]>
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" |
a067558e | 67 | #include "util/string2.h" |
ba6039b6 | 68 | #include "asm/bug.h" |
ddcacfa0 | 69 | |
bd48c63e | 70 | #include <linux/time64.h> |
44b1e60a | 71 | #include <api/fs/fs.h> |
a43783ae | 72 | #include <errno.h> |
9607ad3a | 73 | #include <signal.h> |
1f16c575 | 74 | #include <stdlib.h> |
ddcacfa0 | 75 | #include <sys/prctl.h> |
fd20e811 | 76 | #include <inttypes.h> |
5af52b51 | 77 | #include <locale.h> |
e3b03b6c | 78 | #include <math.h> |
7a8ef4c4 ACM |
79 | #include <sys/types.h> |
80 | #include <sys/stat.h> | |
4208735d | 81 | #include <sys/wait.h> |
7a8ef4c4 | 82 | #include <unistd.h> |
16c8a109 | 83 | |
3d689ed6 ACM |
84 | #include "sane_ctype.h" |
85 | ||
d7470b6a | 86 | #define DEFAULT_SEPARATOR " " |
2cee77c4 DA |
87 | #define CNTR_NOT_SUPPORTED "<not supported>" |
88 | #define CNTR_NOT_COUNTED "<not counted>" | |
d7470b6a | 89 | |
d4f63a47 | 90 | static void print_counters(struct timespec *ts, int argc, const char **argv); |
13370a9b | 91 | |
4cabc3d1 | 92 | /* Default events used for perf stat -T */ |
a454742c JO |
93 | static const char *transaction_attrs = { |
94 | "task-clock," | |
4cabc3d1 AK |
95 | "{" |
96 | "instructions," | |
97 | "cycles," | |
98 | "cpu/cycles-t/," | |
99 | "cpu/tx-start/," | |
100 | "cpu/el-start/," | |
101 | "cpu/cycles-ct/" | |
102 | "}" | |
103 | }; | |
104 | ||
105 | /* More limited version when the CPU does not have all events. */ | |
a454742c JO |
106 | static const char * transaction_limited_attrs = { |
107 | "task-clock," | |
4cabc3d1 AK |
108 | "{" |
109 | "instructions," | |
110 | "cycles," | |
111 | "cpu/cycles-t/," | |
112 | "cpu/tx-start/" | |
113 | "}" | |
114 | }; | |
115 | ||
44b1e60a AK |
116 | static const char * topdown_attrs[] = { |
117 | "topdown-total-slots", | |
118 | "topdown-slots-retired", | |
119 | "topdown-recovery-bubbles", | |
120 | "topdown-fetch-bubbles", | |
121 | "topdown-slots-issued", | |
122 | NULL, | |
123 | }; | |
124 | ||
666e6d48 | 125 | static struct perf_evlist *evsel_list; |
361c99a6 | 126 | |
602ad878 | 127 | static struct target target = { |
77a6f014 NK |
128 | .uid = UINT_MAX, |
129 | }; | |
ddcacfa0 | 130 | |
1e5a2931 JO |
131 | typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu); |
132 | ||
3d632595 | 133 | static int run_count = 1; |
2e6cdf99 | 134 | static bool no_inherit = false; |
d07f0b12 | 135 | static volatile pid_t child_pid = -1; |
c0555642 | 136 | static bool null_run = false; |
2cba3ffb | 137 | static int detailed_run = 0; |
4cabc3d1 | 138 | static bool transaction_run; |
44b1e60a | 139 | static bool topdown_run = false; |
201e0b06 | 140 | static bool big_num = true; |
d7470b6a | 141 | static int big_num_opt = -1; |
d7470b6a SE |
142 | static const char *csv_sep = NULL; |
143 | static bool csv_output = false; | |
43bece79 | 144 | static bool group = false; |
1f16c575 PZ |
145 | static const char *pre_cmd = NULL; |
146 | static const char *post_cmd = NULL; | |
147 | static bool sync_run = false; | |
41191688 | 148 | static unsigned int initial_delay = 0; |
410136f5 | 149 | static unsigned int unit_width = 4; /* strlen("unit") */ |
a7e191c3 | 150 | static bool forever = false; |
54b50916 | 151 | static bool metric_only = false; |
44b1e60a | 152 | static bool force_metric_only = false; |
430daf2d | 153 | static bool no_merge = false; |
13370a9b | 154 | static struct timespec ref_time; |
86ee6e18 | 155 | static struct cpu_map *aggr_map; |
1e5a2931 | 156 | static aggr_get_id_t aggr_get_id; |
e0547311 JO |
157 | static bool append_file; |
158 | static const char *output_name; | |
159 | static int output_fd; | |
02d492e5 | 160 | static int print_free_counters_hint; |
5af52b51 | 161 | |
4979d0c7 JO |
162 | struct perf_stat { |
163 | bool record; | |
164 | struct perf_data_file file; | |
165 | struct perf_session *session; | |
166 | u64 bytes_written; | |
ba6039b6 | 167 | struct perf_tool tool; |
1975d36e JO |
168 | bool maps_allocated; |
169 | struct cpu_map *cpus; | |
170 | struct thread_map *threads; | |
89af4e05 | 171 | enum aggr_mode aggr_mode; |
4979d0c7 JO |
172 | }; |
173 | ||
174 | static struct perf_stat perf_stat; | |
175 | #define STAT_RECORD perf_stat.record | |
176 | ||
60666c63 LW |
177 | static volatile int done = 0; |
178 | ||
421a50f3 JO |
179 | static struct perf_stat_config stat_config = { |
180 | .aggr_mode = AGGR_GLOBAL, | |
711a572e | 181 | .scale = true, |
421a50f3 JO |
182 | }; |
183 | ||
13370a9b SE |
184 | static inline void diff_timespec(struct timespec *r, struct timespec *a, |
185 | struct timespec *b) | |
186 | { | |
187 | r->tv_sec = a->tv_sec - b->tv_sec; | |
188 | if (a->tv_nsec < b->tv_nsec) { | |
310ebb93 | 189 | r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec; |
13370a9b SE |
190 | r->tv_sec--; |
191 | } else { | |
192 | r->tv_nsec = a->tv_nsec - b->tv_nsec ; | |
193 | } | |
194 | } | |
195 | ||
254ecbc7 JO |
196 | static void perf_stat__reset_stats(void) |
197 | { | |
198 | perf_evlist__reset_stats(evsel_list); | |
f87027b9 | 199 | perf_stat__reset_shadow_stats(); |
1eda3b21 JO |
200 | } |
201 | ||
cac21425 | 202 | static int create_perf_stat_counter(struct perf_evsel *evsel) |
ddcacfa0 | 203 | { |
69aad6f1 | 204 | struct perf_event_attr *attr = &evsel->attr; |
727ab04e | 205 | |
711a572e | 206 | if (stat_config.scale) |
a21ca2ca IM |
207 | attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | |
208 | PERF_FORMAT_TOTAL_TIME_RUNNING; | |
ddcacfa0 | 209 | |
5d2cd909 ACM |
210 | attr->inherit = !no_inherit; |
211 | ||
6acd8e92 JO |
212 | /* |
213 | * Some events get initialized with sample_(period/type) set, | |
214 | * like tracepoints. Clear it up for counting. | |
215 | */ | |
216 | attr->sample_period = 0; | |
6db1a5c1 | 217 | |
4979d0c7 JO |
218 | /* |
219 | * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless | |
220 | * while avoiding that older tools show confusing messages. | |
6db1a5c1 JO |
221 | * |
222 | * However for pipe sessions we need to keep it zero, | |
223 | * because script's perf_evsel__check_attr is triggered | |
224 | * by attr->sample_type != 0, and we can't run it on | |
225 | * stat sessions. | |
4979d0c7 | 226 | */ |
6db1a5c1 JO |
227 | if (!(STAT_RECORD && perf_stat.file.is_pipe)) |
228 | attr->sample_type = PERF_SAMPLE_IDENTIFIER; | |
6acd8e92 | 229 | |
67ccdecd JO |
230 | /* |
231 | * Disabling all counters initially, they will be enabled | |
232 | * either manually by us or by kernel via enable_on_exec | |
233 | * set later. | |
234 | */ | |
c8280cec | 235 | if (perf_evsel__is_group_leader(evsel)) { |
67ccdecd JO |
236 | attr->disabled = 1; |
237 | ||
c8280cec JO |
238 | /* |
239 | * In case of initial_delay we enable tracee | |
240 | * events manually. | |
241 | */ | |
242 | if (target__none(&target) && !initial_delay) | |
41191688 | 243 | attr->enable_on_exec = 1; |
ddcacfa0 | 244 | } |
084ab9f8 | 245 | |
c8280cec JO |
246 | if (target__has_cpu(&target)) |
247 | return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel)); | |
248 | ||
594ac61a | 249 | return perf_evsel__open_per_thread(evsel, evsel_list->threads); |
ddcacfa0 IM |
250 | } |
251 | ||
c04f5e5d IM |
252 | /* |
253 | * Does the counter have nsecs as a unit? | |
254 | */ | |
daec78a0 | 255 | static inline int nsec_counter(struct perf_evsel *evsel) |
c04f5e5d | 256 | { |
daec78a0 ACM |
257 | if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || |
258 | perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) | |
c04f5e5d IM |
259 | return 1; |
260 | ||
261 | return 0; | |
262 | } | |
263 | ||
8b99b1a4 JO |
264 | static int process_synthesized_event(struct perf_tool *tool __maybe_unused, |
265 | union perf_event *event, | |
266 | struct perf_sample *sample __maybe_unused, | |
267 | struct machine *machine __maybe_unused) | |
4979d0c7 | 268 | { |
8b99b1a4 | 269 | if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) { |
4979d0c7 JO |
270 | pr_err("failed to write perf data, error: %m\n"); |
271 | return -1; | |
272 | } | |
273 | ||
8b99b1a4 | 274 | perf_stat.bytes_written += event->header.size; |
4979d0c7 JO |
275 | return 0; |
276 | } | |
277 | ||
1975d36e | 278 | static int write_stat_round_event(u64 tm, u64 type) |
7aad0c32 | 279 | { |
1975d36e | 280 | return perf_event__synthesize_stat_round(NULL, tm, type, |
7aad0c32 JO |
281 | process_synthesized_event, |
282 | NULL); | |
283 | } | |
284 | ||
285 | #define WRITE_STAT_ROUND_EVENT(time, interval) \ | |
286 | write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval) | |
287 | ||
5a6ea81b JO |
288 | #define SID(e, x, y) xyarray__entry(e->sample_id, x, y) |
289 | ||
290 | static int | |
291 | perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread, | |
292 | struct perf_counts_values *count) | |
293 | { | |
294 | struct perf_sample_id *sid = SID(counter, cpu, thread); | |
295 | ||
296 | return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count, | |
297 | process_synthesized_event, NULL); | |
298 | } | |
299 | ||
f5b4a9c3 SE |
300 | /* |
301 | * Read out the results of a single counter: | |
302 | * do not aggregate counts across CPUs in system-wide mode | |
303 | */ | |
c52b12ed | 304 | static int read_counter(struct perf_evsel *counter) |
f5b4a9c3 | 305 | { |
9bf1a529 | 306 | int nthreads = thread_map__nr(evsel_list->threads); |
00e727bb MR |
307 | int ncpus, cpu, thread; |
308 | ||
309 | if (target__has_cpu(&target)) | |
310 | ncpus = perf_evsel__nr_cpus(counter); | |
311 | else | |
312 | ncpus = 1; | |
f5b4a9c3 | 313 | |
3b4331d9 SP |
314 | if (!counter->supported) |
315 | return -ENOENT; | |
316 | ||
9bf1a529 JO |
317 | if (counter->system_wide) |
318 | nthreads = 1; | |
319 | ||
320 | for (thread = 0; thread < nthreads; thread++) { | |
321 | for (cpu = 0; cpu < ncpus; cpu++) { | |
3b3eb044 JO |
322 | struct perf_counts_values *count; |
323 | ||
324 | count = perf_counts(counter->counts, cpu, thread); | |
db49a717 SE |
325 | if (perf_evsel__read(counter, cpu, thread, count)) { |
326 | counter->counts->scaled = -1; | |
327 | perf_counts(counter->counts, cpu, thread)->ena = 0; | |
328 | perf_counts(counter->counts, cpu, thread)->run = 0; | |
9bf1a529 | 329 | return -1; |
db49a717 | 330 | } |
5a6ea81b JO |
331 | |
332 | if (STAT_RECORD) { | |
333 | if (perf_evsel__write_stat_event(counter, cpu, thread, count)) { | |
334 | pr_err("failed to write stat event\n"); | |
335 | return -1; | |
336 | } | |
337 | } | |
0b1abbf4 AK |
338 | |
339 | if (verbose > 1) { | |
340 | fprintf(stat_config.output, | |
341 | "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", | |
342 | perf_evsel__name(counter), | |
343 | cpu, | |
344 | count->val, count->ena, count->run); | |
345 | } | |
9bf1a529 | 346 | } |
f5b4a9c3 | 347 | } |
c52b12ed ACM |
348 | |
349 | return 0; | |
2996f5dd IM |
350 | } |
351 | ||
3df33eff | 352 | static void read_counters(void) |
13370a9b | 353 | { |
13370a9b | 354 | struct perf_evsel *counter; |
db49a717 | 355 | int ret; |
13370a9b | 356 | |
e5cadb93 | 357 | evlist__for_each_entry(evsel_list, counter) { |
db49a717 SE |
358 | ret = read_counter(counter); |
359 | if (ret) | |
245bad8e | 360 | pr_debug("failed to read counter %s\n", counter->name); |
3b3eb044 | 361 | |
db49a717 | 362 | if (ret == 0 && perf_stat_process_counter(&stat_config, counter)) |
3b3eb044 | 363 | pr_warning("failed to process counter %s\n", counter->name); |
13370a9b | 364 | } |
106a94a0 JO |
365 | } |
366 | ||
ba411a95 | 367 | static void process_interval(void) |
106a94a0 | 368 | { |
106a94a0 | 369 | struct timespec ts, rs; |
106a94a0 | 370 | |
3df33eff | 371 | read_counters(); |
86ee6e18 | 372 | |
13370a9b SE |
373 | clock_gettime(CLOCK_MONOTONIC, &ts); |
374 | diff_timespec(&rs, &ts, &ref_time); | |
13370a9b | 375 | |
7aad0c32 | 376 | if (STAT_RECORD) { |
bd48c63e | 377 | if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL)) |
7aad0c32 JO |
378 | pr_err("failed to write stat round event\n"); |
379 | } | |
380 | ||
d4f63a47 | 381 | print_counters(&rs, 0, NULL); |
13370a9b SE |
382 | } |
383 | ||
67ccdecd | 384 | static void enable_counters(void) |
41191688 | 385 | { |
67ccdecd | 386 | if (initial_delay) |
310ebb93 | 387 | usleep(initial_delay * USEC_PER_MSEC); |
67ccdecd JO |
388 | |
389 | /* | |
390 | * We need to enable counters only if: | |
391 | * - we don't have tracee (attaching to task or cpu) | |
392 | * - we have initial delay configured | |
393 | */ | |
394 | if (!target__none(&target) || initial_delay) | |
ab46db0a | 395 | perf_evlist__enable(evsel_list); |
41191688 AK |
396 | } |
397 | ||
3df33eff MR |
398 | static void disable_counters(void) |
399 | { | |
400 | /* | |
401 | * If we don't have tracee (attaching to task or cpu), counters may | |
402 | * still be running. To get accurate group ratios, we must stop groups | |
403 | * from counting before reading their constituent counters. | |
404 | */ | |
405 | if (!target__none(&target)) | |
406 | perf_evlist__disable(evsel_list); | |
407 | } | |
408 | ||
f33cbe72 | 409 | static volatile int workload_exec_errno; |
6af206fd ACM |
410 | |
411 | /* | |
412 | * perf_evlist__prepare_workload will send a SIGUSR1 | |
413 | * if the fork fails, since we asked by setting its | |
414 | * want_signal to true. | |
415 | */ | |
f33cbe72 ACM |
416 | static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, |
417 | void *ucontext __maybe_unused) | |
6af206fd | 418 | { |
f33cbe72 | 419 | workload_exec_errno = info->si_value.sival_int; |
6af206fd ACM |
420 | } |
421 | ||
7b60a7e3 JO |
422 | static bool has_unit(struct perf_evsel *counter) |
423 | { | |
424 | return counter->unit && *counter->unit; | |
425 | } | |
426 | ||
427 | static bool has_scale(struct perf_evsel *counter) | |
428 | { | |
429 | return counter->scale != 1; | |
430 | } | |
431 | ||
664c98d4 | 432 | static int perf_stat_synthesize_config(bool is_pipe) |
8b99b1a4 | 433 | { |
7b60a7e3 | 434 | struct perf_evsel *counter; |
8b99b1a4 JO |
435 | int err; |
436 | ||
664c98d4 JO |
437 | if (is_pipe) { |
438 | err = perf_event__synthesize_attrs(NULL, perf_stat.session, | |
439 | process_synthesized_event); | |
440 | if (err < 0) { | |
441 | pr_err("Couldn't synthesize attrs.\n"); | |
442 | return err; | |
443 | } | |
444 | } | |
445 | ||
7b60a7e3 JO |
446 | /* |
447 | * Synthesize other events stuff not carried within | |
448 | * attr event - unit, scale, name | |
449 | */ | |
e5cadb93 | 450 | evlist__for_each_entry(evsel_list, counter) { |
7b60a7e3 JO |
451 | if (!counter->supported) |
452 | continue; | |
453 | ||
454 | /* | |
455 | * Synthesize unit and scale only if it's defined. | |
456 | */ | |
457 | if (has_unit(counter)) { | |
458 | err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event); | |
459 | if (err < 0) { | |
460 | pr_err("Couldn't synthesize evsel unit.\n"); | |
461 | return err; | |
462 | } | |
463 | } | |
464 | ||
465 | if (has_scale(counter)) { | |
466 | err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event); | |
467 | if (err < 0) { | |
468 | pr_err("Couldn't synthesize evsel scale.\n"); | |
469 | return err; | |
470 | } | |
471 | } | |
472 | ||
473 | if (counter->own_cpus) { | |
474 | err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event); | |
475 | if (err < 0) { | |
476 | pr_err("Couldn't synthesize evsel scale.\n"); | |
477 | return err; | |
478 | } | |
479 | } | |
480 | ||
481 | /* | |
482 | * Name is needed only for pipe output, | |
483 | * perf.data carries event names. | |
484 | */ | |
485 | if (is_pipe) { | |
486 | err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event); | |
487 | if (err < 0) { | |
488 | pr_err("Couldn't synthesize evsel name.\n"); | |
489 | return err; | |
490 | } | |
491 | } | |
492 | } | |
493 | ||
8b99b1a4 JO |
494 | err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads, |
495 | process_synthesized_event, | |
496 | NULL); | |
497 | if (err < 0) { | |
498 | pr_err("Couldn't synthesize thread map.\n"); | |
499 | return err; | |
500 | } | |
501 | ||
502 | err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus, | |
503 | process_synthesized_event, NULL); | |
504 | if (err < 0) { | |
505 | pr_err("Couldn't synthesize thread map.\n"); | |
506 | return err; | |
507 | } | |
508 | ||
509 | err = perf_event__synthesize_stat_config(NULL, &stat_config, | |
510 | process_synthesized_event, NULL); | |
511 | if (err < 0) { | |
512 | pr_err("Couldn't synthesize config.\n"); | |
513 | return err; | |
514 | } | |
515 | ||
516 | return 0; | |
517 | } | |
518 | ||
2af4646d JO |
519 | #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y)) |
520 | ||
521 | static int __store_counter_ids(struct perf_evsel *counter, | |
522 | struct cpu_map *cpus, | |
523 | struct thread_map *threads) | |
524 | { | |
525 | int cpu, thread; | |
526 | ||
527 | for (cpu = 0; cpu < cpus->nr; cpu++) { | |
528 | for (thread = 0; thread < threads->nr; thread++) { | |
529 | int fd = FD(counter, cpu, thread); | |
530 | ||
531 | if (perf_evlist__id_add_fd(evsel_list, counter, | |
532 | cpu, thread, fd) < 0) | |
533 | return -1; | |
534 | } | |
535 | } | |
536 | ||
537 | return 0; | |
538 | } | |
539 | ||
540 | static int store_counter_ids(struct perf_evsel *counter) | |
541 | { | |
542 | struct cpu_map *cpus = counter->cpus; | |
543 | struct thread_map *threads = counter->threads; | |
544 | ||
545 | if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr)) | |
546 | return -ENOMEM; | |
547 | ||
548 | return __store_counter_ids(counter, cpus, threads); | |
549 | } | |
550 | ||
acf28922 | 551 | static int __run_perf_stat(int argc, const char **argv) |
42202dd5 | 552 | { |
ec0d3d1f | 553 | int interval = stat_config.interval; |
d6195a6a | 554 | char msg[BUFSIZ]; |
42202dd5 | 555 | unsigned long long t0, t1; |
cac21425 | 556 | struct perf_evsel *counter; |
13370a9b | 557 | struct timespec ts; |
410136f5 | 558 | size_t l; |
42202dd5 | 559 | int status = 0; |
6be2850e | 560 | const bool forks = (argc > 0); |
664c98d4 | 561 | bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false; |
5d8bb1ec | 562 | struct perf_evsel_config_term *err_term; |
42202dd5 | 563 | |
13370a9b | 564 | if (interval) { |
310ebb93 ACM |
565 | ts.tv_sec = interval / USEC_PER_MSEC; |
566 | ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC; | |
13370a9b SE |
567 | } else { |
568 | ts.tv_sec = 1; | |
569 | ts.tv_nsec = 0; | |
570 | } | |
571 | ||
60666c63 | 572 | if (forks) { |
664c98d4 | 573 | if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe, |
735f7e0b | 574 | workload_exec_failed_signal) < 0) { |
acf28922 NK |
575 | perror("failed to prepare workload"); |
576 | return -1; | |
60666c63 | 577 | } |
d20a47e7 | 578 | child_pid = evsel_list->workload.pid; |
051ae7f7 PM |
579 | } |
580 | ||
6a4bb04c | 581 | if (group) |
63dab225 | 582 | perf_evlist__set_leader(evsel_list); |
6a4bb04c | 583 | |
e5cadb93 | 584 | evlist__for_each_entry(evsel_list, counter) { |
42ef8a78 | 585 | try_again: |
cac21425 | 586 | if (create_perf_stat_counter(counter) < 0) { |
979987a5 DA |
587 | /* |
588 | * PPC returns ENXIO for HW counters until 2.6.37 | |
589 | * (behavior changed with commit b0a873e). | |
590 | */ | |
38f6ae1e | 591 | if (errno == EINVAL || errno == ENOSYS || |
979987a5 DA |
592 | errno == ENOENT || errno == EOPNOTSUPP || |
593 | errno == ENXIO) { | |
bb963e16 | 594 | if (verbose > 0) |
c63ca0c0 | 595 | ui__warning("%s event is not supported by the kernel.\n", |
7289f83c | 596 | perf_evsel__name(counter)); |
2cee77c4 | 597 | counter->supported = false; |
cb5ef600 KL |
598 | |
599 | if ((counter->leader != counter) || | |
600 | !(counter->leader->nr_members > 1)) | |
601 | continue; | |
42ef8a78 | 602 | } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) { |
bb963e16 | 603 | if (verbose > 0) |
42ef8a78 ACM |
604 | ui__warning("%s\n", msg); |
605 | goto try_again; | |
606 | } | |
ede70290 | 607 | |
56e52e85 ACM |
608 | perf_evsel__open_strerror(counter, &target, |
609 | errno, msg, sizeof(msg)); | |
610 | ui__error("%s\n", msg); | |
611 | ||
48290609 ACM |
612 | if (child_pid != -1) |
613 | kill(child_pid, SIGTERM); | |
fceda7fe | 614 | |
48290609 ACM |
615 | return -1; |
616 | } | |
2cee77c4 | 617 | counter->supported = true; |
410136f5 SE |
618 | |
619 | l = strlen(counter->unit); | |
620 | if (l > unit_width) | |
621 | unit_width = l; | |
2af4646d JO |
622 | |
623 | if (STAT_RECORD && store_counter_ids(counter)) | |
624 | return -1; | |
084ab9f8 | 625 | } |
42202dd5 | 626 | |
23d4aad4 ACM |
627 | if (perf_evlist__apply_filters(evsel_list, &counter)) { |
628 | error("failed to set filter \"%s\" on event %s with %d (%s)\n", | |
629 | counter->filter, perf_evsel__name(counter), errno, | |
c8b5f2c9 | 630 | str_error_r(errno, msg, sizeof(msg))); |
cfd748ae FW |
631 | return -1; |
632 | } | |
633 | ||
5d8bb1ec MP |
634 | if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) { |
635 | error("failed to set config \"%s\" on event %s with %d (%s)\n", | |
636 | err_term->val.drv_cfg, perf_evsel__name(counter), errno, | |
637 | str_error_r(errno, msg, sizeof(msg))); | |
638 | return -1; | |
639 | } | |
640 | ||
4979d0c7 JO |
641 | if (STAT_RECORD) { |
642 | int err, fd = perf_data_file__fd(&perf_stat.file); | |
643 | ||
664c98d4 JO |
644 | if (is_pipe) { |
645 | err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file)); | |
646 | } else { | |
647 | err = perf_session__write_header(perf_stat.session, evsel_list, | |
648 | fd, false); | |
649 | } | |
650 | ||
4979d0c7 JO |
651 | if (err < 0) |
652 | return err; | |
8b99b1a4 | 653 | |
664c98d4 | 654 | err = perf_stat_synthesize_config(is_pipe); |
8b99b1a4 JO |
655 | if (err < 0) |
656 | return err; | |
4979d0c7 JO |
657 | } |
658 | ||
42202dd5 IM |
659 | /* |
660 | * Enable counters and exec the command: | |
661 | */ | |
662 | t0 = rdclock(); | |
13370a9b | 663 | clock_gettime(CLOCK_MONOTONIC, &ref_time); |
42202dd5 | 664 | |
60666c63 | 665 | if (forks) { |
acf28922 | 666 | perf_evlist__start_workload(evsel_list); |
67ccdecd | 667 | enable_counters(); |
acf28922 | 668 | |
13370a9b SE |
669 | if (interval) { |
670 | while (!waitpid(child_pid, &status, WNOHANG)) { | |
671 | nanosleep(&ts, NULL); | |
ba411a95 | 672 | process_interval(); |
13370a9b SE |
673 | } |
674 | } | |
60666c63 | 675 | wait(&status); |
6af206fd | 676 | |
f33cbe72 | 677 | if (workload_exec_errno) { |
c8b5f2c9 | 678 | const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg)); |
f33cbe72 | 679 | pr_err("Workload failed: %s\n", emsg); |
6af206fd | 680 | return -1; |
f33cbe72 | 681 | } |
6af206fd | 682 | |
33e49ea7 AK |
683 | if (WIFSIGNALED(status)) |
684 | psignal(WTERMSIG(status), argv[0]); | |
60666c63 | 685 | } else { |
67ccdecd | 686 | enable_counters(); |
13370a9b SE |
687 | while (!done) { |
688 | nanosleep(&ts, NULL); | |
689 | if (interval) | |
ba411a95 | 690 | process_interval(); |
13370a9b | 691 | } |
60666c63 | 692 | } |
42202dd5 | 693 | |
3df33eff MR |
694 | disable_counters(); |
695 | ||
42202dd5 IM |
696 | t1 = rdclock(); |
697 | ||
9e9772c4 | 698 | update_stats(&walltime_nsecs_stats, t1 - t0); |
42202dd5 | 699 | |
3df33eff MR |
700 | /* |
701 | * Closing a group leader splits the group, and as we only disable | |
702 | * group leaders, results in remaining events becoming enabled. To | |
703 | * avoid arbitrary skew, we must read all counters before closing any | |
704 | * group leaders. | |
705 | */ | |
706 | read_counters(); | |
707 | perf_evlist__close(evsel_list); | |
c52b12ed | 708 | |
42202dd5 IM |
709 | return WEXITSTATUS(status); |
710 | } | |
711 | ||
41cde476 | 712 | static int run_perf_stat(int argc, const char **argv) |
1f16c575 PZ |
713 | { |
714 | int ret; | |
715 | ||
716 | if (pre_cmd) { | |
717 | ret = system(pre_cmd); | |
718 | if (ret) | |
719 | return ret; | |
720 | } | |
721 | ||
722 | if (sync_run) | |
723 | sync(); | |
724 | ||
725 | ret = __run_perf_stat(argc, argv); | |
726 | if (ret) | |
727 | return ret; | |
728 | ||
729 | if (post_cmd) { | |
730 | ret = system(post_cmd); | |
731 | if (ret) | |
732 | return ret; | |
733 | } | |
734 | ||
735 | return ret; | |
736 | } | |
737 | ||
d73515c0 AK |
738 | static void print_running(u64 run, u64 ena) |
739 | { | |
740 | if (csv_output) { | |
5821522e | 741 | fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f", |
d73515c0 AK |
742 | csv_sep, |
743 | run, | |
744 | csv_sep, | |
745 | ena ? 100.0 * run / ena : 100.0); | |
746 | } else if (run != ena) { | |
5821522e | 747 | fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena); |
d73515c0 AK |
748 | } |
749 | } | |
750 | ||
f99844cb IM |
751 | static void print_noise_pct(double total, double avg) |
752 | { | |
0007ecea | 753 | double pct = rel_stddev_stats(total, avg); |
f99844cb | 754 | |
3ae9a34d | 755 | if (csv_output) |
5821522e | 756 | fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct); |
a1bca6cc | 757 | else if (pct) |
5821522e | 758 | fprintf(stat_config.output, " ( +-%6.2f%% )", pct); |
f99844cb IM |
759 | } |
760 | ||
69aad6f1 | 761 | static void print_noise(struct perf_evsel *evsel, double avg) |
42202dd5 | 762 | { |
581cc8a2 | 763 | struct perf_stat_evsel *ps; |
69aad6f1 | 764 | |
849abde9 PZ |
765 | if (run_count == 1) |
766 | return; | |
767 | ||
69aad6f1 | 768 | ps = evsel->priv; |
f99844cb | 769 | print_noise_pct(stddev_stats(&ps->res_stats[0]), avg); |
42202dd5 IM |
770 | } |
771 | ||
12c08a9f | 772 | static void aggr_printout(struct perf_evsel *evsel, int id, int nr) |
44175b6f | 773 | { |
421a50f3 | 774 | switch (stat_config.aggr_mode) { |
12c08a9f | 775 | case AGGR_CORE: |
5821522e | 776 | fprintf(stat_config.output, "S%d-C%*d%s%*d%s", |
12c08a9f SE |
777 | cpu_map__id_to_socket(id), |
778 | csv_output ? 0 : -8, | |
779 | cpu_map__id_to_cpu(id), | |
780 | csv_sep, | |
781 | csv_output ? 0 : 4, | |
782 | nr, | |
783 | csv_sep); | |
784 | break; | |
86ee6e18 | 785 | case AGGR_SOCKET: |
5821522e | 786 | fprintf(stat_config.output, "S%*d%s%*d%s", |
d7e7a451 | 787 | csv_output ? 0 : -5, |
12c08a9f | 788 | id, |
d7e7a451 SE |
789 | csv_sep, |
790 | csv_output ? 0 : 4, | |
791 | nr, | |
792 | csv_sep); | |
86ee6e18 SE |
793 | break; |
794 | case AGGR_NONE: | |
5821522e | 795 | fprintf(stat_config.output, "CPU%*d%s", |
d7470b6a | 796 | csv_output ? 0 : -4, |
12c08a9f | 797 | perf_evsel__cpus(evsel)->map[id], csv_sep); |
86ee6e18 | 798 | break; |
32b8af82 | 799 | case AGGR_THREAD: |
5821522e | 800 | fprintf(stat_config.output, "%*s-%*d%s", |
32b8af82 JO |
801 | csv_output ? 0 : 16, |
802 | thread_map__comm(evsel->threads, id), | |
803 | csv_output ? 0 : -8, | |
804 | thread_map__pid(evsel->threads, id), | |
805 | csv_sep); | |
806 | break; | |
86ee6e18 | 807 | case AGGR_GLOBAL: |
208df99e | 808 | case AGGR_UNSET: |
86ee6e18 SE |
809 | default: |
810 | break; | |
811 | } | |
812 | } | |
813 | ||
140aeadc AK |
814 | struct outstate { |
815 | FILE *fh; | |
816 | bool newline; | |
f9483392 | 817 | const char *prefix; |
92a61f64 | 818 | int nfields; |
44d49a60 AK |
819 | int id, nr; |
820 | struct perf_evsel *evsel; | |
140aeadc AK |
821 | }; |
822 | ||
823 | #define METRIC_LEN 35 | |
824 | ||
825 | static void new_line_std(void *ctx) | |
826 | { | |
827 | struct outstate *os = ctx; | |
828 | ||
829 | os->newline = true; | |
830 | } | |
831 | ||
832 | static void do_new_line_std(struct outstate *os) | |
833 | { | |
834 | fputc('\n', os->fh); | |
f9483392 | 835 | fputs(os->prefix, os->fh); |
44d49a60 | 836 | aggr_printout(os->evsel, os->id, os->nr); |
140aeadc AK |
837 | if (stat_config.aggr_mode == AGGR_NONE) |
838 | fprintf(os->fh, " "); | |
140aeadc AK |
839 | fprintf(os->fh, " "); |
840 | } | |
841 | ||
842 | static void print_metric_std(void *ctx, const char *color, const char *fmt, | |
843 | const char *unit, double val) | |
844 | { | |
845 | struct outstate *os = ctx; | |
846 | FILE *out = os->fh; | |
847 | int n; | |
848 | bool newline = os->newline; | |
849 | ||
850 | os->newline = false; | |
851 | ||
852 | if (unit == NULL || fmt == NULL) { | |
853 | fprintf(out, "%-*s", METRIC_LEN, ""); | |
854 | return; | |
855 | } | |
856 | ||
857 | if (newline) | |
858 | do_new_line_std(os); | |
859 | ||
860 | n = fprintf(out, " # "); | |
861 | if (color) | |
862 | n += color_fprintf(out, color, fmt, val); | |
863 | else | |
864 | n += fprintf(out, fmt, val); | |
865 | fprintf(out, " %-*s", METRIC_LEN - n - 1, unit); | |
866 | } | |
867 | ||
92a61f64 AK |
868 | static void new_line_csv(void *ctx) |
869 | { | |
870 | struct outstate *os = ctx; | |
871 | int i; | |
872 | ||
873 | fputc('\n', os->fh); | |
874 | if (os->prefix) | |
875 | fprintf(os->fh, "%s%s", os->prefix, csv_sep); | |
44d49a60 | 876 | aggr_printout(os->evsel, os->id, os->nr); |
92a61f64 AK |
877 | for (i = 0; i < os->nfields; i++) |
878 | fputs(csv_sep, os->fh); | |
879 | } | |
880 | ||
881 | static void print_metric_csv(void *ctx, | |
882 | const char *color __maybe_unused, | |
883 | const char *fmt, const char *unit, double val) | |
884 | { | |
885 | struct outstate *os = ctx; | |
886 | FILE *out = os->fh; | |
887 | char buf[64], *vals, *ends; | |
888 | ||
889 | if (unit == NULL || fmt == NULL) { | |
890 | fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep); | |
891 | return; | |
892 | } | |
893 | snprintf(buf, sizeof(buf), fmt, val); | |
b07c40df | 894 | ends = vals = ltrim(buf); |
92a61f64 AK |
895 | while (isdigit(*ends) || *ends == '.') |
896 | ends++; | |
897 | *ends = 0; | |
898 | while (isspace(*unit)) | |
899 | unit++; | |
900 | fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit); | |
901 | } | |
902 | ||
54b50916 AK |
903 | #define METRIC_ONLY_LEN 20 |
904 | ||
905 | /* Filter out some columns that don't work well in metrics only mode */ | |
906 | ||
907 | static bool valid_only_metric(const char *unit) | |
908 | { | |
909 | if (!unit) | |
910 | return false; | |
911 | if (strstr(unit, "/sec") || | |
912 | strstr(unit, "hz") || | |
913 | strstr(unit, "Hz") || | |
914 | strstr(unit, "CPUs utilized")) | |
915 | return false; | |
916 | return true; | |
917 | } | |
918 | ||
919 | static const char *fixunit(char *buf, struct perf_evsel *evsel, | |
920 | const char *unit) | |
921 | { | |
922 | if (!strncmp(unit, "of all", 6)) { | |
923 | snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel), | |
924 | unit); | |
925 | return buf; | |
926 | } | |
927 | return unit; | |
928 | } | |
929 | ||
930 | static void print_metric_only(void *ctx, const char *color, const char *fmt, | |
931 | const char *unit, double val) | |
932 | { | |
933 | struct outstate *os = ctx; | |
934 | FILE *out = os->fh; | |
935 | int n; | |
936 | char buf[1024]; | |
937 | unsigned mlen = METRIC_ONLY_LEN; | |
938 | ||
939 | if (!valid_only_metric(unit)) | |
940 | return; | |
941 | unit = fixunit(buf, os->evsel, unit); | |
942 | if (color) | |
943 | n = color_fprintf(out, color, fmt, val); | |
944 | else | |
945 | n = fprintf(out, fmt, val); | |
946 | if (n > METRIC_ONLY_LEN) | |
947 | n = METRIC_ONLY_LEN; | |
948 | if (mlen < strlen(unit)) | |
949 | mlen = strlen(unit) + 1; | |
950 | fprintf(out, "%*s", mlen - n, ""); | |
951 | } | |
952 | ||
953 | static void print_metric_only_csv(void *ctx, const char *color __maybe_unused, | |
954 | const char *fmt, | |
955 | const char *unit, double val) | |
956 | { | |
957 | struct outstate *os = ctx; | |
958 | FILE *out = os->fh; | |
959 | char buf[64], *vals, *ends; | |
960 | char tbuf[1024]; | |
961 | ||
962 | if (!valid_only_metric(unit)) | |
963 | return; | |
964 | unit = fixunit(tbuf, os->evsel, unit); | |
965 | snprintf(buf, sizeof buf, fmt, val); | |
b07c40df | 966 | ends = vals = ltrim(buf); |
54b50916 AK |
967 | while (isdigit(*ends) || *ends == '.') |
968 | ends++; | |
969 | *ends = 0; | |
970 | fprintf(out, "%s%s", vals, csv_sep); | |
971 | } | |
972 | ||
973 | static void new_line_metric(void *ctx __maybe_unused) | |
974 | { | |
975 | } | |
976 | ||
977 | static void print_metric_header(void *ctx, const char *color __maybe_unused, | |
978 | const char *fmt __maybe_unused, | |
979 | const char *unit, double val __maybe_unused) | |
980 | { | |
981 | struct outstate *os = ctx; | |
982 | char tbuf[1024]; | |
983 | ||
984 | if (!valid_only_metric(unit)) | |
985 | return; | |
986 | unit = fixunit(tbuf, os->evsel, unit); | |
987 | if (csv_output) | |
988 | fprintf(os->fh, "%s%s", unit, csv_sep); | |
989 | else | |
990 | fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit); | |
991 | } | |
992 | ||
da88c7f7 | 993 | static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg) |
86ee6e18 | 994 | { |
5821522e | 995 | FILE *output = stat_config.output; |
310ebb93 | 996 | double msecs = avg / NSEC_PER_MSEC; |
410136f5 | 997 | const char *fmt_v, *fmt_n; |
4bbe5a61 | 998 | char name[25]; |
86ee6e18 | 999 | |
410136f5 SE |
1000 | fmt_v = csv_output ? "%.6f%s" : "%18.6f%s"; |
1001 | fmt_n = csv_output ? "%s" : "%-25s"; | |
1002 | ||
da88c7f7 | 1003 | aggr_printout(evsel, id, nr); |
d7470b6a | 1004 | |
4bbe5a61 DA |
1005 | scnprintf(name, sizeof(name), "%s%s", |
1006 | perf_evsel__name(evsel), csv_output ? "" : " (msec)"); | |
410136f5 SE |
1007 | |
1008 | fprintf(output, fmt_v, msecs, csv_sep); | |
1009 | ||
1010 | if (csv_output) | |
1011 | fprintf(output, "%s%s", evsel->unit, csv_sep); | |
1012 | else | |
1013 | fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep); | |
1014 | ||
1015 | fprintf(output, fmt_n, name); | |
d7470b6a | 1016 | |
023695d9 | 1017 | if (evsel->cgrp) |
4aa9015f | 1018 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); |
44175b6f IM |
1019 | } |
1020 | ||
44d49a60 AK |
1021 | static int first_shadow_cpu(struct perf_evsel *evsel, int id) |
1022 | { | |
1023 | int i; | |
1024 | ||
1025 | if (!aggr_get_id) | |
1026 | return 0; | |
1027 | ||
1028 | if (stat_config.aggr_mode == AGGR_NONE) | |
1029 | return id; | |
1030 | ||
1031 | if (stat_config.aggr_mode == AGGR_GLOBAL) | |
1032 | return 0; | |
1033 | ||
1034 | for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) { | |
1035 | int cpu2 = perf_evsel__cpus(evsel)->map[i]; | |
1036 | ||
1037 | if (aggr_get_id(evsel_list->cpus, cpu2) == id) | |
1038 | return cpu2; | |
1039 | } | |
1040 | return 0; | |
1041 | } | |
1042 | ||
556b1fb7 JO |
1043 | static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg) |
1044 | { | |
5821522e | 1045 | FILE *output = stat_config.output; |
556b1fb7 JO |
1046 | double sc = evsel->scale; |
1047 | const char *fmt; | |
556b1fb7 JO |
1048 | |
1049 | if (csv_output) { | |
e3b03b6c | 1050 | fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s"; |
556b1fb7 JO |
1051 | } else { |
1052 | if (big_num) | |
e3b03b6c | 1053 | fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s"; |
556b1fb7 | 1054 | else |
e3b03b6c | 1055 | fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s"; |
556b1fb7 JO |
1056 | } |
1057 | ||
1058 | aggr_printout(evsel, id, nr); | |
1059 | ||
556b1fb7 JO |
1060 | fprintf(output, fmt, avg, csv_sep); |
1061 | ||
1062 | if (evsel->unit) | |
1063 | fprintf(output, "%-*s%s", | |
1064 | csv_output ? 0 : unit_width, | |
1065 | evsel->unit, csv_sep); | |
1066 | ||
1067 | fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel)); | |
1068 | ||
1069 | if (evsel->cgrp) | |
1070 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); | |
eedfcb4b | 1071 | } |
556b1fb7 | 1072 | |
f9483392 | 1073 | static void printout(int id, int nr, struct perf_evsel *counter, double uval, |
cb110f47 | 1074 | char *prefix, u64 run, u64 ena, double noise) |
eedfcb4b | 1075 | { |
140aeadc | 1076 | struct perf_stat_output_ctx out; |
f9483392 AK |
1077 | struct outstate os = { |
1078 | .fh = stat_config.output, | |
44d49a60 AK |
1079 | .prefix = prefix ? prefix : "", |
1080 | .id = id, | |
1081 | .nr = nr, | |
1082 | .evsel = counter, | |
f9483392 | 1083 | }; |
140aeadc AK |
1084 | print_metric_t pm = print_metric_std; |
1085 | void (*nl)(void *); | |
eedfcb4b | 1086 | |
54b50916 AK |
1087 | if (metric_only) { |
1088 | nl = new_line_metric; | |
1089 | if (csv_output) | |
1090 | pm = print_metric_only_csv; | |
1091 | else | |
1092 | pm = print_metric_only; | |
1093 | } else | |
1094 | nl = new_line_std; | |
eedfcb4b | 1095 | |
54b50916 | 1096 | if (csv_output && !metric_only) { |
92a61f64 AK |
1097 | static int aggr_fields[] = { |
1098 | [AGGR_GLOBAL] = 0, | |
1099 | [AGGR_THREAD] = 1, | |
1100 | [AGGR_NONE] = 1, | |
1101 | [AGGR_SOCKET] = 2, | |
1102 | [AGGR_CORE] = 2, | |
1103 | }; | |
1104 | ||
1105 | pm = print_metric_csv; | |
1106 | nl = new_line_csv; | |
1107 | os.nfields = 3; | |
1108 | os.nfields += aggr_fields[stat_config.aggr_mode]; | |
1109 | if (counter->cgrp) | |
1110 | os.nfields++; | |
1111 | } | |
b002f3bb | 1112 | if (run == 0 || ena == 0 || counter->counts->scaled == -1) { |
54b50916 AK |
1113 | if (metric_only) { |
1114 | pm(&os, NULL, "", "", 0); | |
1115 | return; | |
1116 | } | |
cb110f47 AK |
1117 | aggr_printout(counter, id, nr); |
1118 | ||
1119 | fprintf(stat_config.output, "%*s%s", | |
1120 | csv_output ? 0 : 18, | |
1121 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, | |
1122 | csv_sep); | |
1123 | ||
02d492e5 BP |
1124 | if (counter->supported) |
1125 | print_free_counters_hint = 1; | |
1126 | ||
cb110f47 AK |
1127 | fprintf(stat_config.output, "%-*s%s", |
1128 | csv_output ? 0 : unit_width, | |
1129 | counter->unit, csv_sep); | |
1130 | ||
1131 | fprintf(stat_config.output, "%*s", | |
1132 | csv_output ? 0 : -25, | |
1133 | perf_evsel__name(counter)); | |
1134 | ||
1135 | if (counter->cgrp) | |
1136 | fprintf(stat_config.output, "%s%s", | |
1137 | csv_sep, counter->cgrp->name); | |
1138 | ||
92a61f64 AK |
1139 | if (!csv_output) |
1140 | pm(&os, NULL, NULL, "", 0); | |
1141 | print_noise(counter, noise); | |
cb110f47 | 1142 | print_running(run, ena); |
92a61f64 AK |
1143 | if (csv_output) |
1144 | pm(&os, NULL, NULL, "", 0); | |
cb110f47 AK |
1145 | return; |
1146 | } | |
1147 | ||
54b50916 AK |
1148 | if (metric_only) |
1149 | /* nothing */; | |
1150 | else if (nsec_counter(counter)) | |
eedfcb4b AK |
1151 | nsec_printout(id, nr, counter, uval); |
1152 | else | |
1153 | abs_printout(id, nr, counter, uval); | |
556b1fb7 | 1154 | |
140aeadc AK |
1155 | out.print_metric = pm; |
1156 | out.new_line = nl; | |
1157 | out.ctx = &os; | |
37932c18 | 1158 | out.force_header = false; |
140aeadc | 1159 | |
54b50916 | 1160 | if (csv_output && !metric_only) { |
92a61f64 AK |
1161 | print_noise(counter, noise); |
1162 | print_running(run, ena); | |
1163 | } | |
1164 | ||
1165 | perf_stat__print_shadow_stats(counter, uval, | |
44d49a60 | 1166 | first_shadow_cpu(counter, id), |
140aeadc | 1167 | &out); |
54b50916 | 1168 | if (!csv_output && !metric_only) { |
92a61f64 AK |
1169 | print_noise(counter, noise); |
1170 | print_running(run, ena); | |
1171 | } | |
556b1fb7 JO |
1172 | } |
1173 | ||
44d49a60 AK |
1174 | static void aggr_update_shadow(void) |
1175 | { | |
1176 | int cpu, s2, id, s; | |
1177 | u64 val; | |
1178 | struct perf_evsel *counter; | |
1179 | ||
1180 | for (s = 0; s < aggr_map->nr; s++) { | |
1181 | id = aggr_map->map[s]; | |
e5cadb93 | 1182 | evlist__for_each_entry(evsel_list, counter) { |
44d49a60 AK |
1183 | val = 0; |
1184 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { | |
1185 | s2 = aggr_get_id(evsel_list->cpus, cpu); | |
1186 | if (s2 != id) | |
1187 | continue; | |
1188 | val += perf_counts(counter->counts, cpu, 0)->val; | |
1189 | } | |
1190 | val = val * counter->scale; | |
1191 | perf_stat__update_shadow_stats(counter, &val, | |
1192 | first_shadow_cpu(counter, id)); | |
1193 | } | |
1194 | } | |
1195 | } | |
1196 | ||
430daf2d | 1197 | static void collect_all_aliases(struct perf_evsel *counter, |
fbe51fba AK |
1198 | void (*cb)(struct perf_evsel *counter, void *data, |
1199 | bool first), | |
1200 | void *data) | |
1201 | { | |
430daf2d AK |
1202 | struct perf_evsel *alias; |
1203 | ||
1204 | alias = list_prepare_entry(counter, &(evsel_list->entries), node); | |
1205 | list_for_each_entry_continue (alias, &evsel_list->entries, node) { | |
1206 | if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) || | |
1207 | alias->scale != counter->scale || | |
1208 | alias->cgrp != counter->cgrp || | |
1209 | strcmp(alias->unit, counter->unit) || | |
1210 | nsec_counter(alias) != nsec_counter(counter)) | |
1211 | break; | |
1212 | alias->merged_stat = true; | |
1213 | cb(alias, data, false); | |
1214 | } | |
1215 | } | |
1216 | ||
1217 | static bool collect_data(struct perf_evsel *counter, | |
1218 | void (*cb)(struct perf_evsel *counter, void *data, | |
1219 | bool first), | |
1220 | void *data) | |
1221 | { | |
1222 | if (counter->merged_stat) | |
1223 | return false; | |
fbe51fba | 1224 | cb(counter, data, true); |
430daf2d AK |
1225 | if (!no_merge) |
1226 | collect_all_aliases(counter, cb, data); | |
1227 | return true; | |
fbe51fba AK |
1228 | } |
1229 | ||
1230 | struct aggr_data { | |
1231 | u64 ena, run, val; | |
1232 | int id; | |
1233 | int nr; | |
1234 | int cpu; | |
1235 | }; | |
1236 | ||
1237 | static void aggr_cb(struct perf_evsel *counter, void *data, bool first) | |
1238 | { | |
1239 | struct aggr_data *ad = data; | |
1240 | int cpu, s2; | |
1241 | ||
1242 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { | |
1243 | struct perf_counts_values *counts; | |
1244 | ||
1245 | s2 = aggr_get_id(perf_evsel__cpus(counter), cpu); | |
1246 | if (s2 != ad->id) | |
1247 | continue; | |
1248 | if (first) | |
1249 | ad->nr++; | |
1250 | counts = perf_counts(counter->counts, cpu, 0); | |
b4229e9d AK |
1251 | /* |
1252 | * When any result is bad, make them all to give | |
1253 | * consistent output in interval mode. | |
1254 | */ | |
1255 | if (counts->ena == 0 || counts->run == 0 || | |
1256 | counter->counts->scaled == -1) { | |
1257 | ad->ena = 0; | |
1258 | ad->run = 0; | |
1259 | break; | |
1260 | } | |
fbe51fba AK |
1261 | ad->val += counts->val; |
1262 | ad->ena += counts->ena; | |
1263 | ad->run += counts->run; | |
1264 | } | |
1265 | } | |
1266 | ||
86ee6e18 | 1267 | static void print_aggr(char *prefix) |
d7e7a451 | 1268 | { |
5821522e | 1269 | FILE *output = stat_config.output; |
d7e7a451 | 1270 | struct perf_evsel *counter; |
fbe51fba | 1271 | int s, id, nr; |
410136f5 | 1272 | double uval; |
d7e7a451 | 1273 | u64 ena, run, val; |
54b50916 | 1274 | bool first; |
d7e7a451 | 1275 | |
86ee6e18 | 1276 | if (!(aggr_map || aggr_get_id)) |
d7e7a451 SE |
1277 | return; |
1278 | ||
44d49a60 AK |
1279 | aggr_update_shadow(); |
1280 | ||
54b50916 AK |
1281 | /* |
1282 | * With metric_only everything is on a single line. | |
1283 | * Without each counter has its own line. | |
1284 | */ | |
86ee6e18 | 1285 | for (s = 0; s < aggr_map->nr; s++) { |
fbe51fba | 1286 | struct aggr_data ad; |
54b50916 AK |
1287 | if (prefix && metric_only) |
1288 | fprintf(output, "%s", prefix); | |
1289 | ||
fbe51fba | 1290 | ad.id = id = aggr_map->map[s]; |
54b50916 | 1291 | first = true; |
e5cadb93 | 1292 | evlist__for_each_entry(evsel_list, counter) { |
fbe51fba AK |
1293 | ad.val = ad.ena = ad.run = 0; |
1294 | ad.nr = 0; | |
430daf2d AK |
1295 | if (!collect_data(counter, aggr_cb, &ad)) |
1296 | continue; | |
fbe51fba AK |
1297 | nr = ad.nr; |
1298 | ena = ad.ena; | |
1299 | run = ad.run; | |
1300 | val = ad.val; | |
54b50916 AK |
1301 | if (first && metric_only) { |
1302 | first = false; | |
1303 | aggr_printout(counter, id, nr); | |
1304 | } | |
1305 | if (prefix && !metric_only) | |
d7e7a451 SE |
1306 | fprintf(output, "%s", prefix); |
1307 | ||
410136f5 | 1308 | uval = val * counter->scale; |
cb110f47 | 1309 | printout(id, nr, counter, uval, prefix, run, ena, 1.0); |
54b50916 AK |
1310 | if (!metric_only) |
1311 | fputc('\n', output); | |
d7e7a451 | 1312 | } |
54b50916 AK |
1313 | if (metric_only) |
1314 | fputc('\n', output); | |
d7e7a451 SE |
1315 | } |
1316 | } | |
1317 | ||
32b8af82 JO |
1318 | static void print_aggr_thread(struct perf_evsel *counter, char *prefix) |
1319 | { | |
5821522e | 1320 | FILE *output = stat_config.output; |
32b8af82 JO |
1321 | int nthreads = thread_map__nr(counter->threads); |
1322 | int ncpus = cpu_map__nr(counter->cpus); | |
1323 | int cpu, thread; | |
1324 | double uval; | |
1325 | ||
1326 | for (thread = 0; thread < nthreads; thread++) { | |
1327 | u64 ena = 0, run = 0, val = 0; | |
1328 | ||
1329 | for (cpu = 0; cpu < ncpus; cpu++) { | |
1330 | val += perf_counts(counter->counts, cpu, thread)->val; | |
1331 | ena += perf_counts(counter->counts, cpu, thread)->ena; | |
1332 | run += perf_counts(counter->counts, cpu, thread)->run; | |
1333 | } | |
1334 | ||
1335 | if (prefix) | |
1336 | fprintf(output, "%s", prefix); | |
1337 | ||
1338 | uval = val * counter->scale; | |
cb110f47 | 1339 | printout(thread, 0, counter, uval, prefix, run, ena, 1.0); |
32b8af82 JO |
1340 | fputc('\n', output); |
1341 | } | |
1342 | } | |
1343 | ||
fbe51fba AK |
1344 | struct caggr_data { |
1345 | double avg, avg_enabled, avg_running; | |
1346 | }; | |
1347 | ||
1348 | static void counter_aggr_cb(struct perf_evsel *counter, void *data, | |
1349 | bool first __maybe_unused) | |
1350 | { | |
1351 | struct caggr_data *cd = data; | |
1352 | struct perf_stat_evsel *ps = counter->priv; | |
1353 | ||
1354 | cd->avg += avg_stats(&ps->res_stats[0]); | |
1355 | cd->avg_enabled += avg_stats(&ps->res_stats[1]); | |
1356 | cd->avg_running += avg_stats(&ps->res_stats[2]); | |
1357 | } | |
1358 | ||
2996f5dd IM |
1359 | /* |
1360 | * Print out the results of a single counter: | |
f5b4a9c3 | 1361 | * aggregated counts in system-wide mode |
2996f5dd | 1362 | */ |
13370a9b | 1363 | static void print_counter_aggr(struct perf_evsel *counter, char *prefix) |
2996f5dd | 1364 | { |
5821522e | 1365 | FILE *output = stat_config.output; |
410136f5 | 1366 | double uval; |
fbe51fba | 1367 | struct caggr_data cd = { .avg = 0.0 }; |
d73515c0 | 1368 | |
430daf2d AK |
1369 | if (!collect_data(counter, counter_aggr_cb, &cd)) |
1370 | return; | |
2996f5dd | 1371 | |
54b50916 | 1372 | if (prefix && !metric_only) |
13370a9b SE |
1373 | fprintf(output, "%s", prefix); |
1374 | ||
fbe51fba AK |
1375 | uval = cd.avg * counter->scale; |
1376 | printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled, cd.avg); | |
54b50916 AK |
1377 | if (!metric_only) |
1378 | fprintf(output, "\n"); | |
c04f5e5d IM |
1379 | } |
1380 | ||
fbe51fba AK |
1381 | static void counter_cb(struct perf_evsel *counter, void *data, |
1382 | bool first __maybe_unused) | |
1383 | { | |
1384 | struct aggr_data *ad = data; | |
1385 | ||
1386 | ad->val += perf_counts(counter->counts, ad->cpu, 0)->val; | |
1387 | ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena; | |
1388 | ad->run += perf_counts(counter->counts, ad->cpu, 0)->run; | |
1389 | } | |
1390 | ||
f5b4a9c3 SE |
1391 | /* |
1392 | * Print out the results of a single counter: | |
1393 | * does not use aggregated count in system-wide | |
1394 | */ | |
13370a9b | 1395 | static void print_counter(struct perf_evsel *counter, char *prefix) |
f5b4a9c3 | 1396 | { |
5821522e | 1397 | FILE *output = stat_config.output; |
f5b4a9c3 | 1398 | u64 ena, run, val; |
410136f5 | 1399 | double uval; |
f5b4a9c3 SE |
1400 | int cpu; |
1401 | ||
7ae92e74 | 1402 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { |
fbe51fba AK |
1403 | struct aggr_data ad = { .cpu = cpu }; |
1404 | ||
430daf2d AK |
1405 | if (!collect_data(counter, counter_cb, &ad)) |
1406 | return; | |
fbe51fba AK |
1407 | val = ad.val; |
1408 | ena = ad.ena; | |
1409 | run = ad.run; | |
13370a9b SE |
1410 | |
1411 | if (prefix) | |
1412 | fprintf(output, "%s", prefix); | |
1413 | ||
410136f5 | 1414 | uval = val * counter->scale; |
cb110f47 | 1415 | printout(cpu, 0, counter, uval, prefix, run, ena, 1.0); |
f5b4a9c3 | 1416 | |
4aa9015f | 1417 | fputc('\n', output); |
f5b4a9c3 SE |
1418 | } |
1419 | } | |
1420 | ||
206cab65 AK |
1421 | static void print_no_aggr_metric(char *prefix) |
1422 | { | |
1423 | int cpu; | |
1424 | int nrcpus = 0; | |
1425 | struct perf_evsel *counter; | |
1426 | u64 ena, run, val; | |
1427 | double uval; | |
1428 | ||
1429 | nrcpus = evsel_list->cpus->nr; | |
1430 | for (cpu = 0; cpu < nrcpus; cpu++) { | |
1431 | bool first = true; | |
1432 | ||
1433 | if (prefix) | |
1434 | fputs(prefix, stat_config.output); | |
e5cadb93 | 1435 | evlist__for_each_entry(evsel_list, counter) { |
206cab65 AK |
1436 | if (first) { |
1437 | aggr_printout(counter, cpu, 0); | |
1438 | first = false; | |
1439 | } | |
1440 | val = perf_counts(counter->counts, cpu, 0)->val; | |
1441 | ena = perf_counts(counter->counts, cpu, 0)->ena; | |
1442 | run = perf_counts(counter->counts, cpu, 0)->run; | |
1443 | ||
1444 | uval = val * counter->scale; | |
1445 | printout(cpu, 0, counter, uval, prefix, run, ena, 1.0); | |
1446 | } | |
1447 | fputc('\n', stat_config.output); | |
1448 | } | |
1449 | } | |
1450 | ||
54b50916 AK |
1451 | static int aggr_header_lens[] = { |
1452 | [AGGR_CORE] = 18, | |
1453 | [AGGR_SOCKET] = 12, | |
206cab65 | 1454 | [AGGR_NONE] = 6, |
54b50916 AK |
1455 | [AGGR_THREAD] = 24, |
1456 | [AGGR_GLOBAL] = 0, | |
1457 | }; | |
1458 | ||
c51fd639 AK |
1459 | static const char *aggr_header_csv[] = { |
1460 | [AGGR_CORE] = "core,cpus,", | |
1461 | [AGGR_SOCKET] = "socket,cpus", | |
1462 | [AGGR_NONE] = "cpu,", | |
1463 | [AGGR_THREAD] = "comm-pid,", | |
1464 | [AGGR_GLOBAL] = "" | |
1465 | }; | |
1466 | ||
41c8ca2a | 1467 | static void print_metric_headers(const char *prefix, bool no_indent) |
54b50916 AK |
1468 | { |
1469 | struct perf_stat_output_ctx out; | |
1470 | struct perf_evsel *counter; | |
1471 | struct outstate os = { | |
1472 | .fh = stat_config.output | |
1473 | }; | |
1474 | ||
1475 | if (prefix) | |
1476 | fprintf(stat_config.output, "%s", prefix); | |
1477 | ||
41c8ca2a | 1478 | if (!csv_output && !no_indent) |
54b50916 AK |
1479 | fprintf(stat_config.output, "%*s", |
1480 | aggr_header_lens[stat_config.aggr_mode], ""); | |
c51fd639 AK |
1481 | if (csv_output) { |
1482 | if (stat_config.interval) | |
1483 | fputs("time,", stat_config.output); | |
1484 | fputs(aggr_header_csv[stat_config.aggr_mode], | |
1485 | stat_config.output); | |
1486 | } | |
54b50916 AK |
1487 | |
1488 | /* Print metrics headers only */ | |
e5cadb93 | 1489 | evlist__for_each_entry(evsel_list, counter) { |
54b50916 AK |
1490 | os.evsel = counter; |
1491 | out.ctx = &os; | |
1492 | out.print_metric = print_metric_header; | |
1493 | out.new_line = new_line_metric; | |
37932c18 | 1494 | out.force_header = true; |
54b50916 AK |
1495 | os.evsel = counter; |
1496 | perf_stat__print_shadow_stats(counter, 0, | |
1497 | 0, | |
1498 | &out); | |
1499 | } | |
1500 | fputc('\n', stat_config.output); | |
1501 | } | |
1502 | ||
d4f63a47 JO |
1503 | static void print_interval(char *prefix, struct timespec *ts) |
1504 | { | |
5821522e | 1505 | FILE *output = stat_config.output; |
d4f63a47 JO |
1506 | static int num_print_interval; |
1507 | ||
1508 | sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep); | |
1509 | ||
41c8ca2a | 1510 | if (num_print_interval == 0 && !csv_output) { |
421a50f3 | 1511 | switch (stat_config.aggr_mode) { |
d4f63a47 | 1512 | case AGGR_SOCKET: |
41c8ca2a AK |
1513 | fprintf(output, "# time socket cpus"); |
1514 | if (!metric_only) | |
1515 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
d4f63a47 JO |
1516 | break; |
1517 | case AGGR_CORE: | |
41c8ca2a AK |
1518 | fprintf(output, "# time core cpus"); |
1519 | if (!metric_only) | |
1520 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
d4f63a47 JO |
1521 | break; |
1522 | case AGGR_NONE: | |
41c8ca2a AK |
1523 | fprintf(output, "# time CPU"); |
1524 | if (!metric_only) | |
1525 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
d4f63a47 | 1526 | break; |
32b8af82 | 1527 | case AGGR_THREAD: |
41c8ca2a AK |
1528 | fprintf(output, "# time comm-pid"); |
1529 | if (!metric_only) | |
1530 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
32b8af82 | 1531 | break; |
d4f63a47 JO |
1532 | case AGGR_GLOBAL: |
1533 | default: | |
41c8ca2a AK |
1534 | fprintf(output, "# time"); |
1535 | if (!metric_only) | |
1536 | fprintf(output, " counts %*s events\n", unit_width, "unit"); | |
208df99e JO |
1537 | case AGGR_UNSET: |
1538 | break; | |
d4f63a47 JO |
1539 | } |
1540 | } | |
1541 | ||
41c8ca2a AK |
1542 | if (num_print_interval == 0 && metric_only) |
1543 | print_metric_headers(" ", true); | |
d4f63a47 JO |
1544 | if (++num_print_interval == 25) |
1545 | num_print_interval = 0; | |
1546 | } | |
1547 | ||
1548 | static void print_header(int argc, const char **argv) | |
42202dd5 | 1549 | { |
5821522e | 1550 | FILE *output = stat_config.output; |
69aad6f1 | 1551 | int i; |
42202dd5 | 1552 | |
ddcacfa0 IM |
1553 | fflush(stdout); |
1554 | ||
d7470b6a | 1555 | if (!csv_output) { |
4aa9015f SE |
1556 | fprintf(output, "\n"); |
1557 | fprintf(output, " Performance counter stats for "); | |
62d3b617 DA |
1558 | if (target.system_wide) |
1559 | fprintf(output, "\'system wide"); | |
1560 | else if (target.cpu_list) | |
1561 | fprintf(output, "\'CPU(s) %s", target.cpu_list); | |
602ad878 | 1562 | else if (!target__has_task(&target)) { |
ba6039b6 JO |
1563 | fprintf(output, "\'%s", argv ? argv[0] : "pipe"); |
1564 | for (i = 1; argv && (i < argc); i++) | |
4aa9015f | 1565 | fprintf(output, " %s", argv[i]); |
20f946b4 NK |
1566 | } else if (target.pid) |
1567 | fprintf(output, "process id \'%s", target.pid); | |
d7470b6a | 1568 | else |
20f946b4 | 1569 | fprintf(output, "thread id \'%s", target.tid); |
44db76c8 | 1570 | |
4aa9015f | 1571 | fprintf(output, "\'"); |
d7470b6a | 1572 | if (run_count > 1) |
4aa9015f SE |
1573 | fprintf(output, " (%d runs)", run_count); |
1574 | fprintf(output, ":\n\n"); | |
d7470b6a | 1575 | } |
d4f63a47 JO |
1576 | } |
1577 | ||
1578 | static void print_footer(void) | |
1579 | { | |
5821522e | 1580 | FILE *output = stat_config.output; |
918c7b06 | 1581 | int n; |
5821522e | 1582 | |
d4f63a47 JO |
1583 | if (!null_run) |
1584 | fprintf(output, "\n"); | |
1585 | fprintf(output, " %17.9f seconds time elapsed", | |
310ebb93 | 1586 | avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC); |
d4f63a47 JO |
1587 | if (run_count > 1) { |
1588 | fprintf(output, " "); | |
1589 | print_noise_pct(stddev_stats(&walltime_nsecs_stats), | |
1590 | avg_stats(&walltime_nsecs_stats)); | |
1591 | } | |
1592 | fprintf(output, "\n\n"); | |
02d492e5 | 1593 | |
918c7b06 AK |
1594 | if (print_free_counters_hint && |
1595 | sysctl__read_int("kernel/nmi_watchdog", &n) >= 0 && | |
1596 | n > 0) | |
02d492e5 BP |
1597 | fprintf(output, |
1598 | "Some events weren't counted. Try disabling the NMI watchdog:\n" | |
1599 | " echo 0 > /proc/sys/kernel/nmi_watchdog\n" | |
1600 | " perf stat ...\n" | |
1601 | " echo 1 > /proc/sys/kernel/nmi_watchdog\n"); | |
d4f63a47 JO |
1602 | } |
1603 | ||
1604 | static void print_counters(struct timespec *ts, int argc, const char **argv) | |
1605 | { | |
ec0d3d1f | 1606 | int interval = stat_config.interval; |
d4f63a47 JO |
1607 | struct perf_evsel *counter; |
1608 | char buf[64], *prefix = NULL; | |
1609 | ||
664c98d4 JO |
1610 | /* Do not print anything if we record to the pipe. */ |
1611 | if (STAT_RECORD && perf_stat.file.is_pipe) | |
1612 | return; | |
1613 | ||
d4f63a47 JO |
1614 | if (interval) |
1615 | print_interval(prefix = buf, ts); | |
1616 | else | |
1617 | print_header(argc, argv); | |
2996f5dd | 1618 | |
54b50916 AK |
1619 | if (metric_only) { |
1620 | static int num_print_iv; | |
1621 | ||
41c8ca2a AK |
1622 | if (num_print_iv == 0 && !interval) |
1623 | print_metric_headers(prefix, false); | |
54b50916 AK |
1624 | if (num_print_iv++ == 25) |
1625 | num_print_iv = 0; | |
1626 | if (stat_config.aggr_mode == AGGR_GLOBAL && prefix) | |
1627 | fprintf(stat_config.output, "%s", prefix); | |
1628 | } | |
1629 | ||
421a50f3 | 1630 | switch (stat_config.aggr_mode) { |
12c08a9f | 1631 | case AGGR_CORE: |
86ee6e18 | 1632 | case AGGR_SOCKET: |
d4f63a47 | 1633 | print_aggr(prefix); |
86ee6e18 | 1634 | break; |
32b8af82 | 1635 | case AGGR_THREAD: |
e5cadb93 | 1636 | evlist__for_each_entry(evsel_list, counter) |
32b8af82 JO |
1637 | print_aggr_thread(counter, prefix); |
1638 | break; | |
86ee6e18 | 1639 | case AGGR_GLOBAL: |
e5cadb93 | 1640 | evlist__for_each_entry(evsel_list, counter) |
d4f63a47 | 1641 | print_counter_aggr(counter, prefix); |
54b50916 AK |
1642 | if (metric_only) |
1643 | fputc('\n', stat_config.output); | |
86ee6e18 SE |
1644 | break; |
1645 | case AGGR_NONE: | |
206cab65 AK |
1646 | if (metric_only) |
1647 | print_no_aggr_metric(prefix); | |
1648 | else { | |
e5cadb93 | 1649 | evlist__for_each_entry(evsel_list, counter) |
206cab65 AK |
1650 | print_counter(counter, prefix); |
1651 | } | |
86ee6e18 | 1652 | break; |
208df99e | 1653 | case AGGR_UNSET: |
86ee6e18 SE |
1654 | default: |
1655 | break; | |
f5b4a9c3 | 1656 | } |
ddcacfa0 | 1657 | |
d4f63a47 JO |
1658 | if (!interval && !csv_output) |
1659 | print_footer(); | |
1660 | ||
5821522e | 1661 | fflush(stat_config.output); |
ddcacfa0 IM |
1662 | } |
1663 | ||
f7b7c26e PZ |
1664 | static volatile int signr = -1; |
1665 | ||
5242519b | 1666 | static void skip_signal(int signo) |
ddcacfa0 | 1667 | { |
ec0d3d1f | 1668 | if ((child_pid == -1) || stat_config.interval) |
60666c63 LW |
1669 | done = 1; |
1670 | ||
f7b7c26e | 1671 | signr = signo; |
d07f0b12 SE |
1672 | /* |
1673 | * render child_pid harmless | |
1674 | * won't send SIGTERM to a random | |
1675 | * process in case of race condition | |
1676 | * and fast PID recycling | |
1677 | */ | |
1678 | child_pid = -1; | |
f7b7c26e PZ |
1679 | } |
1680 | ||
1681 | static void sig_atexit(void) | |
1682 | { | |
d07f0b12 SE |
1683 | sigset_t set, oset; |
1684 | ||
1685 | /* | |
1686 | * avoid race condition with SIGCHLD handler | |
1687 | * in skip_signal() which is modifying child_pid | |
1688 | * goal is to avoid send SIGTERM to a random | |
1689 | * process | |
1690 | */ | |
1691 | sigemptyset(&set); | |
1692 | sigaddset(&set, SIGCHLD); | |
1693 | sigprocmask(SIG_BLOCK, &set, &oset); | |
1694 | ||
933da83a CW |
1695 | if (child_pid != -1) |
1696 | kill(child_pid, SIGTERM); | |
1697 | ||
d07f0b12 SE |
1698 | sigprocmask(SIG_SETMASK, &oset, NULL); |
1699 | ||
f7b7c26e PZ |
1700 | if (signr == -1) |
1701 | return; | |
1702 | ||
1703 | signal(signr, SIG_DFL); | |
1704 | kill(getpid(), signr); | |
5242519b IM |
1705 | } |
1706 | ||
1d037ca1 IT |
1707 | static int stat__set_big_num(const struct option *opt __maybe_unused, |
1708 | const char *s __maybe_unused, int unset) | |
d7470b6a SE |
1709 | { |
1710 | big_num_opt = unset ? 0 : 1; | |
1711 | return 0; | |
1712 | } | |
1713 | ||
44b1e60a AK |
1714 | static int enable_metric_only(const struct option *opt __maybe_unused, |
1715 | const char *s __maybe_unused, int unset) | |
1716 | { | |
1717 | force_metric_only = true; | |
1718 | metric_only = !unset; | |
1719 | return 0; | |
1720 | } | |
1721 | ||
e0547311 JO |
1722 | static const struct option stat_options[] = { |
1723 | OPT_BOOLEAN('T', "transaction", &transaction_run, | |
1724 | "hardware transaction statistics"), | |
1725 | OPT_CALLBACK('e', "event", &evsel_list, "event", | |
1726 | "event selector. use 'perf list' to list available events", | |
1727 | parse_events_option), | |
1728 | OPT_CALLBACK(0, "filter", &evsel_list, "filter", | |
1729 | "event filter", parse_filter), | |
1730 | OPT_BOOLEAN('i', "no-inherit", &no_inherit, | |
1731 | "child tasks do not inherit counters"), | |
1732 | OPT_STRING('p', "pid", &target.pid, "pid", | |
1733 | "stat events on existing process id"), | |
1734 | OPT_STRING('t', "tid", &target.tid, "tid", | |
1735 | "stat events on existing thread id"), | |
1736 | OPT_BOOLEAN('a', "all-cpus", &target.system_wide, | |
1737 | "system-wide collection from all CPUs"), | |
1738 | OPT_BOOLEAN('g', "group", &group, | |
1739 | "put the counters into a counter group"), | |
1740 | OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"), | |
1741 | OPT_INCR('v', "verbose", &verbose, | |
1742 | "be more verbose (show counter open errors, etc)"), | |
1743 | OPT_INTEGER('r', "repeat", &run_count, | |
1744 | "repeat command and print average + stddev (max: 100, forever: 0)"), | |
1745 | OPT_BOOLEAN('n', "null", &null_run, | |
1746 | "null run - dont start any counters"), | |
1747 | OPT_INCR('d', "detailed", &detailed_run, | |
1748 | "detailed run - start a lot of events"), | |
1749 | OPT_BOOLEAN('S', "sync", &sync_run, | |
1750 | "call sync() before starting a run"), | |
1751 | OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, | |
1752 | "print large numbers with thousands\' separators", | |
1753 | stat__set_big_num), | |
1754 | OPT_STRING('C', "cpu", &target.cpu_list, "cpu", | |
1755 | "list of cpus to monitor in system-wide"), | |
1756 | OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode, | |
1757 | "disable CPU count aggregation", AGGR_NONE), | |
430daf2d | 1758 | OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"), |
e0547311 JO |
1759 | OPT_STRING('x', "field-separator", &csv_sep, "separator", |
1760 | "print counts with custom separator"), | |
1761 | OPT_CALLBACK('G', "cgroup", &evsel_list, "name", | |
1762 | "monitor event in cgroup name only", parse_cgroups), | |
1763 | OPT_STRING('o', "output", &output_name, "file", "output file name"), | |
1764 | OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), | |
1765 | OPT_INTEGER(0, "log-fd", &output_fd, | |
1766 | "log output to fd, instead of stderr"), | |
1767 | OPT_STRING(0, "pre", &pre_cmd, "command", | |
1768 | "command to run prior to the measured command"), | |
1769 | OPT_STRING(0, "post", &post_cmd, "command", | |
1770 | "command to run after to the measured command"), | |
1771 | OPT_UINTEGER('I', "interval-print", &stat_config.interval, | |
1772 | "print counts at regular interval in ms (>= 10)"), | |
1773 | OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode, | |
1774 | "aggregate counts per processor socket", AGGR_SOCKET), | |
1775 | OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode, | |
1776 | "aggregate counts per physical processor core", AGGR_CORE), | |
1777 | OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode, | |
1778 | "aggregate counts per thread", AGGR_THREAD), | |
1779 | OPT_UINTEGER('D', "delay", &initial_delay, | |
1780 | "ms to wait before starting measurement after program start"), | |
44b1e60a AK |
1781 | OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL, |
1782 | "Only print computed metrics. No raw values", enable_metric_only), | |
1783 | OPT_BOOLEAN(0, "topdown", &topdown_run, | |
1784 | "measure topdown level 1 statistics"), | |
e0547311 JO |
1785 | OPT_END() |
1786 | }; | |
1787 | ||
1fe7a300 JO |
1788 | static int perf_stat__get_socket(struct cpu_map *map, int cpu) |
1789 | { | |
1790 | return cpu_map__get_socket(map, cpu, NULL); | |
1791 | } | |
1792 | ||
1793 | static int perf_stat__get_core(struct cpu_map *map, int cpu) | |
1794 | { | |
1795 | return cpu_map__get_core(map, cpu, NULL); | |
1796 | } | |
1797 | ||
1e5a2931 JO |
1798 | static int cpu_map__get_max(struct cpu_map *map) |
1799 | { | |
1800 | int i, max = -1; | |
1801 | ||
1802 | for (i = 0; i < map->nr; i++) { | |
1803 | if (map->map[i] > max) | |
1804 | max = map->map[i]; | |
1805 | } | |
1806 | ||
1807 | return max; | |
1808 | } | |
1809 | ||
1810 | static struct cpu_map *cpus_aggr_map; | |
1811 | ||
1812 | static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx) | |
1813 | { | |
1814 | int cpu; | |
1815 | ||
1816 | if (idx >= map->nr) | |
1817 | return -1; | |
1818 | ||
1819 | cpu = map->map[idx]; | |
1820 | ||
1821 | if (cpus_aggr_map->map[cpu] == -1) | |
1822 | cpus_aggr_map->map[cpu] = get_id(map, idx); | |
1823 | ||
1824 | return cpus_aggr_map->map[cpu]; | |
1825 | } | |
1826 | ||
1827 | static int perf_stat__get_socket_cached(struct cpu_map *map, int idx) | |
1828 | { | |
1829 | return perf_stat__get_aggr(perf_stat__get_socket, map, idx); | |
1830 | } | |
1831 | ||
1832 | static int perf_stat__get_core_cached(struct cpu_map *map, int idx) | |
1833 | { | |
1834 | return perf_stat__get_aggr(perf_stat__get_core, map, idx); | |
1835 | } | |
1836 | ||
86ee6e18 SE |
1837 | static int perf_stat_init_aggr_mode(void) |
1838 | { | |
1e5a2931 JO |
1839 | int nr; |
1840 | ||
421a50f3 | 1841 | switch (stat_config.aggr_mode) { |
86ee6e18 SE |
1842 | case AGGR_SOCKET: |
1843 | if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) { | |
1844 | perror("cannot build socket map"); | |
1845 | return -1; | |
1846 | } | |
1e5a2931 | 1847 | aggr_get_id = perf_stat__get_socket_cached; |
86ee6e18 | 1848 | break; |
12c08a9f SE |
1849 | case AGGR_CORE: |
1850 | if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) { | |
1851 | perror("cannot build core map"); | |
1852 | return -1; | |
1853 | } | |
1e5a2931 | 1854 | aggr_get_id = perf_stat__get_core_cached; |
12c08a9f | 1855 | break; |
86ee6e18 SE |
1856 | case AGGR_NONE: |
1857 | case AGGR_GLOBAL: | |
32b8af82 | 1858 | case AGGR_THREAD: |
208df99e | 1859 | case AGGR_UNSET: |
86ee6e18 SE |
1860 | default: |
1861 | break; | |
1862 | } | |
1e5a2931 JO |
1863 | |
1864 | /* | |
1865 | * The evsel_list->cpus is the base we operate on, | |
1866 | * taking the highest cpu number to be the size of | |
1867 | * the aggregation translate cpumap. | |
1868 | */ | |
1869 | nr = cpu_map__get_max(evsel_list->cpus); | |
1870 | cpus_aggr_map = cpu_map__empty_new(nr + 1); | |
1871 | return cpus_aggr_map ? 0 : -ENOMEM; | |
86ee6e18 SE |
1872 | } |
1873 | ||
544c2ae7 MH |
1874 | static void perf_stat__exit_aggr_mode(void) |
1875 | { | |
1876 | cpu_map__put(aggr_map); | |
1877 | cpu_map__put(cpus_aggr_map); | |
1878 | aggr_map = NULL; | |
1879 | cpus_aggr_map = NULL; | |
1880 | } | |
1881 | ||
68d702f7 JO |
1882 | static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx) |
1883 | { | |
1884 | int cpu; | |
1885 | ||
1886 | if (idx > map->nr) | |
1887 | return -1; | |
1888 | ||
1889 | cpu = map->map[idx]; | |
1890 | ||
da8a58b5 | 1891 | if (cpu >= env->nr_cpus_avail) |
68d702f7 JO |
1892 | return -1; |
1893 | ||
1894 | return cpu; | |
1895 | } | |
1896 | ||
1897 | static int perf_env__get_socket(struct cpu_map *map, int idx, void *data) | |
1898 | { | |
1899 | struct perf_env *env = data; | |
1900 | int cpu = perf_env__get_cpu(env, map, idx); | |
1901 | ||
1902 | return cpu == -1 ? -1 : env->cpu[cpu].socket_id; | |
1903 | } | |
1904 | ||
1905 | static int perf_env__get_core(struct cpu_map *map, int idx, void *data) | |
1906 | { | |
1907 | struct perf_env *env = data; | |
1908 | int core = -1, cpu = perf_env__get_cpu(env, map, idx); | |
1909 | ||
1910 | if (cpu != -1) { | |
1911 | int socket_id = env->cpu[cpu].socket_id; | |
1912 | ||
1913 | /* | |
1914 | * Encode socket in upper 16 bits | |
1915 | * core_id is relative to socket, and | |
1916 | * we need a global id. So we combine | |
1917 | * socket + core id. | |
1918 | */ | |
1919 | core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff); | |
1920 | } | |
1921 | ||
1922 | return core; | |
1923 | } | |
1924 | ||
1925 | static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus, | |
1926 | struct cpu_map **sockp) | |
1927 | { | |
1928 | return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env); | |
1929 | } | |
1930 | ||
1931 | static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus, | |
1932 | struct cpu_map **corep) | |
1933 | { | |
1934 | return cpu_map__build_map(cpus, corep, perf_env__get_core, env); | |
1935 | } | |
1936 | ||
1937 | static int perf_stat__get_socket_file(struct cpu_map *map, int idx) | |
1938 | { | |
1939 | return perf_env__get_socket(map, idx, &perf_stat.session->header.env); | |
1940 | } | |
1941 | ||
1942 | static int perf_stat__get_core_file(struct cpu_map *map, int idx) | |
1943 | { | |
1944 | return perf_env__get_core(map, idx, &perf_stat.session->header.env); | |
1945 | } | |
1946 | ||
1947 | static int perf_stat_init_aggr_mode_file(struct perf_stat *st) | |
1948 | { | |
1949 | struct perf_env *env = &st->session->header.env; | |
1950 | ||
1951 | switch (stat_config.aggr_mode) { | |
1952 | case AGGR_SOCKET: | |
1953 | if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) { | |
1954 | perror("cannot build socket map"); | |
1955 | return -1; | |
1956 | } | |
1957 | aggr_get_id = perf_stat__get_socket_file; | |
1958 | break; | |
1959 | case AGGR_CORE: | |
1960 | if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) { | |
1961 | perror("cannot build core map"); | |
1962 | return -1; | |
1963 | } | |
1964 | aggr_get_id = perf_stat__get_core_file; | |
1965 | break; | |
1966 | case AGGR_NONE: | |
1967 | case AGGR_GLOBAL: | |
1968 | case AGGR_THREAD: | |
1969 | case AGGR_UNSET: | |
1970 | default: | |
1971 | break; | |
1972 | } | |
1973 | ||
1974 | return 0; | |
1975 | } | |
1976 | ||
44b1e60a AK |
1977 | static int topdown_filter_events(const char **attr, char **str, bool use_group) |
1978 | { | |
1979 | int off = 0; | |
1980 | int i; | |
1981 | int len = 0; | |
1982 | char *s; | |
1983 | ||
1984 | for (i = 0; attr[i]; i++) { | |
1985 | if (pmu_have_event("cpu", attr[i])) { | |
1986 | len += strlen(attr[i]) + 1; | |
1987 | attr[i - off] = attr[i]; | |
1988 | } else | |
1989 | off++; | |
1990 | } | |
1991 | attr[i - off] = NULL; | |
1992 | ||
1993 | *str = malloc(len + 1 + 2); | |
1994 | if (!*str) | |
1995 | return -1; | |
1996 | s = *str; | |
1997 | if (i - off == 0) { | |
1998 | *s = 0; | |
1999 | return 0; | |
2000 | } | |
2001 | if (use_group) | |
2002 | *s++ = '{'; | |
2003 | for (i = 0; attr[i]; i++) { | |
2004 | strcpy(s, attr[i]); | |
2005 | s += strlen(s); | |
2006 | *s++ = ','; | |
2007 | } | |
2008 | if (use_group) { | |
2009 | s[-1] = '}'; | |
2010 | *s = 0; | |
2011 | } else | |
2012 | s[-1] = 0; | |
2013 | return 0; | |
2014 | } | |
2015 | ||
2016 | __weak bool arch_topdown_check_group(bool *warn) | |
2017 | { | |
2018 | *warn = false; | |
2019 | return false; | |
2020 | } | |
2021 | ||
2022 | __weak void arch_topdown_group_warn(void) | |
2023 | { | |
2024 | } | |
2025 | ||
2cba3ffb IM |
2026 | /* |
2027 | * Add default attributes, if there were no attributes specified or | |
2028 | * if -d/--detailed, -d -d or -d -d -d is used: | |
2029 | */ | |
2030 | static int add_default_attributes(void) | |
2031 | { | |
44b1e60a | 2032 | int err; |
9dec4473 | 2033 | struct perf_event_attr default_attrs0[] = { |
b070a547 ACM |
2034 | |
2035 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, | |
2036 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, | |
2037 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, | |
2038 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, | |
2039 | ||
2040 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, | |
9dec4473 AK |
2041 | }; |
2042 | struct perf_event_attr frontend_attrs[] = { | |
b070a547 | 2043 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, |
9dec4473 AK |
2044 | }; |
2045 | struct perf_event_attr backend_attrs[] = { | |
b070a547 | 2046 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, |
9dec4473 AK |
2047 | }; |
2048 | struct perf_event_attr default_attrs1[] = { | |
b070a547 ACM |
2049 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, |
2050 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, | |
2051 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, | |
2052 | ||
2053 | }; | |
2054 | ||
2055 | /* | |
2056 | * Detailed stats (-d), covering the L1 and last level data caches: | |
2057 | */ | |
2058 | struct perf_event_attr detailed_attrs[] = { | |
2059 | ||
2060 | { .type = PERF_TYPE_HW_CACHE, | |
2061 | .config = | |
2062 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
2063 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2064 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2065 | ||
2066 | { .type = PERF_TYPE_HW_CACHE, | |
2067 | .config = | |
2068 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
2069 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2070 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2071 | ||
2072 | { .type = PERF_TYPE_HW_CACHE, | |
2073 | .config = | |
2074 | PERF_COUNT_HW_CACHE_LL << 0 | | |
2075 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2076 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2077 | ||
2078 | { .type = PERF_TYPE_HW_CACHE, | |
2079 | .config = | |
2080 | PERF_COUNT_HW_CACHE_LL << 0 | | |
2081 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2082 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2083 | }; | |
2084 | ||
2085 | /* | |
2086 | * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: | |
2087 | */ | |
2088 | struct perf_event_attr very_detailed_attrs[] = { | |
2089 | ||
2090 | { .type = PERF_TYPE_HW_CACHE, | |
2091 | .config = | |
2092 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
2093 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2094 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2095 | ||
2096 | { .type = PERF_TYPE_HW_CACHE, | |
2097 | .config = | |
2098 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
2099 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2100 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2101 | ||
2102 | { .type = PERF_TYPE_HW_CACHE, | |
2103 | .config = | |
2104 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
2105 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2106 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2107 | ||
2108 | { .type = PERF_TYPE_HW_CACHE, | |
2109 | .config = | |
2110 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
2111 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2112 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2113 | ||
2114 | { .type = PERF_TYPE_HW_CACHE, | |
2115 | .config = | |
2116 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
2117 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2118 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2119 | ||
2120 | { .type = PERF_TYPE_HW_CACHE, | |
2121 | .config = | |
2122 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
2123 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
2124 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2125 | ||
2126 | }; | |
2127 | ||
2128 | /* | |
2129 | * Very, very detailed stats (-d -d -d), adding prefetch events: | |
2130 | */ | |
2131 | struct perf_event_attr very_very_detailed_attrs[] = { | |
2132 | ||
2133 | { .type = PERF_TYPE_HW_CACHE, | |
2134 | .config = | |
2135 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
2136 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
2137 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
2138 | ||
2139 | { .type = PERF_TYPE_HW_CACHE, | |
2140 | .config = | |
2141 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
2142 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
2143 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
2144 | }; | |
2145 | ||
2cba3ffb IM |
2146 | /* Set attrs if no event is selected and !null_run: */ |
2147 | if (null_run) | |
2148 | return 0; | |
2149 | ||
4cabc3d1 | 2150 | if (transaction_run) { |
4cabc3d1 AK |
2151 | if (pmu_have_event("cpu", "cycles-ct") && |
2152 | pmu_have_event("cpu", "el-start")) | |
a454742c | 2153 | err = parse_events(evsel_list, transaction_attrs, NULL); |
4cabc3d1 | 2154 | else |
a454742c JO |
2155 | err = parse_events(evsel_list, transaction_limited_attrs, NULL); |
2156 | if (err) { | |
4cabc3d1 AK |
2157 | fprintf(stderr, "Cannot set up transaction events\n"); |
2158 | return -1; | |
2159 | } | |
2160 | return 0; | |
2161 | } | |
2162 | ||
44b1e60a AK |
2163 | if (topdown_run) { |
2164 | char *str = NULL; | |
2165 | bool warn = false; | |
2166 | ||
2167 | if (stat_config.aggr_mode != AGGR_GLOBAL && | |
2168 | stat_config.aggr_mode != AGGR_CORE) { | |
2169 | pr_err("top down event configuration requires --per-core mode\n"); | |
2170 | return -1; | |
2171 | } | |
2172 | stat_config.aggr_mode = AGGR_CORE; | |
2173 | if (nr_cgroups || !target__has_cpu(&target)) { | |
2174 | pr_err("top down event configuration requires system-wide mode (-a)\n"); | |
2175 | return -1; | |
2176 | } | |
2177 | ||
2178 | if (!force_metric_only) | |
2179 | metric_only = true; | |
2180 | if (topdown_filter_events(topdown_attrs, &str, | |
2181 | arch_topdown_check_group(&warn)) < 0) { | |
2182 | pr_err("Out of memory\n"); | |
2183 | return -1; | |
2184 | } | |
2185 | if (topdown_attrs[0] && str) { | |
2186 | if (warn) | |
2187 | arch_topdown_group_warn(); | |
2188 | err = parse_events(evsel_list, str, NULL); | |
2189 | if (err) { | |
2190 | fprintf(stderr, | |
2191 | "Cannot set up top down events %s: %d\n", | |
2192 | str, err); | |
2193 | free(str); | |
2194 | return -1; | |
2195 | } | |
2196 | } else { | |
2197 | fprintf(stderr, "System does not support topdown\n"); | |
2198 | return -1; | |
2199 | } | |
2200 | free(str); | |
2201 | } | |
2202 | ||
2cba3ffb | 2203 | if (!evsel_list->nr_entries) { |
a1f3d567 NK |
2204 | if (target__has_cpu(&target)) |
2205 | default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK; | |
2206 | ||
9dec4473 AK |
2207 | if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0) |
2208 | return -1; | |
2209 | if (pmu_have_event("cpu", "stalled-cycles-frontend")) { | |
2210 | if (perf_evlist__add_default_attrs(evsel_list, | |
2211 | frontend_attrs) < 0) | |
2212 | return -1; | |
2213 | } | |
2214 | if (pmu_have_event("cpu", "stalled-cycles-backend")) { | |
2215 | if (perf_evlist__add_default_attrs(evsel_list, | |
2216 | backend_attrs) < 0) | |
2217 | return -1; | |
2218 | } | |
2219 | if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0) | |
50d08e47 | 2220 | return -1; |
2cba3ffb IM |
2221 | } |
2222 | ||
2223 | /* Detailed events get appended to the event list: */ | |
2224 | ||
2225 | if (detailed_run < 1) | |
2226 | return 0; | |
2227 | ||
2228 | /* Append detailed run extra attributes: */ | |
79695e1b | 2229 | if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) |
50d08e47 | 2230 | return -1; |
2cba3ffb IM |
2231 | |
2232 | if (detailed_run < 2) | |
2233 | return 0; | |
2234 | ||
2235 | /* Append very detailed run extra attributes: */ | |
79695e1b | 2236 | if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) |
50d08e47 | 2237 | return -1; |
2cba3ffb IM |
2238 | |
2239 | if (detailed_run < 3) | |
2240 | return 0; | |
2241 | ||
2242 | /* Append very, very detailed run extra attributes: */ | |
79695e1b | 2243 | return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); |
2cba3ffb IM |
2244 | } |
2245 | ||
8a59f3cc | 2246 | static const char * const stat_record_usage[] = { |
4979d0c7 JO |
2247 | "perf stat record [<options>]", |
2248 | NULL, | |
2249 | }; | |
2250 | ||
3ba78bd0 JO |
2251 | static void init_features(struct perf_session *session) |
2252 | { | |
2253 | int feat; | |
2254 | ||
2255 | for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++) | |
2256 | perf_header__set_feat(&session->header, feat); | |
2257 | ||
2258 | perf_header__clear_feat(&session->header, HEADER_BUILD_ID); | |
2259 | perf_header__clear_feat(&session->header, HEADER_TRACING_DATA); | |
2260 | perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK); | |
2261 | perf_header__clear_feat(&session->header, HEADER_AUXTRACE); | |
2262 | } | |
2263 | ||
4979d0c7 JO |
2264 | static int __cmd_record(int argc, const char **argv) |
2265 | { | |
2266 | struct perf_session *session; | |
2267 | struct perf_data_file *file = &perf_stat.file; | |
2268 | ||
8a59f3cc | 2269 | argc = parse_options(argc, argv, stat_options, stat_record_usage, |
4979d0c7 JO |
2270 | PARSE_OPT_STOP_AT_NON_OPTION); |
2271 | ||
2272 | if (output_name) | |
2273 | file->path = output_name; | |
2274 | ||
e9d6db8e JO |
2275 | if (run_count != 1 || forever) { |
2276 | pr_err("Cannot use -r option with perf stat record.\n"); | |
2277 | return -1; | |
2278 | } | |
2279 | ||
4979d0c7 JO |
2280 | session = perf_session__new(file, false, NULL); |
2281 | if (session == NULL) { | |
2282 | pr_err("Perf session creation failed.\n"); | |
2283 | return -1; | |
2284 | } | |
2285 | ||
3ba78bd0 JO |
2286 | init_features(session); |
2287 | ||
4979d0c7 JO |
2288 | session->evlist = evsel_list; |
2289 | perf_stat.session = session; | |
2290 | perf_stat.record = true; | |
2291 | return argc; | |
2292 | } | |
2293 | ||
a56f9390 JO |
2294 | static int process_stat_round_event(struct perf_tool *tool __maybe_unused, |
2295 | union perf_event *event, | |
2296 | struct perf_session *session) | |
2297 | { | |
e3b03b6c | 2298 | struct stat_round_event *stat_round = &event->stat_round; |
a56f9390 JO |
2299 | struct perf_evsel *counter; |
2300 | struct timespec tsh, *ts = NULL; | |
2301 | const char **argv = session->header.env.cmdline_argv; | |
2302 | int argc = session->header.env.nr_cmdline; | |
2303 | ||
e5cadb93 | 2304 | evlist__for_each_entry(evsel_list, counter) |
a56f9390 JO |
2305 | perf_stat_process_counter(&stat_config, counter); |
2306 | ||
e3b03b6c AK |
2307 | if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL) |
2308 | update_stats(&walltime_nsecs_stats, stat_round->time); | |
a56f9390 | 2309 | |
e3b03b6c | 2310 | if (stat_config.interval && stat_round->time) { |
bd48c63e ACM |
2311 | tsh.tv_sec = stat_round->time / NSEC_PER_SEC; |
2312 | tsh.tv_nsec = stat_round->time % NSEC_PER_SEC; | |
a56f9390 JO |
2313 | ts = &tsh; |
2314 | } | |
2315 | ||
2316 | print_counters(ts, argc, argv); | |
2317 | return 0; | |
2318 | } | |
2319 | ||
62ba18ba | 2320 | static |
7e6a7998 | 2321 | int process_stat_config_event(struct perf_tool *tool, |
62ba18ba JO |
2322 | union perf_event *event, |
2323 | struct perf_session *session __maybe_unused) | |
2324 | { | |
68d702f7 JO |
2325 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); |
2326 | ||
62ba18ba | 2327 | perf_event__read_stat_config(&stat_config, &event->stat_config); |
68d702f7 | 2328 | |
89af4e05 JO |
2329 | if (cpu_map__empty(st->cpus)) { |
2330 | if (st->aggr_mode != AGGR_UNSET) | |
2331 | pr_warning("warning: processing task data, aggregation mode not set\n"); | |
2332 | return 0; | |
2333 | } | |
2334 | ||
2335 | if (st->aggr_mode != AGGR_UNSET) | |
2336 | stat_config.aggr_mode = st->aggr_mode; | |
2337 | ||
68d702f7 JO |
2338 | if (perf_stat.file.is_pipe) |
2339 | perf_stat_init_aggr_mode(); | |
2340 | else | |
2341 | perf_stat_init_aggr_mode_file(st); | |
2342 | ||
62ba18ba JO |
2343 | return 0; |
2344 | } | |
2345 | ||
1975d36e JO |
2346 | static int set_maps(struct perf_stat *st) |
2347 | { | |
2348 | if (!st->cpus || !st->threads) | |
2349 | return 0; | |
2350 | ||
2351 | if (WARN_ONCE(st->maps_allocated, "stats double allocation\n")) | |
2352 | return -EINVAL; | |
2353 | ||
2354 | perf_evlist__set_maps(evsel_list, st->cpus, st->threads); | |
2355 | ||
2356 | if (perf_evlist__alloc_stats(evsel_list, true)) | |
2357 | return -ENOMEM; | |
2358 | ||
2359 | st->maps_allocated = true; | |
2360 | return 0; | |
2361 | } | |
2362 | ||
2363 | static | |
7e6a7998 | 2364 | int process_thread_map_event(struct perf_tool *tool, |
1975d36e JO |
2365 | union perf_event *event, |
2366 | struct perf_session *session __maybe_unused) | |
2367 | { | |
2368 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); | |
2369 | ||
2370 | if (st->threads) { | |
2371 | pr_warning("Extra thread map event, ignoring.\n"); | |
2372 | return 0; | |
2373 | } | |
2374 | ||
2375 | st->threads = thread_map__new_event(&event->thread_map); | |
2376 | if (!st->threads) | |
2377 | return -ENOMEM; | |
2378 | ||
2379 | return set_maps(st); | |
2380 | } | |
2381 | ||
2382 | static | |
7e6a7998 | 2383 | int process_cpu_map_event(struct perf_tool *tool, |
1975d36e JO |
2384 | union perf_event *event, |
2385 | struct perf_session *session __maybe_unused) | |
2386 | { | |
2387 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); | |
2388 | struct cpu_map *cpus; | |
2389 | ||
2390 | if (st->cpus) { | |
2391 | pr_warning("Extra cpu map event, ignoring.\n"); | |
2392 | return 0; | |
2393 | } | |
2394 | ||
2395 | cpus = cpu_map__new_data(&event->cpu_map.data); | |
2396 | if (!cpus) | |
2397 | return -ENOMEM; | |
2398 | ||
2399 | st->cpus = cpus; | |
2400 | return set_maps(st); | |
2401 | } | |
2402 | ||
8a59f3cc | 2403 | static const char * const stat_report_usage[] = { |
ba6039b6 JO |
2404 | "perf stat report [<options>]", |
2405 | NULL, | |
2406 | }; | |
2407 | ||
2408 | static struct perf_stat perf_stat = { | |
2409 | .tool = { | |
2410 | .attr = perf_event__process_attr, | |
fa6ea781 | 2411 | .event_update = perf_event__process_event_update, |
1975d36e JO |
2412 | .thread_map = process_thread_map_event, |
2413 | .cpu_map = process_cpu_map_event, | |
62ba18ba | 2414 | .stat_config = process_stat_config_event, |
a56f9390 JO |
2415 | .stat = perf_event__process_stat_event, |
2416 | .stat_round = process_stat_round_event, | |
ba6039b6 | 2417 | }, |
89af4e05 | 2418 | .aggr_mode = AGGR_UNSET, |
ba6039b6 JO |
2419 | }; |
2420 | ||
2421 | static int __cmd_report(int argc, const char **argv) | |
2422 | { | |
2423 | struct perf_session *session; | |
2424 | const struct option options[] = { | |
2425 | OPT_STRING('i', "input", &input_name, "file", "input file name"), | |
89af4e05 JO |
2426 | OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode, |
2427 | "aggregate counts per processor socket", AGGR_SOCKET), | |
2428 | OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode, | |
2429 | "aggregate counts per physical processor core", AGGR_CORE), | |
2430 | OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode, | |
2431 | "disable CPU count aggregation", AGGR_NONE), | |
ba6039b6 JO |
2432 | OPT_END() |
2433 | }; | |
2434 | struct stat st; | |
2435 | int ret; | |
2436 | ||
8a59f3cc | 2437 | argc = parse_options(argc, argv, options, stat_report_usage, 0); |
ba6039b6 JO |
2438 | |
2439 | if (!input_name || !strlen(input_name)) { | |
2440 | if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) | |
2441 | input_name = "-"; | |
2442 | else | |
2443 | input_name = "perf.data"; | |
2444 | } | |
2445 | ||
2446 | perf_stat.file.path = input_name; | |
2447 | perf_stat.file.mode = PERF_DATA_MODE_READ; | |
2448 | ||
2449 | session = perf_session__new(&perf_stat.file, false, &perf_stat.tool); | |
2450 | if (session == NULL) | |
2451 | return -1; | |
2452 | ||
2453 | perf_stat.session = session; | |
2454 | stat_config.output = stderr; | |
2455 | evsel_list = session->evlist; | |
2456 | ||
2457 | ret = perf_session__process_events(session); | |
2458 | if (ret) | |
2459 | return ret; | |
2460 | ||
2461 | perf_session__delete(session); | |
2462 | return 0; | |
2463 | } | |
2464 | ||
e3ba76de JO |
2465 | static void setup_system_wide(int forks) |
2466 | { | |
2467 | /* | |
2468 | * Make system wide (-a) the default target if | |
2469 | * no target was specified and one of following | |
2470 | * conditions is met: | |
2471 | * | |
2472 | * - there's no workload specified | |
2473 | * - there is workload specified but all requested | |
2474 | * events are system wide events | |
2475 | */ | |
2476 | if (!target__none(&target)) | |
2477 | return; | |
2478 | ||
2479 | if (!forks) | |
2480 | target.system_wide = true; | |
2481 | else { | |
2482 | struct perf_evsel *counter; | |
2483 | ||
2484 | evlist__for_each_entry(evsel_list, counter) { | |
2485 | if (!counter->system_wide) | |
2486 | return; | |
2487 | } | |
2488 | ||
2489 | if (evsel_list->nr_entries) | |
2490 | target.system_wide = true; | |
2491 | } | |
2492 | } | |
2493 | ||
b0ad8ea6 | 2494 | int cmd_stat(int argc, const char **argv) |
5242519b | 2495 | { |
b070a547 ACM |
2496 | const char * const stat_usage[] = { |
2497 | "perf stat [<options>] [<command>]", | |
2498 | NULL | |
2499 | }; | |
cc03c542 | 2500 | int status = -EINVAL, run_idx; |
4aa9015f | 2501 | const char *mode; |
5821522e | 2502 | FILE *output = stderr; |
ec0d3d1f | 2503 | unsigned int interval; |
ba6039b6 | 2504 | const char * const stat_subcommands[] = { "record", "report" }; |
42202dd5 | 2505 | |
5af52b51 SE |
2506 | setlocale(LC_ALL, ""); |
2507 | ||
334fe7a3 | 2508 | evsel_list = perf_evlist__new(); |
361c99a6 ACM |
2509 | if (evsel_list == NULL) |
2510 | return -ENOMEM; | |
2511 | ||
1669e509 | 2512 | parse_events__shrink_config_terms(); |
4979d0c7 JO |
2513 | argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands, |
2514 | (const char **) stat_usage, | |
2515 | PARSE_OPT_STOP_AT_NON_OPTION); | |
37932c18 | 2516 | perf_stat__collect_metric_expr(evsel_list); |
fb4605ba | 2517 | perf_stat__init_shadow_stats(); |
4979d0c7 | 2518 | |
6edb78a2 JO |
2519 | if (csv_sep) { |
2520 | csv_output = true; | |
2521 | if (!strcmp(csv_sep, "\\t")) | |
2522 | csv_sep = "\t"; | |
2523 | } else | |
2524 | csv_sep = DEFAULT_SEPARATOR; | |
2525 | ||
4979d0c7 JO |
2526 | if (argc && !strncmp(argv[0], "rec", 3)) { |
2527 | argc = __cmd_record(argc, argv); | |
2528 | if (argc < 0) | |
2529 | return -1; | |
ba6039b6 JO |
2530 | } else if (argc && !strncmp(argv[0], "rep", 3)) |
2531 | return __cmd_report(argc, argv); | |
d7470b6a | 2532 | |
ec0d3d1f JO |
2533 | interval = stat_config.interval; |
2534 | ||
4979d0c7 JO |
2535 | /* |
2536 | * For record command the -o is already taken care of. | |
2537 | */ | |
2538 | if (!STAT_RECORD && output_name && strcmp(output_name, "-")) | |
4aa9015f SE |
2539 | output = NULL; |
2540 | ||
56f3bae7 JC |
2541 | if (output_name && output_fd) { |
2542 | fprintf(stderr, "cannot use both --output and --log-fd\n"); | |
e0547311 JO |
2543 | parse_options_usage(stat_usage, stat_options, "o", 1); |
2544 | parse_options_usage(NULL, stat_options, "log-fd", 0); | |
cc03c542 | 2545 | goto out; |
56f3bae7 | 2546 | } |
fc3e4d07 | 2547 | |
54b50916 AK |
2548 | if (metric_only && stat_config.aggr_mode == AGGR_THREAD) { |
2549 | fprintf(stderr, "--metric-only is not supported with --per-thread\n"); | |
2550 | goto out; | |
2551 | } | |
2552 | ||
54b50916 AK |
2553 | if (metric_only && run_count > 1) { |
2554 | fprintf(stderr, "--metric-only is not supported with -r\n"); | |
2555 | goto out; | |
2556 | } | |
2557 | ||
fc3e4d07 SE |
2558 | if (output_fd < 0) { |
2559 | fprintf(stderr, "argument to --log-fd must be a > 0\n"); | |
e0547311 | 2560 | parse_options_usage(stat_usage, stat_options, "log-fd", 0); |
cc03c542 | 2561 | goto out; |
fc3e4d07 SE |
2562 | } |
2563 | ||
4aa9015f SE |
2564 | if (!output) { |
2565 | struct timespec tm; | |
2566 | mode = append_file ? "a" : "w"; | |
2567 | ||
2568 | output = fopen(output_name, mode); | |
2569 | if (!output) { | |
2570 | perror("failed to create output file"); | |
fceda7fe | 2571 | return -1; |
4aa9015f SE |
2572 | } |
2573 | clock_gettime(CLOCK_REALTIME, &tm); | |
2574 | fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); | |
fc3e4d07 | 2575 | } else if (output_fd > 0) { |
56f3bae7 JC |
2576 | mode = append_file ? "a" : "w"; |
2577 | output = fdopen(output_fd, mode); | |
2578 | if (!output) { | |
2579 | perror("Failed opening logfd"); | |
2580 | return -errno; | |
2581 | } | |
4aa9015f SE |
2582 | } |
2583 | ||
5821522e JO |
2584 | stat_config.output = output; |
2585 | ||
d7470b6a SE |
2586 | /* |
2587 | * let the spreadsheet do the pretty-printing | |
2588 | */ | |
2589 | if (csv_output) { | |
61a9f324 | 2590 | /* User explicitly passed -B? */ |
d7470b6a SE |
2591 | if (big_num_opt == 1) { |
2592 | fprintf(stderr, "-B option not supported with -x\n"); | |
e0547311 JO |
2593 | parse_options_usage(stat_usage, stat_options, "B", 1); |
2594 | parse_options_usage(NULL, stat_options, "x", 1); | |
cc03c542 | 2595 | goto out; |
d7470b6a SE |
2596 | } else /* Nope, so disable big number formatting */ |
2597 | big_num = false; | |
2598 | } else if (big_num_opt == 0) /* User passed --no-big-num */ | |
2599 | big_num = false; | |
2600 | ||
e3ba76de | 2601 | setup_system_wide(argc); |
ac3063bd | 2602 | |
a7e191c3 | 2603 | if (run_count < 0) { |
cc03c542 | 2604 | pr_err("Run count must be a positive number\n"); |
e0547311 | 2605 | parse_options_usage(stat_usage, stat_options, "r", 1); |
cc03c542 | 2606 | goto out; |
a7e191c3 FD |
2607 | } else if (run_count == 0) { |
2608 | forever = true; | |
2609 | run_count = 1; | |
2610 | } | |
ddcacfa0 | 2611 | |
421a50f3 | 2612 | if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) { |
32b8af82 JO |
2613 | fprintf(stderr, "The --per-thread option is only available " |
2614 | "when monitoring via -p -t options.\n"); | |
e0547311 JO |
2615 | parse_options_usage(NULL, stat_options, "p", 1); |
2616 | parse_options_usage(NULL, stat_options, "t", 1); | |
32b8af82 JO |
2617 | goto out; |
2618 | } | |
2619 | ||
2620 | /* | |
2621 | * no_aggr, cgroup are for system-wide only | |
2622 | * --per-thread is aggregated per thread, we dont mix it with cpu mode | |
2623 | */ | |
421a50f3 JO |
2624 | if (((stat_config.aggr_mode != AGGR_GLOBAL && |
2625 | stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) && | |
602ad878 | 2626 | !target__has_cpu(&target)) { |
023695d9 SE |
2627 | fprintf(stderr, "both cgroup and no-aggregation " |
2628 | "modes only available in system-wide mode\n"); | |
2629 | ||
e0547311 JO |
2630 | parse_options_usage(stat_usage, stat_options, "G", 1); |
2631 | parse_options_usage(NULL, stat_options, "A", 1); | |
2632 | parse_options_usage(NULL, stat_options, "a", 1); | |
cc03c542 | 2633 | goto out; |
d7e7a451 SE |
2634 | } |
2635 | ||
2cba3ffb IM |
2636 | if (add_default_attributes()) |
2637 | goto out; | |
ddcacfa0 | 2638 | |
602ad878 | 2639 | target__validate(&target); |
5c98d466 | 2640 | |
77a6f014 | 2641 | if (perf_evlist__create_maps(evsel_list, &target) < 0) { |
602ad878 | 2642 | if (target__has_task(&target)) { |
77a6f014 | 2643 | pr_err("Problems finding threads of monitor\n"); |
e0547311 JO |
2644 | parse_options_usage(stat_usage, stat_options, "p", 1); |
2645 | parse_options_usage(NULL, stat_options, "t", 1); | |
602ad878 | 2646 | } else if (target__has_cpu(&target)) { |
77a6f014 | 2647 | perror("failed to parse CPUs map"); |
e0547311 JO |
2648 | parse_options_usage(stat_usage, stat_options, "C", 1); |
2649 | parse_options_usage(NULL, stat_options, "a", 1); | |
cc03c542 NK |
2650 | } |
2651 | goto out; | |
60d567e2 | 2652 | } |
32b8af82 JO |
2653 | |
2654 | /* | |
2655 | * Initialize thread_map with comm names, | |
2656 | * so we could print it out on output. | |
2657 | */ | |
421a50f3 | 2658 | if (stat_config.aggr_mode == AGGR_THREAD) |
32b8af82 JO |
2659 | thread_map__read_comms(evsel_list->threads); |
2660 | ||
13370a9b | 2661 | if (interval && interval < 100) { |
19afd104 KL |
2662 | if (interval < 10) { |
2663 | pr_err("print interval must be >= 10ms\n"); | |
e0547311 | 2664 | parse_options_usage(stat_usage, stat_options, "I", 1); |
19afd104 KL |
2665 | goto out; |
2666 | } else | |
2667 | pr_warning("print interval < 100ms. " | |
2668 | "The overhead percentage could be high in some cases. " | |
2669 | "Please proceed with caution.\n"); | |
13370a9b | 2670 | } |
c45c6ea2 | 2671 | |
d134ffb9 | 2672 | if (perf_evlist__alloc_stats(evsel_list, interval)) |
03ad9747 | 2673 | goto out; |
d6d901c2 | 2674 | |
86ee6e18 | 2675 | if (perf_stat_init_aggr_mode()) |
03ad9747 | 2676 | goto out; |
86ee6e18 | 2677 | |
58d7e993 IM |
2678 | /* |
2679 | * We dont want to block the signals - that would cause | |
2680 | * child tasks to inherit that and Ctrl-C would not work. | |
2681 | * What we want is for Ctrl-C to work in the exec()-ed | |
2682 | * task, but being ignored by perf stat itself: | |
2683 | */ | |
f7b7c26e | 2684 | atexit(sig_atexit); |
a7e191c3 FD |
2685 | if (!forever) |
2686 | signal(SIGINT, skip_signal); | |
13370a9b | 2687 | signal(SIGCHLD, skip_signal); |
58d7e993 IM |
2688 | signal(SIGALRM, skip_signal); |
2689 | signal(SIGABRT, skip_signal); | |
2690 | ||
42202dd5 | 2691 | status = 0; |
a7e191c3 | 2692 | for (run_idx = 0; forever || run_idx < run_count; run_idx++) { |
bb963e16 | 2693 | if (run_count != 1 && verbose > 0) |
4aa9015f SE |
2694 | fprintf(output, "[ perf stat: executing run #%d ... ]\n", |
2695 | run_idx + 1); | |
f9cef0a9 | 2696 | |
42202dd5 | 2697 | status = run_perf_stat(argc, argv); |
a7e191c3 | 2698 | if (forever && status != -1) { |
d4f63a47 | 2699 | print_counters(NULL, argc, argv); |
254ecbc7 | 2700 | perf_stat__reset_stats(); |
a7e191c3 | 2701 | } |
42202dd5 IM |
2702 | } |
2703 | ||
a7e191c3 | 2704 | if (!forever && status != -1 && !interval) |
d4f63a47 | 2705 | print_counters(NULL, argc, argv); |
d134ffb9 | 2706 | |
4979d0c7 JO |
2707 | if (STAT_RECORD) { |
2708 | /* | |
2709 | * We synthesize the kernel mmap record just so that older tools | |
2710 | * don't emit warnings about not being able to resolve symbols | |
2711 | * due to /proc/sys/kernel/kptr_restrict settings and instear provide | |
2712 | * a saner message about no samples being in the perf.data file. | |
2713 | * | |
2714 | * This also serves to suppress a warning about f_header.data.size == 0 | |
8b99b1a4 JO |
2715 | * in header.c at the moment 'perf stat record' gets introduced, which |
2716 | * is not really needed once we start adding the stat specific PERF_RECORD_ | |
2717 | * records, but the need to suppress the kptr_restrict messages in older | |
2718 | * tools remain -acme | |
4979d0c7 JO |
2719 | */ |
2720 | int fd = perf_data_file__fd(&perf_stat.file); | |
2721 | int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat, | |
2722 | process_synthesized_event, | |
2723 | &perf_stat.session->machines.host); | |
2724 | if (err) { | |
2725 | pr_warning("Couldn't synthesize the kernel mmap record, harmless, " | |
2726 | "older tools may produce warnings about this file\n."); | |
2727 | } | |
2728 | ||
7aad0c32 JO |
2729 | if (!interval) { |
2730 | if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL)) | |
2731 | pr_err("failed to write stat round event\n"); | |
2732 | } | |
2733 | ||
664c98d4 JO |
2734 | if (!perf_stat.file.is_pipe) { |
2735 | perf_stat.session->header.data_size += perf_stat.bytes_written; | |
2736 | perf_session__write_header(perf_stat.session, evsel_list, fd, true); | |
2737 | } | |
4979d0c7 JO |
2738 | |
2739 | perf_session__delete(perf_stat.session); | |
2740 | } | |
2741 | ||
544c2ae7 | 2742 | perf_stat__exit_aggr_mode(); |
d134ffb9 | 2743 | perf_evlist__free_stats(evsel_list); |
0015e2e1 ACM |
2744 | out: |
2745 | perf_evlist__delete(evsel_list); | |
42202dd5 | 2746 | return status; |
ddcacfa0 | 2747 | } |