]>
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" |
148be2c1 | 46 | #include "util/util.h" |
5242519b IM |
47 | #include "util/parse-options.h" |
48 | #include "util/parse-events.h" | |
8f28827a | 49 | #include "util/event.h" |
361c99a6 | 50 | #include "util/evlist.h" |
69aad6f1 | 51 | #include "util/evsel.h" |
8f28827a | 52 | #include "util/debug.h" |
a5d243d0 | 53 | #include "util/color.h" |
0007ecea | 54 | #include "util/stat.h" |
60666c63 | 55 | #include "util/header.h" |
a12b51c4 | 56 | #include "util/cpumap.h" |
d6d901c2 | 57 | #include "util/thread.h" |
fd78260b | 58 | #include "util/thread_map.h" |
ddcacfa0 | 59 | |
1f16c575 | 60 | #include <stdlib.h> |
ddcacfa0 | 61 | #include <sys/prctl.h> |
5af52b51 | 62 | #include <locale.h> |
16c8a109 | 63 | |
d7470b6a | 64 | #define DEFAULT_SEPARATOR " " |
2cee77c4 DA |
65 | #define CNTR_NOT_SUPPORTED "<not supported>" |
66 | #define CNTR_NOT_COUNTED "<not counted>" | |
d7470b6a | 67 | |
666e6d48 | 68 | static struct perf_evlist *evsel_list; |
361c99a6 | 69 | |
77a6f014 NK |
70 | static struct perf_target target = { |
71 | .uid = UINT_MAX, | |
72 | }; | |
ddcacfa0 | 73 | |
3d632595 | 74 | static int run_count = 1; |
2e6cdf99 | 75 | static bool no_inherit = false; |
c0555642 | 76 | static bool scale = true; |
f5b4a9c3 | 77 | static bool no_aggr = false; |
933da83a | 78 | static pid_t child_pid = -1; |
c0555642 | 79 | static bool null_run = false; |
2cba3ffb | 80 | static int detailed_run = 0; |
201e0b06 | 81 | static bool big_num = true; |
d7470b6a | 82 | static int big_num_opt = -1; |
d7470b6a SE |
83 | static const char *csv_sep = NULL; |
84 | static bool csv_output = false; | |
43bece79 | 85 | static bool group = false; |
4aa9015f | 86 | static FILE *output = NULL; |
1f16c575 PZ |
87 | static const char *pre_cmd = NULL; |
88 | static const char *post_cmd = NULL; | |
89 | static bool sync_run = false; | |
5af52b51 | 90 | |
60666c63 LW |
91 | static volatile int done = 0; |
92 | ||
69aad6f1 ACM |
93 | struct perf_stat { |
94 | struct stats res_stats[3]; | |
69aad6f1 ACM |
95 | }; |
96 | ||
c52b12ed | 97 | static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel) |
69aad6f1 | 98 | { |
c52b12ed | 99 | evsel->priv = zalloc(sizeof(struct perf_stat)); |
69aad6f1 ACM |
100 | return evsel->priv == NULL ? -ENOMEM : 0; |
101 | } | |
102 | ||
103 | static void perf_evsel__free_stat_priv(struct perf_evsel *evsel) | |
104 | { | |
105 | free(evsel->priv); | |
106 | evsel->priv = NULL; | |
107 | } | |
108 | ||
7ae92e74 YZ |
109 | static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel) |
110 | { | |
111 | return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus; | |
112 | } | |
113 | ||
114 | static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel) | |
115 | { | |
116 | return perf_evsel__cpus(evsel)->nr; | |
117 | } | |
118 | ||
666e6d48 RR |
119 | static struct stats runtime_nsecs_stats[MAX_NR_CPUS]; |
120 | static struct stats runtime_cycles_stats[MAX_NR_CPUS]; | |
121 | static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS]; | |
122 | static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS]; | |
123 | static struct stats runtime_branches_stats[MAX_NR_CPUS]; | |
124 | static struct stats runtime_cacherefs_stats[MAX_NR_CPUS]; | |
125 | static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS]; | |
126 | static struct stats runtime_l1_icache_stats[MAX_NR_CPUS]; | |
127 | static struct stats runtime_ll_cache_stats[MAX_NR_CPUS]; | |
128 | static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS]; | |
129 | static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS]; | |
130 | static struct stats walltime_nsecs_stats; | |
be1ac0d8 | 131 | |
cac21425 | 132 | static int create_perf_stat_counter(struct perf_evsel *evsel) |
ddcacfa0 | 133 | { |
69aad6f1 | 134 | struct perf_event_attr *attr = &evsel->attr; |
5622c07b SE |
135 | bool exclude_guest_missing = false; |
136 | int ret; | |
727ab04e | 137 | |
ddcacfa0 | 138 | if (scale) |
a21ca2ca IM |
139 | attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | |
140 | PERF_FORMAT_TOTAL_TIME_RUNNING; | |
ddcacfa0 | 141 | |
5d2cd909 ACM |
142 | attr->inherit = !no_inherit; |
143 | ||
5622c07b SE |
144 | retry: |
145 | if (exclude_guest_missing) | |
146 | evsel->attr.exclude_guest = evsel->attr.exclude_host = 0; | |
147 | ||
16ee6576 | 148 | if (perf_target__has_cpu(&target)) { |
7ae92e74 | 149 | ret = perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel)); |
5622c07b SE |
150 | if (ret) |
151 | goto check_ret; | |
152 | return 0; | |
153 | } | |
154 | ||
07ac002f ACM |
155 | if (!perf_target__has_task(&target) && |
156 | !perf_evsel__is_group_member(evsel)) { | |
48290609 ACM |
157 | attr->disabled = 1; |
158 | attr->enable_on_exec = 1; | |
ddcacfa0 | 159 | } |
084ab9f8 | 160 | |
6a4bb04c | 161 | ret = perf_evsel__open_per_thread(evsel, evsel_list->threads); |
5622c07b SE |
162 | if (!ret) |
163 | return 0; | |
164 | /* fall through */ | |
165 | check_ret: | |
166 | if (ret && errno == EINVAL) { | |
167 | if (!exclude_guest_missing && | |
168 | (evsel->attr.exclude_guest || evsel->attr.exclude_host)) { | |
169 | pr_debug("Old kernel, cannot exclude " | |
170 | "guest or host samples.\n"); | |
171 | exclude_guest_missing = true; | |
172 | goto retry; | |
173 | } | |
174 | } | |
175 | return ret; | |
ddcacfa0 IM |
176 | } |
177 | ||
c04f5e5d IM |
178 | /* |
179 | * Does the counter have nsecs as a unit? | |
180 | */ | |
daec78a0 | 181 | static inline int nsec_counter(struct perf_evsel *evsel) |
c04f5e5d | 182 | { |
daec78a0 ACM |
183 | if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || |
184 | perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) | |
c04f5e5d IM |
185 | return 1; |
186 | ||
187 | return 0; | |
188 | } | |
189 | ||
dcd9936a IM |
190 | /* |
191 | * Update various tracking values we maintain to print | |
192 | * more semantic information such as miss/hit ratios, | |
193 | * instruction rates, etc: | |
194 | */ | |
195 | static void update_shadow_stats(struct perf_evsel *counter, u64 *count) | |
196 | { | |
197 | if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK)) | |
198 | update_stats(&runtime_nsecs_stats[0], count[0]); | |
199 | else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES)) | |
200 | update_stats(&runtime_cycles_stats[0], count[0]); | |
d3d1e86d IM |
201 | else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) |
202 | update_stats(&runtime_stalled_cycles_front_stats[0], count[0]); | |
129c04cb | 203 | else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND)) |
d3d1e86d | 204 | update_stats(&runtime_stalled_cycles_back_stats[0], count[0]); |
dcd9936a IM |
205 | else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS)) |
206 | update_stats(&runtime_branches_stats[0], count[0]); | |
207 | else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES)) | |
208 | update_stats(&runtime_cacherefs_stats[0], count[0]); | |
8bb6c79f IM |
209 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D)) |
210 | update_stats(&runtime_l1_dcache_stats[0], count[0]); | |
c3305257 IM |
211 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I)) |
212 | update_stats(&runtime_l1_icache_stats[0], count[0]); | |
213 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL)) | |
214 | update_stats(&runtime_ll_cache_stats[0], count[0]); | |
215 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB)) | |
216 | update_stats(&runtime_dtlb_cache_stats[0], count[0]); | |
217 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB)) | |
218 | update_stats(&runtime_itlb_cache_stats[0], count[0]); | |
dcd9936a IM |
219 | } |
220 | ||
c04f5e5d | 221 | /* |
2996f5dd | 222 | * Read out the results of a single counter: |
f5b4a9c3 | 223 | * aggregate counts across CPUs in system-wide mode |
c04f5e5d | 224 | */ |
c52b12ed | 225 | static int read_counter_aggr(struct perf_evsel *counter) |
c04f5e5d | 226 | { |
69aad6f1 | 227 | struct perf_stat *ps = counter->priv; |
c52b12ed ACM |
228 | u64 *count = counter->counts->aggr.values; |
229 | int i; | |
2996f5dd | 230 | |
7ae92e74 | 231 | if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter), |
7e2ed097 | 232 | evsel_list->threads->nr, scale) < 0) |
c52b12ed | 233 | return -1; |
9e9772c4 PZ |
234 | |
235 | for (i = 0; i < 3; i++) | |
69aad6f1 | 236 | update_stats(&ps->res_stats[i], count[i]); |
9e9772c4 PZ |
237 | |
238 | if (verbose) { | |
4aa9015f | 239 | fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", |
7289f83c | 240 | perf_evsel__name(counter), count[0], count[1], count[2]); |
9e9772c4 PZ |
241 | } |
242 | ||
be1ac0d8 IM |
243 | /* |
244 | * Save the full runtime - to allow normalization during printout: | |
245 | */ | |
dcd9936a | 246 | update_shadow_stats(counter, count); |
c52b12ed ACM |
247 | |
248 | return 0; | |
f5b4a9c3 SE |
249 | } |
250 | ||
251 | /* | |
252 | * Read out the results of a single counter: | |
253 | * do not aggregate counts across CPUs in system-wide mode | |
254 | */ | |
c52b12ed | 255 | static int read_counter(struct perf_evsel *counter) |
f5b4a9c3 | 256 | { |
c52b12ed | 257 | u64 *count; |
f5b4a9c3 | 258 | int cpu; |
f5b4a9c3 | 259 | |
7ae92e74 | 260 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { |
c52b12ed ACM |
261 | if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0) |
262 | return -1; | |
f5b4a9c3 | 263 | |
c52b12ed | 264 | count = counter->counts->cpu[cpu].values; |
f5b4a9c3 | 265 | |
dcd9936a | 266 | update_shadow_stats(counter, count); |
f5b4a9c3 | 267 | } |
c52b12ed ACM |
268 | |
269 | return 0; | |
2996f5dd IM |
270 | } |
271 | ||
1f16c575 | 272 | static int __run_perf_stat(int argc __maybe_unused, const char **argv) |
42202dd5 IM |
273 | { |
274 | unsigned long long t0, t1; | |
cac21425 | 275 | struct perf_evsel *counter; |
42202dd5 | 276 | int status = 0; |
051ae7f7 | 277 | int child_ready_pipe[2], go_pipe[2]; |
6be2850e | 278 | const bool forks = (argc > 0); |
051ae7f7 | 279 | char buf; |
42202dd5 | 280 | |
60666c63 | 281 | if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) { |
051ae7f7 | 282 | perror("failed to create pipes"); |
fceda7fe | 283 | return -1; |
051ae7f7 PM |
284 | } |
285 | ||
60666c63 | 286 | if (forks) { |
6be2850e | 287 | if ((child_pid = fork()) < 0) |
60666c63 LW |
288 | perror("failed to fork"); |
289 | ||
6be2850e | 290 | if (!child_pid) { |
60666c63 LW |
291 | close(child_ready_pipe[0]); |
292 | close(go_pipe[1]); | |
293 | fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); | |
294 | ||
295 | /* | |
296 | * Do a dummy execvp to get the PLT entry resolved, | |
297 | * so we avoid the resolver overhead on the real | |
298 | * execvp call. | |
299 | */ | |
300 | execvp("", (char **)argv); | |
301 | ||
302 | /* | |
303 | * Tell the parent we're ready to go | |
304 | */ | |
305 | close(child_ready_pipe[1]); | |
306 | ||
307 | /* | |
308 | * Wait until the parent tells us to go. | |
309 | */ | |
310 | if (read(go_pipe[0], &buf, 1) == -1) | |
311 | perror("unable to read pipe"); | |
312 | ||
313 | execvp(argv[0], (char **)argv); | |
314 | ||
315 | perror(argv[0]); | |
316 | exit(-1); | |
317 | } | |
051ae7f7 | 318 | |
d67356e7 | 319 | if (perf_target__none(&target)) |
7e2ed097 | 320 | evsel_list->threads->map[0] = child_pid; |
d6d901c2 | 321 | |
051ae7f7 | 322 | /* |
60666c63 | 323 | * Wait for the child to be ready to exec. |
051ae7f7 PM |
324 | */ |
325 | close(child_ready_pipe[1]); | |
60666c63 LW |
326 | close(go_pipe[0]); |
327 | if (read(child_ready_pipe[0], &buf, 1) == -1) | |
a92bef0f | 328 | perror("unable to read pipe"); |
60666c63 | 329 | close(child_ready_pipe[0]); |
051ae7f7 PM |
330 | } |
331 | ||
6a4bb04c | 332 | if (group) |
63dab225 | 333 | perf_evlist__set_leader(evsel_list); |
6a4bb04c | 334 | |
361c99a6 | 335 | list_for_each_entry(counter, &evsel_list->entries, node) { |
cac21425 | 336 | if (create_perf_stat_counter(counter) < 0) { |
979987a5 DA |
337 | /* |
338 | * PPC returns ENXIO for HW counters until 2.6.37 | |
339 | * (behavior changed with commit b0a873e). | |
340 | */ | |
38f6ae1e | 341 | if (errno == EINVAL || errno == ENOSYS || |
979987a5 DA |
342 | errno == ENOENT || errno == EOPNOTSUPP || |
343 | errno == ENXIO) { | |
c63ca0c0 DA |
344 | if (verbose) |
345 | ui__warning("%s event is not supported by the kernel.\n", | |
7289f83c | 346 | perf_evsel__name(counter)); |
2cee77c4 | 347 | counter->supported = false; |
ede70290 | 348 | continue; |
c63ca0c0 | 349 | } |
ede70290 IM |
350 | |
351 | if (errno == EPERM || errno == EACCES) { | |
48290609 ACM |
352 | error("You may not have permission to collect %sstats.\n" |
353 | "\t Consider tweaking" | |
354 | " /proc/sys/kernel/perf_event_paranoid or running as root.", | |
20f946b4 | 355 | target.system_wide ? "system-wide " : ""); |
48290609 ACM |
356 | } else { |
357 | error("open_counter returned with %d (%s). " | |
358 | "/bin/dmesg may provide additional information.\n", | |
359 | errno, strerror(errno)); | |
360 | } | |
361 | if (child_pid != -1) | |
362 | kill(child_pid, SIGTERM); | |
fceda7fe DA |
363 | |
364 | pr_err("Not all events could be opened.\n"); | |
48290609 ACM |
365 | return -1; |
366 | } | |
2cee77c4 | 367 | counter->supported = true; |
084ab9f8 | 368 | } |
42202dd5 | 369 | |
1491a632 | 370 | if (perf_evlist__apply_filters(evsel_list)) { |
cfd748ae FW |
371 | error("failed to set filter with %d (%s)\n", errno, |
372 | strerror(errno)); | |
373 | return -1; | |
374 | } | |
375 | ||
42202dd5 IM |
376 | /* |
377 | * Enable counters and exec the command: | |
378 | */ | |
379 | t0 = rdclock(); | |
42202dd5 | 380 | |
60666c63 LW |
381 | if (forks) { |
382 | close(go_pipe[1]); | |
383 | wait(&status); | |
33e49ea7 AK |
384 | if (WIFSIGNALED(status)) |
385 | psignal(WTERMSIG(status), argv[0]); | |
60666c63 | 386 | } else { |
6be2850e | 387 | while(!done) sleep(1); |
60666c63 | 388 | } |
42202dd5 | 389 | |
42202dd5 IM |
390 | t1 = rdclock(); |
391 | ||
9e9772c4 | 392 | update_stats(&walltime_nsecs_stats, t1 - t0); |
42202dd5 | 393 | |
f5b4a9c3 | 394 | if (no_aggr) { |
361c99a6 | 395 | list_for_each_entry(counter, &evsel_list->entries, node) { |
f5b4a9c3 | 396 | read_counter(counter); |
7ae92e74 | 397 | perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1); |
c52b12ed | 398 | } |
f5b4a9c3 | 399 | } else { |
361c99a6 | 400 | list_for_each_entry(counter, &evsel_list->entries, node) { |
f5b4a9c3 | 401 | read_counter_aggr(counter); |
7ae92e74 | 402 | perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), |
7e2ed097 | 403 | evsel_list->threads->nr); |
c52b12ed | 404 | } |
f5b4a9c3 | 405 | } |
c52b12ed | 406 | |
42202dd5 IM |
407 | return WEXITSTATUS(status); |
408 | } | |
409 | ||
1f16c575 PZ |
410 | static int run_perf_stat(int argc __maybe_unused, const char **argv) |
411 | { | |
412 | int ret; | |
413 | ||
414 | if (pre_cmd) { | |
415 | ret = system(pre_cmd); | |
416 | if (ret) | |
417 | return ret; | |
418 | } | |
419 | ||
420 | if (sync_run) | |
421 | sync(); | |
422 | ||
423 | ret = __run_perf_stat(argc, argv); | |
424 | if (ret) | |
425 | return ret; | |
426 | ||
427 | if (post_cmd) { | |
428 | ret = system(post_cmd); | |
429 | if (ret) | |
430 | return ret; | |
431 | } | |
432 | ||
433 | return ret; | |
434 | } | |
435 | ||
f99844cb IM |
436 | static void print_noise_pct(double total, double avg) |
437 | { | |
0007ecea | 438 | double pct = rel_stddev_stats(total, avg); |
f99844cb | 439 | |
3ae9a34d | 440 | if (csv_output) |
4aa9015f | 441 | fprintf(output, "%s%.2f%%", csv_sep, pct); |
a1bca6cc | 442 | else if (pct) |
4aa9015f | 443 | fprintf(output, " ( +-%6.2f%% )", pct); |
f99844cb IM |
444 | } |
445 | ||
69aad6f1 | 446 | static void print_noise(struct perf_evsel *evsel, double avg) |
42202dd5 | 447 | { |
69aad6f1 ACM |
448 | struct perf_stat *ps; |
449 | ||
849abde9 PZ |
450 | if (run_count == 1) |
451 | return; | |
452 | ||
69aad6f1 | 453 | ps = evsel->priv; |
f99844cb | 454 | print_noise_pct(stddev_stats(&ps->res_stats[0]), avg); |
42202dd5 IM |
455 | } |
456 | ||
daec78a0 | 457 | static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg) |
44175b6f | 458 | { |
506d4bc8 | 459 | double msecs = avg / 1e6; |
d7470b6a | 460 | char cpustr[16] = { '\0', }; |
2cba3ffb | 461 | const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s"; |
44175b6f | 462 | |
f5b4a9c3 | 463 | if (no_aggr) |
d7470b6a SE |
464 | sprintf(cpustr, "CPU%*d%s", |
465 | csv_output ? 0 : -4, | |
7ae92e74 | 466 | perf_evsel__cpus(evsel)->map[cpu], csv_sep); |
d7470b6a | 467 | |
7289f83c | 468 | fprintf(output, fmt, cpustr, msecs, csv_sep, perf_evsel__name(evsel)); |
d7470b6a | 469 | |
023695d9 | 470 | if (evsel->cgrp) |
4aa9015f | 471 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); |
023695d9 | 472 | |
d7470b6a SE |
473 | if (csv_output) |
474 | return; | |
44175b6f | 475 | |
daec78a0 | 476 | if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) |
4aa9015f SE |
477 | fprintf(output, " # %8.3f CPUs utilized ", |
478 | avg / avg_stats(&walltime_nsecs_stats)); | |
9dac6a29 NK |
479 | else |
480 | fprintf(output, " "); | |
44175b6f IM |
481 | } |
482 | ||
15e6392f NK |
483 | /* used for get_ratio_color() */ |
484 | enum grc_type { | |
485 | GRC_STALLED_CYCLES_FE, | |
486 | GRC_STALLED_CYCLES_BE, | |
487 | GRC_CACHE_MISSES, | |
488 | GRC_MAX_NR | |
489 | }; | |
490 | ||
491 | static const char *get_ratio_color(enum grc_type type, double ratio) | |
492 | { | |
493 | static const double grc_table[GRC_MAX_NR][3] = { | |
494 | [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 }, | |
495 | [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 }, | |
496 | [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 }, | |
497 | }; | |
498 | const char *color = PERF_COLOR_NORMAL; | |
499 | ||
500 | if (ratio > grc_table[type][0]) | |
501 | color = PERF_COLOR_RED; | |
502 | else if (ratio > grc_table[type][1]) | |
503 | color = PERF_COLOR_MAGENTA; | |
504 | else if (ratio > grc_table[type][2]) | |
505 | color = PERF_COLOR_YELLOW; | |
506 | ||
507 | return color; | |
508 | } | |
509 | ||
1d037ca1 IT |
510 | static void print_stalled_cycles_frontend(int cpu, |
511 | struct perf_evsel *evsel | |
512 | __maybe_unused, double avg) | |
d3d1e86d IM |
513 | { |
514 | double total, ratio = 0.0; | |
515 | const char *color; | |
516 | ||
517 | total = avg_stats(&runtime_cycles_stats[cpu]); | |
518 | ||
519 | if (total) | |
520 | ratio = avg / total * 100.0; | |
521 | ||
15e6392f | 522 | color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio); |
d3d1e86d | 523 | |
4aa9015f SE |
524 | fprintf(output, " # "); |
525 | color_fprintf(output, color, "%6.2f%%", ratio); | |
526 | fprintf(output, " frontend cycles idle "); | |
d3d1e86d IM |
527 | } |
528 | ||
1d037ca1 IT |
529 | static void print_stalled_cycles_backend(int cpu, |
530 | struct perf_evsel *evsel | |
531 | __maybe_unused, double avg) | |
a5d243d0 IM |
532 | { |
533 | double total, ratio = 0.0; | |
534 | const char *color; | |
535 | ||
536 | total = avg_stats(&runtime_cycles_stats[cpu]); | |
537 | ||
538 | if (total) | |
539 | ratio = avg / total * 100.0; | |
540 | ||
15e6392f | 541 | color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio); |
a5d243d0 | 542 | |
4aa9015f SE |
543 | fprintf(output, " # "); |
544 | color_fprintf(output, color, "%6.2f%%", ratio); | |
545 | fprintf(output, " backend cycles idle "); | |
a5d243d0 IM |
546 | } |
547 | ||
1d037ca1 IT |
548 | static void print_branch_misses(int cpu, |
549 | struct perf_evsel *evsel __maybe_unused, | |
550 | double avg) | |
c78df6c1 IM |
551 | { |
552 | double total, ratio = 0.0; | |
553 | const char *color; | |
554 | ||
555 | total = avg_stats(&runtime_branches_stats[cpu]); | |
556 | ||
557 | if (total) | |
558 | ratio = avg / total * 100.0; | |
559 | ||
15e6392f | 560 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c78df6c1 | 561 | |
4aa9015f SE |
562 | fprintf(output, " # "); |
563 | color_fprintf(output, color, "%6.2f%%", ratio); | |
564 | fprintf(output, " of all branches "); | |
c78df6c1 IM |
565 | } |
566 | ||
1d037ca1 IT |
567 | static void print_l1_dcache_misses(int cpu, |
568 | struct perf_evsel *evsel __maybe_unused, | |
569 | double avg) | |
8bb6c79f IM |
570 | { |
571 | double total, ratio = 0.0; | |
572 | const char *color; | |
573 | ||
574 | total = avg_stats(&runtime_l1_dcache_stats[cpu]); | |
575 | ||
576 | if (total) | |
577 | ratio = avg / total * 100.0; | |
578 | ||
15e6392f | 579 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
8bb6c79f | 580 | |
4aa9015f SE |
581 | fprintf(output, " # "); |
582 | color_fprintf(output, color, "%6.2f%%", ratio); | |
583 | fprintf(output, " of all L1-dcache hits "); | |
8bb6c79f IM |
584 | } |
585 | ||
1d037ca1 IT |
586 | static void print_l1_icache_misses(int cpu, |
587 | struct perf_evsel *evsel __maybe_unused, | |
588 | double avg) | |
c3305257 IM |
589 | { |
590 | double total, ratio = 0.0; | |
591 | const char *color; | |
592 | ||
593 | total = avg_stats(&runtime_l1_icache_stats[cpu]); | |
594 | ||
595 | if (total) | |
596 | ratio = avg / total * 100.0; | |
597 | ||
15e6392f | 598 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 599 | |
4aa9015f SE |
600 | fprintf(output, " # "); |
601 | color_fprintf(output, color, "%6.2f%%", ratio); | |
602 | fprintf(output, " of all L1-icache hits "); | |
c3305257 IM |
603 | } |
604 | ||
1d037ca1 IT |
605 | static void print_dtlb_cache_misses(int cpu, |
606 | struct perf_evsel *evsel __maybe_unused, | |
607 | double avg) | |
c3305257 IM |
608 | { |
609 | double total, ratio = 0.0; | |
610 | const char *color; | |
611 | ||
612 | total = avg_stats(&runtime_dtlb_cache_stats[cpu]); | |
613 | ||
614 | if (total) | |
615 | ratio = avg / total * 100.0; | |
616 | ||
15e6392f | 617 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 618 | |
4aa9015f SE |
619 | fprintf(output, " # "); |
620 | color_fprintf(output, color, "%6.2f%%", ratio); | |
621 | fprintf(output, " of all dTLB cache hits "); | |
c3305257 IM |
622 | } |
623 | ||
1d037ca1 IT |
624 | static void print_itlb_cache_misses(int cpu, |
625 | struct perf_evsel *evsel __maybe_unused, | |
626 | double avg) | |
c3305257 IM |
627 | { |
628 | double total, ratio = 0.0; | |
629 | const char *color; | |
630 | ||
631 | total = avg_stats(&runtime_itlb_cache_stats[cpu]); | |
632 | ||
633 | if (total) | |
634 | ratio = avg / total * 100.0; | |
635 | ||
15e6392f | 636 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 637 | |
4aa9015f SE |
638 | fprintf(output, " # "); |
639 | color_fprintf(output, color, "%6.2f%%", ratio); | |
640 | fprintf(output, " of all iTLB cache hits "); | |
c3305257 IM |
641 | } |
642 | ||
1d037ca1 IT |
643 | static void print_ll_cache_misses(int cpu, |
644 | struct perf_evsel *evsel __maybe_unused, | |
645 | double avg) | |
c3305257 IM |
646 | { |
647 | double total, ratio = 0.0; | |
648 | const char *color; | |
649 | ||
650 | total = avg_stats(&runtime_ll_cache_stats[cpu]); | |
651 | ||
652 | if (total) | |
653 | ratio = avg / total * 100.0; | |
654 | ||
15e6392f | 655 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 656 | |
4aa9015f SE |
657 | fprintf(output, " # "); |
658 | color_fprintf(output, color, "%6.2f%%", ratio); | |
659 | fprintf(output, " of all LL-cache hits "); | |
c3305257 IM |
660 | } |
661 | ||
daec78a0 | 662 | static void abs_printout(int cpu, struct perf_evsel *evsel, double avg) |
44175b6f | 663 | { |
c7f7fea3 | 664 | double total, ratio = 0.0; |
f5b4a9c3 | 665 | char cpustr[16] = { '\0', }; |
d7470b6a SE |
666 | const char *fmt; |
667 | ||
668 | if (csv_output) | |
669 | fmt = "%s%.0f%s%s"; | |
670 | else if (big_num) | |
2cba3ffb | 671 | fmt = "%s%'18.0f%s%-25s"; |
d7470b6a | 672 | else |
2cba3ffb | 673 | fmt = "%s%18.0f%s%-25s"; |
f5b4a9c3 SE |
674 | |
675 | if (no_aggr) | |
d7470b6a SE |
676 | sprintf(cpustr, "CPU%*d%s", |
677 | csv_output ? 0 : -4, | |
7ae92e74 | 678 | perf_evsel__cpus(evsel)->map[cpu], csv_sep); |
f5b4a9c3 SE |
679 | else |
680 | cpu = 0; | |
c7f7fea3 | 681 | |
7289f83c | 682 | fprintf(output, fmt, cpustr, avg, csv_sep, perf_evsel__name(evsel)); |
d7470b6a | 683 | |
023695d9 | 684 | if (evsel->cgrp) |
4aa9015f | 685 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); |
023695d9 | 686 | |
d7470b6a SE |
687 | if (csv_output) |
688 | return; | |
44175b6f | 689 | |
daec78a0 | 690 | if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) { |
f5b4a9c3 | 691 | total = avg_stats(&runtime_cycles_stats[cpu]); |
c7f7fea3 IM |
692 | |
693 | if (total) | |
694 | ratio = avg / total; | |
695 | ||
4aa9015f | 696 | fprintf(output, " # %5.2f insns per cycle ", ratio); |
481f988a | 697 | |
d3d1e86d IM |
698 | total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]); |
699 | total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu])); | |
481f988a IM |
700 | |
701 | if (total && avg) { | |
702 | ratio = total / avg; | |
4aa9015f | 703 | fprintf(output, "\n # %5.2f stalled cycles per insn", ratio); |
481f988a IM |
704 | } |
705 | ||
daec78a0 | 706 | } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) && |
f5b4a9c3 | 707 | runtime_branches_stats[cpu].n != 0) { |
c78df6c1 | 708 | print_branch_misses(cpu, evsel, avg); |
8bb6c79f IM |
709 | } else if ( |
710 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
711 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D | | |
712 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
713 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
c6264def | 714 | runtime_l1_dcache_stats[cpu].n != 0) { |
8bb6c79f | 715 | print_l1_dcache_misses(cpu, evsel, avg); |
c3305257 IM |
716 | } else if ( |
717 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
718 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I | | |
719 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
720 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
721 | runtime_l1_icache_stats[cpu].n != 0) { | |
722 | print_l1_icache_misses(cpu, evsel, avg); | |
723 | } else if ( | |
724 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
725 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB | | |
726 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
727 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
728 | runtime_dtlb_cache_stats[cpu].n != 0) { | |
729 | print_dtlb_cache_misses(cpu, evsel, avg); | |
730 | } else if ( | |
731 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
732 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB | | |
733 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
734 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
735 | runtime_itlb_cache_stats[cpu].n != 0) { | |
736 | print_itlb_cache_misses(cpu, evsel, avg); | |
737 | } else if ( | |
738 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
739 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL | | |
740 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
741 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
742 | runtime_ll_cache_stats[cpu].n != 0) { | |
743 | print_ll_cache_misses(cpu, evsel, avg); | |
d58f4c82 IM |
744 | } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) && |
745 | runtime_cacherefs_stats[cpu].n != 0) { | |
746 | total = avg_stats(&runtime_cacherefs_stats[cpu]); | |
747 | ||
748 | if (total) | |
749 | ratio = avg * 100 / total; | |
750 | ||
4aa9015f | 751 | fprintf(output, " # %8.3f %% of all cache refs ", ratio); |
d58f4c82 | 752 | |
d3d1e86d IM |
753 | } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) { |
754 | print_stalled_cycles_frontend(cpu, evsel, avg); | |
129c04cb | 755 | } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) { |
d3d1e86d | 756 | print_stalled_cycles_backend(cpu, evsel, avg); |
481f988a | 757 | } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) { |
f5b4a9c3 | 758 | total = avg_stats(&runtime_nsecs_stats[cpu]); |
c7f7fea3 IM |
759 | |
760 | if (total) | |
481f988a | 761 | ratio = 1.0 * avg / total; |
c7f7fea3 | 762 | |
4aa9015f | 763 | fprintf(output, " # %8.3f GHz ", ratio); |
481f988a | 764 | } else if (runtime_nsecs_stats[cpu].n != 0) { |
5fde2523 NK |
765 | char unit = 'M'; |
766 | ||
481f988a | 767 | total = avg_stats(&runtime_nsecs_stats[cpu]); |
11ba2b85 IM |
768 | |
769 | if (total) | |
481f988a | 770 | ratio = 1000.0 * avg / total; |
5fde2523 NK |
771 | if (ratio < 0.001) { |
772 | ratio *= 1000; | |
773 | unit = 'K'; | |
774 | } | |
11ba2b85 | 775 | |
5fde2523 | 776 | fprintf(output, " # %8.3f %c/sec ", ratio, unit); |
a5d243d0 | 777 | } else { |
4aa9015f | 778 | fprintf(output, " "); |
44175b6f | 779 | } |
44175b6f IM |
780 | } |
781 | ||
2996f5dd IM |
782 | /* |
783 | * Print out the results of a single counter: | |
f5b4a9c3 | 784 | * aggregated counts in system-wide mode |
2996f5dd | 785 | */ |
69aad6f1 | 786 | static void print_counter_aggr(struct perf_evsel *counter) |
2996f5dd | 787 | { |
69aad6f1 ACM |
788 | struct perf_stat *ps = counter->priv; |
789 | double avg = avg_stats(&ps->res_stats[0]); | |
c52b12ed | 790 | int scaled = counter->counts->scaled; |
2996f5dd | 791 | |
2996f5dd | 792 | if (scaled == -1) { |
4aa9015f | 793 | fprintf(output, "%*s%s%*s", |
d7470b6a | 794 | csv_output ? 0 : 18, |
2cee77c4 | 795 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, |
023695d9 SE |
796 | csv_sep, |
797 | csv_output ? 0 : -24, | |
7289f83c | 798 | perf_evsel__name(counter)); |
023695d9 SE |
799 | |
800 | if (counter->cgrp) | |
4aa9015f | 801 | fprintf(output, "%s%s", csv_sep, counter->cgrp->name); |
023695d9 | 802 | |
4aa9015f | 803 | fputc('\n', output); |
2996f5dd IM |
804 | return; |
805 | } | |
c04f5e5d | 806 | |
44175b6f | 807 | if (nsec_counter(counter)) |
f5b4a9c3 | 808 | nsec_printout(-1, counter, avg); |
44175b6f | 809 | else |
f5b4a9c3 | 810 | abs_printout(-1, counter, avg); |
849abde9 | 811 | |
3ae9a34d ZH |
812 | print_noise(counter, avg); |
813 | ||
d7470b6a | 814 | if (csv_output) { |
4aa9015f | 815 | fputc('\n', output); |
d7470b6a SE |
816 | return; |
817 | } | |
818 | ||
506d4bc8 PZ |
819 | if (scaled) { |
820 | double avg_enabled, avg_running; | |
821 | ||
69aad6f1 ACM |
822 | avg_enabled = avg_stats(&ps->res_stats[1]); |
823 | avg_running = avg_stats(&ps->res_stats[2]); | |
d7c29318 | 824 | |
4aa9015f | 825 | fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled); |
506d4bc8 | 826 | } |
4aa9015f | 827 | fprintf(output, "\n"); |
c04f5e5d IM |
828 | } |
829 | ||
f5b4a9c3 SE |
830 | /* |
831 | * Print out the results of a single counter: | |
832 | * does not use aggregated count in system-wide | |
833 | */ | |
69aad6f1 | 834 | static void print_counter(struct perf_evsel *counter) |
f5b4a9c3 SE |
835 | { |
836 | u64 ena, run, val; | |
837 | int cpu; | |
838 | ||
7ae92e74 | 839 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { |
c52b12ed ACM |
840 | val = counter->counts->cpu[cpu].val; |
841 | ena = counter->counts->cpu[cpu].ena; | |
842 | run = counter->counts->cpu[cpu].run; | |
f5b4a9c3 | 843 | if (run == 0 || ena == 0) { |
4aa9015f | 844 | fprintf(output, "CPU%*d%s%*s%s%*s", |
d7470b6a | 845 | csv_output ? 0 : -4, |
7ae92e74 | 846 | perf_evsel__cpus(counter)->map[cpu], csv_sep, |
d7470b6a | 847 | csv_output ? 0 : 18, |
2cee77c4 DA |
848 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, |
849 | csv_sep, | |
023695d9 | 850 | csv_output ? 0 : -24, |
7289f83c | 851 | perf_evsel__name(counter)); |
f5b4a9c3 | 852 | |
023695d9 | 853 | if (counter->cgrp) |
4aa9015f SE |
854 | fprintf(output, "%s%s", |
855 | csv_sep, counter->cgrp->name); | |
023695d9 | 856 | |
4aa9015f | 857 | fputc('\n', output); |
f5b4a9c3 SE |
858 | continue; |
859 | } | |
860 | ||
861 | if (nsec_counter(counter)) | |
862 | nsec_printout(cpu, counter, val); | |
863 | else | |
864 | abs_printout(cpu, counter, val); | |
865 | ||
d7470b6a SE |
866 | if (!csv_output) { |
867 | print_noise(counter, 1.0); | |
f5b4a9c3 | 868 | |
c6264def | 869 | if (run != ena) |
4aa9015f SE |
870 | fprintf(output, " (%.2f%%)", |
871 | 100.0 * run / ena); | |
f5b4a9c3 | 872 | } |
4aa9015f | 873 | fputc('\n', output); |
f5b4a9c3 SE |
874 | } |
875 | } | |
876 | ||
42202dd5 IM |
877 | static void print_stat(int argc, const char **argv) |
878 | { | |
69aad6f1 ACM |
879 | struct perf_evsel *counter; |
880 | int i; | |
42202dd5 | 881 | |
ddcacfa0 IM |
882 | fflush(stdout); |
883 | ||
d7470b6a | 884 | if (!csv_output) { |
4aa9015f SE |
885 | fprintf(output, "\n"); |
886 | fprintf(output, " Performance counter stats for "); | |
aa22dd49 | 887 | if (!perf_target__has_task(&target)) { |
4aa9015f | 888 | fprintf(output, "\'%s", argv[0]); |
d7470b6a | 889 | for (i = 1; i < argc; i++) |
4aa9015f | 890 | fprintf(output, " %s", argv[i]); |
20f946b4 NK |
891 | } else if (target.pid) |
892 | fprintf(output, "process id \'%s", target.pid); | |
d7470b6a | 893 | else |
20f946b4 | 894 | fprintf(output, "thread id \'%s", target.tid); |
44db76c8 | 895 | |
4aa9015f | 896 | fprintf(output, "\'"); |
d7470b6a | 897 | if (run_count > 1) |
4aa9015f SE |
898 | fprintf(output, " (%d runs)", run_count); |
899 | fprintf(output, ":\n\n"); | |
d7470b6a | 900 | } |
2996f5dd | 901 | |
f5b4a9c3 | 902 | if (no_aggr) { |
361c99a6 | 903 | list_for_each_entry(counter, &evsel_list->entries, node) |
f5b4a9c3 SE |
904 | print_counter(counter); |
905 | } else { | |
361c99a6 | 906 | list_for_each_entry(counter, &evsel_list->entries, node) |
f5b4a9c3 SE |
907 | print_counter_aggr(counter); |
908 | } | |
ddcacfa0 | 909 | |
d7470b6a | 910 | if (!csv_output) { |
c3305257 | 911 | if (!null_run) |
4aa9015f SE |
912 | fprintf(output, "\n"); |
913 | fprintf(output, " %17.9f seconds time elapsed", | |
d7470b6a SE |
914 | avg_stats(&walltime_nsecs_stats)/1e9); |
915 | if (run_count > 1) { | |
4aa9015f | 916 | fprintf(output, " "); |
f99844cb IM |
917 | print_noise_pct(stddev_stats(&walltime_nsecs_stats), |
918 | avg_stats(&walltime_nsecs_stats)); | |
d7470b6a | 919 | } |
4aa9015f | 920 | fprintf(output, "\n\n"); |
566747e6 | 921 | } |
ddcacfa0 IM |
922 | } |
923 | ||
f7b7c26e PZ |
924 | static volatile int signr = -1; |
925 | ||
5242519b | 926 | static void skip_signal(int signo) |
ddcacfa0 | 927 | { |
6be2850e | 928 | if(child_pid == -1) |
60666c63 LW |
929 | done = 1; |
930 | ||
f7b7c26e PZ |
931 | signr = signo; |
932 | } | |
933 | ||
934 | static void sig_atexit(void) | |
935 | { | |
933da83a CW |
936 | if (child_pid != -1) |
937 | kill(child_pid, SIGTERM); | |
938 | ||
f7b7c26e PZ |
939 | if (signr == -1) |
940 | return; | |
941 | ||
942 | signal(signr, SIG_DFL); | |
943 | kill(getpid(), signr); | |
5242519b IM |
944 | } |
945 | ||
1d037ca1 IT |
946 | static int stat__set_big_num(const struct option *opt __maybe_unused, |
947 | const char *s __maybe_unused, int unset) | |
d7470b6a SE |
948 | { |
949 | big_num_opt = unset ? 0 : 1; | |
950 | return 0; | |
951 | } | |
952 | ||
2cba3ffb IM |
953 | /* |
954 | * Add default attributes, if there were no attributes specified or | |
955 | * if -d/--detailed, -d -d or -d -d -d is used: | |
956 | */ | |
957 | static int add_default_attributes(void) | |
958 | { | |
b070a547 ACM |
959 | struct perf_event_attr default_attrs[] = { |
960 | ||
961 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, | |
962 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, | |
963 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, | |
964 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, | |
965 | ||
966 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, | |
967 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, | |
968 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, | |
969 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, | |
970 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, | |
971 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, | |
972 | ||
973 | }; | |
974 | ||
975 | /* | |
976 | * Detailed stats (-d), covering the L1 and last level data caches: | |
977 | */ | |
978 | struct perf_event_attr detailed_attrs[] = { | |
979 | ||
980 | { .type = PERF_TYPE_HW_CACHE, | |
981 | .config = | |
982 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
983 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
984 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
985 | ||
986 | { .type = PERF_TYPE_HW_CACHE, | |
987 | .config = | |
988 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
989 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
990 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
991 | ||
992 | { .type = PERF_TYPE_HW_CACHE, | |
993 | .config = | |
994 | PERF_COUNT_HW_CACHE_LL << 0 | | |
995 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
996 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
997 | ||
998 | { .type = PERF_TYPE_HW_CACHE, | |
999 | .config = | |
1000 | PERF_COUNT_HW_CACHE_LL << 0 | | |
1001 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1002 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1003 | }; | |
1004 | ||
1005 | /* | |
1006 | * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: | |
1007 | */ | |
1008 | struct perf_event_attr very_detailed_attrs[] = { | |
1009 | ||
1010 | { .type = PERF_TYPE_HW_CACHE, | |
1011 | .config = | |
1012 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
1013 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1014 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1015 | ||
1016 | { .type = PERF_TYPE_HW_CACHE, | |
1017 | .config = | |
1018 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
1019 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1020 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1021 | ||
1022 | { .type = PERF_TYPE_HW_CACHE, | |
1023 | .config = | |
1024 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1025 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1026 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1027 | ||
1028 | { .type = PERF_TYPE_HW_CACHE, | |
1029 | .config = | |
1030 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1031 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1032 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1033 | ||
1034 | { .type = PERF_TYPE_HW_CACHE, | |
1035 | .config = | |
1036 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
1037 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1038 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1039 | ||
1040 | { .type = PERF_TYPE_HW_CACHE, | |
1041 | .config = | |
1042 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
1043 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1044 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1045 | ||
1046 | }; | |
1047 | ||
1048 | /* | |
1049 | * Very, very detailed stats (-d -d -d), adding prefetch events: | |
1050 | */ | |
1051 | struct perf_event_attr very_very_detailed_attrs[] = { | |
1052 | ||
1053 | { .type = PERF_TYPE_HW_CACHE, | |
1054 | .config = | |
1055 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1056 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
1057 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1058 | ||
1059 | { .type = PERF_TYPE_HW_CACHE, | |
1060 | .config = | |
1061 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1062 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
1063 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1064 | }; | |
1065 | ||
2cba3ffb IM |
1066 | /* Set attrs if no event is selected and !null_run: */ |
1067 | if (null_run) | |
1068 | return 0; | |
1069 | ||
1070 | if (!evsel_list->nr_entries) { | |
79695e1b | 1071 | if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0) |
50d08e47 | 1072 | return -1; |
2cba3ffb IM |
1073 | } |
1074 | ||
1075 | /* Detailed events get appended to the event list: */ | |
1076 | ||
1077 | if (detailed_run < 1) | |
1078 | return 0; | |
1079 | ||
1080 | /* Append detailed run extra attributes: */ | |
79695e1b | 1081 | if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) |
50d08e47 | 1082 | return -1; |
2cba3ffb IM |
1083 | |
1084 | if (detailed_run < 2) | |
1085 | return 0; | |
1086 | ||
1087 | /* Append very detailed run extra attributes: */ | |
79695e1b | 1088 | if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) |
50d08e47 | 1089 | return -1; |
2cba3ffb IM |
1090 | |
1091 | if (detailed_run < 3) | |
1092 | return 0; | |
1093 | ||
1094 | /* Append very, very detailed run extra attributes: */ | |
79695e1b | 1095 | return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); |
2cba3ffb IM |
1096 | } |
1097 | ||
1d037ca1 | 1098 | int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused) |
5242519b | 1099 | { |
1f16c575 | 1100 | bool append_file = false; |
b070a547 ACM |
1101 | int output_fd = 0; |
1102 | const char *output_name = NULL; | |
1103 | const struct option options[] = { | |
1104 | OPT_CALLBACK('e', "event", &evsel_list, "event", | |
1105 | "event selector. use 'perf list' to list available events", | |
1106 | parse_events_option), | |
1107 | OPT_CALLBACK(0, "filter", &evsel_list, "filter", | |
1108 | "event filter", parse_filter), | |
1109 | OPT_BOOLEAN('i', "no-inherit", &no_inherit, | |
1110 | "child tasks do not inherit counters"), | |
1111 | OPT_STRING('p', "pid", &target.pid, "pid", | |
1112 | "stat events on existing process id"), | |
1113 | OPT_STRING('t', "tid", &target.tid, "tid", | |
1114 | "stat events on existing thread id"), | |
1115 | OPT_BOOLEAN('a', "all-cpus", &target.system_wide, | |
1116 | "system-wide collection from all CPUs"), | |
1117 | OPT_BOOLEAN('g', "group", &group, | |
1118 | "put the counters into a counter group"), | |
1119 | OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"), | |
1120 | OPT_INCR('v', "verbose", &verbose, | |
1121 | "be more verbose (show counter open errors, etc)"), | |
1122 | OPT_INTEGER('r', "repeat", &run_count, | |
1123 | "repeat command and print average + stddev (max: 100)"), | |
1124 | OPT_BOOLEAN('n', "null", &null_run, | |
1125 | "null run - dont start any counters"), | |
1126 | OPT_INCR('d', "detailed", &detailed_run, | |
1127 | "detailed run - start a lot of events"), | |
1128 | OPT_BOOLEAN('S', "sync", &sync_run, | |
1129 | "call sync() before starting a run"), | |
1130 | OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, | |
1131 | "print large numbers with thousands\' separators", | |
1132 | stat__set_big_num), | |
1133 | OPT_STRING('C', "cpu", &target.cpu_list, "cpu", | |
1134 | "list of cpus to monitor in system-wide"), | |
1135 | OPT_BOOLEAN('A', "no-aggr", &no_aggr, "disable CPU count aggregation"), | |
1136 | OPT_STRING('x', "field-separator", &csv_sep, "separator", | |
1137 | "print counts with custom separator"), | |
1138 | OPT_CALLBACK('G', "cgroup", &evsel_list, "name", | |
1139 | "monitor event in cgroup name only", parse_cgroups), | |
1140 | OPT_STRING('o', "output", &output_name, "file", "output file name"), | |
1141 | OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), | |
1142 | OPT_INTEGER(0, "log-fd", &output_fd, | |
1143 | "log output to fd, instead of stderr"), | |
1f16c575 PZ |
1144 | OPT_STRING(0, "pre", &pre_cmd, "command", |
1145 | "command to run prior to the measured command"), | |
1146 | OPT_STRING(0, "post", &post_cmd, "command", | |
1147 | "command to run after to the measured command"), | |
b070a547 ACM |
1148 | OPT_END() |
1149 | }; | |
1150 | const char * const stat_usage[] = { | |
1151 | "perf stat [<options>] [<command>]", | |
1152 | NULL | |
1153 | }; | |
69aad6f1 | 1154 | struct perf_evsel *pos; |
b070a547 | 1155 | int status = -ENOMEM, run_idx; |
4aa9015f | 1156 | const char *mode; |
42202dd5 | 1157 | |
5af52b51 SE |
1158 | setlocale(LC_ALL, ""); |
1159 | ||
7e2ed097 | 1160 | evsel_list = perf_evlist__new(NULL, NULL); |
361c99a6 ACM |
1161 | if (evsel_list == NULL) |
1162 | return -ENOMEM; | |
1163 | ||
a0541234 AB |
1164 | argc = parse_options(argc, argv, options, stat_usage, |
1165 | PARSE_OPT_STOP_AT_NON_OPTION); | |
d7470b6a | 1166 | |
4aa9015f SE |
1167 | output = stderr; |
1168 | if (output_name && strcmp(output_name, "-")) | |
1169 | output = NULL; | |
1170 | ||
56f3bae7 JC |
1171 | if (output_name && output_fd) { |
1172 | fprintf(stderr, "cannot use both --output and --log-fd\n"); | |
1173 | usage_with_options(stat_usage, options); | |
1174 | } | |
fc3e4d07 SE |
1175 | |
1176 | if (output_fd < 0) { | |
1177 | fprintf(stderr, "argument to --log-fd must be a > 0\n"); | |
1178 | usage_with_options(stat_usage, options); | |
1179 | } | |
1180 | ||
4aa9015f SE |
1181 | if (!output) { |
1182 | struct timespec tm; | |
1183 | mode = append_file ? "a" : "w"; | |
1184 | ||
1185 | output = fopen(output_name, mode); | |
1186 | if (!output) { | |
1187 | perror("failed to create output file"); | |
fceda7fe | 1188 | return -1; |
4aa9015f SE |
1189 | } |
1190 | clock_gettime(CLOCK_REALTIME, &tm); | |
1191 | fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); | |
fc3e4d07 | 1192 | } else if (output_fd > 0) { |
56f3bae7 JC |
1193 | mode = append_file ? "a" : "w"; |
1194 | output = fdopen(output_fd, mode); | |
1195 | if (!output) { | |
1196 | perror("Failed opening logfd"); | |
1197 | return -errno; | |
1198 | } | |
4aa9015f SE |
1199 | } |
1200 | ||
d4ffd04d | 1201 | if (csv_sep) { |
d7470b6a | 1202 | csv_output = true; |
d4ffd04d JC |
1203 | if (!strcmp(csv_sep, "\\t")) |
1204 | csv_sep = "\t"; | |
1205 | } else | |
d7470b6a SE |
1206 | csv_sep = DEFAULT_SEPARATOR; |
1207 | ||
1208 | /* | |
1209 | * let the spreadsheet do the pretty-printing | |
1210 | */ | |
1211 | if (csv_output) { | |
61a9f324 | 1212 | /* User explicitly passed -B? */ |
d7470b6a SE |
1213 | if (big_num_opt == 1) { |
1214 | fprintf(stderr, "-B option not supported with -x\n"); | |
1215 | usage_with_options(stat_usage, options); | |
1216 | } else /* Nope, so disable big number formatting */ | |
1217 | big_num = false; | |
1218 | } else if (big_num_opt == 0) /* User passed --no-big-num */ | |
1219 | big_num = false; | |
1220 | ||
aa22dd49 | 1221 | if (!argc && !perf_target__has_task(&target)) |
5242519b | 1222 | usage_with_options(stat_usage, options); |
9e9772c4 | 1223 | if (run_count <= 0) |
42202dd5 | 1224 | usage_with_options(stat_usage, options); |
ddcacfa0 | 1225 | |
023695d9 | 1226 | /* no_aggr, cgroup are for system-wide only */ |
aa22dd49 | 1227 | if ((no_aggr || nr_cgroups) && !perf_target__has_cpu(&target)) { |
023695d9 SE |
1228 | fprintf(stderr, "both cgroup and no-aggregation " |
1229 | "modes only available in system-wide mode\n"); | |
1230 | ||
f5b4a9c3 | 1231 | usage_with_options(stat_usage, options); |
023695d9 | 1232 | } |
f5b4a9c3 | 1233 | |
2cba3ffb IM |
1234 | if (add_default_attributes()) |
1235 | goto out; | |
ddcacfa0 | 1236 | |
4bd0f2d2 | 1237 | perf_target__validate(&target); |
5c98d466 | 1238 | |
77a6f014 | 1239 | if (perf_evlist__create_maps(evsel_list, &target) < 0) { |
aa22dd49 | 1240 | if (perf_target__has_task(&target)) |
77a6f014 | 1241 | pr_err("Problems finding threads of monitor\n"); |
aa22dd49 | 1242 | if (perf_target__has_cpu(&target)) |
77a6f014 | 1243 | perror("failed to parse CPUs map"); |
ddcacfa0 | 1244 | |
c45c6ea2 | 1245 | usage_with_options(stat_usage, options); |
60d567e2 ACM |
1246 | return -1; |
1247 | } | |
c45c6ea2 | 1248 | |
361c99a6 | 1249 | list_for_each_entry(pos, &evsel_list->entries, node) { |
c52b12ed | 1250 | if (perf_evsel__alloc_stat_priv(pos) < 0 || |
7ae92e74 | 1251 | perf_evsel__alloc_counts(pos, perf_evsel__nr_cpus(pos)) < 0) |
69aad6f1 | 1252 | goto out_free_fd; |
d6d901c2 ZY |
1253 | } |
1254 | ||
58d7e993 IM |
1255 | /* |
1256 | * We dont want to block the signals - that would cause | |
1257 | * child tasks to inherit that and Ctrl-C would not work. | |
1258 | * What we want is for Ctrl-C to work in the exec()-ed | |
1259 | * task, but being ignored by perf stat itself: | |
1260 | */ | |
f7b7c26e | 1261 | atexit(sig_atexit); |
58d7e993 IM |
1262 | signal(SIGINT, skip_signal); |
1263 | signal(SIGALRM, skip_signal); | |
1264 | signal(SIGABRT, skip_signal); | |
1265 | ||
42202dd5 IM |
1266 | status = 0; |
1267 | for (run_idx = 0; run_idx < run_count; run_idx++) { | |
1268 | if (run_count != 1 && verbose) | |
4aa9015f SE |
1269 | fprintf(output, "[ perf stat: executing run #%d ... ]\n", |
1270 | run_idx + 1); | |
f9cef0a9 | 1271 | |
42202dd5 IM |
1272 | status = run_perf_stat(argc, argv); |
1273 | } | |
1274 | ||
084ab9f8 ACM |
1275 | if (status != -1) |
1276 | print_stat(argc, argv); | |
69aad6f1 | 1277 | out_free_fd: |
361c99a6 | 1278 | list_for_each_entry(pos, &evsel_list->entries, node) |
69aad6f1 | 1279 | perf_evsel__free_stat_priv(pos); |
7e2ed097 | 1280 | perf_evlist__delete_maps(evsel_list); |
0015e2e1 ACM |
1281 | out: |
1282 | perf_evlist__delete(evsel_list); | |
42202dd5 | 1283 | return status; |
ddcacfa0 | 1284 | } |