |
2063 | return -1; | |
2064 | } | |
2065 | ||
2066 | if (!force_metric_only) | |
2067 | metric_only = true; | |
2068 | if (topdown_filter_events(topdown_attrs, &str, | |
2069 | arch_topdown_check_group(&warn)) < 0) { | |
2070 | pr_err("Out of memory\n"); | |
2071 | return -1; | |
2072 | } | |
2073 | if (topdown_attrs[0] && str) { | |
2074 | if (warn) | |
2075 | arch_topdown_group_warn(); | |
2076 | err = parse_events(evsel_list, str, NULL); | |
2077 | if (err) { | |
2078 | fprintf(stderr, | |
2079 | "Cannot set up top down events %s: %d\n", | |
2080 | str, err); | |
2081 | free(str); | |
2082 | return -1; | |
2083 | } | |
2084 | } else { | |
2085 | fprintf(stderr, "System does not support topdown\n"); | |
2086 | return -1; | |
2087 | } | |
2088 | free(str); | |
2089 | } | |
2090 | ||
2cba3ffb | 2091 | if (!evsel_list->nr_entries) { |
a1f3d567 NK |
2092 | if (target__has_cpu(&target)) |
2093 | default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK; | |
2094 | ||
9dec4473 AK |
2095 | if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0) |
2096 | return -1; | |
2097 | if (pmu_have_event("cpu", "stalled-cycles-frontend")) { | |
2098 | if (perf_evlist__add_default_attrs(evsel_list, | |
2099 | frontend_attrs) < 0) | |
2100 | return -1; | |
2101 | } | |
2102 | if (pmu_have_event("cpu", "stalled-cycles-backend")) { | |
2103 | if (perf_evlist__add_default_attrs(evsel_list, | |
2104 | backend_attrs) < 0) | |
2105 | return -1; | |
2106 | } | |
2107 | if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0) | |
50d08e47 | 2108 | return -1; |
2cba3ffb IM |
2109 | } |
2110 | ||
2111 | /* Detailed events get appended to the event list: */ | |
2112 | ||
2113 | if (detailed_run < 1) | |
2114 | return 0; | |
2115 | ||
2116 | /* Append detailed run extra attributes: */ | |
79695e1b | 2117 | if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) |
50d08e47 | 2118 | return -1; |
2cba3ffb IM |
2119 | |
2120 | if (detailed_run < 2) | |
2121 | return 0; | |
2122 | ||
2123 | /* Append very detailed run extra attributes: */ | |
79695e1b | 2124 | if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) |
50d08e47 | 2125 | return -1; |
2cba3ffb IM |
2126 | |
2127 | if (detailed_run < 3) | |
2128 | return 0; | |
2129 | ||
2130 | /* Append very, very detailed run extra attributes: */ | |
79695e1b | 2131 | return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); |
2cba3ffb IM |
2132 | } |
2133 | ||
8a59f3cc | 2134 | static const char * const stat_record_usage[] = { |
4979d0c7 JO |
2135 | "perf stat record [<options>]", |
2136 | NULL, | |
2137 | }; | |
2138 | ||
3ba78bd0 JO |
2139 | static void init_features(struct perf_session *session) |
2140 | { | |
2141 | int feat; | |
2142 | ||
2143 | for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++) | |
2144 | perf_header__set_feat(&session->header, feat); | |
2145 | ||
2146 | perf_header__clear_feat(&session->header, HEADER_BUILD_ID); | |
2147 | perf_header__clear_feat(&session->header, HEADER_TRACING_DATA); | |
2148 | perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK); | |
2149 | perf_header__clear_feat(&session->header, HEADER_AUXTRACE); | |
2150 | } | |
2151 | ||
4979d0c7 JO |
2152 | static int __cmd_record(int argc, const char **argv) |
2153 | { | |
2154 | struct perf_session *session; | |
2155 | struct perf_data_file *file = &perf_stat.file; | |
2156 | ||
8a59f3cc | 2157 | argc = parse_options(argc, argv, stat_options, stat_record_usage, |
4979d0c7 JO |
2158 | PARSE_OPT_STOP_AT_NON_OPTION); |
2159 | ||
2160 | if (output_name) | |
2161 | file->path = output_name; | |
2162 | ||
e9d6db8e JO |
2163 | if (run_count != 1 || forever) { |
2164 | pr_err("Cannot use -r option with perf stat record.\n"); | |
2165 | return -1; | |
2166 | } | |
2167 | ||
4979d0c7 JO |
2168 | session = perf_session__new(file, false, NULL); |
2169 | if (session == NULL) { | |
2170 | pr_err("Perf session creation failed.\n"); | |
2171 | return -1; | |
2172 | } | |
2173 | ||
3ba78bd0 JO |
2174 | init_features(session); |
2175 | ||
4979d0c7 JO |
2176 | session->evlist = evsel_list; |
2177 | perf_stat.session = session; | |
2178 | perf_stat.record = true; | |
2179 | return argc; | |
2180 | } | |
2181 | ||
a56f9390 JO |
2182 | static int process_stat_round_event(struct perf_tool *tool __maybe_unused, |
2183 | union perf_event *event, | |
2184 | struct perf_session *session) | |
2185 | { | |
e3b03b6c | 2186 | struct stat_round_event *stat_round = &event->stat_round; |
a56f9390 JO |
2187 | struct perf_evsel *counter; |
2188 | struct timespec tsh, *ts = NULL; | |
2189 | const char **argv = session->header.env.cmdline_argv; | |
2190 | int argc = session->header.env.nr_cmdline; | |
2191 | ||
e5cadb93 | 2192 | evlist__for_each_entry(evsel_list, counter) |
a56f9390 JO |
2193 | perf_stat_process_counter(&stat_config, counter); |
2194 | ||
e3b03b6c AK |
2195 | if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL) |
2196 | update_stats(&walltime_nsecs_stats, stat_round->time); | |
a56f9390 | 2197 | |
e3b03b6c | 2198 | if (stat_config.interval && stat_round->time) { |
bd48c63e ACM |
2199 | tsh.tv_sec = stat_round->time / NSEC_PER_SEC; |
2200 | tsh.tv_nsec = stat_round->time % NSEC_PER_SEC; | |
a56f9390 JO |
2201 | ts = &tsh; |
2202 | } | |
2203 | ||
2204 | print_counters(ts, argc, argv); | |
2205 | return 0; | |
2206 | } | |
2207 | ||
62ba18ba | 2208 | static |
7e6a7998 | 2209 | int process_stat_config_event(struct perf_tool *tool, |
62ba18ba JO |
2210 | union perf_event *event, |
2211 | struct perf_session *session __maybe_unused) | |
2212 | { | |
68d702f7 JO |
2213 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); |
2214 | ||
62ba18ba | 2215 | perf_event__read_stat_config(&stat_config, &event->stat_config); |
68d702f7 | 2216 | |
89af4e05 JO |
2217 | if (cpu_map__empty(st->cpus)) { |
2218 | if (st->aggr_mode != AGGR_UNSET) | |
2219 | pr_warning("warning: processing task data, aggregation mode not set\n"); | |
2220 | return 0; | |
2221 | } | |
2222 | ||
2223 | if (st->aggr_mode != AGGR_UNSET) | |
2224 | stat_config.aggr_mode = st->aggr_mode; | |
2225 | ||
68d702f7 JO |
2226 | if (perf_stat.file.is_pipe) |
2227 | perf_stat_init_aggr_mode(); | |
2228 | else | |
2229 | perf_stat_init_aggr_mode_file(st); | |
2230 | ||
62ba18ba JO |
2231 | return 0; |
2232 | } | |
2233 | ||
1975d36e JO |
2234 | static int set_maps(struct perf_stat *st) |
2235 | { | |
2236 | if (!st->cpus || !st->threads) | |
2237 | return 0; | |
2238 | ||
2239 | if (WARN_ONCE(st->maps_allocated, "stats double allocation\n")) | |
2240 | return -EINVAL; | |
2241 | ||
2242 | perf_evlist__set_maps(evsel_list, st->cpus, st->threads); | |
2243 | ||
2244 | if (perf_evlist__alloc_stats(evsel_list, true)) | |
2245 | return -ENOMEM; | |
2246 | ||
2247 | st->maps_allocated = true; | |
2248 | return 0; | |
2249 | } | |
2250 | ||
2251 | static | |
7e6a7998 | 2252 | int process_thread_map_event(struct perf_tool *tool, |
1975d36e JO |
2253 | union perf_event *event, |
2254 | struct perf_session *session __maybe_unused) | |
2255 | { | |
2256 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); | |
2257 | ||
2258 | if (st->threads) { | |
2259 | pr_warning("Extra thread map event, ignoring.\n"); | |
2260 | return 0; | |
2261 | } | |
2262 | ||
2263 | st->threads = thread_map__new_event(&event->thread_map); | |
2264 | if (!st->threads) | |
2265 | return -ENOMEM; | |
2266 | ||
2267 | return set_maps(st); | |
2268 | } | |
2269 | ||
2270 | static | |
7e6a7998 | 2271 | int process_cpu_map_event(struct perf_tool *tool, |
1975d36e JO |
2272 | union perf_event *event, |
2273 | struct perf_session *session __maybe_unused) | |
2274 | { | |
2275 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); | |
2276 | struct cpu_map *cpus; | |
2277 | ||
2278 | if (st->cpus) { | |
2279 | pr_warning("Extra cpu map event, ignoring.\n"); | |
2280 | return 0; | |
2281 | } | |
2282 | ||
2283 | cpus = cpu_map__new_data(&event->cpu_map.data); | |
2284 | if (!cpus) | |
2285 | return -ENOMEM; | |
2286 | ||
2287 | st->cpus = cpus; | |
2288 | return set_maps(st); | |
2289 | } | |
2290 | ||
8a59f3cc | 2291 | static const char * const stat_report_usage[] = { |
ba6039b6 JO |
2292 | "perf stat report [<options>]", |
2293 | NULL, | |
2294 | }; | |
2295 | ||
2296 | static struct perf_stat perf_stat = { | |
2297 | .tool = { | |
2298 | .attr = perf_event__process_attr, | |
fa6ea781 | 2299 | .event_update = perf_event__process_event_update, |
1975d36e JO |
2300 | .thread_map = process_thread_map_event, |
2301 | .cpu_map = process_cpu_map_event, | |
62ba18ba | 2302 | .stat_config = process_stat_config_event, |
a56f9390 JO |
2303 | .stat = perf_event__process_stat_event, |
2304 | .stat_round = process_stat_round_event, | |
ba6039b6 | 2305 | }, |
89af4e05 | 2306 | .aggr_mode = AGGR_UNSET, |
ba6039b6 JO |
2307 | }; |
2308 | ||
2309 | static int __cmd_report(int argc, const char **argv) | |
2310 | { | |
2311 | struct perf_session *session; | |
2312 | const struct option options[] = { | |
2313 | OPT_STRING('i', "input", &input_name, "file", "input file name"), | |
89af4e05 JO |
2314 | OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode, |
2315 | "aggregate counts per processor socket", AGGR_SOCKET), | |
2316 | OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode, | |
2317 | "aggregate counts per physical processor core", AGGR_CORE), | |
2318 | OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode, | |
2319 | "disable CPU count aggregation", AGGR_NONE), | |
ba6039b6 JO |
2320 | OPT_END() |
2321 | }; | |
2322 | struct stat st; | |
2323 | int ret; | |
2324 | ||
8a59f3cc | 2325 | argc = parse_options(argc, argv, options, stat_report_usage, 0); |
ba6039b6 JO |
2326 | |
2327 | if (!input_name || !strlen(input_name)) { | |
2328 | if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) | |
2329 | input_name = "-"; | |
2330 | else | |
2331 | input_name = "perf.data"; | |
2332 | } | |
2333 | ||
2334 | perf_stat.file.path = input_name; | |
2335 | perf_stat.file.mode = PERF_DATA_MODE_READ; | |
2336 | ||
2337 | session = perf_session__new(&perf_stat.file, false, &perf_stat.tool); | |
2338 | if (session == NULL) | |
2339 | return -1; | |
2340 | ||
2341 | perf_stat.session = session; | |
2342 | stat_config.output = stderr; | |
2343 | evsel_list = session->evlist; | |
2344 | ||
2345 | ret = perf_session__process_events(session); | |
2346 | if (ret) | |
2347 | return ret; | |
2348 | ||
2349 | perf_session__delete(session); | |
2350 | return 0; | |
2351 | } | |
2352 | ||
e3ba76de JO |
2353 | static void setup_system_wide(int forks) |
2354 | { | |
2355 | /* | |
2356 | * Make system wide (-a) the default target if | |
2357 | * no target was specified and one of following | |
2358 | * conditions is met: | |
2359 | * | |
2360 | * - there's no workload specified | |
2361 | * - there is workload specified but all requested | |
2362 | * events are system wide events | |
2363 | */ | |
2364 | if (!target__none(&target)) | |
2365 | return; | |
2366 | ||
2367 | if (!forks) | |
2368 | target.system_wide = true; | |
2369 | else { | |
2370 | struct perf_evsel *counter; | |
2371 | ||
2372 | evlist__for_each_entry(evsel_list, counter) { | |
2373 | if (!counter->system_wide) | |
2374 | return; | |
2375 | } | |
2376 | ||
2377 | if (evsel_list->nr_entries) | |
2378 | target.system_wide = true; | |
2379 | } | |
2380 | } | |
2381 | ||
1d037ca1 | 2382 | int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused) |
5242519b | 2383 | { |
b070a547 ACM |
2384 | const char * const stat_usage[] = { |
2385 | "perf stat [<options>] [<command>]", | |
2386 | NULL | |
2387 | }; | |
cc03c542 | 2388 | int status = -EINVAL, run_idx; |
4aa9015f | 2389 | const char *mode; |
5821522e | 2390 | FILE *output = stderr; |
ec0d3d1f | 2391 | unsigned int interval; |
ba6039b6 | 2392 | const char * const stat_subcommands[] = { "record", "report" }; |
42202dd5 | 2393 | |
5af52b51 SE |
2394 | setlocale(LC_ALL, ""); |
2395 | ||
334fe7a3 | 2396 | evsel_list = perf_evlist__new(); |
361c99a6 ACM |
2397 | if (evsel_list == NULL) |
2398 | return -ENOMEM; | |
2399 | ||
1669e509 | 2400 | parse_events__shrink_config_terms(); |
4979d0c7 JO |
2401 | argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands, |
2402 | (const char **) stat_usage, | |
2403 | PARSE_OPT_STOP_AT_NON_OPTION); | |
fb4605ba | 2404 | perf_stat__init_shadow_stats(); |
4979d0c7 | 2405 | |
6edb78a2 JO |
2406 | if (csv_sep) { |
2407 | csv_output = true; | |
2408 | if (!strcmp(csv_sep, "\\t")) | |
2409 | csv_sep = "\t"; | |
2410 | } else | |
2411 | csv_sep = DEFAULT_SEPARATOR; | |
2412 | ||
4979d0c7 JO |
2413 | if (argc && !strncmp(argv[0], "rec", 3)) { |
2414 | argc = __cmd_record(argc, argv); | |
2415 | if (argc < 0) | |
2416 | return -1; | |
ba6039b6 JO |
2417 | } else if (argc && !strncmp(argv[0], "rep", 3)) |
2418 | return __cmd_report(argc, argv); | |
d7470b6a | 2419 | |
ec0d3d1f JO |
2420 | interval = stat_config.interval; |
2421 | ||
4979d0c7 JO |
2422 | /* |
2423 | * For record command the -o is already taken care of. | |
2424 | */ | |
2425 | if (!STAT_RECORD && output_name && strcmp(output_name, "-")) | |
4aa9015f SE |
2426 | output = NULL; |
2427 | ||
56f3bae7 JC |
2428 | if (output_name && output_fd) { |
2429 | fprintf(stderr, "cannot use both --output and --log-fd\n"); | |
e0547311 JO |
2430 | parse_options_usage(stat_usage, stat_options, "o", 1); |
2431 | parse_options_usage(NULL, stat_options, "log-fd", 0); | |
cc03c542 | 2432 | goto out; |
56f3bae7 | 2433 | } |
fc3e4d07 | 2434 | |
54b50916 AK |
2435 | if (metric_only && stat_config.aggr_mode == AGGR_THREAD) { |
2436 | fprintf(stderr, "--metric-only is not supported with --per-thread\n"); | |
2437 | goto out; | |
2438 | } | |
2439 | ||
54b50916 AK |
2440 | if (metric_only && run_count > 1) { |
2441 | fprintf(stderr, "--metric-only is not supported with -r\n"); | |
2442 | goto out; | |
2443 | } | |
2444 | ||
fc3e4d07 SE |
2445 | if (output_fd < 0) { |
2446 | fprintf(stderr, "argument to --log-fd must be a > 0\n"); | |
e0547311 | 2447 | parse_options_usage(stat_usage, stat_options, "log-fd", 0); |
cc03c542 | 2448 | goto out; |
fc3e4d07 SE |
2449 | } |
2450 | ||
4aa9015f SE |
2451 | if (!output) { |
2452 | struct timespec tm; | |
2453 | mode = append_file ? "a" : "w"; | |
2454 | ||
2455 | output = fopen(output_name, mode); | |
2456 | if (!output) { | |
2457 | perror("failed to create output file"); | |
fceda7fe | 2458 | return -1; |
4aa9015f SE |
2459 | } |
2460 | clock_gettime(CLOCK_REALTIME, &tm); | |
2461 | fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); | |
fc3e4d07 | 2462 | } else if (output_fd > 0) { |
56f3bae7 JC |
2463 | mode = append_file ? "a" : "w"; |
2464 | output = fdopen(output_fd, mode); | |
2465 | if (!output) { | |
2466 | perror("Failed opening logfd"); | |
2467 | return -errno; | |
2468 | } | |
4aa9015f SE |
2469 | } |
2470 | ||
5821522e JO |
2471 | stat_config.output = output; |
2472 | ||
d7470b6a SE |
2473 | /* |
2474 | * let the spreadsheet do the pretty-printing | |
2475 | */ | |
2476 | if (csv_output) { | |
61a9f324 | 2477 | /* User explicitly passed -B? */ |
d7470b6a SE |
2478 | if (big_num_opt == 1) { |
2479 | fprintf(stderr, "-B option not supported with -x\n"); | |
e0547311 JO |
2480 | parse_options_usage(stat_usage, stat_options, "B", 1); |
2481 | parse_options_usage(NULL, stat_options, "x", 1); | |
cc03c542 | 2482 | goto out; |
d7470b6a SE |
2483 | } else /* Nope, so disable big number formatting */ |
2484 | big_num = false; | |
2485 | } else if (big_num_opt == 0) /* User passed --no-big-num */ | |
2486 | big_num = false; | |
2487 | ||
e3ba76de | 2488 | setup_system_wide(argc); |
ac3063bd | 2489 | |
a7e191c3 | 2490 | if (run_count < 0) { |
cc03c542 | 2491 | pr_err("Run count must be a positive number\n"); |
e0547311 | 2492 | parse_options_usage(stat_usage, stat_options, "r", 1); |
cc03c542 | 2493 | goto out; |
a7e191c3 FD |
2494 | } else if (run_count == 0) { |
2495 | forever = true; | |
2496 | run_count = 1; | |
2497 | } | |
ddcacfa0 | 2498 | |
421a50f3 | 2499 | if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) { |
32b8af82 JO |
2500 | fprintf(stderr, "The --per-thread option is only available " |
2501 | "when monitoring via -p -t options.\n"); | |
e0547311 JO |
2502 | parse_options_usage(NULL, stat_options, "p", 1); |
2503 | parse_options_usage(NULL, stat_options, "t", 1); | |
32b8af82 JO |
2504 | goto out; |
2505 | } | |
2506 | ||
2507 | /* | |
2508 | * no_aggr, cgroup are for system-wide only | |
2509 | * --per-thread is aggregated per thread, we dont mix it with cpu mode | |
2510 | */ | |
421a50f3 JO |
2511 | if (((stat_config.aggr_mode != AGGR_GLOBAL && |
2512 | stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) && | |
602ad878 | 2513 | !target__has_cpu(&target)) { |
023695d9 SE |
2514 | fprintf(stderr, "both cgroup and no-aggregation " |
2515 | "modes only available in system-wide mode\n"); | |
2516 | ||
e0547311 JO |
2517 | parse_options_usage(stat_usage, stat_options, "G", 1); |
2518 | parse_options_usage(NULL, stat_options, "A", 1); | |
2519 | parse_options_usage(NULL, stat_options, "a", 1); | |
cc03c542 | 2520 | goto out; |
d7e7a451 SE |
2521 | } |
2522 | ||
2cba3ffb IM |
2523 | if (add_default_attributes()) |
2524 | goto out; | |
ddcacfa0 | 2525 | |
602ad878 | 2526 | target__validate(&target); |
5c98d466 | 2527 | |
77a6f014 | 2528 | if (perf_evlist__create_maps(evsel_list, &target) < 0) { |
602ad878 | 2529 | if (target__has_task(&target)) { |
77a6f014 | 2530 | pr_err("Problems finding threads of monitor\n"); |
e0547311 JO |
2531 | parse_options_usage(stat_usage, stat_options, "p", 1); |
2532 | parse_options_usage(NULL, stat_options, "t", 1); | |
602ad878 | 2533 | } else if (target__has_cpu(&target)) { |
77a6f014 | 2534 | perror("failed to parse CPUs map"); |
e0547311 JO |
2535 | parse_options_usage(stat_usage, stat_options, "C", 1); |
2536 | parse_options_usage(NULL, stat_options, "a", 1); | |
cc03c542 NK |
2537 | } |
2538 | goto out; | |
60d567e2 | 2539 | } |
32b8af82 JO |
2540 | |
2541 | /* | |
2542 | * Initialize thread_map with comm names, | |
2543 | * so we could print it out on output. | |
2544 | */ | |
421a50f3 | 2545 | if (stat_config.aggr_mode == AGGR_THREAD) |
32b8af82 JO |
2546 | thread_map__read_comms(evsel_list->threads); |
2547 | ||
13370a9b | 2548 | if (interval && interval < 100) { |
19afd104 KL |
2549 | if (interval < 10) { |
2550 | pr_err("print interval must be >= 10ms\n"); | |
e0547311 | 2551 | parse_options_usage(stat_usage, stat_options, "I", 1); |
19afd104 KL |
2552 | goto out; |
2553 | } else | |
2554 | pr_warning("print interval < 100ms. " | |
2555 | "The overhead percentage could be high in some cases. " | |
2556 | "Please proceed with caution.\n"); | |
13370a9b | 2557 | } |
c45c6ea2 | 2558 | |
d134ffb9 | 2559 | if (perf_evlist__alloc_stats(evsel_list, interval)) |
03ad9747 | 2560 | goto out; |
d6d901c2 | 2561 | |
86ee6e18 | 2562 | if (perf_stat_init_aggr_mode()) |
03ad9747 | 2563 | goto out; |
86ee6e18 | 2564 | |
58d7e993 IM |
2565 | /* |
2566 | * We dont want to block the signals - that would cause | |
2567 | * child tasks to inherit that and Ctrl-C would not work. | |
2568 | * What we want is for Ctrl-C to work in the exec()-ed | |
2569 | * task, but being ignored by perf stat itself: | |
2570 | */ | |
f7b7c26e | 2571 | atexit(sig_atexit); |
a7e191c3 FD |
2572 | if (!forever) |
2573 | signal(SIGINT, skip_signal); | |
13370a9b | 2574 | signal(SIGCHLD, skip_signal); |
58d7e993 IM |
2575 | signal(SIGALRM, skip_signal); |
2576 | signal(SIGABRT, skip_signal); | |
2577 | ||
42202dd5 | 2578 | status = 0; |
a7e191c3 | 2579 | for (run_idx = 0; forever || run_idx < run_count; run_idx++) { |
bb963e16 | 2580 | if (run_count != 1 && verbose > 0) |
4aa9015f SE |
2581 | fprintf(output, "[ perf stat: executing run #%d ... ]\n", |
2582 | run_idx + 1); | |
f9cef0a9 | 2583 | |
42202dd5 | 2584 | status = run_perf_stat(argc, argv); |
a7e191c3 | 2585 | if (forever && status != -1) { |
d4f63a47 | 2586 | print_counters(NULL, argc, argv); |
254ecbc7 | 2587 | perf_stat__reset_stats(); |
a7e191c3 | 2588 | } |
42202dd5 IM |
2589 | } |
2590 | ||
a7e191c3 | 2591 | if (!forever && status != -1 && !interval) |
d4f63a47 | 2592 | print_counters(NULL, argc, argv); |
d134ffb9 | 2593 | |
4979d0c7 JO |
2594 | if (STAT_RECORD) { |
2595 | /* | |
2596 | * We synthesize the kernel mmap record just so that older tools | |
2597 | * don't emit warnings about not being able to resolve symbols | |
2598 | * due to /proc/sys/kernel/kptr_restrict settings and instear provide | |
2599 | * a saner message about no samples being in the perf.data file. | |
2600 | * | |
2601 | * This also serves to suppress a warning about f_header.data.size == 0 | |
8b99b1a4 JO |
2602 | * in header.c at the moment 'perf stat record' gets introduced, which |
2603 | * is not really needed once we start adding the stat specific PERF_RECORD_ | |
2604 | * records, but the need to suppress the kptr_restrict messages in older | |
2605 | * tools remain -acme | |
4979d0c7 JO |
2606 | */ |
2607 | int fd = perf_data_file__fd(&perf_stat.file); | |
2608 | int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat, | |
2609 | process_synthesized_event, | |
2610 | &perf_stat.session->machines.host); | |
2611 | if (err) { | |
2612 | pr_warning("Couldn't synthesize the kernel mmap record, harmless, " | |
2613 | "older tools may produce warnings about this file\n."); | |
2614 | } | |
2615 | ||
7aad0c32 JO |
2616 | if (!interval) { |
2617 | if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL)) | |
2618 | pr_err("failed to write stat round event\n"); | |
2619 | } | |
2620 | ||
664c98d4 JO |
2621 | if (!perf_stat.file.is_pipe) { |
2622 | perf_stat.session->header.data_size += perf_stat.bytes_written; | |
2623 | perf_session__write_header(perf_stat.session, evsel_list, fd, true); | |
2624 | } | |
4979d0c7 JO |
2625 | |
2626 | perf_session__delete(perf_stat.session); | |
2627 | } | |
2628 | ||
544c2ae7 | 2629 | perf_stat__exit_aggr_mode(); |
d134ffb9 | 2630 | perf_evlist__free_stats(evsel_list); |
0015e2e1 ACM |
2631 | out: |
2632 | perf_evlist__delete(evsel_list); | |
42202dd5 | 2633 | return status; |
ddcacfa0 | 2634 | } |