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" |
5242519b IM |
48 | #include "util/parse-options.h" |
49 | #include "util/parse-events.h" | |
4cabc3d1 | 50 | #include "util/pmu.h" |
8f28827a | 51 | #include "util/event.h" |
361c99a6 | 52 | #include "util/evlist.h" |
69aad6f1 | 53 | #include "util/evsel.h" |
8f28827a | 54 | #include "util/debug.h" |
a5d243d0 | 55 | #include "util/color.h" |
0007ecea | 56 | #include "util/stat.h" |
60666c63 | 57 | #include "util/header.h" |
a12b51c4 | 58 | #include "util/cpumap.h" |
d6d901c2 | 59 | #include "util/thread.h" |
fd78260b | 60 | #include "util/thread_map.h" |
d809560b | 61 | #include "util/counts.h" |
ddcacfa0 | 62 | |
1f16c575 | 63 | #include <stdlib.h> |
ddcacfa0 | 64 | #include <sys/prctl.h> |
5af52b51 | 65 | #include <locale.h> |
16c8a109 | 66 | |
d7470b6a | 67 | #define DEFAULT_SEPARATOR " " |
2cee77c4 DA |
68 | #define CNTR_NOT_SUPPORTED "<not supported>" |
69 | #define CNTR_NOT_COUNTED "<not counted>" | |
d7470b6a | 70 | |
d4f63a47 | 71 | static void print_counters(struct timespec *ts, int argc, const char **argv); |
13370a9b | 72 | |
4cabc3d1 | 73 | /* Default events used for perf stat -T */ |
a454742c JO |
74 | static const char *transaction_attrs = { |
75 | "task-clock," | |
4cabc3d1 AK |
76 | "{" |
77 | "instructions," | |
78 | "cycles," | |
79 | "cpu/cycles-t/," | |
80 | "cpu/tx-start/," | |
81 | "cpu/el-start/," | |
82 | "cpu/cycles-ct/" | |
83 | "}" | |
84 | }; | |
85 | ||
86 | /* More limited version when the CPU does not have all events. */ | |
a454742c JO |
87 | static const char * transaction_limited_attrs = { |
88 | "task-clock," | |
4cabc3d1 AK |
89 | "{" |
90 | "instructions," | |
91 | "cycles," | |
92 | "cpu/cycles-t/," | |
93 | "cpu/tx-start/" | |
94 | "}" | |
95 | }; | |
96 | ||
666e6d48 | 97 | static struct perf_evlist *evsel_list; |
361c99a6 | 98 | |
602ad878 | 99 | static struct target target = { |
77a6f014 NK |
100 | .uid = UINT_MAX, |
101 | }; | |
ddcacfa0 | 102 | |
1e5a2931 JO |
103 | typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu); |
104 | ||
3d632595 | 105 | static int run_count = 1; |
2e6cdf99 | 106 | static bool no_inherit = false; |
d07f0b12 | 107 | static volatile pid_t child_pid = -1; |
c0555642 | 108 | static bool null_run = false; |
2cba3ffb | 109 | static int detailed_run = 0; |
4cabc3d1 | 110 | static bool transaction_run; |
201e0b06 | 111 | static bool big_num = true; |
d7470b6a | 112 | static int big_num_opt = -1; |
d7470b6a SE |
113 | static const char *csv_sep = NULL; |
114 | static bool csv_output = false; | |
43bece79 | 115 | static bool group = false; |
1f16c575 PZ |
116 | static const char *pre_cmd = NULL; |
117 | static const char *post_cmd = NULL; | |
118 | static bool sync_run = false; | |
41191688 | 119 | static unsigned int initial_delay = 0; |
410136f5 | 120 | static unsigned int unit_width = 4; /* strlen("unit") */ |
a7e191c3 | 121 | static bool forever = false; |
13370a9b | 122 | static struct timespec ref_time; |
86ee6e18 | 123 | static struct cpu_map *aggr_map; |
1e5a2931 | 124 | static aggr_get_id_t aggr_get_id; |
e0547311 JO |
125 | static bool append_file; |
126 | static const char *output_name; | |
127 | static int output_fd; | |
5af52b51 | 128 | |
60666c63 LW |
129 | static volatile int done = 0; |
130 | ||
421a50f3 JO |
131 | static struct perf_stat_config stat_config = { |
132 | .aggr_mode = AGGR_GLOBAL, | |
711a572e | 133 | .scale = true, |
421a50f3 JO |
134 | }; |
135 | ||
13370a9b SE |
136 | static inline void diff_timespec(struct timespec *r, struct timespec *a, |
137 | struct timespec *b) | |
138 | { | |
139 | r->tv_sec = a->tv_sec - b->tv_sec; | |
140 | if (a->tv_nsec < b->tv_nsec) { | |
141 | r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec; | |
142 | r->tv_sec--; | |
143 | } else { | |
144 | r->tv_nsec = a->tv_nsec - b->tv_nsec ; | |
145 | } | |
146 | } | |
147 | ||
254ecbc7 JO |
148 | static void perf_stat__reset_stats(void) |
149 | { | |
150 | perf_evlist__reset_stats(evsel_list); | |
f87027b9 | 151 | perf_stat__reset_shadow_stats(); |
1eda3b21 JO |
152 | } |
153 | ||
cac21425 | 154 | static int create_perf_stat_counter(struct perf_evsel *evsel) |
ddcacfa0 | 155 | { |
69aad6f1 | 156 | struct perf_event_attr *attr = &evsel->attr; |
727ab04e | 157 | |
711a572e | 158 | if (stat_config.scale) |
a21ca2ca IM |
159 | attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | |
160 | PERF_FORMAT_TOTAL_TIME_RUNNING; | |
ddcacfa0 | 161 | |
5d2cd909 ACM |
162 | attr->inherit = !no_inherit; |
163 | ||
602ad878 | 164 | if (target__has_cpu(&target)) |
594ac61a | 165 | return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel)); |
5622c07b | 166 | |
602ad878 | 167 | if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) { |
48290609 | 168 | attr->disabled = 1; |
41191688 AK |
169 | if (!initial_delay) |
170 | attr->enable_on_exec = 1; | |
ddcacfa0 | 171 | } |
084ab9f8 | 172 | |
594ac61a | 173 | return perf_evsel__open_per_thread(evsel, evsel_list->threads); |
ddcacfa0 IM |
174 | } |
175 | ||
c04f5e5d IM |
176 | /* |
177 | * Does the counter have nsecs as a unit? | |
178 | */ | |
daec78a0 | 179 | static inline int nsec_counter(struct perf_evsel *evsel) |
c04f5e5d | 180 | { |
daec78a0 ACM |
181 | if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || |
182 | perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) | |
c04f5e5d IM |
183 | return 1; |
184 | ||
185 | return 0; | |
186 | } | |
187 | ||
f5b4a9c3 SE |
188 | /* |
189 | * Read out the results of a single counter: | |
190 | * do not aggregate counts across CPUs in system-wide mode | |
191 | */ | |
c52b12ed | 192 | static int read_counter(struct perf_evsel *counter) |
f5b4a9c3 | 193 | { |
9bf1a529 JO |
194 | int nthreads = thread_map__nr(evsel_list->threads); |
195 | int ncpus = perf_evsel__nr_cpus(counter); | |
196 | int cpu, thread; | |
f5b4a9c3 | 197 | |
3b4331d9 SP |
198 | if (!counter->supported) |
199 | return -ENOENT; | |
200 | ||
9bf1a529 JO |
201 | if (counter->system_wide) |
202 | nthreads = 1; | |
203 | ||
204 | for (thread = 0; thread < nthreads; thread++) { | |
205 | for (cpu = 0; cpu < ncpus; cpu++) { | |
3b3eb044 JO |
206 | struct perf_counts_values *count; |
207 | ||
208 | count = perf_counts(counter->counts, cpu, thread); | |
209 | if (perf_evsel__read(counter, cpu, thread, count)) | |
9bf1a529 JO |
210 | return -1; |
211 | } | |
f5b4a9c3 | 212 | } |
c52b12ed ACM |
213 | |
214 | return 0; | |
2996f5dd IM |
215 | } |
216 | ||
5fc472a6 | 217 | static void read_counters(bool close_counters) |
13370a9b | 218 | { |
13370a9b | 219 | struct perf_evsel *counter; |
13370a9b | 220 | |
106a94a0 | 221 | evlist__for_each(evsel_list, counter) { |
3b3eb044 | 222 | if (read_counter(counter)) |
245bad8e | 223 | pr_debug("failed to read counter %s\n", counter->name); |
3b3eb044 | 224 | |
f80010eb | 225 | if (perf_stat_process_counter(&stat_config, counter)) |
3b3eb044 | 226 | pr_warning("failed to process counter %s\n", counter->name); |
106a94a0 | 227 | |
5fc472a6 | 228 | if (close_counters) { |
106a94a0 JO |
229 | perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), |
230 | thread_map__nr(evsel_list->threads)); | |
13370a9b SE |
231 | } |
232 | } | |
106a94a0 JO |
233 | } |
234 | ||
ba411a95 | 235 | static void process_interval(void) |
106a94a0 | 236 | { |
106a94a0 | 237 | struct timespec ts, rs; |
106a94a0 JO |
238 | |
239 | read_counters(false); | |
86ee6e18 | 240 | |
13370a9b SE |
241 | clock_gettime(CLOCK_MONOTONIC, &ts); |
242 | diff_timespec(&rs, &ts, &ref_time); | |
13370a9b | 243 | |
d4f63a47 | 244 | print_counters(&rs, 0, NULL); |
13370a9b SE |
245 | } |
246 | ||
41191688 AK |
247 | static void handle_initial_delay(void) |
248 | { | |
249 | struct perf_evsel *counter; | |
250 | ||
251 | if (initial_delay) { | |
252 | const int ncpus = cpu_map__nr(evsel_list->cpus), | |
253 | nthreads = thread_map__nr(evsel_list->threads); | |
254 | ||
255 | usleep(initial_delay * 1000); | |
0050f7aa | 256 | evlist__for_each(evsel_list, counter) |
41191688 AK |
257 | perf_evsel__enable(counter, ncpus, nthreads); |
258 | } | |
259 | } | |
260 | ||
f33cbe72 | 261 | static volatile int workload_exec_errno; |
6af206fd ACM |
262 | |
263 | /* | |
264 | * perf_evlist__prepare_workload will send a SIGUSR1 | |
265 | * if the fork fails, since we asked by setting its | |
266 | * want_signal to true. | |
267 | */ | |
f33cbe72 ACM |
268 | static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, |
269 | void *ucontext __maybe_unused) | |
6af206fd | 270 | { |
f33cbe72 | 271 | workload_exec_errno = info->si_value.sival_int; |
6af206fd ACM |
272 | } |
273 | ||
acf28922 | 274 | static int __run_perf_stat(int argc, const char **argv) |
42202dd5 | 275 | { |
ec0d3d1f | 276 | int interval = stat_config.interval; |
56e52e85 | 277 | char msg[512]; |
42202dd5 | 278 | unsigned long long t0, t1; |
cac21425 | 279 | struct perf_evsel *counter; |
13370a9b | 280 | struct timespec ts; |
410136f5 | 281 | size_t l; |
42202dd5 | 282 | int status = 0; |
6be2850e | 283 | const bool forks = (argc > 0); |
42202dd5 | 284 | |
13370a9b SE |
285 | if (interval) { |
286 | ts.tv_sec = interval / 1000; | |
287 | ts.tv_nsec = (interval % 1000) * 1000000; | |
288 | } else { | |
289 | ts.tv_sec = 1; | |
290 | ts.tv_nsec = 0; | |
291 | } | |
292 | ||
60666c63 | 293 | if (forks) { |
735f7e0b ACM |
294 | if (perf_evlist__prepare_workload(evsel_list, &target, argv, false, |
295 | workload_exec_failed_signal) < 0) { | |
acf28922 NK |
296 | perror("failed to prepare workload"); |
297 | return -1; | |
60666c63 | 298 | } |
d20a47e7 | 299 | child_pid = evsel_list->workload.pid; |
051ae7f7 PM |
300 | } |
301 | ||
6a4bb04c | 302 | if (group) |
63dab225 | 303 | perf_evlist__set_leader(evsel_list); |
6a4bb04c | 304 | |
0050f7aa | 305 | evlist__for_each(evsel_list, counter) { |
cac21425 | 306 | if (create_perf_stat_counter(counter) < 0) { |
979987a5 DA |
307 | /* |
308 | * PPC returns ENXIO for HW counters until 2.6.37 | |
309 | * (behavior changed with commit b0a873e). | |
310 | */ | |
38f6ae1e | 311 | if (errno == EINVAL || errno == ENOSYS || |
979987a5 DA |
312 | errno == ENOENT || errno == EOPNOTSUPP || |
313 | errno == ENXIO) { | |
c63ca0c0 DA |
314 | if (verbose) |
315 | ui__warning("%s event is not supported by the kernel.\n", | |
7289f83c | 316 | perf_evsel__name(counter)); |
2cee77c4 | 317 | counter->supported = false; |
cb5ef600 KL |
318 | |
319 | if ((counter->leader != counter) || | |
320 | !(counter->leader->nr_members > 1)) | |
321 | continue; | |
c63ca0c0 | 322 | } |
ede70290 | 323 | |
56e52e85 ACM |
324 | perf_evsel__open_strerror(counter, &target, |
325 | errno, msg, sizeof(msg)); | |
326 | ui__error("%s\n", msg); | |
327 | ||
48290609 ACM |
328 | if (child_pid != -1) |
329 | kill(child_pid, SIGTERM); | |
fceda7fe | 330 | |
48290609 ACM |
331 | return -1; |
332 | } | |
2cee77c4 | 333 | counter->supported = true; |
410136f5 SE |
334 | |
335 | l = strlen(counter->unit); | |
336 | if (l > unit_width) | |
337 | unit_width = l; | |
084ab9f8 | 338 | } |
42202dd5 | 339 | |
23d4aad4 ACM |
340 | if (perf_evlist__apply_filters(evsel_list, &counter)) { |
341 | error("failed to set filter \"%s\" on event %s with %d (%s)\n", | |
342 | counter->filter, perf_evsel__name(counter), errno, | |
759e612b | 343 | strerror_r(errno, msg, sizeof(msg))); |
cfd748ae FW |
344 | return -1; |
345 | } | |
346 | ||
42202dd5 IM |
347 | /* |
348 | * Enable counters and exec the command: | |
349 | */ | |
350 | t0 = rdclock(); | |
13370a9b | 351 | clock_gettime(CLOCK_MONOTONIC, &ref_time); |
42202dd5 | 352 | |
60666c63 | 353 | if (forks) { |
acf28922 | 354 | perf_evlist__start_workload(evsel_list); |
41191688 | 355 | handle_initial_delay(); |
acf28922 | 356 | |
13370a9b SE |
357 | if (interval) { |
358 | while (!waitpid(child_pid, &status, WNOHANG)) { | |
359 | nanosleep(&ts, NULL); | |
ba411a95 | 360 | process_interval(); |
13370a9b SE |
361 | } |
362 | } | |
60666c63 | 363 | wait(&status); |
6af206fd | 364 | |
f33cbe72 ACM |
365 | if (workload_exec_errno) { |
366 | const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg)); | |
367 | pr_err("Workload failed: %s\n", emsg); | |
6af206fd | 368 | return -1; |
f33cbe72 | 369 | } |
6af206fd | 370 | |
33e49ea7 AK |
371 | if (WIFSIGNALED(status)) |
372 | psignal(WTERMSIG(status), argv[0]); | |
60666c63 | 373 | } else { |
41191688 | 374 | handle_initial_delay(); |
13370a9b SE |
375 | while (!done) { |
376 | nanosleep(&ts, NULL); | |
377 | if (interval) | |
ba411a95 | 378 | process_interval(); |
13370a9b | 379 | } |
60666c63 | 380 | } |
42202dd5 | 381 | |
42202dd5 IM |
382 | t1 = rdclock(); |
383 | ||
9e9772c4 | 384 | update_stats(&walltime_nsecs_stats, t1 - t0); |
42202dd5 | 385 | |
106a94a0 | 386 | read_counters(true); |
c52b12ed | 387 | |
42202dd5 IM |
388 | return WEXITSTATUS(status); |
389 | } | |
390 | ||
41cde476 | 391 | static int run_perf_stat(int argc, const char **argv) |
1f16c575 PZ |
392 | { |
393 | int ret; | |
394 | ||
395 | if (pre_cmd) { | |
396 | ret = system(pre_cmd); | |
397 | if (ret) | |
398 | return ret; | |
399 | } | |
400 | ||
401 | if (sync_run) | |
402 | sync(); | |
403 | ||
404 | ret = __run_perf_stat(argc, argv); | |
405 | if (ret) | |
406 | return ret; | |
407 | ||
408 | if (post_cmd) { | |
409 | ret = system(post_cmd); | |
410 | if (ret) | |
411 | return ret; | |
412 | } | |
413 | ||
414 | return ret; | |
415 | } | |
416 | ||
d73515c0 AK |
417 | static void print_running(u64 run, u64 ena) |
418 | { | |
419 | if (csv_output) { | |
5821522e | 420 | fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f", |
d73515c0 AK |
421 | csv_sep, |
422 | run, | |
423 | csv_sep, | |
424 | ena ? 100.0 * run / ena : 100.0); | |
425 | } else if (run != ena) { | |
5821522e | 426 | fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena); |
d73515c0 AK |
427 | } |
428 | } | |
429 | ||
f99844cb IM |
430 | static void print_noise_pct(double total, double avg) |
431 | { | |
0007ecea | 432 | double pct = rel_stddev_stats(total, avg); |
f99844cb | 433 | |
3ae9a34d | 434 | if (csv_output) |
5821522e | 435 | fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct); |
a1bca6cc | 436 | else if (pct) |
5821522e | 437 | fprintf(stat_config.output, " ( +-%6.2f%% )", pct); |
f99844cb IM |
438 | } |
439 | ||
69aad6f1 | 440 | static void print_noise(struct perf_evsel *evsel, double avg) |
42202dd5 | 441 | { |
581cc8a2 | 442 | struct perf_stat_evsel *ps; |
69aad6f1 | 443 | |
849abde9 PZ |
444 | if (run_count == 1) |
445 | return; | |
446 | ||
69aad6f1 | 447 | ps = evsel->priv; |
f99844cb | 448 | print_noise_pct(stddev_stats(&ps->res_stats[0]), avg); |
42202dd5 IM |
449 | } |
450 | ||
12c08a9f | 451 | static void aggr_printout(struct perf_evsel *evsel, int id, int nr) |
44175b6f | 452 | { |
421a50f3 | 453 | switch (stat_config.aggr_mode) { |
12c08a9f | 454 | case AGGR_CORE: |
5821522e | 455 | fprintf(stat_config.output, "S%d-C%*d%s%*d%s", |
12c08a9f SE |
456 | cpu_map__id_to_socket(id), |
457 | csv_output ? 0 : -8, | |
458 | cpu_map__id_to_cpu(id), | |
459 | csv_sep, | |
460 | csv_output ? 0 : 4, | |
461 | nr, | |
462 | csv_sep); | |
463 | break; | |
86ee6e18 | 464 | case AGGR_SOCKET: |
5821522e | 465 | fprintf(stat_config.output, "S%*d%s%*d%s", |
d7e7a451 | 466 | csv_output ? 0 : -5, |
12c08a9f | 467 | id, |
d7e7a451 SE |
468 | csv_sep, |
469 | csv_output ? 0 : 4, | |
470 | nr, | |
471 | csv_sep); | |
86ee6e18 SE |
472 | break; |
473 | case AGGR_NONE: | |
5821522e | 474 | fprintf(stat_config.output, "CPU%*d%s", |
d7470b6a | 475 | csv_output ? 0 : -4, |
12c08a9f | 476 | perf_evsel__cpus(evsel)->map[id], csv_sep); |
86ee6e18 | 477 | break; |
32b8af82 | 478 | case AGGR_THREAD: |
5821522e | 479 | fprintf(stat_config.output, "%*s-%*d%s", |
32b8af82 JO |
480 | csv_output ? 0 : 16, |
481 | thread_map__comm(evsel->threads, id), | |
482 | csv_output ? 0 : -8, | |
483 | thread_map__pid(evsel->threads, id), | |
484 | csv_sep); | |
485 | break; | |
86ee6e18 | 486 | case AGGR_GLOBAL: |
208df99e | 487 | case AGGR_UNSET: |
86ee6e18 SE |
488 | default: |
489 | break; | |
490 | } | |
491 | } | |
492 | ||
da88c7f7 | 493 | static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg) |
86ee6e18 | 494 | { |
5821522e | 495 | FILE *output = stat_config.output; |
86ee6e18 | 496 | double msecs = avg / 1e6; |
410136f5 | 497 | const char *fmt_v, *fmt_n; |
4bbe5a61 | 498 | char name[25]; |
86ee6e18 | 499 | |
410136f5 SE |
500 | fmt_v = csv_output ? "%.6f%s" : "%18.6f%s"; |
501 | fmt_n = csv_output ? "%s" : "%-25s"; | |
502 | ||
da88c7f7 | 503 | aggr_printout(evsel, id, nr); |
d7470b6a | 504 | |
4bbe5a61 DA |
505 | scnprintf(name, sizeof(name), "%s%s", |
506 | perf_evsel__name(evsel), csv_output ? "" : " (msec)"); | |
410136f5 SE |
507 | |
508 | fprintf(output, fmt_v, msecs, csv_sep); | |
509 | ||
510 | if (csv_output) | |
511 | fprintf(output, "%s%s", evsel->unit, csv_sep); | |
512 | else | |
513 | fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep); | |
514 | ||
515 | fprintf(output, fmt_n, name); | |
d7470b6a | 516 | |
023695d9 | 517 | if (evsel->cgrp) |
4aa9015f | 518 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); |
44175b6f IM |
519 | } |
520 | ||
556b1fb7 JO |
521 | static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg) |
522 | { | |
5821522e | 523 | FILE *output = stat_config.output; |
556b1fb7 JO |
524 | double sc = evsel->scale; |
525 | const char *fmt; | |
556b1fb7 JO |
526 | |
527 | if (csv_output) { | |
528 | fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s"; | |
529 | } else { | |
530 | if (big_num) | |
531 | fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s"; | |
532 | else | |
533 | fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s"; | |
534 | } | |
535 | ||
536 | aggr_printout(evsel, id, nr); | |
537 | ||
556b1fb7 JO |
538 | fprintf(output, fmt, avg, csv_sep); |
539 | ||
540 | if (evsel->unit) | |
541 | fprintf(output, "%-*s%s", | |
542 | csv_output ? 0 : unit_width, | |
543 | evsel->unit, csv_sep); | |
544 | ||
545 | fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel)); | |
546 | ||
547 | if (evsel->cgrp) | |
548 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); | |
eedfcb4b | 549 | } |
556b1fb7 | 550 | |
eedfcb4b AK |
551 | static void printout(int id, int nr, struct perf_evsel *counter, double uval) |
552 | { | |
553 | int cpu = cpu_map__id_to_cpu(id); | |
554 | ||
555 | if (stat_config.aggr_mode == AGGR_GLOBAL) | |
556 | cpu = 0; | |
557 | ||
558 | if (nsec_counter(counter)) | |
559 | nsec_printout(id, nr, counter, uval); | |
560 | else | |
561 | abs_printout(id, nr, counter, uval); | |
556b1fb7 | 562 | |
eedfcb4b AK |
563 | if (!csv_output && !stat_config.interval) |
564 | perf_stat__print_shadow_stats(stat_config.output, counter, | |
565 | uval, cpu, | |
566 | stat_config.aggr_mode); | |
556b1fb7 JO |
567 | } |
568 | ||
86ee6e18 | 569 | static void print_aggr(char *prefix) |
d7e7a451 | 570 | { |
5821522e | 571 | FILE *output = stat_config.output; |
d7e7a451 | 572 | struct perf_evsel *counter; |
601083cf | 573 | int cpu, s, s2, id, nr; |
410136f5 | 574 | double uval; |
d7e7a451 | 575 | u64 ena, run, val; |
d7e7a451 | 576 | |
86ee6e18 | 577 | if (!(aggr_map || aggr_get_id)) |
d7e7a451 SE |
578 | return; |
579 | ||
86ee6e18 SE |
580 | for (s = 0; s < aggr_map->nr; s++) { |
581 | id = aggr_map->map[s]; | |
0050f7aa | 582 | evlist__for_each(evsel_list, counter) { |
d7e7a451 SE |
583 | val = ena = run = 0; |
584 | nr = 0; | |
585 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { | |
601083cf | 586 | s2 = aggr_get_id(perf_evsel__cpus(counter), cpu); |
86ee6e18 | 587 | if (s2 != id) |
d7e7a451 | 588 | continue; |
a6fa0038 JO |
589 | val += perf_counts(counter->counts, cpu, 0)->val; |
590 | ena += perf_counts(counter->counts, cpu, 0)->ena; | |
591 | run += perf_counts(counter->counts, cpu, 0)->run; | |
d7e7a451 SE |
592 | nr++; |
593 | } | |
594 | if (prefix) | |
595 | fprintf(output, "%s", prefix); | |
596 | ||
597 | if (run == 0 || ena == 0) { | |
582ec082 | 598 | aggr_printout(counter, id, nr); |
86ee6e18 | 599 | |
410136f5 | 600 | fprintf(output, "%*s%s", |
d7e7a451 SE |
601 | csv_output ? 0 : 18, |
602 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, | |
410136f5 SE |
603 | csv_sep); |
604 | ||
605 | fprintf(output, "%-*s%s", | |
606 | csv_output ? 0 : unit_width, | |
607 | counter->unit, csv_sep); | |
608 | ||
609 | fprintf(output, "%*s", | |
610 | csv_output ? 0 : -25, | |
d7e7a451 | 611 | perf_evsel__name(counter)); |
86ee6e18 | 612 | |
d7e7a451 SE |
613 | if (counter->cgrp) |
614 | fprintf(output, "%s%s", | |
615 | csv_sep, counter->cgrp->name); | |
616 | ||
d73515c0 | 617 | print_running(run, ena); |
d7e7a451 SE |
618 | fputc('\n', output); |
619 | continue; | |
620 | } | |
410136f5 | 621 | uval = val * counter->scale; |
eedfcb4b | 622 | printout(id, nr, counter, uval); |
d73515c0 | 623 | if (!csv_output) |
d7e7a451 SE |
624 | print_noise(counter, 1.0); |
625 | ||
d73515c0 | 626 | print_running(run, ena); |
d7e7a451 SE |
627 | fputc('\n', output); |
628 | } | |
629 | } | |
630 | } | |
631 | ||
32b8af82 JO |
632 | static void print_aggr_thread(struct perf_evsel *counter, char *prefix) |
633 | { | |
5821522e | 634 | FILE *output = stat_config.output; |
32b8af82 JO |
635 | int nthreads = thread_map__nr(counter->threads); |
636 | int ncpus = cpu_map__nr(counter->cpus); | |
637 | int cpu, thread; | |
638 | double uval; | |
639 | ||
640 | for (thread = 0; thread < nthreads; thread++) { | |
641 | u64 ena = 0, run = 0, val = 0; | |
642 | ||
643 | for (cpu = 0; cpu < ncpus; cpu++) { | |
644 | val += perf_counts(counter->counts, cpu, thread)->val; | |
645 | ena += perf_counts(counter->counts, cpu, thread)->ena; | |
646 | run += perf_counts(counter->counts, cpu, thread)->run; | |
647 | } | |
648 | ||
649 | if (prefix) | |
650 | fprintf(output, "%s", prefix); | |
651 | ||
652 | uval = val * counter->scale; | |
eedfcb4b | 653 | printout(thread, 0, counter, uval); |
32b8af82 JO |
654 | |
655 | if (!csv_output) | |
656 | print_noise(counter, 1.0); | |
657 | ||
658 | print_running(run, ena); | |
659 | fputc('\n', output); | |
660 | } | |
661 | } | |
662 | ||
2996f5dd IM |
663 | /* |
664 | * Print out the results of a single counter: | |
f5b4a9c3 | 665 | * aggregated counts in system-wide mode |
2996f5dd | 666 | */ |
13370a9b | 667 | static void print_counter_aggr(struct perf_evsel *counter, char *prefix) |
2996f5dd | 668 | { |
5821522e | 669 | FILE *output = stat_config.output; |
581cc8a2 | 670 | struct perf_stat_evsel *ps = counter->priv; |
69aad6f1 | 671 | double avg = avg_stats(&ps->res_stats[0]); |
c52b12ed | 672 | int scaled = counter->counts->scaled; |
410136f5 | 673 | double uval; |
d73515c0 AK |
674 | double avg_enabled, avg_running; |
675 | ||
676 | avg_enabled = avg_stats(&ps->res_stats[1]); | |
677 | avg_running = avg_stats(&ps->res_stats[2]); | |
2996f5dd | 678 | |
13370a9b SE |
679 | if (prefix) |
680 | fprintf(output, "%s", prefix); | |
681 | ||
3b4331d9 | 682 | if (scaled == -1 || !counter->supported) { |
410136f5 | 683 | fprintf(output, "%*s%s", |
d7470b6a | 684 | csv_output ? 0 : 18, |
2cee77c4 | 685 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, |
410136f5 SE |
686 | csv_sep); |
687 | fprintf(output, "%-*s%s", | |
688 | csv_output ? 0 : unit_width, | |
689 | counter->unit, csv_sep); | |
690 | fprintf(output, "%*s", | |
691 | csv_output ? 0 : -25, | |
7289f83c | 692 | perf_evsel__name(counter)); |
023695d9 SE |
693 | |
694 | if (counter->cgrp) | |
4aa9015f | 695 | fprintf(output, "%s%s", csv_sep, counter->cgrp->name); |
023695d9 | 696 | |
d73515c0 | 697 | print_running(avg_running, avg_enabled); |
4aa9015f | 698 | fputc('\n', output); |
2996f5dd IM |
699 | return; |
700 | } | |
c04f5e5d | 701 | |
410136f5 | 702 | uval = avg * counter->scale; |
eedfcb4b | 703 | printout(-1, 0, counter, uval); |
849abde9 | 704 | |
3ae9a34d ZH |
705 | print_noise(counter, avg); |
706 | ||
d73515c0 | 707 | print_running(avg_running, avg_enabled); |
4aa9015f | 708 | fprintf(output, "\n"); |
c04f5e5d IM |
709 | } |
710 | ||
f5b4a9c3 SE |
711 | /* |
712 | * Print out the results of a single counter: | |
713 | * does not use aggregated count in system-wide | |
714 | */ | |
13370a9b | 715 | static void print_counter(struct perf_evsel *counter, char *prefix) |
f5b4a9c3 | 716 | { |
5821522e | 717 | FILE *output = stat_config.output; |
f5b4a9c3 | 718 | u64 ena, run, val; |
410136f5 | 719 | double uval; |
f5b4a9c3 SE |
720 | int cpu; |
721 | ||
7ae92e74 | 722 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { |
a6fa0038 JO |
723 | val = perf_counts(counter->counts, cpu, 0)->val; |
724 | ena = perf_counts(counter->counts, cpu, 0)->ena; | |
725 | run = perf_counts(counter->counts, cpu, 0)->run; | |
13370a9b SE |
726 | |
727 | if (prefix) | |
728 | fprintf(output, "%s", prefix); | |
729 | ||
f5b4a9c3 | 730 | if (run == 0 || ena == 0) { |
410136f5 | 731 | fprintf(output, "CPU%*d%s%*s%s", |
d7470b6a | 732 | csv_output ? 0 : -4, |
7ae92e74 | 733 | perf_evsel__cpus(counter)->map[cpu], csv_sep, |
d7470b6a | 734 | csv_output ? 0 : 18, |
2cee77c4 | 735 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, |
410136f5 SE |
736 | csv_sep); |
737 | ||
738 | fprintf(output, "%-*s%s", | |
739 | csv_output ? 0 : unit_width, | |
740 | counter->unit, csv_sep); | |
741 | ||
742 | fprintf(output, "%*s", | |
743 | csv_output ? 0 : -25, | |
744 | perf_evsel__name(counter)); | |
f5b4a9c3 | 745 | |
023695d9 | 746 | if (counter->cgrp) |
4aa9015f SE |
747 | fprintf(output, "%s%s", |
748 | csv_sep, counter->cgrp->name); | |
023695d9 | 749 | |
d73515c0 | 750 | print_running(run, ena); |
4aa9015f | 751 | fputc('\n', output); |
f5b4a9c3 SE |
752 | continue; |
753 | } | |
754 | ||
410136f5 | 755 | uval = val * counter->scale; |
eedfcb4b | 756 | printout(cpu, 0, counter, uval); |
d73515c0 | 757 | if (!csv_output) |
d7470b6a | 758 | print_noise(counter, 1.0); |
d73515c0 | 759 | print_running(run, ena); |
f5b4a9c3 | 760 | |
4aa9015f | 761 | fputc('\n', output); |
f5b4a9c3 SE |
762 | } |
763 | } | |
764 | ||
d4f63a47 JO |
765 | static void print_interval(char *prefix, struct timespec *ts) |
766 | { | |
5821522e | 767 | FILE *output = stat_config.output; |
d4f63a47 JO |
768 | static int num_print_interval; |
769 | ||
770 | sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep); | |
771 | ||
772 | if (num_print_interval == 0 && !csv_output) { | |
421a50f3 | 773 | switch (stat_config.aggr_mode) { |
d4f63a47 JO |
774 | case AGGR_SOCKET: |
775 | fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit"); | |
776 | break; | |
777 | case AGGR_CORE: | |
778 | fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit"); | |
779 | break; | |
780 | case AGGR_NONE: | |
781 | fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit"); | |
782 | break; | |
32b8af82 JO |
783 | case AGGR_THREAD: |
784 | fprintf(output, "# time comm-pid counts %*s events\n", unit_width, "unit"); | |
785 | break; | |
d4f63a47 JO |
786 | case AGGR_GLOBAL: |
787 | default: | |
788 | fprintf(output, "# time counts %*s events\n", unit_width, "unit"); | |
208df99e JO |
789 | case AGGR_UNSET: |
790 | break; | |
d4f63a47 JO |
791 | } |
792 | } | |
793 | ||
794 | if (++num_print_interval == 25) | |
795 | num_print_interval = 0; | |
796 | } | |
797 | ||
798 | static void print_header(int argc, const char **argv) | |
42202dd5 | 799 | { |
5821522e | 800 | FILE *output = stat_config.output; |
69aad6f1 | 801 | int i; |
42202dd5 | 802 | |
ddcacfa0 IM |
803 | fflush(stdout); |
804 | ||
d7470b6a | 805 | if (!csv_output) { |
4aa9015f SE |
806 | fprintf(output, "\n"); |
807 | fprintf(output, " Performance counter stats for "); | |
62d3b617 DA |
808 | if (target.system_wide) |
809 | fprintf(output, "\'system wide"); | |
810 | else if (target.cpu_list) | |
811 | fprintf(output, "\'CPU(s) %s", target.cpu_list); | |
602ad878 | 812 | else if (!target__has_task(&target)) { |
4aa9015f | 813 | fprintf(output, "\'%s", argv[0]); |
d7470b6a | 814 | for (i = 1; i < argc; i++) |
4aa9015f | 815 | fprintf(output, " %s", argv[i]); |
20f946b4 NK |
816 | } else if (target.pid) |
817 | fprintf(output, "process id \'%s", target.pid); | |
d7470b6a | 818 | else |
20f946b4 | 819 | fprintf(output, "thread id \'%s", target.tid); |
44db76c8 | 820 | |
4aa9015f | 821 | fprintf(output, "\'"); |
d7470b6a | 822 | if (run_count > 1) |
4aa9015f SE |
823 | fprintf(output, " (%d runs)", run_count); |
824 | fprintf(output, ":\n\n"); | |
d7470b6a | 825 | } |
d4f63a47 JO |
826 | } |
827 | ||
828 | static void print_footer(void) | |
829 | { | |
5821522e JO |
830 | FILE *output = stat_config.output; |
831 | ||
d4f63a47 JO |
832 | if (!null_run) |
833 | fprintf(output, "\n"); | |
834 | fprintf(output, " %17.9f seconds time elapsed", | |
835 | avg_stats(&walltime_nsecs_stats)/1e9); | |
836 | if (run_count > 1) { | |
837 | fprintf(output, " "); | |
838 | print_noise_pct(stddev_stats(&walltime_nsecs_stats), | |
839 | avg_stats(&walltime_nsecs_stats)); | |
840 | } | |
841 | fprintf(output, "\n\n"); | |
842 | } | |
843 | ||
844 | static void print_counters(struct timespec *ts, int argc, const char **argv) | |
845 | { | |
ec0d3d1f | 846 | int interval = stat_config.interval; |
d4f63a47 JO |
847 | struct perf_evsel *counter; |
848 | char buf[64], *prefix = NULL; | |
849 | ||
850 | if (interval) | |
851 | print_interval(prefix = buf, ts); | |
852 | else | |
853 | print_header(argc, argv); | |
2996f5dd | 854 | |
421a50f3 | 855 | switch (stat_config.aggr_mode) { |
12c08a9f | 856 | case AGGR_CORE: |
86ee6e18 | 857 | case AGGR_SOCKET: |
d4f63a47 | 858 | print_aggr(prefix); |
86ee6e18 | 859 | break; |
32b8af82 JO |
860 | case AGGR_THREAD: |
861 | evlist__for_each(evsel_list, counter) | |
862 | print_aggr_thread(counter, prefix); | |
863 | break; | |
86ee6e18 | 864 | case AGGR_GLOBAL: |
0050f7aa | 865 | evlist__for_each(evsel_list, counter) |
d4f63a47 | 866 | print_counter_aggr(counter, prefix); |
86ee6e18 SE |
867 | break; |
868 | case AGGR_NONE: | |
0050f7aa | 869 | evlist__for_each(evsel_list, counter) |
d4f63a47 | 870 | print_counter(counter, prefix); |
86ee6e18 | 871 | break; |
208df99e | 872 | case AGGR_UNSET: |
86ee6e18 SE |
873 | default: |
874 | break; | |
f5b4a9c3 | 875 | } |
ddcacfa0 | 876 | |
d4f63a47 JO |
877 | if (!interval && !csv_output) |
878 | print_footer(); | |
879 | ||
5821522e | 880 | fflush(stat_config.output); |
ddcacfa0 IM |
881 | } |
882 | ||
f7b7c26e PZ |
883 | static volatile int signr = -1; |
884 | ||
5242519b | 885 | static void skip_signal(int signo) |
ddcacfa0 | 886 | { |
ec0d3d1f | 887 | if ((child_pid == -1) || stat_config.interval) |
60666c63 LW |
888 | done = 1; |
889 | ||
f7b7c26e | 890 | signr = signo; |
d07f0b12 SE |
891 | /* |
892 | * render child_pid harmless | |
893 | * won't send SIGTERM to a random | |
894 | * process in case of race condition | |
895 | * and fast PID recycling | |
896 | */ | |
897 | child_pid = -1; | |
f7b7c26e PZ |
898 | } |
899 | ||
900 | static void sig_atexit(void) | |
901 | { | |
d07f0b12 SE |
902 | sigset_t set, oset; |
903 | ||
904 | /* | |
905 | * avoid race condition with SIGCHLD handler | |
906 | * in skip_signal() which is modifying child_pid | |
907 | * goal is to avoid send SIGTERM to a random | |
908 | * process | |
909 | */ | |
910 | sigemptyset(&set); | |
911 | sigaddset(&set, SIGCHLD); | |
912 | sigprocmask(SIG_BLOCK, &set, &oset); | |
913 | ||
933da83a CW |
914 | if (child_pid != -1) |
915 | kill(child_pid, SIGTERM); | |
916 | ||
d07f0b12 SE |
917 | sigprocmask(SIG_SETMASK, &oset, NULL); |
918 | ||
f7b7c26e PZ |
919 | if (signr == -1) |
920 | return; | |
921 | ||
922 | signal(signr, SIG_DFL); | |
923 | kill(getpid(), signr); | |
5242519b IM |
924 | } |
925 | ||
1d037ca1 IT |
926 | static int stat__set_big_num(const struct option *opt __maybe_unused, |
927 | const char *s __maybe_unused, int unset) | |
d7470b6a SE |
928 | { |
929 | big_num_opt = unset ? 0 : 1; | |
930 | return 0; | |
931 | } | |
932 | ||
e0547311 JO |
933 | static const struct option stat_options[] = { |
934 | OPT_BOOLEAN('T', "transaction", &transaction_run, | |
935 | "hardware transaction statistics"), | |
936 | OPT_CALLBACK('e', "event", &evsel_list, "event", | |
937 | "event selector. use 'perf list' to list available events", | |
938 | parse_events_option), | |
939 | OPT_CALLBACK(0, "filter", &evsel_list, "filter", | |
940 | "event filter", parse_filter), | |
941 | OPT_BOOLEAN('i', "no-inherit", &no_inherit, | |
942 | "child tasks do not inherit counters"), | |
943 | OPT_STRING('p', "pid", &target.pid, "pid", | |
944 | "stat events on existing process id"), | |
945 | OPT_STRING('t', "tid", &target.tid, "tid", | |
946 | "stat events on existing thread id"), | |
947 | OPT_BOOLEAN('a', "all-cpus", &target.system_wide, | |
948 | "system-wide collection from all CPUs"), | |
949 | OPT_BOOLEAN('g', "group", &group, | |
950 | "put the counters into a counter group"), | |
951 | OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"), | |
952 | OPT_INCR('v', "verbose", &verbose, | |
953 | "be more verbose (show counter open errors, etc)"), | |
954 | OPT_INTEGER('r', "repeat", &run_count, | |
955 | "repeat command and print average + stddev (max: 100, forever: 0)"), | |
956 | OPT_BOOLEAN('n', "null", &null_run, | |
957 | "null run - dont start any counters"), | |
958 | OPT_INCR('d', "detailed", &detailed_run, | |
959 | "detailed run - start a lot of events"), | |
960 | OPT_BOOLEAN('S', "sync", &sync_run, | |
961 | "call sync() before starting a run"), | |
962 | OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, | |
963 | "print large numbers with thousands\' separators", | |
964 | stat__set_big_num), | |
965 | OPT_STRING('C', "cpu", &target.cpu_list, "cpu", | |
966 | "list of cpus to monitor in system-wide"), | |
967 | OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode, | |
968 | "disable CPU count aggregation", AGGR_NONE), | |
969 | OPT_STRING('x', "field-separator", &csv_sep, "separator", | |
970 | "print counts with custom separator"), | |
971 | OPT_CALLBACK('G', "cgroup", &evsel_list, "name", | |
972 | "monitor event in cgroup name only", parse_cgroups), | |
973 | OPT_STRING('o', "output", &output_name, "file", "output file name"), | |
974 | OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), | |
975 | OPT_INTEGER(0, "log-fd", &output_fd, | |
976 | "log output to fd, instead of stderr"), | |
977 | OPT_STRING(0, "pre", &pre_cmd, "command", | |
978 | "command to run prior to the measured command"), | |
979 | OPT_STRING(0, "post", &post_cmd, "command", | |
980 | "command to run after to the measured command"), | |
981 | OPT_UINTEGER('I', "interval-print", &stat_config.interval, | |
982 | "print counts at regular interval in ms (>= 10)"), | |
983 | OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode, | |
984 | "aggregate counts per processor socket", AGGR_SOCKET), | |
985 | OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode, | |
986 | "aggregate counts per physical processor core", AGGR_CORE), | |
987 | OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode, | |
988 | "aggregate counts per thread", AGGR_THREAD), | |
989 | OPT_UINTEGER('D', "delay", &initial_delay, | |
990 | "ms to wait before starting measurement after program start"), | |
991 | OPT_END() | |
992 | }; | |
993 | ||
1fe7a300 JO |
994 | static int perf_stat__get_socket(struct cpu_map *map, int cpu) |
995 | { | |
996 | return cpu_map__get_socket(map, cpu, NULL); | |
997 | } | |
998 | ||
999 | static int perf_stat__get_core(struct cpu_map *map, int cpu) | |
1000 | { | |
1001 | return cpu_map__get_core(map, cpu, NULL); | |
1002 | } | |
1003 | ||
1e5a2931 JO |
1004 | static int cpu_map__get_max(struct cpu_map *map) |
1005 | { | |
1006 | int i, max = -1; | |
1007 | ||
1008 | for (i = 0; i < map->nr; i++) { | |
1009 | if (map->map[i] > max) | |
1010 | max = map->map[i]; | |
1011 | } | |
1012 | ||
1013 | return max; | |
1014 | } | |
1015 | ||
1016 | static struct cpu_map *cpus_aggr_map; | |
1017 | ||
1018 | static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx) | |
1019 | { | |
1020 | int cpu; | |
1021 | ||
1022 | if (idx >= map->nr) | |
1023 | return -1; | |
1024 | ||
1025 | cpu = map->map[idx]; | |
1026 | ||
1027 | if (cpus_aggr_map->map[cpu] == -1) | |
1028 | cpus_aggr_map->map[cpu] = get_id(map, idx); | |
1029 | ||
1030 | return cpus_aggr_map->map[cpu]; | |
1031 | } | |
1032 | ||
1033 | static int perf_stat__get_socket_cached(struct cpu_map *map, int idx) | |
1034 | { | |
1035 | return perf_stat__get_aggr(perf_stat__get_socket, map, idx); | |
1036 | } | |
1037 | ||
1038 | static int perf_stat__get_core_cached(struct cpu_map *map, int idx) | |
1039 | { | |
1040 | return perf_stat__get_aggr(perf_stat__get_core, map, idx); | |
1041 | } | |
1042 | ||
86ee6e18 SE |
1043 | static int perf_stat_init_aggr_mode(void) |
1044 | { | |
1e5a2931 JO |
1045 | int nr; |
1046 | ||
421a50f3 | 1047 | switch (stat_config.aggr_mode) { |
86ee6e18 SE |
1048 | case AGGR_SOCKET: |
1049 | if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) { | |
1050 | perror("cannot build socket map"); | |
1051 | return -1; | |
1052 | } | |
1e5a2931 | 1053 | aggr_get_id = perf_stat__get_socket_cached; |
86ee6e18 | 1054 | break; |
12c08a9f SE |
1055 | case AGGR_CORE: |
1056 | if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) { | |
1057 | perror("cannot build core map"); | |
1058 | return -1; | |
1059 | } | |
1e5a2931 | 1060 | aggr_get_id = perf_stat__get_core_cached; |
12c08a9f | 1061 | break; |
86ee6e18 SE |
1062 | case AGGR_NONE: |
1063 | case AGGR_GLOBAL: | |
32b8af82 | 1064 | case AGGR_THREAD: |
208df99e | 1065 | case AGGR_UNSET: |
86ee6e18 SE |
1066 | default: |
1067 | break; | |
1068 | } | |
1e5a2931 JO |
1069 | |
1070 | /* | |
1071 | * The evsel_list->cpus is the base we operate on, | |
1072 | * taking the highest cpu number to be the size of | |
1073 | * the aggregation translate cpumap. | |
1074 | */ | |
1075 | nr = cpu_map__get_max(evsel_list->cpus); | |
1076 | cpus_aggr_map = cpu_map__empty_new(nr + 1); | |
1077 | return cpus_aggr_map ? 0 : -ENOMEM; | |
86ee6e18 SE |
1078 | } |
1079 | ||
2cba3ffb IM |
1080 | /* |
1081 | * Add default attributes, if there were no attributes specified or | |
1082 | * if -d/--detailed, -d -d or -d -d -d is used: | |
1083 | */ | |
1084 | static int add_default_attributes(void) | |
1085 | { | |
b070a547 ACM |
1086 | struct perf_event_attr default_attrs[] = { |
1087 | ||
1088 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, | |
1089 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, | |
1090 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, | |
1091 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, | |
1092 | ||
1093 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, | |
1094 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, | |
1095 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, | |
1096 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, | |
1097 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, | |
1098 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, | |
1099 | ||
1100 | }; | |
1101 | ||
1102 | /* | |
1103 | * Detailed stats (-d), covering the L1 and last level data caches: | |
1104 | */ | |
1105 | struct perf_event_attr detailed_attrs[] = { | |
1106 | ||
1107 | { .type = PERF_TYPE_HW_CACHE, | |
1108 | .config = | |
1109 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1110 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1111 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1112 | ||
1113 | { .type = PERF_TYPE_HW_CACHE, | |
1114 | .config = | |
1115 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1116 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1117 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1118 | ||
1119 | { .type = PERF_TYPE_HW_CACHE, | |
1120 | .config = | |
1121 | PERF_COUNT_HW_CACHE_LL << 0 | | |
1122 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1123 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1124 | ||
1125 | { .type = PERF_TYPE_HW_CACHE, | |
1126 | .config = | |
1127 | PERF_COUNT_HW_CACHE_LL << 0 | | |
1128 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1129 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1130 | }; | |
1131 | ||
1132 | /* | |
1133 | * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: | |
1134 | */ | |
1135 | struct perf_event_attr very_detailed_attrs[] = { | |
1136 | ||
1137 | { .type = PERF_TYPE_HW_CACHE, | |
1138 | .config = | |
1139 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
1140 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1141 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1142 | ||
1143 | { .type = PERF_TYPE_HW_CACHE, | |
1144 | .config = | |
1145 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
1146 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1147 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1148 | ||
1149 | { .type = PERF_TYPE_HW_CACHE, | |
1150 | .config = | |
1151 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1152 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1153 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1154 | ||
1155 | { .type = PERF_TYPE_HW_CACHE, | |
1156 | .config = | |
1157 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1158 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1159 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1160 | ||
1161 | { .type = PERF_TYPE_HW_CACHE, | |
1162 | .config = | |
1163 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
1164 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1165 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1166 | ||
1167 | { .type = PERF_TYPE_HW_CACHE, | |
1168 | .config = | |
1169 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
1170 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1171 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1172 | ||
1173 | }; | |
1174 | ||
1175 | /* | |
1176 | * Very, very detailed stats (-d -d -d), adding prefetch events: | |
1177 | */ | |
1178 | struct perf_event_attr very_very_detailed_attrs[] = { | |
1179 | ||
1180 | { .type = PERF_TYPE_HW_CACHE, | |
1181 | .config = | |
1182 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1183 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
1184 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1185 | ||
1186 | { .type = PERF_TYPE_HW_CACHE, | |
1187 | .config = | |
1188 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1189 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
1190 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1191 | }; | |
1192 | ||
2cba3ffb IM |
1193 | /* Set attrs if no event is selected and !null_run: */ |
1194 | if (null_run) | |
1195 | return 0; | |
1196 | ||
4cabc3d1 AK |
1197 | if (transaction_run) { |
1198 | int err; | |
1199 | if (pmu_have_event("cpu", "cycles-ct") && | |
1200 | pmu_have_event("cpu", "el-start")) | |
a454742c | 1201 | err = parse_events(evsel_list, transaction_attrs, NULL); |
4cabc3d1 | 1202 | else |
a454742c JO |
1203 | err = parse_events(evsel_list, transaction_limited_attrs, NULL); |
1204 | if (err) { | |
4cabc3d1 AK |
1205 | fprintf(stderr, "Cannot set up transaction events\n"); |
1206 | return -1; | |
1207 | } | |
1208 | return 0; | |
1209 | } | |
1210 | ||
2cba3ffb | 1211 | if (!evsel_list->nr_entries) { |
79695e1b | 1212 | if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0) |
50d08e47 | 1213 | return -1; |
2cba3ffb IM |
1214 | } |
1215 | ||
1216 | /* Detailed events get appended to the event list: */ | |
1217 | ||
1218 | if (detailed_run < 1) | |
1219 | return 0; | |
1220 | ||
1221 | /* Append detailed run extra attributes: */ | |
79695e1b | 1222 | if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) |
50d08e47 | 1223 | return -1; |
2cba3ffb IM |
1224 | |
1225 | if (detailed_run < 2) | |
1226 | return 0; | |
1227 | ||
1228 | /* Append very detailed run extra attributes: */ | |
79695e1b | 1229 | if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) |
50d08e47 | 1230 | return -1; |
2cba3ffb IM |
1231 | |
1232 | if (detailed_run < 3) | |
1233 | return 0; | |
1234 | ||
1235 | /* Append very, very detailed run extra attributes: */ | |
79695e1b | 1236 | return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); |
2cba3ffb IM |
1237 | } |
1238 | ||
1d037ca1 | 1239 | int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused) |
5242519b | 1240 | { |
b070a547 ACM |
1241 | const char * const stat_usage[] = { |
1242 | "perf stat [<options>] [<command>]", | |
1243 | NULL | |
1244 | }; | |
cc03c542 | 1245 | int status = -EINVAL, run_idx; |
4aa9015f | 1246 | const char *mode; |
5821522e | 1247 | FILE *output = stderr; |
ec0d3d1f | 1248 | unsigned int interval; |
42202dd5 | 1249 | |
5af52b51 SE |
1250 | setlocale(LC_ALL, ""); |
1251 | ||
334fe7a3 | 1252 | evsel_list = perf_evlist__new(); |
361c99a6 ACM |
1253 | if (evsel_list == NULL) |
1254 | return -ENOMEM; | |
1255 | ||
e0547311 | 1256 | argc = parse_options(argc, argv, stat_options, stat_usage, |
a0541234 | 1257 | PARSE_OPT_STOP_AT_NON_OPTION); |
d7470b6a | 1258 | |
ec0d3d1f JO |
1259 | interval = stat_config.interval; |
1260 | ||
4aa9015f SE |
1261 | if (output_name && strcmp(output_name, "-")) |
1262 | output = NULL; | |
1263 | ||
56f3bae7 JC |
1264 | if (output_name && output_fd) { |
1265 | fprintf(stderr, "cannot use both --output and --log-fd\n"); | |
e0547311 JO |
1266 | parse_options_usage(stat_usage, stat_options, "o", 1); |
1267 | parse_options_usage(NULL, stat_options, "log-fd", 0); | |
cc03c542 | 1268 | goto out; |
56f3bae7 | 1269 | } |
fc3e4d07 SE |
1270 | |
1271 | if (output_fd < 0) { | |
1272 | fprintf(stderr, "argument to --log-fd must be a > 0\n"); | |
e0547311 | 1273 | parse_options_usage(stat_usage, stat_options, "log-fd", 0); |
cc03c542 | 1274 | goto out; |
fc3e4d07 SE |
1275 | } |
1276 | ||
4aa9015f SE |
1277 | if (!output) { |
1278 | struct timespec tm; | |
1279 | mode = append_file ? "a" : "w"; | |
1280 | ||
1281 | output = fopen(output_name, mode); | |
1282 | if (!output) { | |
1283 | perror("failed to create output file"); | |
fceda7fe | 1284 | return -1; |
4aa9015f SE |
1285 | } |
1286 | clock_gettime(CLOCK_REALTIME, &tm); | |
1287 | fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); | |
fc3e4d07 | 1288 | } else if (output_fd > 0) { |
56f3bae7 JC |
1289 | mode = append_file ? "a" : "w"; |
1290 | output = fdopen(output_fd, mode); | |
1291 | if (!output) { | |
1292 | perror("Failed opening logfd"); | |
1293 | return -errno; | |
1294 | } | |
4aa9015f SE |
1295 | } |
1296 | ||
5821522e JO |
1297 | stat_config.output = output; |
1298 | ||
d4ffd04d | 1299 | if (csv_sep) { |
d7470b6a | 1300 | csv_output = true; |
d4ffd04d JC |
1301 | if (!strcmp(csv_sep, "\\t")) |
1302 | csv_sep = "\t"; | |
1303 | } else | |
d7470b6a SE |
1304 | csv_sep = DEFAULT_SEPARATOR; |
1305 | ||
1306 | /* | |
1307 | * let the spreadsheet do the pretty-printing | |
1308 | */ | |
1309 | if (csv_output) { | |
61a9f324 | 1310 | /* User explicitly passed -B? */ |
d7470b6a SE |
1311 | if (big_num_opt == 1) { |
1312 | fprintf(stderr, "-B option not supported with -x\n"); | |
e0547311 JO |
1313 | parse_options_usage(stat_usage, stat_options, "B", 1); |
1314 | parse_options_usage(NULL, stat_options, "x", 1); | |
cc03c542 | 1315 | goto out; |
d7470b6a SE |
1316 | } else /* Nope, so disable big number formatting */ |
1317 | big_num = false; | |
1318 | } else if (big_num_opt == 0) /* User passed --no-big-num */ | |
1319 | big_num = false; | |
1320 | ||
602ad878 | 1321 | if (!argc && target__none(&target)) |
e0547311 | 1322 | usage_with_options(stat_usage, stat_options); |
ac3063bd | 1323 | |
a7e191c3 | 1324 | if (run_count < 0) { |
cc03c542 | 1325 | pr_err("Run count must be a positive number\n"); |
e0547311 | 1326 | parse_options_usage(stat_usage, stat_options, "r", 1); |
cc03c542 | 1327 | goto out; |
a7e191c3 FD |
1328 | } else if (run_count == 0) { |
1329 | forever = true; | |
1330 | run_count = 1; | |
1331 | } | |
ddcacfa0 | 1332 | |
421a50f3 | 1333 | if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) { |
32b8af82 JO |
1334 | fprintf(stderr, "The --per-thread option is only available " |
1335 | "when monitoring via -p -t options.\n"); | |
e0547311 JO |
1336 | parse_options_usage(NULL, stat_options, "p", 1); |
1337 | parse_options_usage(NULL, stat_options, "t", 1); | |
32b8af82 JO |
1338 | goto out; |
1339 | } | |
1340 | ||
1341 | /* | |
1342 | * no_aggr, cgroup are for system-wide only | |
1343 | * --per-thread is aggregated per thread, we dont mix it with cpu mode | |
1344 | */ | |
421a50f3 JO |
1345 | if (((stat_config.aggr_mode != AGGR_GLOBAL && |
1346 | stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) && | |
602ad878 | 1347 | !target__has_cpu(&target)) { |
023695d9 SE |
1348 | fprintf(stderr, "both cgroup and no-aggregation " |
1349 | "modes only available in system-wide mode\n"); | |
1350 | ||
e0547311 JO |
1351 | parse_options_usage(stat_usage, stat_options, "G", 1); |
1352 | parse_options_usage(NULL, stat_options, "A", 1); | |
1353 | parse_options_usage(NULL, stat_options, "a", 1); | |
cc03c542 | 1354 | goto out; |
d7e7a451 SE |
1355 | } |
1356 | ||
2cba3ffb IM |
1357 | if (add_default_attributes()) |
1358 | goto out; | |
ddcacfa0 | 1359 | |
602ad878 | 1360 | target__validate(&target); |
5c98d466 | 1361 | |
77a6f014 | 1362 | if (perf_evlist__create_maps(evsel_list, &target) < 0) { |
602ad878 | 1363 | if (target__has_task(&target)) { |
77a6f014 | 1364 | pr_err("Problems finding threads of monitor\n"); |
e0547311 JO |
1365 | parse_options_usage(stat_usage, stat_options, "p", 1); |
1366 | parse_options_usage(NULL, stat_options, "t", 1); | |
602ad878 | 1367 | } else if (target__has_cpu(&target)) { |
77a6f014 | 1368 | perror("failed to parse CPUs map"); |
e0547311 JO |
1369 | parse_options_usage(stat_usage, stat_options, "C", 1); |
1370 | parse_options_usage(NULL, stat_options, "a", 1); | |
cc03c542 NK |
1371 | } |
1372 | goto out; | |
60d567e2 | 1373 | } |
32b8af82 JO |
1374 | |
1375 | /* | |
1376 | * Initialize thread_map with comm names, | |
1377 | * so we could print it out on output. | |
1378 | */ | |
421a50f3 | 1379 | if (stat_config.aggr_mode == AGGR_THREAD) |
32b8af82 JO |
1380 | thread_map__read_comms(evsel_list->threads); |
1381 | ||
13370a9b | 1382 | if (interval && interval < 100) { |
19afd104 KL |
1383 | if (interval < 10) { |
1384 | pr_err("print interval must be >= 10ms\n"); | |
e0547311 | 1385 | parse_options_usage(stat_usage, stat_options, "I", 1); |
19afd104 KL |
1386 | goto out; |
1387 | } else | |
1388 | pr_warning("print interval < 100ms. " | |
1389 | "The overhead percentage could be high in some cases. " | |
1390 | "Please proceed with caution.\n"); | |
13370a9b | 1391 | } |
c45c6ea2 | 1392 | |
d134ffb9 | 1393 | if (perf_evlist__alloc_stats(evsel_list, interval)) |
03ad9747 | 1394 | goto out; |
d6d901c2 | 1395 | |
86ee6e18 | 1396 | if (perf_stat_init_aggr_mode()) |
03ad9747 | 1397 | goto out; |
86ee6e18 | 1398 | |
58d7e993 IM |
1399 | /* |
1400 | * We dont want to block the signals - that would cause | |
1401 | * child tasks to inherit that and Ctrl-C would not work. | |
1402 | * What we want is for Ctrl-C to work in the exec()-ed | |
1403 | * task, but being ignored by perf stat itself: | |
1404 | */ | |
f7b7c26e | 1405 | atexit(sig_atexit); |
a7e191c3 FD |
1406 | if (!forever) |
1407 | signal(SIGINT, skip_signal); | |
13370a9b | 1408 | signal(SIGCHLD, skip_signal); |
58d7e993 IM |
1409 | signal(SIGALRM, skip_signal); |
1410 | signal(SIGABRT, skip_signal); | |
1411 | ||
42202dd5 | 1412 | status = 0; |
a7e191c3 | 1413 | for (run_idx = 0; forever || run_idx < run_count; run_idx++) { |
42202dd5 | 1414 | if (run_count != 1 && verbose) |
4aa9015f SE |
1415 | fprintf(output, "[ perf stat: executing run #%d ... ]\n", |
1416 | run_idx + 1); | |
f9cef0a9 | 1417 | |
42202dd5 | 1418 | status = run_perf_stat(argc, argv); |
a7e191c3 | 1419 | if (forever && status != -1) { |
d4f63a47 | 1420 | print_counters(NULL, argc, argv); |
254ecbc7 | 1421 | perf_stat__reset_stats(); |
a7e191c3 | 1422 | } |
42202dd5 IM |
1423 | } |
1424 | ||
a7e191c3 | 1425 | if (!forever && status != -1 && !interval) |
d4f63a47 | 1426 | print_counters(NULL, argc, argv); |
d134ffb9 ACM |
1427 | |
1428 | perf_evlist__free_stats(evsel_list); | |
0015e2e1 ACM |
1429 | out: |
1430 | perf_evlist__delete(evsel_list); | |
42202dd5 | 1431 | return status; |
ddcacfa0 | 1432 | } |