]>
Commit | Line | Data |
---|---|---|
ddcacfa0 | 1 | /* |
bf9e1876 IM |
2 | * builtin-stat.c |
3 | * | |
4 | * Builtin stat command: Give a precise performance counters summary | |
5 | * overview about any workload, CPU or specific PID. | |
6 | * | |
7 | * Sample output: | |
ddcacfa0 | 8 | |
2cba3ffb | 9 | $ perf stat ./hackbench 10 |
ddcacfa0 | 10 | |
2cba3ffb | 11 | Time: 0.118 |
ddcacfa0 | 12 | |
2cba3ffb | 13 | Performance counter stats for './hackbench 10': |
ddcacfa0 | 14 | |
2cba3ffb IM |
15 | 1708.761321 task-clock # 11.037 CPUs utilized |
16 | 41,190 context-switches # 0.024 M/sec | |
17 | 6,735 CPU-migrations # 0.004 M/sec | |
18 | 17,318 page-faults # 0.010 M/sec | |
19 | 5,205,202,243 cycles # 3.046 GHz | |
20 | 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle | |
21 | 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle | |
22 | 2,603,501,247 instructions # 0.50 insns per cycle | |
23 | # 1.48 stalled cycles per insn | |
24 | 484,357,498 branches # 283.455 M/sec | |
25 | 6,388,934 branch-misses # 1.32% of all branches | |
26 | ||
27 | 0.154822978 seconds time elapsed | |
ddcacfa0 | 28 | |
5242519b | 29 | * |
2cba3ffb | 30 | * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com> |
5242519b IM |
31 | * |
32 | * Improvements and fixes by: | |
33 | * | |
34 | * Arjan van de Ven <arjan@linux.intel.com> | |
35 | * Yanmin Zhang <yanmin.zhang@intel.com> | |
36 | * Wu Fengguang <fengguang.wu@intel.com> | |
37 | * Mike Galbraith <efault@gmx.de> | |
38 | * Paul Mackerras <paulus@samba.org> | |
6e750a8f | 39 | * Jaswinder Singh Rajput <jaswinder@kernel.org> |
5242519b IM |
40 | * |
41 | * Released under the GPL v2. (and only v2, not any later version) | |
ddcacfa0 IM |
42 | */ |
43 | ||
1a482f38 | 44 | #include "perf.h" |
16f762a2 | 45 | #include "builtin.h" |
f14d5707 | 46 | #include "util/cgroup.h" |
148be2c1 | 47 | #include "util/util.h" |
4b6ab94e | 48 | #include <subcmd/parse-options.h> |
5242519b | 49 | #include "util/parse-events.h" |
4cabc3d1 | 50 | #include "util/pmu.h" |
8f28827a | 51 | #include "util/event.h" |
361c99a6 | 52 | #include "util/evlist.h" |
69aad6f1 | 53 | #include "util/evsel.h" |
8f28827a | 54 | #include "util/debug.h" |
a5d243d0 | 55 | #include "util/color.h" |
0007ecea | 56 | #include "util/stat.h" |
60666c63 | 57 | #include "util/header.h" |
a12b51c4 | 58 | #include "util/cpumap.h" |
d6d901c2 | 59 | #include "util/thread.h" |
fd78260b | 60 | #include "util/thread_map.h" |
d809560b | 61 | #include "util/counts.h" |
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 | ||
6acd8e92 JO |
164 | /* |
165 | * Some events get initialized with sample_(period/type) set, | |
166 | * like tracepoints. Clear it up for counting. | |
167 | */ | |
168 | attr->sample_period = 0; | |
169 | attr->sample_type = 0; | |
170 | ||
67ccdecd JO |
171 | /* |
172 | * Disabling all counters initially, they will be enabled | |
173 | * either manually by us or by kernel via enable_on_exec | |
174 | * set later. | |
175 | */ | |
c8280cec | 176 | if (perf_evsel__is_group_leader(evsel)) { |
67ccdecd JO |
177 | attr->disabled = 1; |
178 | ||
c8280cec JO |
179 | /* |
180 | * In case of initial_delay we enable tracee | |
181 | * events manually. | |
182 | */ | |
183 | if (target__none(&target) && !initial_delay) | |
41191688 | 184 | attr->enable_on_exec = 1; |
ddcacfa0 | 185 | } |
084ab9f8 | 186 | |
c8280cec JO |
187 | if (target__has_cpu(&target)) |
188 | return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel)); | |
189 | ||
594ac61a | 190 | return perf_evsel__open_per_thread(evsel, evsel_list->threads); |
ddcacfa0 IM |
191 | } |
192 | ||
c04f5e5d IM |
193 | /* |
194 | * Does the counter have nsecs as a unit? | |
195 | */ | |
daec78a0 | 196 | static inline int nsec_counter(struct perf_evsel *evsel) |
c04f5e5d | 197 | { |
daec78a0 ACM |
198 | if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || |
199 | perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) | |
c04f5e5d IM |
200 | return 1; |
201 | ||
202 | return 0; | |
203 | } | |
204 | ||
f5b4a9c3 SE |
205 | /* |
206 | * Read out the results of a single counter: | |
207 | * do not aggregate counts across CPUs in system-wide mode | |
208 | */ | |
c52b12ed | 209 | static int read_counter(struct perf_evsel *counter) |
f5b4a9c3 | 210 | { |
9bf1a529 JO |
211 | int nthreads = thread_map__nr(evsel_list->threads); |
212 | int ncpus = perf_evsel__nr_cpus(counter); | |
213 | int cpu, thread; | |
f5b4a9c3 | 214 | |
3b4331d9 SP |
215 | if (!counter->supported) |
216 | return -ENOENT; | |
217 | ||
9bf1a529 JO |
218 | if (counter->system_wide) |
219 | nthreads = 1; | |
220 | ||
221 | for (thread = 0; thread < nthreads; thread++) { | |
222 | for (cpu = 0; cpu < ncpus; cpu++) { | |
3b3eb044 JO |
223 | struct perf_counts_values *count; |
224 | ||
225 | count = perf_counts(counter->counts, cpu, thread); | |
226 | if (perf_evsel__read(counter, cpu, thread, count)) | |
9bf1a529 JO |
227 | return -1; |
228 | } | |
f5b4a9c3 | 229 | } |
c52b12ed ACM |
230 | |
231 | return 0; | |
2996f5dd IM |
232 | } |
233 | ||
5fc472a6 | 234 | static void read_counters(bool close_counters) |
13370a9b | 235 | { |
13370a9b | 236 | struct perf_evsel *counter; |
13370a9b | 237 | |
106a94a0 | 238 | evlist__for_each(evsel_list, counter) { |
3b3eb044 | 239 | if (read_counter(counter)) |
245bad8e | 240 | pr_debug("failed to read counter %s\n", counter->name); |
3b3eb044 | 241 | |
f80010eb | 242 | if (perf_stat_process_counter(&stat_config, counter)) |
3b3eb044 | 243 | pr_warning("failed to process counter %s\n", counter->name); |
106a94a0 | 244 | |
5fc472a6 | 245 | if (close_counters) { |
106a94a0 JO |
246 | perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), |
247 | thread_map__nr(evsel_list->threads)); | |
13370a9b SE |
248 | } |
249 | } | |
106a94a0 JO |
250 | } |
251 | ||
ba411a95 | 252 | static void process_interval(void) |
106a94a0 | 253 | { |
106a94a0 | 254 | struct timespec ts, rs; |
106a94a0 JO |
255 | |
256 | read_counters(false); | |
86ee6e18 | 257 | |
13370a9b SE |
258 | clock_gettime(CLOCK_MONOTONIC, &ts); |
259 | diff_timespec(&rs, &ts, &ref_time); | |
13370a9b | 260 | |
d4f63a47 | 261 | print_counters(&rs, 0, NULL); |
13370a9b SE |
262 | } |
263 | ||
67ccdecd | 264 | static void enable_counters(void) |
41191688 | 265 | { |
67ccdecd | 266 | if (initial_delay) |
41191688 | 267 | usleep(initial_delay * 1000); |
67ccdecd JO |
268 | |
269 | /* | |
270 | * We need to enable counters only if: | |
271 | * - we don't have tracee (attaching to task or cpu) | |
272 | * - we have initial delay configured | |
273 | */ | |
274 | if (!target__none(&target) || initial_delay) | |
ab46db0a | 275 | perf_evlist__enable(evsel_list); |
41191688 AK |
276 | } |
277 | ||
f33cbe72 | 278 | static volatile int workload_exec_errno; |
6af206fd ACM |
279 | |
280 | /* | |
281 | * perf_evlist__prepare_workload will send a SIGUSR1 | |
282 | * if the fork fails, since we asked by setting its | |
283 | * want_signal to true. | |
284 | */ | |
f33cbe72 ACM |
285 | static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, |
286 | void *ucontext __maybe_unused) | |
6af206fd | 287 | { |
f33cbe72 | 288 | workload_exec_errno = info->si_value.sival_int; |
6af206fd ACM |
289 | } |
290 | ||
acf28922 | 291 | static int __run_perf_stat(int argc, const char **argv) |
42202dd5 | 292 | { |
ec0d3d1f | 293 | int interval = stat_config.interval; |
56e52e85 | 294 | char msg[512]; |
42202dd5 | 295 | unsigned long long t0, t1; |
cac21425 | 296 | struct perf_evsel *counter; |
13370a9b | 297 | struct timespec ts; |
410136f5 | 298 | size_t l; |
42202dd5 | 299 | int status = 0; |
6be2850e | 300 | const bool forks = (argc > 0); |
42202dd5 | 301 | |
13370a9b SE |
302 | if (interval) { |
303 | ts.tv_sec = interval / 1000; | |
304 | ts.tv_nsec = (interval % 1000) * 1000000; | |
305 | } else { | |
306 | ts.tv_sec = 1; | |
307 | ts.tv_nsec = 0; | |
308 | } | |
309 | ||
60666c63 | 310 | if (forks) { |
735f7e0b ACM |
311 | if (perf_evlist__prepare_workload(evsel_list, &target, argv, false, |
312 | workload_exec_failed_signal) < 0) { | |
acf28922 NK |
313 | perror("failed to prepare workload"); |
314 | return -1; | |
60666c63 | 315 | } |
d20a47e7 | 316 | child_pid = evsel_list->workload.pid; |
051ae7f7 PM |
317 | } |
318 | ||
6a4bb04c | 319 | if (group) |
63dab225 | 320 | perf_evlist__set_leader(evsel_list); |
6a4bb04c | 321 | |
0050f7aa | 322 | evlist__for_each(evsel_list, counter) { |
cac21425 | 323 | if (create_perf_stat_counter(counter) < 0) { |
979987a5 DA |
324 | /* |
325 | * PPC returns ENXIO for HW counters until 2.6.37 | |
326 | * (behavior changed with commit b0a873e). | |
327 | */ | |
38f6ae1e | 328 | if (errno == EINVAL || errno == ENOSYS || |
979987a5 DA |
329 | errno == ENOENT || errno == EOPNOTSUPP || |
330 | errno == ENXIO) { | |
c63ca0c0 DA |
331 | if (verbose) |
332 | ui__warning("%s event is not supported by the kernel.\n", | |
7289f83c | 333 | perf_evsel__name(counter)); |
2cee77c4 | 334 | counter->supported = false; |
cb5ef600 KL |
335 | |
336 | if ((counter->leader != counter) || | |
337 | !(counter->leader->nr_members > 1)) | |
338 | continue; | |
c63ca0c0 | 339 | } |
ede70290 | 340 | |
56e52e85 ACM |
341 | perf_evsel__open_strerror(counter, &target, |
342 | errno, msg, sizeof(msg)); | |
343 | ui__error("%s\n", msg); | |
344 | ||
48290609 ACM |
345 | if (child_pid != -1) |
346 | kill(child_pid, SIGTERM); | |
fceda7fe | 347 | |
48290609 ACM |
348 | return -1; |
349 | } | |
2cee77c4 | 350 | counter->supported = true; |
410136f5 SE |
351 | |
352 | l = strlen(counter->unit); | |
353 | if (l > unit_width) | |
354 | unit_width = l; | |
084ab9f8 | 355 | } |
42202dd5 | 356 | |
23d4aad4 ACM |
357 | if (perf_evlist__apply_filters(evsel_list, &counter)) { |
358 | error("failed to set filter \"%s\" on event %s with %d (%s)\n", | |
359 | counter->filter, perf_evsel__name(counter), errno, | |
759e612b | 360 | strerror_r(errno, msg, sizeof(msg))); |
cfd748ae FW |
361 | return -1; |
362 | } | |
363 | ||
42202dd5 IM |
364 | /* |
365 | * Enable counters and exec the command: | |
366 | */ | |
367 | t0 = rdclock(); | |
13370a9b | 368 | clock_gettime(CLOCK_MONOTONIC, &ref_time); |
42202dd5 | 369 | |
60666c63 | 370 | if (forks) { |
acf28922 | 371 | perf_evlist__start_workload(evsel_list); |
67ccdecd | 372 | enable_counters(); |
acf28922 | 373 | |
13370a9b SE |
374 | if (interval) { |
375 | while (!waitpid(child_pid, &status, WNOHANG)) { | |
376 | nanosleep(&ts, NULL); | |
ba411a95 | 377 | process_interval(); |
13370a9b SE |
378 | } |
379 | } | |
60666c63 | 380 | wait(&status); |
6af206fd | 381 | |
f33cbe72 ACM |
382 | if (workload_exec_errno) { |
383 | const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg)); | |
384 | pr_err("Workload failed: %s\n", emsg); | |
6af206fd | 385 | return -1; |
f33cbe72 | 386 | } |
6af206fd | 387 | |
33e49ea7 AK |
388 | if (WIFSIGNALED(status)) |
389 | psignal(WTERMSIG(status), argv[0]); | |
60666c63 | 390 | } else { |
67ccdecd | 391 | enable_counters(); |
13370a9b SE |
392 | while (!done) { |
393 | nanosleep(&ts, NULL); | |
394 | if (interval) | |
ba411a95 | 395 | process_interval(); |
13370a9b | 396 | } |
60666c63 | 397 | } |
42202dd5 | 398 | |
42202dd5 IM |
399 | t1 = rdclock(); |
400 | ||
9e9772c4 | 401 | update_stats(&walltime_nsecs_stats, t1 - t0); |
42202dd5 | 402 | |
106a94a0 | 403 | read_counters(true); |
c52b12ed | 404 | |
42202dd5 IM |
405 | return WEXITSTATUS(status); |
406 | } | |
407 | ||
41cde476 | 408 | static int run_perf_stat(int argc, const char **argv) |
1f16c575 PZ |
409 | { |
410 | int ret; | |
411 | ||
412 | if (pre_cmd) { | |
413 | ret = system(pre_cmd); | |
414 | if (ret) | |
415 | return ret; | |
416 | } | |
417 | ||
418 | if (sync_run) | |
419 | sync(); | |
420 | ||
421 | ret = __run_perf_stat(argc, argv); | |
422 | if (ret) | |
423 | return ret; | |
424 | ||
425 | if (post_cmd) { | |
426 | ret = system(post_cmd); | |
427 | if (ret) | |
428 | return ret; | |
429 | } | |
430 | ||
431 | return ret; | |
432 | } | |
433 | ||
d73515c0 AK |
434 | static void print_running(u64 run, u64 ena) |
435 | { | |
436 | if (csv_output) { | |
5821522e | 437 | fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f", |
d73515c0 AK |
438 | csv_sep, |
439 | run, | |
440 | csv_sep, | |
441 | ena ? 100.0 * run / ena : 100.0); | |
442 | } else if (run != ena) { | |
5821522e | 443 | fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena); |
d73515c0 AK |
444 | } |
445 | } | |
446 | ||
f99844cb IM |
447 | static void print_noise_pct(double total, double avg) |
448 | { | |
0007ecea | 449 | double pct = rel_stddev_stats(total, avg); |
f99844cb | 450 | |
3ae9a34d | 451 | if (csv_output) |
5821522e | 452 | fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct); |
a1bca6cc | 453 | else if (pct) |
5821522e | 454 | fprintf(stat_config.output, " ( +-%6.2f%% )", pct); |
f99844cb IM |
455 | } |
456 | ||
69aad6f1 | 457 | static void print_noise(struct perf_evsel *evsel, double avg) |
42202dd5 | 458 | { |
581cc8a2 | 459 | struct perf_stat_evsel *ps; |
69aad6f1 | 460 | |
849abde9 PZ |
461 | if (run_count == 1) |
462 | return; | |
463 | ||
69aad6f1 | 464 | ps = evsel->priv; |
f99844cb | 465 | print_noise_pct(stddev_stats(&ps->res_stats[0]), avg); |
42202dd5 IM |
466 | } |
467 | ||
12c08a9f | 468 | static void aggr_printout(struct perf_evsel *evsel, int id, int nr) |
44175b6f | 469 | { |
421a50f3 | 470 | switch (stat_config.aggr_mode) { |
12c08a9f | 471 | case AGGR_CORE: |
5821522e | 472 | fprintf(stat_config.output, "S%d-C%*d%s%*d%s", |
12c08a9f SE |
473 | cpu_map__id_to_socket(id), |
474 | csv_output ? 0 : -8, | |
475 | cpu_map__id_to_cpu(id), | |
476 | csv_sep, | |
477 | csv_output ? 0 : 4, | |
478 | nr, | |
479 | csv_sep); | |
480 | break; | |
86ee6e18 | 481 | case AGGR_SOCKET: |
5821522e | 482 | fprintf(stat_config.output, "S%*d%s%*d%s", |
d7e7a451 | 483 | csv_output ? 0 : -5, |
12c08a9f | 484 | id, |
d7e7a451 SE |
485 | csv_sep, |
486 | csv_output ? 0 : 4, | |
487 | nr, | |
488 | csv_sep); | |
86ee6e18 SE |
489 | break; |
490 | case AGGR_NONE: | |
5821522e | 491 | fprintf(stat_config.output, "CPU%*d%s", |
d7470b6a | 492 | csv_output ? 0 : -4, |
12c08a9f | 493 | perf_evsel__cpus(evsel)->map[id], csv_sep); |
86ee6e18 | 494 | break; |
32b8af82 | 495 | case AGGR_THREAD: |
5821522e | 496 | fprintf(stat_config.output, "%*s-%*d%s", |
32b8af82 JO |
497 | csv_output ? 0 : 16, |
498 | thread_map__comm(evsel->threads, id), | |
499 | csv_output ? 0 : -8, | |
500 | thread_map__pid(evsel->threads, id), | |
501 | csv_sep); | |
502 | break; | |
86ee6e18 | 503 | case AGGR_GLOBAL: |
208df99e | 504 | case AGGR_UNSET: |
86ee6e18 SE |
505 | default: |
506 | break; | |
507 | } | |
508 | } | |
509 | ||
da88c7f7 | 510 | static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg) |
86ee6e18 | 511 | { |
5821522e | 512 | FILE *output = stat_config.output; |
86ee6e18 | 513 | double msecs = avg / 1e6; |
410136f5 | 514 | const char *fmt_v, *fmt_n; |
4bbe5a61 | 515 | char name[25]; |
86ee6e18 | 516 | |
410136f5 SE |
517 | fmt_v = csv_output ? "%.6f%s" : "%18.6f%s"; |
518 | fmt_n = csv_output ? "%s" : "%-25s"; | |
519 | ||
da88c7f7 | 520 | aggr_printout(evsel, id, nr); |
d7470b6a | 521 | |
4bbe5a61 DA |
522 | scnprintf(name, sizeof(name), "%s%s", |
523 | perf_evsel__name(evsel), csv_output ? "" : " (msec)"); | |
410136f5 SE |
524 | |
525 | fprintf(output, fmt_v, msecs, csv_sep); | |
526 | ||
527 | if (csv_output) | |
528 | fprintf(output, "%s%s", evsel->unit, csv_sep); | |
529 | else | |
530 | fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep); | |
531 | ||
532 | fprintf(output, fmt_n, name); | |
d7470b6a | 533 | |
023695d9 | 534 | if (evsel->cgrp) |
4aa9015f | 535 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); |
44175b6f IM |
536 | } |
537 | ||
556b1fb7 JO |
538 | static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg) |
539 | { | |
5821522e | 540 | FILE *output = stat_config.output; |
556b1fb7 JO |
541 | double sc = evsel->scale; |
542 | const char *fmt; | |
556b1fb7 JO |
543 | |
544 | if (csv_output) { | |
545 | fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s"; | |
546 | } else { | |
547 | if (big_num) | |
548 | fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s"; | |
549 | else | |
550 | fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s"; | |
551 | } | |
552 | ||
553 | aggr_printout(evsel, id, nr); | |
554 | ||
556b1fb7 JO |
555 | fprintf(output, fmt, avg, csv_sep); |
556 | ||
557 | if (evsel->unit) | |
558 | fprintf(output, "%-*s%s", | |
559 | csv_output ? 0 : unit_width, | |
560 | evsel->unit, csv_sep); | |
561 | ||
562 | fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel)); | |
563 | ||
564 | if (evsel->cgrp) | |
565 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); | |
eedfcb4b | 566 | } |
556b1fb7 | 567 | |
eedfcb4b AK |
568 | static void printout(int id, int nr, struct perf_evsel *counter, double uval) |
569 | { | |
570 | int cpu = cpu_map__id_to_cpu(id); | |
571 | ||
572 | if (stat_config.aggr_mode == AGGR_GLOBAL) | |
573 | cpu = 0; | |
574 | ||
575 | if (nsec_counter(counter)) | |
576 | nsec_printout(id, nr, counter, uval); | |
577 | else | |
578 | abs_printout(id, nr, counter, uval); | |
556b1fb7 | 579 | |
eedfcb4b AK |
580 | if (!csv_output && !stat_config.interval) |
581 | perf_stat__print_shadow_stats(stat_config.output, counter, | |
582 | uval, cpu, | |
583 | stat_config.aggr_mode); | |
556b1fb7 JO |
584 | } |
585 | ||
86ee6e18 | 586 | static void print_aggr(char *prefix) |
d7e7a451 | 587 | { |
5821522e | 588 | FILE *output = stat_config.output; |
d7e7a451 | 589 | struct perf_evsel *counter; |
601083cf | 590 | int cpu, s, s2, id, nr; |
410136f5 | 591 | double uval; |
d7e7a451 | 592 | u64 ena, run, val; |
d7e7a451 | 593 | |
86ee6e18 | 594 | if (!(aggr_map || aggr_get_id)) |
d7e7a451 SE |
595 | return; |
596 | ||
86ee6e18 SE |
597 | for (s = 0; s < aggr_map->nr; s++) { |
598 | id = aggr_map->map[s]; | |
0050f7aa | 599 | evlist__for_each(evsel_list, counter) { |
d7e7a451 SE |
600 | val = ena = run = 0; |
601 | nr = 0; | |
602 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { | |
601083cf | 603 | s2 = aggr_get_id(perf_evsel__cpus(counter), cpu); |
86ee6e18 | 604 | if (s2 != id) |
d7e7a451 | 605 | continue; |
a6fa0038 JO |
606 | val += perf_counts(counter->counts, cpu, 0)->val; |
607 | ena += perf_counts(counter->counts, cpu, 0)->ena; | |
608 | run += perf_counts(counter->counts, cpu, 0)->run; | |
d7e7a451 SE |
609 | nr++; |
610 | } | |
611 | if (prefix) | |
612 | fprintf(output, "%s", prefix); | |
613 | ||
614 | if (run == 0 || ena == 0) { | |
582ec082 | 615 | aggr_printout(counter, id, nr); |
86ee6e18 | 616 | |
410136f5 | 617 | fprintf(output, "%*s%s", |
d7e7a451 SE |
618 | csv_output ? 0 : 18, |
619 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, | |
410136f5 SE |
620 | csv_sep); |
621 | ||
622 | fprintf(output, "%-*s%s", | |
623 | csv_output ? 0 : unit_width, | |
624 | counter->unit, csv_sep); | |
625 | ||
626 | fprintf(output, "%*s", | |
627 | csv_output ? 0 : -25, | |
d7e7a451 | 628 | perf_evsel__name(counter)); |
86ee6e18 | 629 | |
d7e7a451 SE |
630 | if (counter->cgrp) |
631 | fprintf(output, "%s%s", | |
632 | csv_sep, counter->cgrp->name); | |
633 | ||
d73515c0 | 634 | print_running(run, ena); |
d7e7a451 SE |
635 | fputc('\n', output); |
636 | continue; | |
637 | } | |
410136f5 | 638 | uval = val * counter->scale; |
eedfcb4b | 639 | printout(id, nr, counter, uval); |
d73515c0 | 640 | if (!csv_output) |
d7e7a451 SE |
641 | print_noise(counter, 1.0); |
642 | ||
d73515c0 | 643 | print_running(run, ena); |
d7e7a451 SE |
644 | fputc('\n', output); |
645 | } | |
646 | } | |
647 | } | |
648 | ||
32b8af82 JO |
649 | static void print_aggr_thread(struct perf_evsel *counter, char *prefix) |
650 | { | |
5821522e | 651 | FILE *output = stat_config.output; |
32b8af82 JO |
652 | int nthreads = thread_map__nr(counter->threads); |
653 | int ncpus = cpu_map__nr(counter->cpus); | |
654 | int cpu, thread; | |
655 | double uval; | |
656 | ||
657 | for (thread = 0; thread < nthreads; thread++) { | |
658 | u64 ena = 0, run = 0, val = 0; | |
659 | ||
660 | for (cpu = 0; cpu < ncpus; cpu++) { | |
661 | val += perf_counts(counter->counts, cpu, thread)->val; | |
662 | ena += perf_counts(counter->counts, cpu, thread)->ena; | |
663 | run += perf_counts(counter->counts, cpu, thread)->run; | |
664 | } | |
665 | ||
666 | if (prefix) | |
667 | fprintf(output, "%s", prefix); | |
668 | ||
669 | uval = val * counter->scale; | |
eedfcb4b | 670 | printout(thread, 0, counter, uval); |
32b8af82 JO |
671 | |
672 | if (!csv_output) | |
673 | print_noise(counter, 1.0); | |
674 | ||
675 | print_running(run, ena); | |
676 | fputc('\n', output); | |
677 | } | |
678 | } | |
679 | ||
2996f5dd IM |
680 | /* |
681 | * Print out the results of a single counter: | |
f5b4a9c3 | 682 | * aggregated counts in system-wide mode |
2996f5dd | 683 | */ |
13370a9b | 684 | static void print_counter_aggr(struct perf_evsel *counter, char *prefix) |
2996f5dd | 685 | { |
5821522e | 686 | FILE *output = stat_config.output; |
581cc8a2 | 687 | struct perf_stat_evsel *ps = counter->priv; |
69aad6f1 | 688 | double avg = avg_stats(&ps->res_stats[0]); |
c52b12ed | 689 | int scaled = counter->counts->scaled; |
410136f5 | 690 | double uval; |
d73515c0 AK |
691 | double avg_enabled, avg_running; |
692 | ||
693 | avg_enabled = avg_stats(&ps->res_stats[1]); | |
694 | avg_running = avg_stats(&ps->res_stats[2]); | |
2996f5dd | 695 | |
13370a9b SE |
696 | if (prefix) |
697 | fprintf(output, "%s", prefix); | |
698 | ||
3b4331d9 | 699 | if (scaled == -1 || !counter->supported) { |
410136f5 | 700 | fprintf(output, "%*s%s", |
d7470b6a | 701 | csv_output ? 0 : 18, |
2cee77c4 | 702 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, |
410136f5 SE |
703 | csv_sep); |
704 | fprintf(output, "%-*s%s", | |
705 | csv_output ? 0 : unit_width, | |
706 | counter->unit, csv_sep); | |
707 | fprintf(output, "%*s", | |
708 | csv_output ? 0 : -25, | |
7289f83c | 709 | perf_evsel__name(counter)); |
023695d9 SE |
710 | |
711 | if (counter->cgrp) | |
4aa9015f | 712 | fprintf(output, "%s%s", csv_sep, counter->cgrp->name); |
023695d9 | 713 | |
d73515c0 | 714 | print_running(avg_running, avg_enabled); |
4aa9015f | 715 | fputc('\n', output); |
2996f5dd IM |
716 | return; |
717 | } | |
c04f5e5d | 718 | |
410136f5 | 719 | uval = avg * counter->scale; |
eedfcb4b | 720 | printout(-1, 0, counter, uval); |
849abde9 | 721 | |
3ae9a34d ZH |
722 | print_noise(counter, avg); |
723 | ||
d73515c0 | 724 | print_running(avg_running, avg_enabled); |
4aa9015f | 725 | fprintf(output, "\n"); |
c04f5e5d IM |
726 | } |
727 | ||
f5b4a9c3 SE |
728 | /* |
729 | * Print out the results of a single counter: | |
730 | * does not use aggregated count in system-wide | |
731 | */ | |
13370a9b | 732 | static void print_counter(struct perf_evsel *counter, char *prefix) |
f5b4a9c3 | 733 | { |
5821522e | 734 | FILE *output = stat_config.output; |
f5b4a9c3 | 735 | u64 ena, run, val; |
410136f5 | 736 | double uval; |
f5b4a9c3 SE |
737 | int cpu; |
738 | ||
7ae92e74 | 739 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { |
a6fa0038 JO |
740 | val = perf_counts(counter->counts, cpu, 0)->val; |
741 | ena = perf_counts(counter->counts, cpu, 0)->ena; | |
742 | run = perf_counts(counter->counts, cpu, 0)->run; | |
13370a9b SE |
743 | |
744 | if (prefix) | |
745 | fprintf(output, "%s", prefix); | |
746 | ||
f5b4a9c3 | 747 | if (run == 0 || ena == 0) { |
410136f5 | 748 | fprintf(output, "CPU%*d%s%*s%s", |
d7470b6a | 749 | csv_output ? 0 : -4, |
7ae92e74 | 750 | perf_evsel__cpus(counter)->map[cpu], csv_sep, |
d7470b6a | 751 | csv_output ? 0 : 18, |
2cee77c4 | 752 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, |
410136f5 SE |
753 | csv_sep); |
754 | ||
755 | fprintf(output, "%-*s%s", | |
756 | csv_output ? 0 : unit_width, | |
757 | counter->unit, csv_sep); | |
758 | ||
759 | fprintf(output, "%*s", | |
760 | csv_output ? 0 : -25, | |
761 | perf_evsel__name(counter)); | |
f5b4a9c3 | 762 | |
023695d9 | 763 | if (counter->cgrp) |
4aa9015f SE |
764 | fprintf(output, "%s%s", |
765 | csv_sep, counter->cgrp->name); | |
023695d9 | 766 | |
d73515c0 | 767 | print_running(run, ena); |
4aa9015f | 768 | fputc('\n', output); |
f5b4a9c3 SE |
769 | continue; |
770 | } | |
771 | ||
410136f5 | 772 | uval = val * counter->scale; |
eedfcb4b | 773 | printout(cpu, 0, counter, uval); |
d73515c0 | 774 | if (!csv_output) |
d7470b6a | 775 | print_noise(counter, 1.0); |
d73515c0 | 776 | print_running(run, ena); |
f5b4a9c3 | 777 | |
4aa9015f | 778 | fputc('\n', output); |
f5b4a9c3 SE |
779 | } |
780 | } | |
781 | ||
d4f63a47 JO |
782 | static void print_interval(char *prefix, struct timespec *ts) |
783 | { | |
5821522e | 784 | FILE *output = stat_config.output; |
d4f63a47 JO |
785 | static int num_print_interval; |
786 | ||
787 | sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep); | |
788 | ||
789 | if (num_print_interval == 0 && !csv_output) { | |
421a50f3 | 790 | switch (stat_config.aggr_mode) { |
d4f63a47 JO |
791 | case AGGR_SOCKET: |
792 | fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit"); | |
793 | break; | |
794 | case AGGR_CORE: | |
795 | fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit"); | |
796 | break; | |
797 | case AGGR_NONE: | |
798 | fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit"); | |
799 | break; | |
32b8af82 JO |
800 | case AGGR_THREAD: |
801 | fprintf(output, "# time comm-pid counts %*s events\n", unit_width, "unit"); | |
802 | break; | |
d4f63a47 JO |
803 | case AGGR_GLOBAL: |
804 | default: | |
805 | fprintf(output, "# time counts %*s events\n", unit_width, "unit"); | |
208df99e JO |
806 | case AGGR_UNSET: |
807 | break; | |
d4f63a47 JO |
808 | } |
809 | } | |
810 | ||
811 | if (++num_print_interval == 25) | |
812 | num_print_interval = 0; | |
813 | } | |
814 | ||
815 | static void print_header(int argc, const char **argv) | |
42202dd5 | 816 | { |
5821522e | 817 | FILE *output = stat_config.output; |
69aad6f1 | 818 | int i; |
42202dd5 | 819 | |
ddcacfa0 IM |
820 | fflush(stdout); |
821 | ||
d7470b6a | 822 | if (!csv_output) { |
4aa9015f SE |
823 | fprintf(output, "\n"); |
824 | fprintf(output, " Performance counter stats for "); | |
62d3b617 DA |
825 | if (target.system_wide) |
826 | fprintf(output, "\'system wide"); | |
827 | else if (target.cpu_list) | |
828 | fprintf(output, "\'CPU(s) %s", target.cpu_list); | |
602ad878 | 829 | else if (!target__has_task(&target)) { |
4aa9015f | 830 | fprintf(output, "\'%s", argv[0]); |
d7470b6a | 831 | for (i = 1; i < argc; i++) |
4aa9015f | 832 | fprintf(output, " %s", argv[i]); |
20f946b4 NK |
833 | } else if (target.pid) |
834 | fprintf(output, "process id \'%s", target.pid); | |
d7470b6a | 835 | else |
20f946b4 | 836 | fprintf(output, "thread id \'%s", target.tid); |
44db76c8 | 837 | |
4aa9015f | 838 | fprintf(output, "\'"); |
d7470b6a | 839 | if (run_count > 1) |
4aa9015f SE |
840 | fprintf(output, " (%d runs)", run_count); |
841 | fprintf(output, ":\n\n"); | |
d7470b6a | 842 | } |
d4f63a47 JO |
843 | } |
844 | ||
845 | static void print_footer(void) | |
846 | { | |
5821522e JO |
847 | FILE *output = stat_config.output; |
848 | ||
d4f63a47 JO |
849 | if (!null_run) |
850 | fprintf(output, "\n"); | |
851 | fprintf(output, " %17.9f seconds time elapsed", | |
852 | avg_stats(&walltime_nsecs_stats)/1e9); | |
853 | if (run_count > 1) { | |
854 | fprintf(output, " "); | |
855 | print_noise_pct(stddev_stats(&walltime_nsecs_stats), | |
856 | avg_stats(&walltime_nsecs_stats)); | |
857 | } | |
858 | fprintf(output, "\n\n"); | |
859 | } | |
860 | ||
861 | static void print_counters(struct timespec *ts, int argc, const char **argv) | |
862 | { | |
ec0d3d1f | 863 | int interval = stat_config.interval; |
d4f63a47 JO |
864 | struct perf_evsel *counter; |
865 | char buf[64], *prefix = NULL; | |
866 | ||
867 | if (interval) | |
868 | print_interval(prefix = buf, ts); | |
869 | else | |
870 | print_header(argc, argv); | |
2996f5dd | 871 | |
421a50f3 | 872 | switch (stat_config.aggr_mode) { |
12c08a9f | 873 | case AGGR_CORE: |
86ee6e18 | 874 | case AGGR_SOCKET: |
d4f63a47 | 875 | print_aggr(prefix); |
86ee6e18 | 876 | break; |
32b8af82 JO |
877 | case AGGR_THREAD: |
878 | evlist__for_each(evsel_list, counter) | |
879 | print_aggr_thread(counter, prefix); | |
880 | break; | |
86ee6e18 | 881 | case AGGR_GLOBAL: |
0050f7aa | 882 | evlist__for_each(evsel_list, counter) |
d4f63a47 | 883 | print_counter_aggr(counter, prefix); |
86ee6e18 SE |
884 | break; |
885 | case AGGR_NONE: | |
0050f7aa | 886 | evlist__for_each(evsel_list, counter) |
d4f63a47 | 887 | print_counter(counter, prefix); |
86ee6e18 | 888 | break; |
208df99e | 889 | case AGGR_UNSET: |
86ee6e18 SE |
890 | default: |
891 | break; | |
f5b4a9c3 | 892 | } |
ddcacfa0 | 893 | |
d4f63a47 JO |
894 | if (!interval && !csv_output) |
895 | print_footer(); | |
896 | ||
5821522e | 897 | fflush(stat_config.output); |
ddcacfa0 IM |
898 | } |
899 | ||
f7b7c26e PZ |
900 | static volatile int signr = -1; |
901 | ||
5242519b | 902 | static void skip_signal(int signo) |
ddcacfa0 | 903 | { |
ec0d3d1f | 904 | if ((child_pid == -1) || stat_config.interval) |
60666c63 LW |
905 | done = 1; |
906 | ||
f7b7c26e | 907 | signr = signo; |
d07f0b12 SE |
908 | /* |
909 | * render child_pid harmless | |
910 | * won't send SIGTERM to a random | |
911 | * process in case of race condition | |
912 | * and fast PID recycling | |
913 | */ | |
914 | child_pid = -1; | |
f7b7c26e PZ |
915 | } |
916 | ||
917 | static void sig_atexit(void) | |
918 | { | |
d07f0b12 SE |
919 | sigset_t set, oset; |
920 | ||
921 | /* | |
922 | * avoid race condition with SIGCHLD handler | |
923 | * in skip_signal() which is modifying child_pid | |
924 | * goal is to avoid send SIGTERM to a random | |
925 | * process | |
926 | */ | |
927 | sigemptyset(&set); | |
928 | sigaddset(&set, SIGCHLD); | |
929 | sigprocmask(SIG_BLOCK, &set, &oset); | |
930 | ||
933da83a CW |
931 | if (child_pid != -1) |
932 | kill(child_pid, SIGTERM); | |
933 | ||
d07f0b12 SE |
934 | sigprocmask(SIG_SETMASK, &oset, NULL); |
935 | ||
f7b7c26e PZ |
936 | if (signr == -1) |
937 | return; | |
938 | ||
939 | signal(signr, SIG_DFL); | |
940 | kill(getpid(), signr); | |
5242519b IM |
941 | } |
942 | ||
1d037ca1 IT |
943 | static int stat__set_big_num(const struct option *opt __maybe_unused, |
944 | const char *s __maybe_unused, int unset) | |
d7470b6a SE |
945 | { |
946 | big_num_opt = unset ? 0 : 1; | |
947 | return 0; | |
948 | } | |
949 | ||
e0547311 JO |
950 | static const struct option stat_options[] = { |
951 | OPT_BOOLEAN('T', "transaction", &transaction_run, | |
952 | "hardware transaction statistics"), | |
953 | OPT_CALLBACK('e', "event", &evsel_list, "event", | |
954 | "event selector. use 'perf list' to list available events", | |
955 | parse_events_option), | |
956 | OPT_CALLBACK(0, "filter", &evsel_list, "filter", | |
957 | "event filter", parse_filter), | |
958 | OPT_BOOLEAN('i', "no-inherit", &no_inherit, | |
959 | "child tasks do not inherit counters"), | |
960 | OPT_STRING('p', "pid", &target.pid, "pid", | |
961 | "stat events on existing process id"), | |
962 | OPT_STRING('t', "tid", &target.tid, "tid", | |
963 | "stat events on existing thread id"), | |
964 | OPT_BOOLEAN('a', "all-cpus", &target.system_wide, | |
965 | "system-wide collection from all CPUs"), | |
966 | OPT_BOOLEAN('g', "group", &group, | |
967 | "put the counters into a counter group"), | |
968 | OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"), | |
969 | OPT_INCR('v', "verbose", &verbose, | |
970 | "be more verbose (show counter open errors, etc)"), | |
971 | OPT_INTEGER('r', "repeat", &run_count, | |
972 | "repeat command and print average + stddev (max: 100, forever: 0)"), | |
973 | OPT_BOOLEAN('n', "null", &null_run, | |
974 | "null run - dont start any counters"), | |
975 | OPT_INCR('d', "detailed", &detailed_run, | |
976 | "detailed run - start a lot of events"), | |
977 | OPT_BOOLEAN('S', "sync", &sync_run, | |
978 | "call sync() before starting a run"), | |
979 | OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, | |
980 | "print large numbers with thousands\' separators", | |
981 | stat__set_big_num), | |
982 | OPT_STRING('C', "cpu", &target.cpu_list, "cpu", | |
983 | "list of cpus to monitor in system-wide"), | |
984 | OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode, | |
985 | "disable CPU count aggregation", AGGR_NONE), | |
986 | OPT_STRING('x', "field-separator", &csv_sep, "separator", | |
987 | "print counts with custom separator"), | |
988 | OPT_CALLBACK('G', "cgroup", &evsel_list, "name", | |
989 | "monitor event in cgroup name only", parse_cgroups), | |
990 | OPT_STRING('o', "output", &output_name, "file", "output file name"), | |
991 | OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), | |
992 | OPT_INTEGER(0, "log-fd", &output_fd, | |
993 | "log output to fd, instead of stderr"), | |
994 | OPT_STRING(0, "pre", &pre_cmd, "command", | |
995 | "command to run prior to the measured command"), | |
996 | OPT_STRING(0, "post", &post_cmd, "command", | |
997 | "command to run after to the measured command"), | |
998 | OPT_UINTEGER('I', "interval-print", &stat_config.interval, | |
999 | "print counts at regular interval in ms (>= 10)"), | |
1000 | OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode, | |
1001 | "aggregate counts per processor socket", AGGR_SOCKET), | |
1002 | OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode, | |
1003 | "aggregate counts per physical processor core", AGGR_CORE), | |
1004 | OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode, | |
1005 | "aggregate counts per thread", AGGR_THREAD), | |
1006 | OPT_UINTEGER('D', "delay", &initial_delay, | |
1007 | "ms to wait before starting measurement after program start"), | |
1008 | OPT_END() | |
1009 | }; | |
1010 | ||
1fe7a300 JO |
1011 | static int perf_stat__get_socket(struct cpu_map *map, int cpu) |
1012 | { | |
1013 | return cpu_map__get_socket(map, cpu, NULL); | |
1014 | } | |
1015 | ||
1016 | static int perf_stat__get_core(struct cpu_map *map, int cpu) | |
1017 | { | |
1018 | return cpu_map__get_core(map, cpu, NULL); | |
1019 | } | |
1020 | ||
1e5a2931 JO |
1021 | static int cpu_map__get_max(struct cpu_map *map) |
1022 | { | |
1023 | int i, max = -1; | |
1024 | ||
1025 | for (i = 0; i < map->nr; i++) { | |
1026 | if (map->map[i] > max) | |
1027 | max = map->map[i]; | |
1028 | } | |
1029 | ||
1030 | return max; | |
1031 | } | |
1032 | ||
1033 | static struct cpu_map *cpus_aggr_map; | |
1034 | ||
1035 | static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx) | |
1036 | { | |
1037 | int cpu; | |
1038 | ||
1039 | if (idx >= map->nr) | |
1040 | return -1; | |
1041 | ||
1042 | cpu = map->map[idx]; | |
1043 | ||
1044 | if (cpus_aggr_map->map[cpu] == -1) | |
1045 | cpus_aggr_map->map[cpu] = get_id(map, idx); | |
1046 | ||
1047 | return cpus_aggr_map->map[cpu]; | |
1048 | } | |
1049 | ||
1050 | static int perf_stat__get_socket_cached(struct cpu_map *map, int idx) | |
1051 | { | |
1052 | return perf_stat__get_aggr(perf_stat__get_socket, map, idx); | |
1053 | } | |
1054 | ||
1055 | static int perf_stat__get_core_cached(struct cpu_map *map, int idx) | |
1056 | { | |
1057 | return perf_stat__get_aggr(perf_stat__get_core, map, idx); | |
1058 | } | |
1059 | ||
86ee6e18 SE |
1060 | static int perf_stat_init_aggr_mode(void) |
1061 | { | |
1e5a2931 JO |
1062 | int nr; |
1063 | ||
421a50f3 | 1064 | switch (stat_config.aggr_mode) { |
86ee6e18 SE |
1065 | case AGGR_SOCKET: |
1066 | if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) { | |
1067 | perror("cannot build socket map"); | |
1068 | return -1; | |
1069 | } | |
1e5a2931 | 1070 | aggr_get_id = perf_stat__get_socket_cached; |
86ee6e18 | 1071 | break; |
12c08a9f SE |
1072 | case AGGR_CORE: |
1073 | if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) { | |
1074 | perror("cannot build core map"); | |
1075 | return -1; | |
1076 | } | |
1e5a2931 | 1077 | aggr_get_id = perf_stat__get_core_cached; |
12c08a9f | 1078 | break; |
86ee6e18 SE |
1079 | case AGGR_NONE: |
1080 | case AGGR_GLOBAL: | |
32b8af82 | 1081 | case AGGR_THREAD: |
208df99e | 1082 | case AGGR_UNSET: |
86ee6e18 SE |
1083 | default: |
1084 | break; | |
1085 | } | |
1e5a2931 JO |
1086 | |
1087 | /* | |
1088 | * The evsel_list->cpus is the base we operate on, | |
1089 | * taking the highest cpu number to be the size of | |
1090 | * the aggregation translate cpumap. | |
1091 | */ | |
1092 | nr = cpu_map__get_max(evsel_list->cpus); | |
1093 | cpus_aggr_map = cpu_map__empty_new(nr + 1); | |
1094 | return cpus_aggr_map ? 0 : -ENOMEM; | |
86ee6e18 SE |
1095 | } |
1096 | ||
544c2ae7 MH |
1097 | static void perf_stat__exit_aggr_mode(void) |
1098 | { | |
1099 | cpu_map__put(aggr_map); | |
1100 | cpu_map__put(cpus_aggr_map); | |
1101 | aggr_map = NULL; | |
1102 | cpus_aggr_map = NULL; | |
1103 | } | |
1104 | ||
2cba3ffb IM |
1105 | /* |
1106 | * Add default attributes, if there were no attributes specified or | |
1107 | * if -d/--detailed, -d -d or -d -d -d is used: | |
1108 | */ | |
1109 | static int add_default_attributes(void) | |
1110 | { | |
b070a547 ACM |
1111 | struct perf_event_attr default_attrs[] = { |
1112 | ||
1113 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, | |
1114 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, | |
1115 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, | |
1116 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, | |
1117 | ||
1118 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, | |
1119 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, | |
1120 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, | |
1121 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, | |
1122 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, | |
1123 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, | |
1124 | ||
1125 | }; | |
1126 | ||
1127 | /* | |
1128 | * Detailed stats (-d), covering the L1 and last level data caches: | |
1129 | */ | |
1130 | struct perf_event_attr detailed_attrs[] = { | |
1131 | ||
1132 | { .type = PERF_TYPE_HW_CACHE, | |
1133 | .config = | |
1134 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1135 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1136 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1137 | ||
1138 | { .type = PERF_TYPE_HW_CACHE, | |
1139 | .config = | |
1140 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1141 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1142 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1143 | ||
1144 | { .type = PERF_TYPE_HW_CACHE, | |
1145 | .config = | |
1146 | PERF_COUNT_HW_CACHE_LL << 0 | | |
1147 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1148 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1149 | ||
1150 | { .type = PERF_TYPE_HW_CACHE, | |
1151 | .config = | |
1152 | PERF_COUNT_HW_CACHE_LL << 0 | | |
1153 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1154 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1155 | }; | |
1156 | ||
1157 | /* | |
1158 | * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: | |
1159 | */ | |
1160 | struct perf_event_attr very_detailed_attrs[] = { | |
1161 | ||
1162 | { .type = PERF_TYPE_HW_CACHE, | |
1163 | .config = | |
1164 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
1165 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1166 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1167 | ||
1168 | { .type = PERF_TYPE_HW_CACHE, | |
1169 | .config = | |
1170 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
1171 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1172 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1173 | ||
1174 | { .type = PERF_TYPE_HW_CACHE, | |
1175 | .config = | |
1176 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1177 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1178 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1179 | ||
1180 | { .type = PERF_TYPE_HW_CACHE, | |
1181 | .config = | |
1182 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1183 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1184 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1185 | ||
1186 | { .type = PERF_TYPE_HW_CACHE, | |
1187 | .config = | |
1188 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
1189 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1190 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1191 | ||
1192 | { .type = PERF_TYPE_HW_CACHE, | |
1193 | .config = | |
1194 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
1195 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1196 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1197 | ||
1198 | }; | |
1199 | ||
1200 | /* | |
1201 | * Very, very detailed stats (-d -d -d), adding prefetch events: | |
1202 | */ | |
1203 | struct perf_event_attr very_very_detailed_attrs[] = { | |
1204 | ||
1205 | { .type = PERF_TYPE_HW_CACHE, | |
1206 | .config = | |
1207 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1208 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
1209 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1210 | ||
1211 | { .type = PERF_TYPE_HW_CACHE, | |
1212 | .config = | |
1213 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1214 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
1215 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1216 | }; | |
1217 | ||
2cba3ffb IM |
1218 | /* Set attrs if no event is selected and !null_run: */ |
1219 | if (null_run) | |
1220 | return 0; | |
1221 | ||
4cabc3d1 AK |
1222 | if (transaction_run) { |
1223 | int err; | |
1224 | if (pmu_have_event("cpu", "cycles-ct") && | |
1225 | pmu_have_event("cpu", "el-start")) | |
a454742c | 1226 | err = parse_events(evsel_list, transaction_attrs, NULL); |
4cabc3d1 | 1227 | else |
a454742c JO |
1228 | err = parse_events(evsel_list, transaction_limited_attrs, NULL); |
1229 | if (err) { | |
4cabc3d1 AK |
1230 | fprintf(stderr, "Cannot set up transaction events\n"); |
1231 | return -1; | |
1232 | } | |
1233 | return 0; | |
1234 | } | |
1235 | ||
2cba3ffb | 1236 | if (!evsel_list->nr_entries) { |
79695e1b | 1237 | if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0) |
50d08e47 | 1238 | return -1; |
2cba3ffb IM |
1239 | } |
1240 | ||
1241 | /* Detailed events get appended to the event list: */ | |
1242 | ||
1243 | if (detailed_run < 1) | |
1244 | return 0; | |
1245 | ||
1246 | /* Append detailed run extra attributes: */ | |
79695e1b | 1247 | if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) |
50d08e47 | 1248 | return -1; |
2cba3ffb IM |
1249 | |
1250 | if (detailed_run < 2) | |
1251 | return 0; | |
1252 | ||
1253 | /* Append very detailed run extra attributes: */ | |
79695e1b | 1254 | if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) |
50d08e47 | 1255 | return -1; |
2cba3ffb IM |
1256 | |
1257 | if (detailed_run < 3) | |
1258 | return 0; | |
1259 | ||
1260 | /* Append very, very detailed run extra attributes: */ | |
79695e1b | 1261 | return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); |
2cba3ffb IM |
1262 | } |
1263 | ||
1d037ca1 | 1264 | int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused) |
5242519b | 1265 | { |
b070a547 ACM |
1266 | const char * const stat_usage[] = { |
1267 | "perf stat [<options>] [<command>]", | |
1268 | NULL | |
1269 | }; | |
cc03c542 | 1270 | int status = -EINVAL, run_idx; |
4aa9015f | 1271 | const char *mode; |
5821522e | 1272 | FILE *output = stderr; |
ec0d3d1f | 1273 | unsigned int interval; |
42202dd5 | 1274 | |
5af52b51 SE |
1275 | setlocale(LC_ALL, ""); |
1276 | ||
334fe7a3 | 1277 | evsel_list = perf_evlist__new(); |
361c99a6 ACM |
1278 | if (evsel_list == NULL) |
1279 | return -ENOMEM; | |
1280 | ||
e0547311 | 1281 | argc = parse_options(argc, argv, stat_options, stat_usage, |
a0541234 | 1282 | PARSE_OPT_STOP_AT_NON_OPTION); |
d7470b6a | 1283 | |
ec0d3d1f JO |
1284 | interval = stat_config.interval; |
1285 | ||
4aa9015f SE |
1286 | if (output_name && strcmp(output_name, "-")) |
1287 | output = NULL; | |
1288 | ||
56f3bae7 JC |
1289 | if (output_name && output_fd) { |
1290 | fprintf(stderr, "cannot use both --output and --log-fd\n"); | |
e0547311 JO |
1291 | parse_options_usage(stat_usage, stat_options, "o", 1); |
1292 | parse_options_usage(NULL, stat_options, "log-fd", 0); | |
cc03c542 | 1293 | goto out; |
56f3bae7 | 1294 | } |
fc3e4d07 SE |
1295 | |
1296 | if (output_fd < 0) { | |
1297 | fprintf(stderr, "argument to --log-fd must be a > 0\n"); | |
e0547311 | 1298 | parse_options_usage(stat_usage, stat_options, "log-fd", 0); |
cc03c542 | 1299 | goto out; |
fc3e4d07 SE |
1300 | } |
1301 | ||
4aa9015f SE |
1302 | if (!output) { |
1303 | struct timespec tm; | |
1304 | mode = append_file ? "a" : "w"; | |
1305 | ||
1306 | output = fopen(output_name, mode); | |
1307 | if (!output) { | |
1308 | perror("failed to create output file"); | |
fceda7fe | 1309 | return -1; |
4aa9015f SE |
1310 | } |
1311 | clock_gettime(CLOCK_REALTIME, &tm); | |
1312 | fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); | |
fc3e4d07 | 1313 | } else if (output_fd > 0) { |
56f3bae7 JC |
1314 | mode = append_file ? "a" : "w"; |
1315 | output = fdopen(output_fd, mode); | |
1316 | if (!output) { | |
1317 | perror("Failed opening logfd"); | |
1318 | return -errno; | |
1319 | } | |
4aa9015f SE |
1320 | } |
1321 | ||
5821522e JO |
1322 | stat_config.output = output; |
1323 | ||
d4ffd04d | 1324 | if (csv_sep) { |
d7470b6a | 1325 | csv_output = true; |
d4ffd04d JC |
1326 | if (!strcmp(csv_sep, "\\t")) |
1327 | csv_sep = "\t"; | |
1328 | } else | |
d7470b6a SE |
1329 | csv_sep = DEFAULT_SEPARATOR; |
1330 | ||
1331 | /* | |
1332 | * let the spreadsheet do the pretty-printing | |
1333 | */ | |
1334 | if (csv_output) { | |
61a9f324 | 1335 | /* User explicitly passed -B? */ |
d7470b6a SE |
1336 | if (big_num_opt == 1) { |
1337 | fprintf(stderr, "-B option not supported with -x\n"); | |
e0547311 JO |
1338 | parse_options_usage(stat_usage, stat_options, "B", 1); |
1339 | parse_options_usage(NULL, stat_options, "x", 1); | |
cc03c542 | 1340 | goto out; |
d7470b6a SE |
1341 | } else /* Nope, so disable big number formatting */ |
1342 | big_num = false; | |
1343 | } else if (big_num_opt == 0) /* User passed --no-big-num */ | |
1344 | big_num = false; | |
1345 | ||
602ad878 | 1346 | if (!argc && target__none(&target)) |
e0547311 | 1347 | usage_with_options(stat_usage, stat_options); |
ac3063bd | 1348 | |
a7e191c3 | 1349 | if (run_count < 0) { |
cc03c542 | 1350 | pr_err("Run count must be a positive number\n"); |
e0547311 | 1351 | parse_options_usage(stat_usage, stat_options, "r", 1); |
cc03c542 | 1352 | goto out; |
a7e191c3 FD |
1353 | } else if (run_count == 0) { |
1354 | forever = true; | |
1355 | run_count = 1; | |
1356 | } | |
ddcacfa0 | 1357 | |
421a50f3 | 1358 | if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) { |
32b8af82 JO |
1359 | fprintf(stderr, "The --per-thread option is only available " |
1360 | "when monitoring via -p -t options.\n"); | |
e0547311 JO |
1361 | parse_options_usage(NULL, stat_options, "p", 1); |
1362 | parse_options_usage(NULL, stat_options, "t", 1); | |
32b8af82 JO |
1363 | goto out; |
1364 | } | |
1365 | ||
1366 | /* | |
1367 | * no_aggr, cgroup are for system-wide only | |
1368 | * --per-thread is aggregated per thread, we dont mix it with cpu mode | |
1369 | */ | |
421a50f3 JO |
1370 | if (((stat_config.aggr_mode != AGGR_GLOBAL && |
1371 | stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) && | |
602ad878 | 1372 | !target__has_cpu(&target)) { |
023695d9 SE |
1373 | fprintf(stderr, "both cgroup and no-aggregation " |
1374 | "modes only available in system-wide mode\n"); | |
1375 | ||
e0547311 JO |
1376 | parse_options_usage(stat_usage, stat_options, "G", 1); |
1377 | parse_options_usage(NULL, stat_options, "A", 1); | |
1378 | parse_options_usage(NULL, stat_options, "a", 1); | |
cc03c542 | 1379 | goto out; |
d7e7a451 SE |
1380 | } |
1381 | ||
2cba3ffb IM |
1382 | if (add_default_attributes()) |
1383 | goto out; | |
ddcacfa0 | 1384 | |
602ad878 | 1385 | target__validate(&target); |
5c98d466 | 1386 | |
77a6f014 | 1387 | if (perf_evlist__create_maps(evsel_list, &target) < 0) { |
602ad878 | 1388 | if (target__has_task(&target)) { |
77a6f014 | 1389 | pr_err("Problems finding threads of monitor\n"); |
e0547311 JO |
1390 | parse_options_usage(stat_usage, stat_options, "p", 1); |
1391 | parse_options_usage(NULL, stat_options, "t", 1); | |
602ad878 | 1392 | } else if (target__has_cpu(&target)) { |
77a6f014 | 1393 | perror("failed to parse CPUs map"); |
e0547311 JO |
1394 | parse_options_usage(stat_usage, stat_options, "C", 1); |
1395 | parse_options_usage(NULL, stat_options, "a", 1); | |
cc03c542 NK |
1396 | } |
1397 | goto out; | |
60d567e2 | 1398 | } |
32b8af82 JO |
1399 | |
1400 | /* | |
1401 | * Initialize thread_map with comm names, | |
1402 | * so we could print it out on output. | |
1403 | */ | |
421a50f3 | 1404 | if (stat_config.aggr_mode == AGGR_THREAD) |
32b8af82 JO |
1405 | thread_map__read_comms(evsel_list->threads); |
1406 | ||
13370a9b | 1407 | if (interval && interval < 100) { |
19afd104 KL |
1408 | if (interval < 10) { |
1409 | pr_err("print interval must be >= 10ms\n"); | |
e0547311 | 1410 | parse_options_usage(stat_usage, stat_options, "I", 1); |
19afd104 KL |
1411 | goto out; |
1412 | } else | |
1413 | pr_warning("print interval < 100ms. " | |
1414 | "The overhead percentage could be high in some cases. " | |
1415 | "Please proceed with caution.\n"); | |
13370a9b | 1416 | } |
c45c6ea2 | 1417 | |
d134ffb9 | 1418 | if (perf_evlist__alloc_stats(evsel_list, interval)) |
03ad9747 | 1419 | goto out; |
d6d901c2 | 1420 | |
86ee6e18 | 1421 | if (perf_stat_init_aggr_mode()) |
03ad9747 | 1422 | goto out; |
86ee6e18 | 1423 | |
58d7e993 IM |
1424 | /* |
1425 | * We dont want to block the signals - that would cause | |
1426 | * child tasks to inherit that and Ctrl-C would not work. | |
1427 | * What we want is for Ctrl-C to work in the exec()-ed | |
1428 | * task, but being ignored by perf stat itself: | |
1429 | */ | |
f7b7c26e | 1430 | atexit(sig_atexit); |
a7e191c3 FD |
1431 | if (!forever) |
1432 | signal(SIGINT, skip_signal); | |
13370a9b | 1433 | signal(SIGCHLD, skip_signal); |
58d7e993 IM |
1434 | signal(SIGALRM, skip_signal); |
1435 | signal(SIGABRT, skip_signal); | |
1436 | ||
42202dd5 | 1437 | status = 0; |
a7e191c3 | 1438 | for (run_idx = 0; forever || run_idx < run_count; run_idx++) { |
42202dd5 | 1439 | if (run_count != 1 && verbose) |
4aa9015f SE |
1440 | fprintf(output, "[ perf stat: executing run #%d ... ]\n", |
1441 | run_idx + 1); | |
f9cef0a9 | 1442 | |
42202dd5 | 1443 | status = run_perf_stat(argc, argv); |
a7e191c3 | 1444 | if (forever && status != -1) { |
d4f63a47 | 1445 | print_counters(NULL, argc, argv); |
254ecbc7 | 1446 | perf_stat__reset_stats(); |
a7e191c3 | 1447 | } |
42202dd5 IM |
1448 | } |
1449 | ||
a7e191c3 | 1450 | if (!forever && status != -1 && !interval) |
d4f63a47 | 1451 | print_counters(NULL, argc, argv); |
d134ffb9 | 1452 | |
544c2ae7 | 1453 | perf_stat__exit_aggr_mode(); |
d134ffb9 | 1454 | perf_evlist__free_stats(evsel_list); |
0015e2e1 ACM |
1455 | out: |
1456 | perf_evlist__delete(evsel_list); | |
42202dd5 | 1457 | return status; |
ddcacfa0 | 1458 | } |