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 | |
ddcacfa0 | 60 | #include <sys/prctl.h> |
5af52b51 | 61 | #include <locale.h> |
16c8a109 | 62 | |
d7470b6a | 63 | #define DEFAULT_SEPARATOR " " |
2cee77c4 DA |
64 | #define CNTR_NOT_SUPPORTED "<not supported>" |
65 | #define CNTR_NOT_COUNTED "<not counted>" | |
d7470b6a | 66 | |
666e6d48 | 67 | static struct perf_evlist *evsel_list; |
361c99a6 | 68 | |
77a6f014 NK |
69 | static struct perf_target target = { |
70 | .uid = UINT_MAX, | |
71 | }; | |
ddcacfa0 | 72 | |
3d632595 | 73 | static int run_count = 1; |
2e6cdf99 | 74 | static bool no_inherit = false; |
c0555642 | 75 | static bool scale = true; |
f5b4a9c3 | 76 | static bool no_aggr = false; |
933da83a | 77 | static pid_t child_pid = -1; |
c0555642 | 78 | static bool null_run = false; |
2cba3ffb | 79 | static int detailed_run = 0; |
201e0b06 | 80 | static bool big_num = true; |
d7470b6a | 81 | static int big_num_opt = -1; |
d7470b6a SE |
82 | static const char *csv_sep = NULL; |
83 | static bool csv_output = false; | |
43bece79 | 84 | static bool group = false; |
4aa9015f | 85 | static FILE *output = NULL; |
5af52b51 | 86 | |
60666c63 LW |
87 | static volatile int done = 0; |
88 | ||
69aad6f1 ACM |
89 | struct perf_stat { |
90 | struct stats res_stats[3]; | |
69aad6f1 ACM |
91 | }; |
92 | ||
c52b12ed | 93 | static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel) |
69aad6f1 | 94 | { |
c52b12ed | 95 | evsel->priv = zalloc(sizeof(struct perf_stat)); |
69aad6f1 ACM |
96 | return evsel->priv == NULL ? -ENOMEM : 0; |
97 | } | |
98 | ||
99 | static void perf_evsel__free_stat_priv(struct perf_evsel *evsel) | |
100 | { | |
101 | free(evsel->priv); | |
102 | evsel->priv = NULL; | |
103 | } | |
104 | ||
7ae92e74 YZ |
105 | static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel) |
106 | { | |
107 | return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus; | |
108 | } | |
109 | ||
110 | static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel) | |
111 | { | |
112 | return perf_evsel__cpus(evsel)->nr; | |
113 | } | |
114 | ||
666e6d48 RR |
115 | static struct stats runtime_nsecs_stats[MAX_NR_CPUS]; |
116 | static struct stats runtime_cycles_stats[MAX_NR_CPUS]; | |
117 | static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS]; | |
118 | static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS]; | |
119 | static struct stats runtime_branches_stats[MAX_NR_CPUS]; | |
120 | static struct stats runtime_cacherefs_stats[MAX_NR_CPUS]; | |
121 | static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS]; | |
122 | static struct stats runtime_l1_icache_stats[MAX_NR_CPUS]; | |
123 | static struct stats runtime_ll_cache_stats[MAX_NR_CPUS]; | |
124 | static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS]; | |
125 | static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS]; | |
126 | static struct stats walltime_nsecs_stats; | |
be1ac0d8 | 127 | |
727ab04e ACM |
128 | static int create_perf_stat_counter(struct perf_evsel *evsel, |
129 | struct perf_evsel *first) | |
ddcacfa0 | 130 | { |
69aad6f1 | 131 | struct perf_event_attr *attr = &evsel->attr; |
5622c07b SE |
132 | bool exclude_guest_missing = false; |
133 | int ret; | |
727ab04e | 134 | |
ddcacfa0 | 135 | if (scale) |
a21ca2ca IM |
136 | attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | |
137 | PERF_FORMAT_TOTAL_TIME_RUNNING; | |
ddcacfa0 | 138 | |
5d2cd909 ACM |
139 | attr->inherit = !no_inherit; |
140 | ||
5622c07b SE |
141 | retry: |
142 | if (exclude_guest_missing) | |
143 | evsel->attr.exclude_guest = evsel->attr.exclude_host = 0; | |
144 | ||
16ee6576 | 145 | if (perf_target__has_cpu(&target)) { |
7ae92e74 | 146 | ret = perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel)); |
5622c07b SE |
147 | if (ret) |
148 | goto check_ret; | |
149 | return 0; | |
150 | } | |
151 | ||
aa22dd49 | 152 | if (!perf_target__has_task(&target) && (!group || evsel == first)) { |
48290609 ACM |
153 | attr->disabled = 1; |
154 | attr->enable_on_exec = 1; | |
ddcacfa0 | 155 | } |
084ab9f8 | 156 | |
6a4bb04c | 157 | ret = perf_evsel__open_per_thread(evsel, evsel_list->threads); |
5622c07b SE |
158 | if (!ret) |
159 | return 0; | |
160 | /* fall through */ | |
161 | check_ret: | |
162 | if (ret && errno == EINVAL) { | |
163 | if (!exclude_guest_missing && | |
164 | (evsel->attr.exclude_guest || evsel->attr.exclude_host)) { | |
165 | pr_debug("Old kernel, cannot exclude " | |
166 | "guest or host samples.\n"); | |
167 | exclude_guest_missing = true; | |
168 | goto retry; | |
169 | } | |
170 | } | |
171 | return ret; | |
ddcacfa0 IM |
172 | } |
173 | ||
c04f5e5d IM |
174 | /* |
175 | * Does the counter have nsecs as a unit? | |
176 | */ | |
daec78a0 | 177 | static inline int nsec_counter(struct perf_evsel *evsel) |
c04f5e5d | 178 | { |
daec78a0 ACM |
179 | if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || |
180 | perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) | |
c04f5e5d IM |
181 | return 1; |
182 | ||
183 | return 0; | |
184 | } | |
185 | ||
dcd9936a IM |
186 | /* |
187 | * Update various tracking values we maintain to print | |
188 | * more semantic information such as miss/hit ratios, | |
189 | * instruction rates, etc: | |
190 | */ | |
191 | static void update_shadow_stats(struct perf_evsel *counter, u64 *count) | |
192 | { | |
193 | if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK)) | |
194 | update_stats(&runtime_nsecs_stats[0], count[0]); | |
195 | else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES)) | |
196 | update_stats(&runtime_cycles_stats[0], count[0]); | |
d3d1e86d IM |
197 | else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) |
198 | update_stats(&runtime_stalled_cycles_front_stats[0], count[0]); | |
129c04cb | 199 | else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND)) |
d3d1e86d | 200 | update_stats(&runtime_stalled_cycles_back_stats[0], count[0]); |
dcd9936a IM |
201 | else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS)) |
202 | update_stats(&runtime_branches_stats[0], count[0]); | |
203 | else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES)) | |
204 | update_stats(&runtime_cacherefs_stats[0], count[0]); | |
8bb6c79f IM |
205 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D)) |
206 | update_stats(&runtime_l1_dcache_stats[0], count[0]); | |
c3305257 IM |
207 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I)) |
208 | update_stats(&runtime_l1_icache_stats[0], count[0]); | |
209 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL)) | |
210 | update_stats(&runtime_ll_cache_stats[0], count[0]); | |
211 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB)) | |
212 | update_stats(&runtime_dtlb_cache_stats[0], count[0]); | |
213 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB)) | |
214 | update_stats(&runtime_itlb_cache_stats[0], count[0]); | |
dcd9936a IM |
215 | } |
216 | ||
c04f5e5d | 217 | /* |
2996f5dd | 218 | * Read out the results of a single counter: |
f5b4a9c3 | 219 | * aggregate counts across CPUs in system-wide mode |
c04f5e5d | 220 | */ |
c52b12ed | 221 | static int read_counter_aggr(struct perf_evsel *counter) |
c04f5e5d | 222 | { |
69aad6f1 | 223 | struct perf_stat *ps = counter->priv; |
c52b12ed ACM |
224 | u64 *count = counter->counts->aggr.values; |
225 | int i; | |
2996f5dd | 226 | |
7ae92e74 | 227 | if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter), |
7e2ed097 | 228 | evsel_list->threads->nr, scale) < 0) |
c52b12ed | 229 | return -1; |
9e9772c4 PZ |
230 | |
231 | for (i = 0; i < 3; i++) | |
69aad6f1 | 232 | update_stats(&ps->res_stats[i], count[i]); |
9e9772c4 PZ |
233 | |
234 | if (verbose) { | |
4aa9015f | 235 | fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", |
7289f83c | 236 | perf_evsel__name(counter), count[0], count[1], count[2]); |
9e9772c4 PZ |
237 | } |
238 | ||
be1ac0d8 IM |
239 | /* |
240 | * Save the full runtime - to allow normalization during printout: | |
241 | */ | |
dcd9936a | 242 | update_shadow_stats(counter, count); |
c52b12ed ACM |
243 | |
244 | return 0; | |
f5b4a9c3 SE |
245 | } |
246 | ||
247 | /* | |
248 | * Read out the results of a single counter: | |
249 | * do not aggregate counts across CPUs in system-wide mode | |
250 | */ | |
c52b12ed | 251 | static int read_counter(struct perf_evsel *counter) |
f5b4a9c3 | 252 | { |
c52b12ed | 253 | u64 *count; |
f5b4a9c3 | 254 | int cpu; |
f5b4a9c3 | 255 | |
7ae92e74 | 256 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { |
c52b12ed ACM |
257 | if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0) |
258 | return -1; | |
f5b4a9c3 | 259 | |
c52b12ed | 260 | count = counter->counts->cpu[cpu].values; |
f5b4a9c3 | 261 | |
dcd9936a | 262 | update_shadow_stats(counter, count); |
f5b4a9c3 | 263 | } |
c52b12ed ACM |
264 | |
265 | return 0; | |
2996f5dd IM |
266 | } |
267 | ||
1d037ca1 | 268 | static int run_perf_stat(int argc __maybe_unused, const char **argv) |
42202dd5 IM |
269 | { |
270 | unsigned long long t0, t1; | |
727ab04e | 271 | struct perf_evsel *counter, *first; |
42202dd5 | 272 | int status = 0; |
051ae7f7 | 273 | int child_ready_pipe[2], go_pipe[2]; |
6be2850e | 274 | const bool forks = (argc > 0); |
051ae7f7 | 275 | char buf; |
42202dd5 | 276 | |
60666c63 | 277 | if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) { |
051ae7f7 | 278 | perror("failed to create pipes"); |
fceda7fe | 279 | return -1; |
051ae7f7 PM |
280 | } |
281 | ||
60666c63 | 282 | if (forks) { |
6be2850e | 283 | if ((child_pid = fork()) < 0) |
60666c63 LW |
284 | perror("failed to fork"); |
285 | ||
6be2850e | 286 | if (!child_pid) { |
60666c63 LW |
287 | close(child_ready_pipe[0]); |
288 | close(go_pipe[1]); | |
289 | fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); | |
290 | ||
291 | /* | |
292 | * Do a dummy execvp to get the PLT entry resolved, | |
293 | * so we avoid the resolver overhead on the real | |
294 | * execvp call. | |
295 | */ | |
296 | execvp("", (char **)argv); | |
297 | ||
298 | /* | |
299 | * Tell the parent we're ready to go | |
300 | */ | |
301 | close(child_ready_pipe[1]); | |
302 | ||
303 | /* | |
304 | * Wait until the parent tells us to go. | |
305 | */ | |
306 | if (read(go_pipe[0], &buf, 1) == -1) | |
307 | perror("unable to read pipe"); | |
308 | ||
309 | execvp(argv[0], (char **)argv); | |
310 | ||
311 | perror(argv[0]); | |
312 | exit(-1); | |
313 | } | |
051ae7f7 | 314 | |
d67356e7 | 315 | if (perf_target__none(&target)) |
7e2ed097 | 316 | evsel_list->threads->map[0] = child_pid; |
d6d901c2 | 317 | |
051ae7f7 | 318 | /* |
60666c63 | 319 | * Wait for the child to be ready to exec. |
051ae7f7 PM |
320 | */ |
321 | close(child_ready_pipe[1]); | |
60666c63 LW |
322 | close(go_pipe[0]); |
323 | if (read(child_ready_pipe[0], &buf, 1) == -1) | |
a92bef0f | 324 | perror("unable to read pipe"); |
60666c63 | 325 | close(child_ready_pipe[0]); |
051ae7f7 PM |
326 | } |
327 | ||
6a4bb04c | 328 | if (group) |
63dab225 | 329 | perf_evlist__set_leader(evsel_list); |
6a4bb04c | 330 | |
0c21f736 | 331 | first = perf_evlist__first(evsel_list); |
727ab04e | 332 | |
361c99a6 | 333 | list_for_each_entry(counter, &evsel_list->entries, node) { |
727ab04e | 334 | if (create_perf_stat_counter(counter, first) < 0) { |
979987a5 DA |
335 | /* |
336 | * PPC returns ENXIO for HW counters until 2.6.37 | |
337 | * (behavior changed with commit b0a873e). | |
338 | */ | |
38f6ae1e | 339 | if (errno == EINVAL || errno == ENOSYS || |
979987a5 DA |
340 | errno == ENOENT || errno == EOPNOTSUPP || |
341 | errno == ENXIO) { | |
c63ca0c0 DA |
342 | if (verbose) |
343 | ui__warning("%s event is not supported by the kernel.\n", | |
7289f83c | 344 | perf_evsel__name(counter)); |
2cee77c4 | 345 | counter->supported = false; |
ede70290 | 346 | continue; |
c63ca0c0 | 347 | } |
ede70290 IM |
348 | |
349 | if (errno == EPERM || errno == EACCES) { | |
48290609 ACM |
350 | error("You may not have permission to collect %sstats.\n" |
351 | "\t Consider tweaking" | |
352 | " /proc/sys/kernel/perf_event_paranoid or running as root.", | |
20f946b4 | 353 | target.system_wide ? "system-wide " : ""); |
48290609 ACM |
354 | } else { |
355 | error("open_counter returned with %d (%s). " | |
356 | "/bin/dmesg may provide additional information.\n", | |
357 | errno, strerror(errno)); | |
358 | } | |
359 | if (child_pid != -1) | |
360 | kill(child_pid, SIGTERM); | |
fceda7fe DA |
361 | |
362 | pr_err("Not all events could be opened.\n"); | |
48290609 ACM |
363 | return -1; |
364 | } | |
2cee77c4 | 365 | counter->supported = true; |
084ab9f8 | 366 | } |
42202dd5 | 367 | |
1491a632 | 368 | if (perf_evlist__apply_filters(evsel_list)) { |
cfd748ae FW |
369 | error("failed to set filter with %d (%s)\n", errno, |
370 | strerror(errno)); | |
371 | return -1; | |
372 | } | |
373 | ||
42202dd5 IM |
374 | /* |
375 | * Enable counters and exec the command: | |
376 | */ | |
377 | t0 = rdclock(); | |
42202dd5 | 378 | |
60666c63 LW |
379 | if (forks) { |
380 | close(go_pipe[1]); | |
381 | wait(&status); | |
33e49ea7 AK |
382 | if (WIFSIGNALED(status)) |
383 | psignal(WTERMSIG(status), argv[0]); | |
60666c63 | 384 | } else { |
6be2850e | 385 | while(!done) sleep(1); |
60666c63 | 386 | } |
42202dd5 | 387 | |
42202dd5 IM |
388 | t1 = rdclock(); |
389 | ||
9e9772c4 | 390 | update_stats(&walltime_nsecs_stats, t1 - t0); |
42202dd5 | 391 | |
f5b4a9c3 | 392 | if (no_aggr) { |
361c99a6 | 393 | list_for_each_entry(counter, &evsel_list->entries, node) { |
f5b4a9c3 | 394 | read_counter(counter); |
7ae92e74 | 395 | perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1); |
c52b12ed | 396 | } |
f5b4a9c3 | 397 | } else { |
361c99a6 | 398 | list_for_each_entry(counter, &evsel_list->entries, node) { |
f5b4a9c3 | 399 | read_counter_aggr(counter); |
7ae92e74 | 400 | perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), |
7e2ed097 | 401 | evsel_list->threads->nr); |
c52b12ed | 402 | } |
f5b4a9c3 | 403 | } |
c52b12ed | 404 | |
42202dd5 IM |
405 | return WEXITSTATUS(status); |
406 | } | |
407 | ||
f99844cb IM |
408 | static void print_noise_pct(double total, double avg) |
409 | { | |
0007ecea | 410 | double pct = rel_stddev_stats(total, avg); |
f99844cb | 411 | |
3ae9a34d | 412 | if (csv_output) |
4aa9015f | 413 | fprintf(output, "%s%.2f%%", csv_sep, pct); |
a1bca6cc | 414 | else if (pct) |
4aa9015f | 415 | fprintf(output, " ( +-%6.2f%% )", pct); |
f99844cb IM |
416 | } |
417 | ||
69aad6f1 | 418 | static void print_noise(struct perf_evsel *evsel, double avg) |
42202dd5 | 419 | { |
69aad6f1 ACM |
420 | struct perf_stat *ps; |
421 | ||
849abde9 PZ |
422 | if (run_count == 1) |
423 | return; | |
424 | ||
69aad6f1 | 425 | ps = evsel->priv; |
f99844cb | 426 | print_noise_pct(stddev_stats(&ps->res_stats[0]), avg); |
42202dd5 IM |
427 | } |
428 | ||
daec78a0 | 429 | static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg) |
44175b6f | 430 | { |
506d4bc8 | 431 | double msecs = avg / 1e6; |
d7470b6a | 432 | char cpustr[16] = { '\0', }; |
2cba3ffb | 433 | const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s"; |
44175b6f | 434 | |
f5b4a9c3 | 435 | if (no_aggr) |
d7470b6a SE |
436 | sprintf(cpustr, "CPU%*d%s", |
437 | csv_output ? 0 : -4, | |
7ae92e74 | 438 | perf_evsel__cpus(evsel)->map[cpu], csv_sep); |
d7470b6a | 439 | |
7289f83c | 440 | fprintf(output, fmt, cpustr, msecs, csv_sep, perf_evsel__name(evsel)); |
d7470b6a | 441 | |
023695d9 | 442 | if (evsel->cgrp) |
4aa9015f | 443 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); |
023695d9 | 444 | |
d7470b6a SE |
445 | if (csv_output) |
446 | return; | |
44175b6f | 447 | |
daec78a0 | 448 | if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) |
4aa9015f SE |
449 | fprintf(output, " # %8.3f CPUs utilized ", |
450 | avg / avg_stats(&walltime_nsecs_stats)); | |
9dac6a29 NK |
451 | else |
452 | fprintf(output, " "); | |
44175b6f IM |
453 | } |
454 | ||
15e6392f NK |
455 | /* used for get_ratio_color() */ |
456 | enum grc_type { | |
457 | GRC_STALLED_CYCLES_FE, | |
458 | GRC_STALLED_CYCLES_BE, | |
459 | GRC_CACHE_MISSES, | |
460 | GRC_MAX_NR | |
461 | }; | |
462 | ||
463 | static const char *get_ratio_color(enum grc_type type, double ratio) | |
464 | { | |
465 | static const double grc_table[GRC_MAX_NR][3] = { | |
466 | [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 }, | |
467 | [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 }, | |
468 | [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 }, | |
469 | }; | |
470 | const char *color = PERF_COLOR_NORMAL; | |
471 | ||
472 | if (ratio > grc_table[type][0]) | |
473 | color = PERF_COLOR_RED; | |
474 | else if (ratio > grc_table[type][1]) | |
475 | color = PERF_COLOR_MAGENTA; | |
476 | else if (ratio > grc_table[type][2]) | |
477 | color = PERF_COLOR_YELLOW; | |
478 | ||
479 | return color; | |
480 | } | |
481 | ||
1d037ca1 IT |
482 | static void print_stalled_cycles_frontend(int cpu, |
483 | struct perf_evsel *evsel | |
484 | __maybe_unused, double avg) | |
d3d1e86d IM |
485 | { |
486 | double total, ratio = 0.0; | |
487 | const char *color; | |
488 | ||
489 | total = avg_stats(&runtime_cycles_stats[cpu]); | |
490 | ||
491 | if (total) | |
492 | ratio = avg / total * 100.0; | |
493 | ||
15e6392f | 494 | color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio); |
d3d1e86d | 495 | |
4aa9015f SE |
496 | fprintf(output, " # "); |
497 | color_fprintf(output, color, "%6.2f%%", ratio); | |
498 | fprintf(output, " frontend cycles idle "); | |
d3d1e86d IM |
499 | } |
500 | ||
1d037ca1 IT |
501 | static void print_stalled_cycles_backend(int cpu, |
502 | struct perf_evsel *evsel | |
503 | __maybe_unused, double avg) | |
a5d243d0 IM |
504 | { |
505 | double total, ratio = 0.0; | |
506 | const char *color; | |
507 | ||
508 | total = avg_stats(&runtime_cycles_stats[cpu]); | |
509 | ||
510 | if (total) | |
511 | ratio = avg / total * 100.0; | |
512 | ||
15e6392f | 513 | color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio); |
a5d243d0 | 514 | |
4aa9015f SE |
515 | fprintf(output, " # "); |
516 | color_fprintf(output, color, "%6.2f%%", ratio); | |
517 | fprintf(output, " backend cycles idle "); | |
a5d243d0 IM |
518 | } |
519 | ||
1d037ca1 IT |
520 | static void print_branch_misses(int cpu, |
521 | struct perf_evsel *evsel __maybe_unused, | |
522 | double avg) | |
c78df6c1 IM |
523 | { |
524 | double total, ratio = 0.0; | |
525 | const char *color; | |
526 | ||
527 | total = avg_stats(&runtime_branches_stats[cpu]); | |
528 | ||
529 | if (total) | |
530 | ratio = avg / total * 100.0; | |
531 | ||
15e6392f | 532 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c78df6c1 | 533 | |
4aa9015f SE |
534 | fprintf(output, " # "); |
535 | color_fprintf(output, color, "%6.2f%%", ratio); | |
536 | fprintf(output, " of all branches "); | |
c78df6c1 IM |
537 | } |
538 | ||
1d037ca1 IT |
539 | static void print_l1_dcache_misses(int cpu, |
540 | struct perf_evsel *evsel __maybe_unused, | |
541 | double avg) | |
8bb6c79f IM |
542 | { |
543 | double total, ratio = 0.0; | |
544 | const char *color; | |
545 | ||
546 | total = avg_stats(&runtime_l1_dcache_stats[cpu]); | |
547 | ||
548 | if (total) | |
549 | ratio = avg / total * 100.0; | |
550 | ||
15e6392f | 551 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
8bb6c79f | 552 | |
4aa9015f SE |
553 | fprintf(output, " # "); |
554 | color_fprintf(output, color, "%6.2f%%", ratio); | |
555 | fprintf(output, " of all L1-dcache hits "); | |
8bb6c79f IM |
556 | } |
557 | ||
1d037ca1 IT |
558 | static void print_l1_icache_misses(int cpu, |
559 | struct perf_evsel *evsel __maybe_unused, | |
560 | double avg) | |
c3305257 IM |
561 | { |
562 | double total, ratio = 0.0; | |
563 | const char *color; | |
564 | ||
565 | total = avg_stats(&runtime_l1_icache_stats[cpu]); | |
566 | ||
567 | if (total) | |
568 | ratio = avg / total * 100.0; | |
569 | ||
15e6392f | 570 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 571 | |
4aa9015f SE |
572 | fprintf(output, " # "); |
573 | color_fprintf(output, color, "%6.2f%%", ratio); | |
574 | fprintf(output, " of all L1-icache hits "); | |
c3305257 IM |
575 | } |
576 | ||
1d037ca1 IT |
577 | static void print_dtlb_cache_misses(int cpu, |
578 | struct perf_evsel *evsel __maybe_unused, | |
579 | double avg) | |
c3305257 IM |
580 | { |
581 | double total, ratio = 0.0; | |
582 | const char *color; | |
583 | ||
584 | total = avg_stats(&runtime_dtlb_cache_stats[cpu]); | |
585 | ||
586 | if (total) | |
587 | ratio = avg / total * 100.0; | |
588 | ||
15e6392f | 589 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 590 | |
4aa9015f SE |
591 | fprintf(output, " # "); |
592 | color_fprintf(output, color, "%6.2f%%", ratio); | |
593 | fprintf(output, " of all dTLB cache hits "); | |
c3305257 IM |
594 | } |
595 | ||
1d037ca1 IT |
596 | static void print_itlb_cache_misses(int cpu, |
597 | struct perf_evsel *evsel __maybe_unused, | |
598 | double avg) | |
c3305257 IM |
599 | { |
600 | double total, ratio = 0.0; | |
601 | const char *color; | |
602 | ||
603 | total = avg_stats(&runtime_itlb_cache_stats[cpu]); | |
604 | ||
605 | if (total) | |
606 | ratio = avg / total * 100.0; | |
607 | ||
15e6392f | 608 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 609 | |
4aa9015f SE |
610 | fprintf(output, " # "); |
611 | color_fprintf(output, color, "%6.2f%%", ratio); | |
612 | fprintf(output, " of all iTLB cache hits "); | |
c3305257 IM |
613 | } |
614 | ||
1d037ca1 IT |
615 | static void print_ll_cache_misses(int cpu, |
616 | struct perf_evsel *evsel __maybe_unused, | |
617 | double avg) | |
c3305257 IM |
618 | { |
619 | double total, ratio = 0.0; | |
620 | const char *color; | |
621 | ||
622 | total = avg_stats(&runtime_ll_cache_stats[cpu]); | |
623 | ||
624 | if (total) | |
625 | ratio = avg / total * 100.0; | |
626 | ||
15e6392f | 627 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 628 | |
4aa9015f SE |
629 | fprintf(output, " # "); |
630 | color_fprintf(output, color, "%6.2f%%", ratio); | |
631 | fprintf(output, " of all LL-cache hits "); | |
c3305257 IM |
632 | } |
633 | ||
daec78a0 | 634 | static void abs_printout(int cpu, struct perf_evsel *evsel, double avg) |
44175b6f | 635 | { |
c7f7fea3 | 636 | double total, ratio = 0.0; |
f5b4a9c3 | 637 | char cpustr[16] = { '\0', }; |
d7470b6a SE |
638 | const char *fmt; |
639 | ||
640 | if (csv_output) | |
641 | fmt = "%s%.0f%s%s"; | |
642 | else if (big_num) | |
2cba3ffb | 643 | fmt = "%s%'18.0f%s%-25s"; |
d7470b6a | 644 | else |
2cba3ffb | 645 | fmt = "%s%18.0f%s%-25s"; |
f5b4a9c3 SE |
646 | |
647 | if (no_aggr) | |
d7470b6a SE |
648 | sprintf(cpustr, "CPU%*d%s", |
649 | csv_output ? 0 : -4, | |
7ae92e74 | 650 | perf_evsel__cpus(evsel)->map[cpu], csv_sep); |
f5b4a9c3 SE |
651 | else |
652 | cpu = 0; | |
c7f7fea3 | 653 | |
7289f83c | 654 | fprintf(output, fmt, cpustr, avg, csv_sep, perf_evsel__name(evsel)); |
d7470b6a | 655 | |
023695d9 | 656 | if (evsel->cgrp) |
4aa9015f | 657 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); |
023695d9 | 658 | |
d7470b6a SE |
659 | if (csv_output) |
660 | return; | |
44175b6f | 661 | |
daec78a0 | 662 | if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) { |
f5b4a9c3 | 663 | total = avg_stats(&runtime_cycles_stats[cpu]); |
c7f7fea3 IM |
664 | |
665 | if (total) | |
666 | ratio = avg / total; | |
667 | ||
4aa9015f | 668 | fprintf(output, " # %5.2f insns per cycle ", ratio); |
481f988a | 669 | |
d3d1e86d IM |
670 | total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]); |
671 | total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu])); | |
481f988a IM |
672 | |
673 | if (total && avg) { | |
674 | ratio = total / avg; | |
4aa9015f | 675 | fprintf(output, "\n # %5.2f stalled cycles per insn", ratio); |
481f988a IM |
676 | } |
677 | ||
daec78a0 | 678 | } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) && |
f5b4a9c3 | 679 | runtime_branches_stats[cpu].n != 0) { |
c78df6c1 | 680 | print_branch_misses(cpu, evsel, avg); |
8bb6c79f IM |
681 | } else if ( |
682 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
683 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D | | |
684 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
685 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
c6264def | 686 | runtime_l1_dcache_stats[cpu].n != 0) { |
8bb6c79f | 687 | print_l1_dcache_misses(cpu, evsel, avg); |
c3305257 IM |
688 | } else if ( |
689 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
690 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I | | |
691 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
692 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
693 | runtime_l1_icache_stats[cpu].n != 0) { | |
694 | print_l1_icache_misses(cpu, evsel, avg); | |
695 | } else if ( | |
696 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
697 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB | | |
698 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
699 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
700 | runtime_dtlb_cache_stats[cpu].n != 0) { | |
701 | print_dtlb_cache_misses(cpu, evsel, avg); | |
702 | } else if ( | |
703 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
704 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB | | |
705 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
706 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
707 | runtime_itlb_cache_stats[cpu].n != 0) { | |
708 | print_itlb_cache_misses(cpu, evsel, avg); | |
709 | } else if ( | |
710 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
711 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL | | |
712 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
713 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
714 | runtime_ll_cache_stats[cpu].n != 0) { | |
715 | print_ll_cache_misses(cpu, evsel, avg); | |
d58f4c82 IM |
716 | } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) && |
717 | runtime_cacherefs_stats[cpu].n != 0) { | |
718 | total = avg_stats(&runtime_cacherefs_stats[cpu]); | |
719 | ||
720 | if (total) | |
721 | ratio = avg * 100 / total; | |
722 | ||
4aa9015f | 723 | fprintf(output, " # %8.3f %% of all cache refs ", ratio); |
d58f4c82 | 724 | |
d3d1e86d IM |
725 | } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) { |
726 | print_stalled_cycles_frontend(cpu, evsel, avg); | |
129c04cb | 727 | } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) { |
d3d1e86d | 728 | print_stalled_cycles_backend(cpu, evsel, avg); |
481f988a | 729 | } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) { |
f5b4a9c3 | 730 | total = avg_stats(&runtime_nsecs_stats[cpu]); |
c7f7fea3 IM |
731 | |
732 | if (total) | |
481f988a | 733 | ratio = 1.0 * avg / total; |
c7f7fea3 | 734 | |
4aa9015f | 735 | fprintf(output, " # %8.3f GHz ", ratio); |
481f988a | 736 | } else if (runtime_nsecs_stats[cpu].n != 0) { |
5fde2523 NK |
737 | char unit = 'M'; |
738 | ||
481f988a | 739 | total = avg_stats(&runtime_nsecs_stats[cpu]); |
11ba2b85 IM |
740 | |
741 | if (total) | |
481f988a | 742 | ratio = 1000.0 * avg / total; |
5fde2523 NK |
743 | if (ratio < 0.001) { |
744 | ratio *= 1000; | |
745 | unit = 'K'; | |
746 | } | |
11ba2b85 | 747 | |
5fde2523 | 748 | fprintf(output, " # %8.3f %c/sec ", ratio, unit); |
a5d243d0 | 749 | } else { |
4aa9015f | 750 | fprintf(output, " "); |
44175b6f | 751 | } |
44175b6f IM |
752 | } |
753 | ||
2996f5dd IM |
754 | /* |
755 | * Print out the results of a single counter: | |
f5b4a9c3 | 756 | * aggregated counts in system-wide mode |
2996f5dd | 757 | */ |
69aad6f1 | 758 | static void print_counter_aggr(struct perf_evsel *counter) |
2996f5dd | 759 | { |
69aad6f1 ACM |
760 | struct perf_stat *ps = counter->priv; |
761 | double avg = avg_stats(&ps->res_stats[0]); | |
c52b12ed | 762 | int scaled = counter->counts->scaled; |
2996f5dd | 763 | |
2996f5dd | 764 | if (scaled == -1) { |
4aa9015f | 765 | fprintf(output, "%*s%s%*s", |
d7470b6a | 766 | csv_output ? 0 : 18, |
2cee77c4 | 767 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, |
023695d9 SE |
768 | csv_sep, |
769 | csv_output ? 0 : -24, | |
7289f83c | 770 | perf_evsel__name(counter)); |
023695d9 SE |
771 | |
772 | if (counter->cgrp) | |
4aa9015f | 773 | fprintf(output, "%s%s", csv_sep, counter->cgrp->name); |
023695d9 | 774 | |
4aa9015f | 775 | fputc('\n', output); |
2996f5dd IM |
776 | return; |
777 | } | |
c04f5e5d | 778 | |
44175b6f | 779 | if (nsec_counter(counter)) |
f5b4a9c3 | 780 | nsec_printout(-1, counter, avg); |
44175b6f | 781 | else |
f5b4a9c3 | 782 | abs_printout(-1, counter, avg); |
849abde9 | 783 | |
3ae9a34d ZH |
784 | print_noise(counter, avg); |
785 | ||
d7470b6a | 786 | if (csv_output) { |
4aa9015f | 787 | fputc('\n', output); |
d7470b6a SE |
788 | return; |
789 | } | |
790 | ||
506d4bc8 PZ |
791 | if (scaled) { |
792 | double avg_enabled, avg_running; | |
793 | ||
69aad6f1 ACM |
794 | avg_enabled = avg_stats(&ps->res_stats[1]); |
795 | avg_running = avg_stats(&ps->res_stats[2]); | |
d7c29318 | 796 | |
4aa9015f | 797 | fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled); |
506d4bc8 | 798 | } |
4aa9015f | 799 | fprintf(output, "\n"); |
c04f5e5d IM |
800 | } |
801 | ||
f5b4a9c3 SE |
802 | /* |
803 | * Print out the results of a single counter: | |
804 | * does not use aggregated count in system-wide | |
805 | */ | |
69aad6f1 | 806 | static void print_counter(struct perf_evsel *counter) |
f5b4a9c3 SE |
807 | { |
808 | u64 ena, run, val; | |
809 | int cpu; | |
810 | ||
7ae92e74 | 811 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { |
c52b12ed ACM |
812 | val = counter->counts->cpu[cpu].val; |
813 | ena = counter->counts->cpu[cpu].ena; | |
814 | run = counter->counts->cpu[cpu].run; | |
f5b4a9c3 | 815 | if (run == 0 || ena == 0) { |
4aa9015f | 816 | fprintf(output, "CPU%*d%s%*s%s%*s", |
d7470b6a | 817 | csv_output ? 0 : -4, |
7ae92e74 | 818 | perf_evsel__cpus(counter)->map[cpu], csv_sep, |
d7470b6a | 819 | csv_output ? 0 : 18, |
2cee77c4 DA |
820 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, |
821 | csv_sep, | |
023695d9 | 822 | csv_output ? 0 : -24, |
7289f83c | 823 | perf_evsel__name(counter)); |
f5b4a9c3 | 824 | |
023695d9 | 825 | if (counter->cgrp) |
4aa9015f SE |
826 | fprintf(output, "%s%s", |
827 | csv_sep, counter->cgrp->name); | |
023695d9 | 828 | |
4aa9015f | 829 | fputc('\n', output); |
f5b4a9c3 SE |
830 | continue; |
831 | } | |
832 | ||
833 | if (nsec_counter(counter)) | |
834 | nsec_printout(cpu, counter, val); | |
835 | else | |
836 | abs_printout(cpu, counter, val); | |
837 | ||
d7470b6a SE |
838 | if (!csv_output) { |
839 | print_noise(counter, 1.0); | |
f5b4a9c3 | 840 | |
c6264def | 841 | if (run != ena) |
4aa9015f SE |
842 | fprintf(output, " (%.2f%%)", |
843 | 100.0 * run / ena); | |
f5b4a9c3 | 844 | } |
4aa9015f | 845 | fputc('\n', output); |
f5b4a9c3 SE |
846 | } |
847 | } | |
848 | ||
42202dd5 IM |
849 | static void print_stat(int argc, const char **argv) |
850 | { | |
69aad6f1 ACM |
851 | struct perf_evsel *counter; |
852 | int i; | |
42202dd5 | 853 | |
ddcacfa0 IM |
854 | fflush(stdout); |
855 | ||
d7470b6a | 856 | if (!csv_output) { |
4aa9015f SE |
857 | fprintf(output, "\n"); |
858 | fprintf(output, " Performance counter stats for "); | |
aa22dd49 | 859 | if (!perf_target__has_task(&target)) { |
4aa9015f | 860 | fprintf(output, "\'%s", argv[0]); |
d7470b6a | 861 | for (i = 1; i < argc; i++) |
4aa9015f | 862 | fprintf(output, " %s", argv[i]); |
20f946b4 NK |
863 | } else if (target.pid) |
864 | fprintf(output, "process id \'%s", target.pid); | |
d7470b6a | 865 | else |
20f946b4 | 866 | fprintf(output, "thread id \'%s", target.tid); |
44db76c8 | 867 | |
4aa9015f | 868 | fprintf(output, "\'"); |
d7470b6a | 869 | if (run_count > 1) |
4aa9015f SE |
870 | fprintf(output, " (%d runs)", run_count); |
871 | fprintf(output, ":\n\n"); | |
d7470b6a | 872 | } |
2996f5dd | 873 | |
f5b4a9c3 | 874 | if (no_aggr) { |
361c99a6 | 875 | list_for_each_entry(counter, &evsel_list->entries, node) |
f5b4a9c3 SE |
876 | print_counter(counter); |
877 | } else { | |
361c99a6 | 878 | list_for_each_entry(counter, &evsel_list->entries, node) |
f5b4a9c3 SE |
879 | print_counter_aggr(counter); |
880 | } | |
ddcacfa0 | 881 | |
d7470b6a | 882 | if (!csv_output) { |
c3305257 | 883 | if (!null_run) |
4aa9015f SE |
884 | fprintf(output, "\n"); |
885 | fprintf(output, " %17.9f seconds time elapsed", | |
d7470b6a SE |
886 | avg_stats(&walltime_nsecs_stats)/1e9); |
887 | if (run_count > 1) { | |
4aa9015f | 888 | fprintf(output, " "); |
f99844cb IM |
889 | print_noise_pct(stddev_stats(&walltime_nsecs_stats), |
890 | avg_stats(&walltime_nsecs_stats)); | |
d7470b6a | 891 | } |
4aa9015f | 892 | fprintf(output, "\n\n"); |
566747e6 | 893 | } |
ddcacfa0 IM |
894 | } |
895 | ||
f7b7c26e PZ |
896 | static volatile int signr = -1; |
897 | ||
5242519b | 898 | static void skip_signal(int signo) |
ddcacfa0 | 899 | { |
6be2850e | 900 | if(child_pid == -1) |
60666c63 LW |
901 | done = 1; |
902 | ||
f7b7c26e PZ |
903 | signr = signo; |
904 | } | |
905 | ||
906 | static void sig_atexit(void) | |
907 | { | |
933da83a CW |
908 | if (child_pid != -1) |
909 | kill(child_pid, SIGTERM); | |
910 | ||
f7b7c26e PZ |
911 | if (signr == -1) |
912 | return; | |
913 | ||
914 | signal(signr, SIG_DFL); | |
915 | kill(getpid(), signr); | |
5242519b IM |
916 | } |
917 | ||
1d037ca1 IT |
918 | static int stat__set_big_num(const struct option *opt __maybe_unused, |
919 | const char *s __maybe_unused, int unset) | |
d7470b6a SE |
920 | { |
921 | big_num_opt = unset ? 0 : 1; | |
922 | return 0; | |
923 | } | |
924 | ||
2cba3ffb IM |
925 | /* |
926 | * Add default attributes, if there were no attributes specified or | |
927 | * if -d/--detailed, -d -d or -d -d -d is used: | |
928 | */ | |
929 | static int add_default_attributes(void) | |
930 | { | |
b070a547 ACM |
931 | struct perf_event_attr default_attrs[] = { |
932 | ||
933 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, | |
934 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, | |
935 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, | |
936 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, | |
937 | ||
938 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, | |
939 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, | |
940 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, | |
941 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, | |
942 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, | |
943 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, | |
944 | ||
945 | }; | |
946 | ||
947 | /* | |
948 | * Detailed stats (-d), covering the L1 and last level data caches: | |
949 | */ | |
950 | struct perf_event_attr detailed_attrs[] = { | |
951 | ||
952 | { .type = PERF_TYPE_HW_CACHE, | |
953 | .config = | |
954 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
955 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
956 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
957 | ||
958 | { .type = PERF_TYPE_HW_CACHE, | |
959 | .config = | |
960 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
961 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
962 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
963 | ||
964 | { .type = PERF_TYPE_HW_CACHE, | |
965 | .config = | |
966 | PERF_COUNT_HW_CACHE_LL << 0 | | |
967 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
968 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
969 | ||
970 | { .type = PERF_TYPE_HW_CACHE, | |
971 | .config = | |
972 | PERF_COUNT_HW_CACHE_LL << 0 | | |
973 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
974 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
975 | }; | |
976 | ||
977 | /* | |
978 | * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: | |
979 | */ | |
980 | struct perf_event_attr very_detailed_attrs[] = { | |
981 | ||
982 | { .type = PERF_TYPE_HW_CACHE, | |
983 | .config = | |
984 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
985 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
986 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
987 | ||
988 | { .type = PERF_TYPE_HW_CACHE, | |
989 | .config = | |
990 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
991 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
992 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
993 | ||
994 | { .type = PERF_TYPE_HW_CACHE, | |
995 | .config = | |
996 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
997 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
998 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
999 | ||
1000 | { .type = PERF_TYPE_HW_CACHE, | |
1001 | .config = | |
1002 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1003 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1004 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1005 | ||
1006 | { .type = PERF_TYPE_HW_CACHE, | |
1007 | .config = | |
1008 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
1009 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1010 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1011 | ||
1012 | { .type = PERF_TYPE_HW_CACHE, | |
1013 | .config = | |
1014 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
1015 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1016 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1017 | ||
1018 | }; | |
1019 | ||
1020 | /* | |
1021 | * Very, very detailed stats (-d -d -d), adding prefetch events: | |
1022 | */ | |
1023 | struct perf_event_attr very_very_detailed_attrs[] = { | |
1024 | ||
1025 | { .type = PERF_TYPE_HW_CACHE, | |
1026 | .config = | |
1027 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1028 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
1029 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1030 | ||
1031 | { .type = PERF_TYPE_HW_CACHE, | |
1032 | .config = | |
1033 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1034 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
1035 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1036 | }; | |
1037 | ||
2cba3ffb IM |
1038 | /* Set attrs if no event is selected and !null_run: */ |
1039 | if (null_run) | |
1040 | return 0; | |
1041 | ||
1042 | if (!evsel_list->nr_entries) { | |
79695e1b | 1043 | if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0) |
50d08e47 | 1044 | return -1; |
2cba3ffb IM |
1045 | } |
1046 | ||
1047 | /* Detailed events get appended to the event list: */ | |
1048 | ||
1049 | if (detailed_run < 1) | |
1050 | return 0; | |
1051 | ||
1052 | /* Append detailed run extra attributes: */ | |
79695e1b | 1053 | if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) |
50d08e47 | 1054 | return -1; |
2cba3ffb IM |
1055 | |
1056 | if (detailed_run < 2) | |
1057 | return 0; | |
1058 | ||
1059 | /* Append very detailed run extra attributes: */ | |
79695e1b | 1060 | if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) |
50d08e47 | 1061 | return -1; |
2cba3ffb IM |
1062 | |
1063 | if (detailed_run < 3) | |
1064 | return 0; | |
1065 | ||
1066 | /* Append very, very detailed run extra attributes: */ | |
79695e1b | 1067 | return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); |
2cba3ffb IM |
1068 | } |
1069 | ||
1d037ca1 | 1070 | int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused) |
5242519b | 1071 | { |
b070a547 ACM |
1072 | bool append_file = false, |
1073 | sync_run = false; | |
1074 | int output_fd = 0; | |
1075 | const char *output_name = NULL; | |
1076 | const struct option options[] = { | |
1077 | OPT_CALLBACK('e', "event", &evsel_list, "event", | |
1078 | "event selector. use 'perf list' to list available events", | |
1079 | parse_events_option), | |
1080 | OPT_CALLBACK(0, "filter", &evsel_list, "filter", | |
1081 | "event filter", parse_filter), | |
1082 | OPT_BOOLEAN('i', "no-inherit", &no_inherit, | |
1083 | "child tasks do not inherit counters"), | |
1084 | OPT_STRING('p', "pid", &target.pid, "pid", | |
1085 | "stat events on existing process id"), | |
1086 | OPT_STRING('t', "tid", &target.tid, "tid", | |
1087 | "stat events on existing thread id"), | |
1088 | OPT_BOOLEAN('a', "all-cpus", &target.system_wide, | |
1089 | "system-wide collection from all CPUs"), | |
1090 | OPT_BOOLEAN('g', "group", &group, | |
1091 | "put the counters into a counter group"), | |
1092 | OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"), | |
1093 | OPT_INCR('v', "verbose", &verbose, | |
1094 | "be more verbose (show counter open errors, etc)"), | |
1095 | OPT_INTEGER('r', "repeat", &run_count, | |
1096 | "repeat command and print average + stddev (max: 100)"), | |
1097 | OPT_BOOLEAN('n', "null", &null_run, | |
1098 | "null run - dont start any counters"), | |
1099 | OPT_INCR('d', "detailed", &detailed_run, | |
1100 | "detailed run - start a lot of events"), | |
1101 | OPT_BOOLEAN('S', "sync", &sync_run, | |
1102 | "call sync() before starting a run"), | |
1103 | OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, | |
1104 | "print large numbers with thousands\' separators", | |
1105 | stat__set_big_num), | |
1106 | OPT_STRING('C', "cpu", &target.cpu_list, "cpu", | |
1107 | "list of cpus to monitor in system-wide"), | |
1108 | OPT_BOOLEAN('A', "no-aggr", &no_aggr, "disable CPU count aggregation"), | |
1109 | OPT_STRING('x', "field-separator", &csv_sep, "separator", | |
1110 | "print counts with custom separator"), | |
1111 | OPT_CALLBACK('G', "cgroup", &evsel_list, "name", | |
1112 | "monitor event in cgroup name only", parse_cgroups), | |
1113 | OPT_STRING('o', "output", &output_name, "file", "output file name"), | |
1114 | OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), | |
1115 | OPT_INTEGER(0, "log-fd", &output_fd, | |
1116 | "log output to fd, instead of stderr"), | |
1117 | OPT_END() | |
1118 | }; | |
1119 | const char * const stat_usage[] = { | |
1120 | "perf stat [<options>] [<command>]", | |
1121 | NULL | |
1122 | }; | |
69aad6f1 | 1123 | struct perf_evsel *pos; |
b070a547 | 1124 | int status = -ENOMEM, run_idx; |
4aa9015f | 1125 | const char *mode; |
42202dd5 | 1126 | |
5af52b51 SE |
1127 | setlocale(LC_ALL, ""); |
1128 | ||
7e2ed097 | 1129 | evsel_list = perf_evlist__new(NULL, NULL); |
361c99a6 ACM |
1130 | if (evsel_list == NULL) |
1131 | return -ENOMEM; | |
1132 | ||
a0541234 AB |
1133 | argc = parse_options(argc, argv, options, stat_usage, |
1134 | PARSE_OPT_STOP_AT_NON_OPTION); | |
d7470b6a | 1135 | |
4aa9015f SE |
1136 | output = stderr; |
1137 | if (output_name && strcmp(output_name, "-")) | |
1138 | output = NULL; | |
1139 | ||
56f3bae7 JC |
1140 | if (output_name && output_fd) { |
1141 | fprintf(stderr, "cannot use both --output and --log-fd\n"); | |
1142 | usage_with_options(stat_usage, options); | |
1143 | } | |
fc3e4d07 SE |
1144 | |
1145 | if (output_fd < 0) { | |
1146 | fprintf(stderr, "argument to --log-fd must be a > 0\n"); | |
1147 | usage_with_options(stat_usage, options); | |
1148 | } | |
1149 | ||
4aa9015f SE |
1150 | if (!output) { |
1151 | struct timespec tm; | |
1152 | mode = append_file ? "a" : "w"; | |
1153 | ||
1154 | output = fopen(output_name, mode); | |
1155 | if (!output) { | |
1156 | perror("failed to create output file"); | |
fceda7fe | 1157 | return -1; |
4aa9015f SE |
1158 | } |
1159 | clock_gettime(CLOCK_REALTIME, &tm); | |
1160 | fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); | |
fc3e4d07 | 1161 | } else if (output_fd > 0) { |
56f3bae7 JC |
1162 | mode = append_file ? "a" : "w"; |
1163 | output = fdopen(output_fd, mode); | |
1164 | if (!output) { | |
1165 | perror("Failed opening logfd"); | |
1166 | return -errno; | |
1167 | } | |
4aa9015f SE |
1168 | } |
1169 | ||
d4ffd04d | 1170 | if (csv_sep) { |
d7470b6a | 1171 | csv_output = true; |
d4ffd04d JC |
1172 | if (!strcmp(csv_sep, "\\t")) |
1173 | csv_sep = "\t"; | |
1174 | } else | |
d7470b6a SE |
1175 | csv_sep = DEFAULT_SEPARATOR; |
1176 | ||
1177 | /* | |
1178 | * let the spreadsheet do the pretty-printing | |
1179 | */ | |
1180 | if (csv_output) { | |
61a9f324 | 1181 | /* User explicitly passed -B? */ |
d7470b6a SE |
1182 | if (big_num_opt == 1) { |
1183 | fprintf(stderr, "-B option not supported with -x\n"); | |
1184 | usage_with_options(stat_usage, options); | |
1185 | } else /* Nope, so disable big number formatting */ | |
1186 | big_num = false; | |
1187 | } else if (big_num_opt == 0) /* User passed --no-big-num */ | |
1188 | big_num = false; | |
1189 | ||
aa22dd49 | 1190 | if (!argc && !perf_target__has_task(&target)) |
5242519b | 1191 | usage_with_options(stat_usage, options); |
9e9772c4 | 1192 | if (run_count <= 0) |
42202dd5 | 1193 | usage_with_options(stat_usage, options); |
ddcacfa0 | 1194 | |
023695d9 | 1195 | /* no_aggr, cgroup are for system-wide only */ |
aa22dd49 | 1196 | if ((no_aggr || nr_cgroups) && !perf_target__has_cpu(&target)) { |
023695d9 SE |
1197 | fprintf(stderr, "both cgroup and no-aggregation " |
1198 | "modes only available in system-wide mode\n"); | |
1199 | ||
f5b4a9c3 | 1200 | usage_with_options(stat_usage, options); |
023695d9 | 1201 | } |
f5b4a9c3 | 1202 | |
2cba3ffb IM |
1203 | if (add_default_attributes()) |
1204 | goto out; | |
ddcacfa0 | 1205 | |
4bd0f2d2 | 1206 | perf_target__validate(&target); |
5c98d466 | 1207 | |
77a6f014 | 1208 | if (perf_evlist__create_maps(evsel_list, &target) < 0) { |
aa22dd49 | 1209 | if (perf_target__has_task(&target)) |
77a6f014 | 1210 | pr_err("Problems finding threads of monitor\n"); |
aa22dd49 | 1211 | if (perf_target__has_cpu(&target)) |
77a6f014 | 1212 | perror("failed to parse CPUs map"); |
ddcacfa0 | 1213 | |
c45c6ea2 | 1214 | usage_with_options(stat_usage, options); |
60d567e2 ACM |
1215 | return -1; |
1216 | } | |
c45c6ea2 | 1217 | |
361c99a6 | 1218 | list_for_each_entry(pos, &evsel_list->entries, node) { |
c52b12ed | 1219 | if (perf_evsel__alloc_stat_priv(pos) < 0 || |
7ae92e74 | 1220 | perf_evsel__alloc_counts(pos, perf_evsel__nr_cpus(pos)) < 0) |
69aad6f1 | 1221 | goto out_free_fd; |
d6d901c2 ZY |
1222 | } |
1223 | ||
58d7e993 IM |
1224 | /* |
1225 | * We dont want to block the signals - that would cause | |
1226 | * child tasks to inherit that and Ctrl-C would not work. | |
1227 | * What we want is for Ctrl-C to work in the exec()-ed | |
1228 | * task, but being ignored by perf stat itself: | |
1229 | */ | |
f7b7c26e | 1230 | atexit(sig_atexit); |
58d7e993 IM |
1231 | signal(SIGINT, skip_signal); |
1232 | signal(SIGALRM, skip_signal); | |
1233 | signal(SIGABRT, skip_signal); | |
1234 | ||
42202dd5 IM |
1235 | status = 0; |
1236 | for (run_idx = 0; run_idx < run_count; run_idx++) { | |
1237 | if (run_count != 1 && verbose) | |
4aa9015f SE |
1238 | fprintf(output, "[ perf stat: executing run #%d ... ]\n", |
1239 | run_idx + 1); | |
f9cef0a9 IM |
1240 | |
1241 | if (sync_run) | |
1242 | sync(); | |
1243 | ||
42202dd5 IM |
1244 | status = run_perf_stat(argc, argv); |
1245 | } | |
1246 | ||
084ab9f8 ACM |
1247 | if (status != -1) |
1248 | print_stat(argc, argv); | |
69aad6f1 | 1249 | out_free_fd: |
361c99a6 | 1250 | list_for_each_entry(pos, &evsel_list->entries, node) |
69aad6f1 | 1251 | perf_evsel__free_stat_priv(pos); |
7e2ed097 | 1252 | perf_evlist__delete_maps(evsel_list); |
0015e2e1 ACM |
1253 | out: |
1254 | perf_evlist__delete(evsel_list); | |
42202dd5 | 1255 | return status; |
ddcacfa0 | 1256 | } |