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 | |
bf9e1876 IM |
9 | $ perf stat ~/hackbench 10 |
10 | Time: 0.104 | |
ddcacfa0 | 11 | |
bf9e1876 | 12 | Performance counter stats for '/home/mingo/hackbench': |
ddcacfa0 | 13 | |
bf9e1876 IM |
14 | 1255.538611 task clock ticks # 10.143 CPU utilization factor |
15 | 54011 context switches # 0.043 M/sec | |
16 | 385 CPU migrations # 0.000 M/sec | |
17 | 17755 pagefaults # 0.014 M/sec | |
18 | 3808323185 CPU cycles # 3033.219 M/sec | |
19 | 1575111190 instructions # 1254.530 M/sec | |
20 | 17367895 cache references # 13.833 M/sec | |
21 | 7674421 cache misses # 6.112 M/sec | |
ddcacfa0 | 22 | |
bf9e1876 | 23 | Wall-clock time elapsed: 123.786620 msecs |
ddcacfa0 | 24 | |
5242519b IM |
25 | * |
26 | * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com> | |
27 | * | |
28 | * Improvements and fixes by: | |
29 | * | |
30 | * Arjan van de Ven <arjan@linux.intel.com> | |
31 | * Yanmin Zhang <yanmin.zhang@intel.com> | |
32 | * Wu Fengguang <fengguang.wu@intel.com> | |
33 | * Mike Galbraith <efault@gmx.de> | |
34 | * Paul Mackerras <paulus@samba.org> | |
6e750a8f | 35 | * Jaswinder Singh Rajput <jaswinder@kernel.org> |
5242519b IM |
36 | * |
37 | * Released under the GPL v2. (and only v2, not any later version) | |
ddcacfa0 IM |
38 | */ |
39 | ||
1a482f38 | 40 | #include "perf.h" |
16f762a2 | 41 | #include "builtin.h" |
148be2c1 | 42 | #include "util/util.h" |
5242519b IM |
43 | #include "util/parse-options.h" |
44 | #include "util/parse-events.h" | |
8f28827a | 45 | #include "util/event.h" |
361c99a6 | 46 | #include "util/evlist.h" |
69aad6f1 | 47 | #include "util/evsel.h" |
8f28827a | 48 | #include "util/debug.h" |
60666c63 | 49 | #include "util/header.h" |
a12b51c4 | 50 | #include "util/cpumap.h" |
d6d901c2 | 51 | #include "util/thread.h" |
fd78260b | 52 | #include "util/thread_map.h" |
ddcacfa0 | 53 | |
ddcacfa0 | 54 | #include <sys/prctl.h> |
42202dd5 | 55 | #include <math.h> |
5af52b51 | 56 | #include <locale.h> |
16c8a109 | 57 | |
d7470b6a SE |
58 | #define DEFAULT_SEPARATOR " " |
59 | ||
cdd6c482 | 60 | static struct perf_event_attr default_attrs[] = { |
ddcacfa0 | 61 | |
56aab464 IM |
62 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, |
63 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, | |
64 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, | |
65 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, | |
f4dbfa8f | 66 | |
56aab464 IM |
67 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, |
68 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, | |
56aab464 IM |
69 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, |
70 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, | |
dd86e72a IM |
71 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES }, |
72 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES }, | |
f4dbfa8f | 73 | |
ddcacfa0 | 74 | }; |
5242519b | 75 | |
361c99a6 ACM |
76 | struct perf_evlist *evsel_list; |
77 | ||
c0555642 | 78 | static bool system_wide = false; |
60d567e2 | 79 | static struct cpu_map *cpus; |
3d632595 | 80 | static int run_idx = 0; |
ddcacfa0 | 81 | |
3d632595 | 82 | static int run_count = 1; |
2e6cdf99 | 83 | static bool no_inherit = false; |
c0555642 | 84 | static bool scale = true; |
f5b4a9c3 | 85 | static bool no_aggr = false; |
933da83a | 86 | static pid_t target_pid = -1; |
d6d901c2 | 87 | static pid_t target_tid = -1; |
5c98d466 | 88 | static struct thread_map *threads; |
933da83a | 89 | static pid_t child_pid = -1; |
c0555642 | 90 | static bool null_run = false; |
201e0b06 | 91 | static bool big_num = true; |
d7470b6a | 92 | static int big_num_opt = -1; |
c45c6ea2 | 93 | static const char *cpu_list; |
d7470b6a SE |
94 | static const char *csv_sep = NULL; |
95 | static bool csv_output = false; | |
5af52b51 | 96 | |
60666c63 LW |
97 | static volatile int done = 0; |
98 | ||
506d4bc8 PZ |
99 | struct stats |
100 | { | |
8a02631a | 101 | double n, mean, M2; |
506d4bc8 | 102 | }; |
42202dd5 | 103 | |
69aad6f1 ACM |
104 | struct perf_stat { |
105 | struct stats res_stats[3]; | |
69aad6f1 ACM |
106 | }; |
107 | ||
c52b12ed | 108 | static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel) |
69aad6f1 | 109 | { |
c52b12ed | 110 | evsel->priv = zalloc(sizeof(struct perf_stat)); |
69aad6f1 ACM |
111 | return evsel->priv == NULL ? -ENOMEM : 0; |
112 | } | |
113 | ||
114 | static void perf_evsel__free_stat_priv(struct perf_evsel *evsel) | |
115 | { | |
116 | free(evsel->priv); | |
117 | evsel->priv = NULL; | |
118 | } | |
119 | ||
9e9772c4 PZ |
120 | static void update_stats(struct stats *stats, u64 val) |
121 | { | |
8a02631a | 122 | double delta; |
9e9772c4 | 123 | |
8a02631a PZ |
124 | stats->n++; |
125 | delta = val - stats->mean; | |
126 | stats->mean += delta / stats->n; | |
127 | stats->M2 += delta*(val - stats->mean); | |
9e9772c4 PZ |
128 | } |
129 | ||
506d4bc8 PZ |
130 | static double avg_stats(struct stats *stats) |
131 | { | |
8a02631a | 132 | return stats->mean; |
506d4bc8 | 133 | } |
42202dd5 | 134 | |
506d4bc8 | 135 | /* |
63d40deb PZ |
136 | * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance |
137 | * | |
8a02631a PZ |
138 | * (\Sum n_i^2) - ((\Sum n_i)^2)/n |
139 | * s^2 = ------------------------------- | |
140 | * n - 1 | |
63d40deb PZ |
141 | * |
142 | * http://en.wikipedia.org/wiki/Stddev | |
143 | * | |
144 | * The std dev of the mean is related to the std dev by: | |
145 | * | |
146 | * s | |
147 | * s_mean = ------- | |
148 | * sqrt(n) | |
149 | * | |
506d4bc8 PZ |
150 | */ |
151 | static double stddev_stats(struct stats *stats) | |
152 | { | |
8a02631a PZ |
153 | double variance = stats->M2 / (stats->n - 1); |
154 | double variance_mean = variance / stats->n; | |
42202dd5 | 155 | |
63d40deb | 156 | return sqrt(variance_mean); |
506d4bc8 | 157 | } |
42202dd5 | 158 | |
f5b4a9c3 SE |
159 | struct stats runtime_nsecs_stats[MAX_NR_CPUS]; |
160 | struct stats runtime_cycles_stats[MAX_NR_CPUS]; | |
161 | struct stats runtime_branches_stats[MAX_NR_CPUS]; | |
506d4bc8 | 162 | struct stats walltime_nsecs_stats; |
be1ac0d8 | 163 | |
48290609 | 164 | static int create_perf_stat_counter(struct perf_evsel *evsel) |
ddcacfa0 | 165 | { |
69aad6f1 | 166 | struct perf_event_attr *attr = &evsel->attr; |
16c8a109 | 167 | |
ddcacfa0 | 168 | if (scale) |
a21ca2ca IM |
169 | attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | |
170 | PERF_FORMAT_TOTAL_TIME_RUNNING; | |
ddcacfa0 | 171 | |
48290609 | 172 | if (system_wide) |
9d04f178 | 173 | return perf_evsel__open_per_cpu(evsel, cpus, false, false); |
48290609 ACM |
174 | |
175 | attr->inherit = !no_inherit; | |
176 | if (target_pid == -1 && target_tid == -1) { | |
177 | attr->disabled = 1; | |
178 | attr->enable_on_exec = 1; | |
ddcacfa0 | 179 | } |
084ab9f8 | 180 | |
9d04f178 | 181 | return perf_evsel__open_per_thread(evsel, threads, false, false); |
ddcacfa0 IM |
182 | } |
183 | ||
c04f5e5d IM |
184 | /* |
185 | * Does the counter have nsecs as a unit? | |
186 | */ | |
daec78a0 | 187 | static inline int nsec_counter(struct perf_evsel *evsel) |
c04f5e5d | 188 | { |
daec78a0 ACM |
189 | if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || |
190 | perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) | |
c04f5e5d IM |
191 | return 1; |
192 | ||
193 | return 0; | |
194 | } | |
195 | ||
196 | /* | |
2996f5dd | 197 | * Read out the results of a single counter: |
f5b4a9c3 | 198 | * aggregate counts across CPUs in system-wide mode |
c04f5e5d | 199 | */ |
c52b12ed | 200 | static int read_counter_aggr(struct perf_evsel *counter) |
c04f5e5d | 201 | { |
69aad6f1 | 202 | struct perf_stat *ps = counter->priv; |
c52b12ed ACM |
203 | u64 *count = counter->counts->aggr.values; |
204 | int i; | |
2996f5dd | 205 | |
5c98d466 | 206 | if (__perf_evsel__read(counter, cpus->nr, threads->nr, scale) < 0) |
c52b12ed | 207 | return -1; |
9e9772c4 PZ |
208 | |
209 | for (i = 0; i < 3; i++) | |
69aad6f1 | 210 | update_stats(&ps->res_stats[i], count[i]); |
9e9772c4 PZ |
211 | |
212 | if (verbose) { | |
9486aa38 ACM |
213 | fprintf(stderr, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", |
214 | event_name(counter), count[0], count[1], count[2]); | |
9e9772c4 PZ |
215 | } |
216 | ||
be1ac0d8 IM |
217 | /* |
218 | * Save the full runtime - to allow normalization during printout: | |
219 | */ | |
daec78a0 | 220 | if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK)) |
f5b4a9c3 | 221 | update_stats(&runtime_nsecs_stats[0], count[0]); |
daec78a0 | 222 | if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES)) |
f5b4a9c3 | 223 | update_stats(&runtime_cycles_stats[0], count[0]); |
daec78a0 | 224 | if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS)) |
f5b4a9c3 | 225 | update_stats(&runtime_branches_stats[0], count[0]); |
c52b12ed ACM |
226 | |
227 | return 0; | |
f5b4a9c3 SE |
228 | } |
229 | ||
230 | /* | |
231 | * Read out the results of a single counter: | |
232 | * do not aggregate counts across CPUs in system-wide mode | |
233 | */ | |
c52b12ed | 234 | static int read_counter(struct perf_evsel *counter) |
f5b4a9c3 | 235 | { |
c52b12ed | 236 | u64 *count; |
f5b4a9c3 | 237 | int cpu; |
f5b4a9c3 | 238 | |
60d567e2 | 239 | for (cpu = 0; cpu < cpus->nr; cpu++) { |
c52b12ed ACM |
240 | if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0) |
241 | return -1; | |
f5b4a9c3 | 242 | |
c52b12ed | 243 | count = counter->counts->cpu[cpu].values; |
f5b4a9c3 | 244 | |
daec78a0 | 245 | if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK)) |
f5b4a9c3 | 246 | update_stats(&runtime_nsecs_stats[cpu], count[0]); |
daec78a0 | 247 | if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES)) |
f5b4a9c3 | 248 | update_stats(&runtime_cycles_stats[cpu], count[0]); |
daec78a0 | 249 | if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS)) |
f5b4a9c3 SE |
250 | update_stats(&runtime_branches_stats[cpu], count[0]); |
251 | } | |
c52b12ed ACM |
252 | |
253 | return 0; | |
2996f5dd IM |
254 | } |
255 | ||
f37a291c | 256 | static int run_perf_stat(int argc __used, const char **argv) |
42202dd5 IM |
257 | { |
258 | unsigned long long t0, t1; | |
69aad6f1 | 259 | struct perf_evsel *counter; |
42202dd5 | 260 | int status = 0; |
051ae7f7 | 261 | int child_ready_pipe[2], go_pipe[2]; |
6be2850e | 262 | const bool forks = (argc > 0); |
051ae7f7 | 263 | char buf; |
42202dd5 | 264 | |
60666c63 | 265 | if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) { |
051ae7f7 PM |
266 | perror("failed to create pipes"); |
267 | exit(1); | |
268 | } | |
269 | ||
60666c63 | 270 | if (forks) { |
6be2850e | 271 | if ((child_pid = fork()) < 0) |
60666c63 LW |
272 | perror("failed to fork"); |
273 | ||
6be2850e | 274 | if (!child_pid) { |
60666c63 LW |
275 | close(child_ready_pipe[0]); |
276 | close(go_pipe[1]); | |
277 | fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); | |
278 | ||
279 | /* | |
280 | * Do a dummy execvp to get the PLT entry resolved, | |
281 | * so we avoid the resolver overhead on the real | |
282 | * execvp call. | |
283 | */ | |
284 | execvp("", (char **)argv); | |
285 | ||
286 | /* | |
287 | * Tell the parent we're ready to go | |
288 | */ | |
289 | close(child_ready_pipe[1]); | |
290 | ||
291 | /* | |
292 | * Wait until the parent tells us to go. | |
293 | */ | |
294 | if (read(go_pipe[0], &buf, 1) == -1) | |
295 | perror("unable to read pipe"); | |
296 | ||
297 | execvp(argv[0], (char **)argv); | |
298 | ||
299 | perror(argv[0]); | |
300 | exit(-1); | |
301 | } | |
051ae7f7 | 302 | |
d6d901c2 | 303 | if (target_tid == -1 && target_pid == -1 && !system_wide) |
5c98d466 | 304 | threads->map[0] = child_pid; |
d6d901c2 | 305 | |
051ae7f7 | 306 | /* |
60666c63 | 307 | * Wait for the child to be ready to exec. |
051ae7f7 PM |
308 | */ |
309 | close(child_ready_pipe[1]); | |
60666c63 LW |
310 | close(go_pipe[0]); |
311 | if (read(child_ready_pipe[0], &buf, 1) == -1) | |
a92bef0f | 312 | perror("unable to read pipe"); |
60666c63 | 313 | close(child_ready_pipe[0]); |
051ae7f7 PM |
314 | } |
315 | ||
361c99a6 | 316 | list_for_each_entry(counter, &evsel_list->entries, node) { |
48290609 ACM |
317 | if (create_perf_stat_counter(counter) < 0) { |
318 | if (errno == -EPERM || errno == -EACCES) { | |
319 | error("You may not have permission to collect %sstats.\n" | |
320 | "\t Consider tweaking" | |
321 | " /proc/sys/kernel/perf_event_paranoid or running as root.", | |
322 | system_wide ? "system-wide " : ""); | |
5a3446bc DA |
323 | } else if (errno == ENOENT) { |
324 | error("%s event is not supported. ", event_name(counter)); | |
48290609 ACM |
325 | } else { |
326 | error("open_counter returned with %d (%s). " | |
327 | "/bin/dmesg may provide additional information.\n", | |
328 | errno, strerror(errno)); | |
329 | } | |
330 | if (child_pid != -1) | |
331 | kill(child_pid, SIGTERM); | |
332 | die("Not all events could be opened.\n"); | |
333 | return -1; | |
334 | } | |
084ab9f8 | 335 | } |
42202dd5 IM |
336 | |
337 | /* | |
338 | * Enable counters and exec the command: | |
339 | */ | |
340 | t0 = rdclock(); | |
42202dd5 | 341 | |
60666c63 LW |
342 | if (forks) { |
343 | close(go_pipe[1]); | |
344 | wait(&status); | |
345 | } else { | |
6be2850e | 346 | while(!done) sleep(1); |
60666c63 | 347 | } |
42202dd5 | 348 | |
42202dd5 IM |
349 | t1 = rdclock(); |
350 | ||
9e9772c4 | 351 | update_stats(&walltime_nsecs_stats, t1 - t0); |
42202dd5 | 352 | |
f5b4a9c3 | 353 | if (no_aggr) { |
361c99a6 | 354 | list_for_each_entry(counter, &evsel_list->entries, node) { |
f5b4a9c3 | 355 | read_counter(counter); |
60d567e2 | 356 | perf_evsel__close_fd(counter, cpus->nr, 1); |
c52b12ed | 357 | } |
f5b4a9c3 | 358 | } else { |
361c99a6 | 359 | list_for_each_entry(counter, &evsel_list->entries, node) { |
f5b4a9c3 | 360 | read_counter_aggr(counter); |
5c98d466 | 361 | perf_evsel__close_fd(counter, cpus->nr, threads->nr); |
c52b12ed | 362 | } |
f5b4a9c3 | 363 | } |
c52b12ed | 364 | |
42202dd5 IM |
365 | return WEXITSTATUS(status); |
366 | } | |
367 | ||
69aad6f1 | 368 | static void print_noise(struct perf_evsel *evsel, double avg) |
42202dd5 | 369 | { |
69aad6f1 ACM |
370 | struct perf_stat *ps; |
371 | ||
849abde9 PZ |
372 | if (run_count == 1) |
373 | return; | |
374 | ||
69aad6f1 | 375 | ps = evsel->priv; |
849abde9 | 376 | fprintf(stderr, " ( +- %7.3f%% )", |
69aad6f1 | 377 | 100 * stddev_stats(&ps->res_stats[0]) / avg); |
42202dd5 IM |
378 | } |
379 | ||
daec78a0 | 380 | static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg) |
44175b6f | 381 | { |
506d4bc8 | 382 | double msecs = avg / 1e6; |
d7470b6a SE |
383 | char cpustr[16] = { '\0', }; |
384 | const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s"; | |
44175b6f | 385 | |
f5b4a9c3 | 386 | if (no_aggr) |
d7470b6a SE |
387 | sprintf(cpustr, "CPU%*d%s", |
388 | csv_output ? 0 : -4, | |
60d567e2 | 389 | cpus->map[cpu], csv_sep); |
d7470b6a | 390 | |
daec78a0 | 391 | fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel)); |
d7470b6a SE |
392 | |
393 | if (csv_output) | |
394 | return; | |
44175b6f | 395 | |
daec78a0 | 396 | if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) |
506d4bc8 PZ |
397 | fprintf(stderr, " # %10.3f CPUs ", |
398 | avg / avg_stats(&walltime_nsecs_stats)); | |
44175b6f IM |
399 | } |
400 | ||
daec78a0 | 401 | static void abs_printout(int cpu, struct perf_evsel *evsel, double avg) |
44175b6f | 402 | { |
c7f7fea3 | 403 | double total, ratio = 0.0; |
f5b4a9c3 | 404 | char cpustr[16] = { '\0', }; |
d7470b6a SE |
405 | const char *fmt; |
406 | ||
407 | if (csv_output) | |
408 | fmt = "%s%.0f%s%s"; | |
409 | else if (big_num) | |
410 | fmt = "%s%'18.0f%s%-24s"; | |
411 | else | |
412 | fmt = "%s%18.0f%s%-24s"; | |
f5b4a9c3 SE |
413 | |
414 | if (no_aggr) | |
d7470b6a SE |
415 | sprintf(cpustr, "CPU%*d%s", |
416 | csv_output ? 0 : -4, | |
60d567e2 | 417 | cpus->map[cpu], csv_sep); |
f5b4a9c3 SE |
418 | else |
419 | cpu = 0; | |
c7f7fea3 | 420 | |
daec78a0 | 421 | fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel)); |
d7470b6a SE |
422 | |
423 | if (csv_output) | |
424 | return; | |
44175b6f | 425 | |
daec78a0 | 426 | if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) { |
f5b4a9c3 | 427 | total = avg_stats(&runtime_cycles_stats[cpu]); |
c7f7fea3 IM |
428 | |
429 | if (total) | |
430 | ratio = avg / total; | |
431 | ||
432 | fprintf(stderr, " # %10.3f IPC ", ratio); | |
daec78a0 | 433 | } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) && |
f5b4a9c3 SE |
434 | runtime_branches_stats[cpu].n != 0) { |
435 | total = avg_stats(&runtime_branches_stats[cpu]); | |
11018201 AB |
436 | |
437 | if (total) | |
438 | ratio = avg * 100 / total; | |
439 | ||
dd86e72a | 440 | fprintf(stderr, " # %10.3f %% ", ratio); |
11018201 | 441 | |
f5b4a9c3 SE |
442 | } else if (runtime_nsecs_stats[cpu].n != 0) { |
443 | total = avg_stats(&runtime_nsecs_stats[cpu]); | |
c7f7fea3 IM |
444 | |
445 | if (total) | |
446 | ratio = 1000.0 * avg / total; | |
447 | ||
448 | fprintf(stderr, " # %10.3f M/sec", ratio); | |
44175b6f | 449 | } |
44175b6f IM |
450 | } |
451 | ||
2996f5dd IM |
452 | /* |
453 | * Print out the results of a single counter: | |
f5b4a9c3 | 454 | * aggregated counts in system-wide mode |
2996f5dd | 455 | */ |
69aad6f1 | 456 | static void print_counter_aggr(struct perf_evsel *counter) |
2996f5dd | 457 | { |
69aad6f1 ACM |
458 | struct perf_stat *ps = counter->priv; |
459 | double avg = avg_stats(&ps->res_stats[0]); | |
c52b12ed | 460 | int scaled = counter->counts->scaled; |
2996f5dd | 461 | |
2996f5dd | 462 | if (scaled == -1) { |
d7470b6a SE |
463 | fprintf(stderr, "%*s%s%-24s\n", |
464 | csv_output ? 0 : 18, | |
465 | "<not counted>", csv_sep, event_name(counter)); | |
2996f5dd IM |
466 | return; |
467 | } | |
c04f5e5d | 468 | |
44175b6f | 469 | if (nsec_counter(counter)) |
f5b4a9c3 | 470 | nsec_printout(-1, counter, avg); |
44175b6f | 471 | else |
f5b4a9c3 | 472 | abs_printout(-1, counter, avg); |
849abde9 | 473 | |
d7470b6a SE |
474 | if (csv_output) { |
475 | fputc('\n', stderr); | |
476 | return; | |
477 | } | |
478 | ||
849abde9 | 479 | print_noise(counter, avg); |
506d4bc8 PZ |
480 | |
481 | if (scaled) { | |
482 | double avg_enabled, avg_running; | |
483 | ||
69aad6f1 ACM |
484 | avg_enabled = avg_stats(&ps->res_stats[1]); |
485 | avg_running = avg_stats(&ps->res_stats[2]); | |
d7c29318 | 486 | |
210ad39f | 487 | fprintf(stderr, " (scaled from %.2f%%)", |
506d4bc8 PZ |
488 | 100 * avg_running / avg_enabled); |
489 | } | |
44175b6f | 490 | |
c04f5e5d IM |
491 | fprintf(stderr, "\n"); |
492 | } | |
493 | ||
f5b4a9c3 SE |
494 | /* |
495 | * Print out the results of a single counter: | |
496 | * does not use aggregated count in system-wide | |
497 | */ | |
69aad6f1 | 498 | static void print_counter(struct perf_evsel *counter) |
f5b4a9c3 SE |
499 | { |
500 | u64 ena, run, val; | |
501 | int cpu; | |
502 | ||
60d567e2 | 503 | for (cpu = 0; cpu < cpus->nr; cpu++) { |
c52b12ed ACM |
504 | val = counter->counts->cpu[cpu].val; |
505 | ena = counter->counts->cpu[cpu].ena; | |
506 | run = counter->counts->cpu[cpu].run; | |
f5b4a9c3 | 507 | if (run == 0 || ena == 0) { |
d7470b6a SE |
508 | fprintf(stderr, "CPU%*d%s%*s%s%-24s", |
509 | csv_output ? 0 : -4, | |
60d567e2 | 510 | cpus->map[cpu], csv_sep, |
d7470b6a SE |
511 | csv_output ? 0 : 18, |
512 | "<not counted>", csv_sep, | |
513 | event_name(counter)); | |
f5b4a9c3 SE |
514 | |
515 | fprintf(stderr, "\n"); | |
516 | continue; | |
517 | } | |
518 | ||
519 | if (nsec_counter(counter)) | |
520 | nsec_printout(cpu, counter, val); | |
521 | else | |
522 | abs_printout(cpu, counter, val); | |
523 | ||
d7470b6a SE |
524 | if (!csv_output) { |
525 | print_noise(counter, 1.0); | |
f5b4a9c3 | 526 | |
d7470b6a SE |
527 | if (run != ena) { |
528 | fprintf(stderr, " (scaled from %.2f%%)", | |
f5b4a9c3 | 529 | 100.0 * run / ena); |
d7470b6a | 530 | } |
f5b4a9c3 SE |
531 | } |
532 | fprintf(stderr, "\n"); | |
533 | } | |
534 | } | |
535 | ||
42202dd5 IM |
536 | static void print_stat(int argc, const char **argv) |
537 | { | |
69aad6f1 ACM |
538 | struct perf_evsel *counter; |
539 | int i; | |
42202dd5 | 540 | |
ddcacfa0 IM |
541 | fflush(stdout); |
542 | ||
d7470b6a SE |
543 | if (!csv_output) { |
544 | fprintf(stderr, "\n"); | |
545 | fprintf(stderr, " Performance counter stats for "); | |
546 | if(target_pid == -1 && target_tid == -1) { | |
547 | fprintf(stderr, "\'%s", argv[0]); | |
548 | for (i = 1; i < argc; i++) | |
549 | fprintf(stderr, " %s", argv[i]); | |
550 | } else if (target_pid != -1) | |
551 | fprintf(stderr, "process id \'%d", target_pid); | |
552 | else | |
553 | fprintf(stderr, "thread id \'%d", target_tid); | |
44db76c8 | 554 | |
d7470b6a SE |
555 | fprintf(stderr, "\'"); |
556 | if (run_count > 1) | |
557 | fprintf(stderr, " (%d runs)", run_count); | |
558 | fprintf(stderr, ":\n\n"); | |
559 | } | |
2996f5dd | 560 | |
f5b4a9c3 | 561 | if (no_aggr) { |
361c99a6 | 562 | list_for_each_entry(counter, &evsel_list->entries, node) |
f5b4a9c3 SE |
563 | print_counter(counter); |
564 | } else { | |
361c99a6 | 565 | list_for_each_entry(counter, &evsel_list->entries, node) |
f5b4a9c3 SE |
566 | print_counter_aggr(counter); |
567 | } | |
ddcacfa0 | 568 | |
d7470b6a SE |
569 | if (!csv_output) { |
570 | fprintf(stderr, "\n"); | |
571 | fprintf(stderr, " %18.9f seconds time elapsed", | |
572 | avg_stats(&walltime_nsecs_stats)/1e9); | |
573 | if (run_count > 1) { | |
574 | fprintf(stderr, " ( +- %7.3f%% )", | |
506d4bc8 PZ |
575 | 100*stddev_stats(&walltime_nsecs_stats) / |
576 | avg_stats(&walltime_nsecs_stats)); | |
d7470b6a SE |
577 | } |
578 | fprintf(stderr, "\n\n"); | |
566747e6 | 579 | } |
ddcacfa0 IM |
580 | } |
581 | ||
f7b7c26e PZ |
582 | static volatile int signr = -1; |
583 | ||
5242519b | 584 | static void skip_signal(int signo) |
ddcacfa0 | 585 | { |
6be2850e | 586 | if(child_pid == -1) |
60666c63 LW |
587 | done = 1; |
588 | ||
f7b7c26e PZ |
589 | signr = signo; |
590 | } | |
591 | ||
592 | static void sig_atexit(void) | |
593 | { | |
933da83a CW |
594 | if (child_pid != -1) |
595 | kill(child_pid, SIGTERM); | |
596 | ||
f7b7c26e PZ |
597 | if (signr == -1) |
598 | return; | |
599 | ||
600 | signal(signr, SIG_DFL); | |
601 | kill(getpid(), signr); | |
5242519b IM |
602 | } |
603 | ||
604 | static const char * const stat_usage[] = { | |
60666c63 | 605 | "perf stat [<options>] [<command>]", |
5242519b IM |
606 | NULL |
607 | }; | |
608 | ||
d7470b6a SE |
609 | static int stat__set_big_num(const struct option *opt __used, |
610 | const char *s __used, int unset) | |
611 | { | |
612 | big_num_opt = unset ? 0 : 1; | |
613 | return 0; | |
614 | } | |
615 | ||
5242519b | 616 | static const struct option options[] = { |
361c99a6 | 617 | OPT_CALLBACK('e', "event", &evsel_list, "event", |
86847b62 TG |
618 | "event selector. use 'perf list' to list available events", |
619 | parse_events), | |
2e6cdf99 SE |
620 | OPT_BOOLEAN('i', "no-inherit", &no_inherit, |
621 | "child tasks do not inherit counters"), | |
5242519b | 622 | OPT_INTEGER('p', "pid", &target_pid, |
d6d901c2 ZY |
623 | "stat events on existing process id"), |
624 | OPT_INTEGER('t', "tid", &target_tid, | |
625 | "stat events on existing thread id"), | |
5242519b | 626 | OPT_BOOLEAN('a', "all-cpus", &system_wide, |
3d632595 | 627 | "system-wide collection from all CPUs"), |
b26bc5a7 | 628 | OPT_BOOLEAN('c', "scale", &scale, |
3d632595 | 629 | "scale/normalize counters"), |
c0555642 | 630 | OPT_INCR('v', "verbose", &verbose, |
743ee1f8 | 631 | "be more verbose (show counter open errors, etc)"), |
42202dd5 IM |
632 | OPT_INTEGER('r', "repeat", &run_count, |
633 | "repeat command and print average + stddev (max: 100)"), | |
0cfb7a13 IM |
634 | OPT_BOOLEAN('n', "null", &null_run, |
635 | "null run - dont start any counters"), | |
d7470b6a SE |
636 | OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, |
637 | "print large numbers with thousands\' separators", | |
638 | stat__set_big_num), | |
c45c6ea2 SE |
639 | OPT_STRING('C', "cpu", &cpu_list, "cpu", |
640 | "list of cpus to monitor in system-wide"), | |
f5b4a9c3 SE |
641 | OPT_BOOLEAN('A', "no-aggr", &no_aggr, |
642 | "disable CPU count aggregation"), | |
d7470b6a SE |
643 | OPT_STRING('x', "field-separator", &csv_sep, "separator", |
644 | "print counts with custom separator"), | |
5242519b IM |
645 | OPT_END() |
646 | }; | |
647 | ||
f37a291c | 648 | int cmd_stat(int argc, const char **argv, const char *prefix __used) |
5242519b | 649 | { |
69aad6f1 ACM |
650 | struct perf_evsel *pos; |
651 | int status = -ENOMEM; | |
42202dd5 | 652 | |
5af52b51 SE |
653 | setlocale(LC_ALL, ""); |
654 | ||
361c99a6 ACM |
655 | evsel_list = perf_evlist__new(); |
656 | if (evsel_list == NULL) | |
657 | return -ENOMEM; | |
658 | ||
a0541234 AB |
659 | argc = parse_options(argc, argv, options, stat_usage, |
660 | PARSE_OPT_STOP_AT_NON_OPTION); | |
d7470b6a SE |
661 | |
662 | if (csv_sep) | |
663 | csv_output = true; | |
664 | else | |
665 | csv_sep = DEFAULT_SEPARATOR; | |
666 | ||
667 | /* | |
668 | * let the spreadsheet do the pretty-printing | |
669 | */ | |
670 | if (csv_output) { | |
671 | /* User explicitely passed -B? */ | |
672 | if (big_num_opt == 1) { | |
673 | fprintf(stderr, "-B option not supported with -x\n"); | |
674 | usage_with_options(stat_usage, options); | |
675 | } else /* Nope, so disable big number formatting */ | |
676 | big_num = false; | |
677 | } else if (big_num_opt == 0) /* User passed --no-big-num */ | |
678 | big_num = false; | |
679 | ||
d6d901c2 | 680 | if (!argc && target_pid == -1 && target_tid == -1) |
5242519b | 681 | usage_with_options(stat_usage, options); |
9e9772c4 | 682 | if (run_count <= 0) |
42202dd5 | 683 | usage_with_options(stat_usage, options); |
ddcacfa0 | 684 | |
f5b4a9c3 SE |
685 | /* no_aggr is for system-wide only */ |
686 | if (no_aggr && !system_wide) | |
687 | usage_with_options(stat_usage, options); | |
688 | ||
c3043569 | 689 | /* Set attrs and nr_counters if no event is selected and !null_run */ |
361c99a6 | 690 | if (!null_run && !evsel_list->nr_entries) { |
69aad6f1 ACM |
691 | size_t c; |
692 | ||
69aad6f1 | 693 | for (c = 0; c < ARRAY_SIZE(default_attrs); ++c) { |
361c99a6 | 694 | pos = perf_evsel__new(&default_attrs[c], c); |
69aad6f1 ACM |
695 | if (pos == NULL) |
696 | goto out; | |
361c99a6 | 697 | perf_evlist__add(evsel_list, pos); |
69aad6f1 | 698 | } |
c3043569 | 699 | } |
ddcacfa0 | 700 | |
5c98d466 ACM |
701 | if (target_pid != -1) |
702 | target_tid = target_pid; | |
703 | ||
704 | threads = thread_map__new(target_pid, target_tid); | |
705 | if (threads == NULL) { | |
706 | pr_err("Problems finding threads of monitor\n"); | |
707 | usage_with_options(stat_usage, options); | |
708 | } | |
709 | ||
a12b51c4 | 710 | if (system_wide) |
60d567e2 | 711 | cpus = cpu_map__new(cpu_list); |
a12b51c4 | 712 | else |
60d567e2 | 713 | cpus = cpu_map__dummy_new(); |
ddcacfa0 | 714 | |
60d567e2 ACM |
715 | if (cpus == NULL) { |
716 | perror("failed to parse CPUs map"); | |
c45c6ea2 | 717 | usage_with_options(stat_usage, options); |
60d567e2 ACM |
718 | return -1; |
719 | } | |
c45c6ea2 | 720 | |
361c99a6 | 721 | list_for_each_entry(pos, &evsel_list->entries, node) { |
c52b12ed | 722 | if (perf_evsel__alloc_stat_priv(pos) < 0 || |
60d567e2 | 723 | perf_evsel__alloc_counts(pos, cpus->nr) < 0 || |
5c98d466 | 724 | perf_evsel__alloc_fd(pos, cpus->nr, threads->nr) < 0) |
69aad6f1 | 725 | goto out_free_fd; |
d6d901c2 ZY |
726 | } |
727 | ||
58d7e993 IM |
728 | /* |
729 | * We dont want to block the signals - that would cause | |
730 | * child tasks to inherit that and Ctrl-C would not work. | |
731 | * What we want is for Ctrl-C to work in the exec()-ed | |
732 | * task, but being ignored by perf stat itself: | |
733 | */ | |
f7b7c26e | 734 | atexit(sig_atexit); |
58d7e993 IM |
735 | signal(SIGINT, skip_signal); |
736 | signal(SIGALRM, skip_signal); | |
737 | signal(SIGABRT, skip_signal); | |
738 | ||
42202dd5 IM |
739 | status = 0; |
740 | for (run_idx = 0; run_idx < run_count; run_idx++) { | |
741 | if (run_count != 1 && verbose) | |
3d632595 | 742 | fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1); |
42202dd5 IM |
743 | status = run_perf_stat(argc, argv); |
744 | } | |
745 | ||
084ab9f8 ACM |
746 | if (status != -1) |
747 | print_stat(argc, argv); | |
69aad6f1 | 748 | out_free_fd: |
361c99a6 | 749 | list_for_each_entry(pos, &evsel_list->entries, node) |
69aad6f1 | 750 | perf_evsel__free_stat_priv(pos); |
361c99a6 | 751 | perf_evlist__delete(evsel_list); |
69aad6f1 | 752 | out: |
5c98d466 ACM |
753 | thread_map__delete(threads); |
754 | threads = NULL; | |
42202dd5 | 755 | return status; |
ddcacfa0 | 756 | } |