]>
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" |
60666c63 | 54 | #include "util/header.h" |
a12b51c4 | 55 | #include "util/cpumap.h" |
d6d901c2 | 56 | #include "util/thread.h" |
fd78260b | 57 | #include "util/thread_map.h" |
ddcacfa0 | 58 | |
ddcacfa0 | 59 | #include <sys/prctl.h> |
42202dd5 | 60 | #include <math.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 | |
cdd6c482 | 67 | static struct perf_event_attr default_attrs[] = { |
ddcacfa0 | 68 | |
56aab464 IM |
69 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, |
70 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, | |
71 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, | |
72 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, | |
f4dbfa8f | 73 | |
56aab464 | 74 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, |
129c04cb IM |
75 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, |
76 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, | |
56aab464 | 77 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, |
56aab464 IM |
78 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, |
79 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, | |
f4dbfa8f | 80 | |
ddcacfa0 | 81 | }; |
5242519b | 82 | |
c6264def | 83 | /* |
2cba3ffb | 84 | * Detailed stats (-d), covering the L1 and last level data caches: |
c6264def IM |
85 | */ |
86 | static struct perf_event_attr detailed_attrs[] = { | |
87 | ||
c6264def IM |
88 | { .type = PERF_TYPE_HW_CACHE, |
89 | .config = | |
90 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
91 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
92 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
93 | ||
94 | { .type = PERF_TYPE_HW_CACHE, | |
95 | .config = | |
96 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
97 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
98 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
99 | ||
100 | { .type = PERF_TYPE_HW_CACHE, | |
101 | .config = | |
102 | PERF_COUNT_HW_CACHE_LL << 0 | | |
103 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
104 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
105 | ||
106 | { .type = PERF_TYPE_HW_CACHE, | |
107 | .config = | |
108 | PERF_COUNT_HW_CACHE_LL << 0 | | |
109 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
110 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
111 | }; | |
112 | ||
2cba3ffb IM |
113 | /* |
114 | * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: | |
115 | */ | |
116 | static struct perf_event_attr very_detailed_attrs[] = { | |
117 | ||
118 | { .type = PERF_TYPE_HW_CACHE, | |
119 | .config = | |
120 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
121 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
122 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
123 | ||
124 | { .type = PERF_TYPE_HW_CACHE, | |
125 | .config = | |
126 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
127 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
128 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
129 | ||
130 | { .type = PERF_TYPE_HW_CACHE, | |
131 | .config = | |
132 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
133 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
134 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
135 | ||
136 | { .type = PERF_TYPE_HW_CACHE, | |
137 | .config = | |
138 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
139 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
140 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
141 | ||
142 | { .type = PERF_TYPE_HW_CACHE, | |
143 | .config = | |
144 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
145 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
146 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
147 | ||
148 | { .type = PERF_TYPE_HW_CACHE, | |
149 | .config = | |
150 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
151 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
152 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
153 | ||
154 | }; | |
155 | ||
156 | /* | |
157 | * Very, very detailed stats (-d -d -d), adding prefetch events: | |
158 | */ | |
159 | static struct perf_event_attr very_very_detailed_attrs[] = { | |
160 | ||
161 | { .type = PERF_TYPE_HW_CACHE, | |
162 | .config = | |
163 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
164 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
165 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
166 | ||
167 | { .type = PERF_TYPE_HW_CACHE, | |
168 | .config = | |
169 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
170 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
171 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
172 | }; | |
173 | ||
174 | ||
175 | ||
666e6d48 | 176 | static struct perf_evlist *evsel_list; |
361c99a6 | 177 | |
77a6f014 NK |
178 | static struct perf_target target = { |
179 | .uid = UINT_MAX, | |
180 | }; | |
ddcacfa0 | 181 | |
20f946b4 | 182 | static int run_idx = 0; |
3d632595 | 183 | static int run_count = 1; |
2e6cdf99 | 184 | static bool no_inherit = false; |
c0555642 | 185 | static bool scale = true; |
f5b4a9c3 | 186 | static bool no_aggr = false; |
933da83a | 187 | static pid_t child_pid = -1; |
c0555642 | 188 | static bool null_run = false; |
2cba3ffb | 189 | static int detailed_run = 0; |
f9cef0a9 | 190 | static bool sync_run = false; |
201e0b06 | 191 | static bool big_num = true; |
d7470b6a | 192 | static int big_num_opt = -1; |
d7470b6a SE |
193 | static const char *csv_sep = NULL; |
194 | static bool csv_output = false; | |
43bece79 | 195 | static bool group = false; |
4aa9015f SE |
196 | static const char *output_name = NULL; |
197 | static FILE *output = NULL; | |
56f3bae7 | 198 | static int output_fd; |
5af52b51 | 199 | |
60666c63 LW |
200 | static volatile int done = 0; |
201 | ||
506d4bc8 PZ |
202 | struct stats |
203 | { | |
8a02631a | 204 | double n, mean, M2; |
506d4bc8 | 205 | }; |
42202dd5 | 206 | |
69aad6f1 ACM |
207 | struct perf_stat { |
208 | struct stats res_stats[3]; | |
69aad6f1 ACM |
209 | }; |
210 | ||
c52b12ed | 211 | static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel) |
69aad6f1 | 212 | { |
c52b12ed | 213 | evsel->priv = zalloc(sizeof(struct perf_stat)); |
69aad6f1 ACM |
214 | return evsel->priv == NULL ? -ENOMEM : 0; |
215 | } | |
216 | ||
217 | static void perf_evsel__free_stat_priv(struct perf_evsel *evsel) | |
218 | { | |
219 | free(evsel->priv); | |
220 | evsel->priv = NULL; | |
221 | } | |
222 | ||
9e9772c4 PZ |
223 | static void update_stats(struct stats *stats, u64 val) |
224 | { | |
8a02631a | 225 | double delta; |
9e9772c4 | 226 | |
8a02631a PZ |
227 | stats->n++; |
228 | delta = val - stats->mean; | |
229 | stats->mean += delta / stats->n; | |
230 | stats->M2 += delta*(val - stats->mean); | |
9e9772c4 PZ |
231 | } |
232 | ||
506d4bc8 PZ |
233 | static double avg_stats(struct stats *stats) |
234 | { | |
8a02631a | 235 | return stats->mean; |
506d4bc8 | 236 | } |
42202dd5 | 237 | |
506d4bc8 | 238 | /* |
63d40deb PZ |
239 | * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance |
240 | * | |
8a02631a PZ |
241 | * (\Sum n_i^2) - ((\Sum n_i)^2)/n |
242 | * s^2 = ------------------------------- | |
243 | * n - 1 | |
63d40deb PZ |
244 | * |
245 | * http://en.wikipedia.org/wiki/Stddev | |
246 | * | |
247 | * The std dev of the mean is related to the std dev by: | |
248 | * | |
249 | * s | |
250 | * s_mean = ------- | |
251 | * sqrt(n) | |
252 | * | |
506d4bc8 PZ |
253 | */ |
254 | static double stddev_stats(struct stats *stats) | |
255 | { | |
19f47402 JC |
256 | double variance, variance_mean; |
257 | ||
258 | if (!stats->n) | |
259 | return 0.0; | |
260 | ||
261 | variance = stats->M2 / (stats->n - 1); | |
262 | variance_mean = variance / stats->n; | |
42202dd5 | 263 | |
63d40deb | 264 | return sqrt(variance_mean); |
506d4bc8 | 265 | } |
42202dd5 | 266 | |
666e6d48 RR |
267 | static struct stats runtime_nsecs_stats[MAX_NR_CPUS]; |
268 | static struct stats runtime_cycles_stats[MAX_NR_CPUS]; | |
269 | static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS]; | |
270 | static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS]; | |
271 | static struct stats runtime_branches_stats[MAX_NR_CPUS]; | |
272 | static struct stats runtime_cacherefs_stats[MAX_NR_CPUS]; | |
273 | static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS]; | |
274 | static struct stats runtime_l1_icache_stats[MAX_NR_CPUS]; | |
275 | static struct stats runtime_ll_cache_stats[MAX_NR_CPUS]; | |
276 | static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS]; | |
277 | static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS]; | |
278 | static struct stats walltime_nsecs_stats; | |
be1ac0d8 | 279 | |
727ab04e ACM |
280 | static int create_perf_stat_counter(struct perf_evsel *evsel, |
281 | struct perf_evsel *first) | |
ddcacfa0 | 282 | { |
69aad6f1 | 283 | struct perf_event_attr *attr = &evsel->attr; |
727ab04e | 284 | struct xyarray *group_fd = NULL; |
5622c07b SE |
285 | bool exclude_guest_missing = false; |
286 | int ret; | |
727ab04e ACM |
287 | |
288 | if (group && evsel != first) | |
289 | group_fd = first->fd; | |
16c8a109 | 290 | |
ddcacfa0 | 291 | if (scale) |
a21ca2ca IM |
292 | attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | |
293 | PERF_FORMAT_TOTAL_TIME_RUNNING; | |
ddcacfa0 | 294 | |
5d2cd909 ACM |
295 | attr->inherit = !no_inherit; |
296 | ||
5622c07b SE |
297 | retry: |
298 | if (exclude_guest_missing) | |
299 | evsel->attr.exclude_guest = evsel->attr.exclude_host = 0; | |
300 | ||
16ee6576 | 301 | if (perf_target__has_cpu(&target)) { |
5622c07b | 302 | ret = perf_evsel__open_per_cpu(evsel, evsel_list->cpus, |
16ee6576 | 303 | group, group_fd); |
5622c07b SE |
304 | if (ret) |
305 | goto check_ret; | |
306 | return 0; | |
307 | } | |
308 | ||
aa22dd49 | 309 | if (!perf_target__has_task(&target) && (!group || evsel == first)) { |
48290609 ACM |
310 | attr->disabled = 1; |
311 | attr->enable_on_exec = 1; | |
ddcacfa0 | 312 | } |
084ab9f8 | 313 | |
5622c07b SE |
314 | ret = perf_evsel__open_per_thread(evsel, evsel_list->threads, |
315 | group, group_fd); | |
316 | if (!ret) | |
317 | return 0; | |
318 | /* fall through */ | |
319 | check_ret: | |
320 | if (ret && errno == EINVAL) { | |
321 | if (!exclude_guest_missing && | |
322 | (evsel->attr.exclude_guest || evsel->attr.exclude_host)) { | |
323 | pr_debug("Old kernel, cannot exclude " | |
324 | "guest or host samples.\n"); | |
325 | exclude_guest_missing = true; | |
326 | goto retry; | |
327 | } | |
328 | } | |
329 | return ret; | |
ddcacfa0 IM |
330 | } |
331 | ||
c04f5e5d IM |
332 | /* |
333 | * Does the counter have nsecs as a unit? | |
334 | */ | |
daec78a0 | 335 | static inline int nsec_counter(struct perf_evsel *evsel) |
c04f5e5d | 336 | { |
daec78a0 ACM |
337 | if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || |
338 | perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) | |
c04f5e5d IM |
339 | return 1; |
340 | ||
341 | return 0; | |
342 | } | |
343 | ||
dcd9936a IM |
344 | /* |
345 | * Update various tracking values we maintain to print | |
346 | * more semantic information such as miss/hit ratios, | |
347 | * instruction rates, etc: | |
348 | */ | |
349 | static void update_shadow_stats(struct perf_evsel *counter, u64 *count) | |
350 | { | |
351 | if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK)) | |
352 | update_stats(&runtime_nsecs_stats[0], count[0]); | |
353 | else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES)) | |
354 | update_stats(&runtime_cycles_stats[0], count[0]); | |
d3d1e86d IM |
355 | else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) |
356 | update_stats(&runtime_stalled_cycles_front_stats[0], count[0]); | |
129c04cb | 357 | else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND)) |
d3d1e86d | 358 | update_stats(&runtime_stalled_cycles_back_stats[0], count[0]); |
dcd9936a IM |
359 | else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS)) |
360 | update_stats(&runtime_branches_stats[0], count[0]); | |
361 | else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES)) | |
362 | update_stats(&runtime_cacherefs_stats[0], count[0]); | |
8bb6c79f IM |
363 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D)) |
364 | update_stats(&runtime_l1_dcache_stats[0], count[0]); | |
c3305257 IM |
365 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I)) |
366 | update_stats(&runtime_l1_icache_stats[0], count[0]); | |
367 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL)) | |
368 | update_stats(&runtime_ll_cache_stats[0], count[0]); | |
369 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB)) | |
370 | update_stats(&runtime_dtlb_cache_stats[0], count[0]); | |
371 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB)) | |
372 | update_stats(&runtime_itlb_cache_stats[0], count[0]); | |
dcd9936a IM |
373 | } |
374 | ||
c04f5e5d | 375 | /* |
2996f5dd | 376 | * Read out the results of a single counter: |
f5b4a9c3 | 377 | * aggregate counts across CPUs in system-wide mode |
c04f5e5d | 378 | */ |
c52b12ed | 379 | static int read_counter_aggr(struct perf_evsel *counter) |
c04f5e5d | 380 | { |
69aad6f1 | 381 | struct perf_stat *ps = counter->priv; |
c52b12ed ACM |
382 | u64 *count = counter->counts->aggr.values; |
383 | int i; | |
2996f5dd | 384 | |
7e2ed097 ACM |
385 | if (__perf_evsel__read(counter, evsel_list->cpus->nr, |
386 | evsel_list->threads->nr, scale) < 0) | |
c52b12ed | 387 | return -1; |
9e9772c4 PZ |
388 | |
389 | for (i = 0; i < 3; i++) | |
69aad6f1 | 390 | update_stats(&ps->res_stats[i], count[i]); |
9e9772c4 PZ |
391 | |
392 | if (verbose) { | |
4aa9015f | 393 | fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", |
9486aa38 | 394 | event_name(counter), count[0], count[1], count[2]); |
9e9772c4 PZ |
395 | } |
396 | ||
be1ac0d8 IM |
397 | /* |
398 | * Save the full runtime - to allow normalization during printout: | |
399 | */ | |
dcd9936a | 400 | update_shadow_stats(counter, count); |
c52b12ed ACM |
401 | |
402 | return 0; | |
f5b4a9c3 SE |
403 | } |
404 | ||
405 | /* | |
406 | * Read out the results of a single counter: | |
407 | * do not aggregate counts across CPUs in system-wide mode | |
408 | */ | |
c52b12ed | 409 | static int read_counter(struct perf_evsel *counter) |
f5b4a9c3 | 410 | { |
c52b12ed | 411 | u64 *count; |
f5b4a9c3 | 412 | int cpu; |
f5b4a9c3 | 413 | |
7e2ed097 | 414 | for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) { |
c52b12ed ACM |
415 | if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0) |
416 | return -1; | |
f5b4a9c3 | 417 | |
c52b12ed | 418 | count = counter->counts->cpu[cpu].values; |
f5b4a9c3 | 419 | |
dcd9936a | 420 | update_shadow_stats(counter, count); |
f5b4a9c3 | 421 | } |
c52b12ed ACM |
422 | |
423 | return 0; | |
2996f5dd IM |
424 | } |
425 | ||
f37a291c | 426 | static int run_perf_stat(int argc __used, const char **argv) |
42202dd5 IM |
427 | { |
428 | unsigned long long t0, t1; | |
727ab04e | 429 | struct perf_evsel *counter, *first; |
42202dd5 | 430 | int status = 0; |
051ae7f7 | 431 | int child_ready_pipe[2], go_pipe[2]; |
6be2850e | 432 | const bool forks = (argc > 0); |
051ae7f7 | 433 | char buf; |
42202dd5 | 434 | |
60666c63 | 435 | if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) { |
051ae7f7 PM |
436 | perror("failed to create pipes"); |
437 | exit(1); | |
438 | } | |
439 | ||
60666c63 | 440 | if (forks) { |
6be2850e | 441 | if ((child_pid = fork()) < 0) |
60666c63 LW |
442 | perror("failed to fork"); |
443 | ||
6be2850e | 444 | if (!child_pid) { |
60666c63 LW |
445 | close(child_ready_pipe[0]); |
446 | close(go_pipe[1]); | |
447 | fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); | |
448 | ||
449 | /* | |
450 | * Do a dummy execvp to get the PLT entry resolved, | |
451 | * so we avoid the resolver overhead on the real | |
452 | * execvp call. | |
453 | */ | |
454 | execvp("", (char **)argv); | |
455 | ||
456 | /* | |
457 | * Tell the parent we're ready to go | |
458 | */ | |
459 | close(child_ready_pipe[1]); | |
460 | ||
461 | /* | |
462 | * Wait until the parent tells us to go. | |
463 | */ | |
464 | if (read(go_pipe[0], &buf, 1) == -1) | |
465 | perror("unable to read pipe"); | |
466 | ||
467 | execvp(argv[0], (char **)argv); | |
468 | ||
469 | perror(argv[0]); | |
470 | exit(-1); | |
471 | } | |
051ae7f7 | 472 | |
d67356e7 | 473 | if (perf_target__none(&target)) |
7e2ed097 | 474 | evsel_list->threads->map[0] = child_pid; |
d6d901c2 | 475 | |
051ae7f7 | 476 | /* |
60666c63 | 477 | * Wait for the child to be ready to exec. |
051ae7f7 PM |
478 | */ |
479 | close(child_ready_pipe[1]); | |
60666c63 LW |
480 | close(go_pipe[0]); |
481 | if (read(child_ready_pipe[0], &buf, 1) == -1) | |
a92bef0f | 482 | perror("unable to read pipe"); |
60666c63 | 483 | close(child_ready_pipe[0]); |
051ae7f7 PM |
484 | } |
485 | ||
727ab04e ACM |
486 | first = list_entry(evsel_list->entries.next, struct perf_evsel, node); |
487 | ||
361c99a6 | 488 | list_for_each_entry(counter, &evsel_list->entries, node) { |
727ab04e | 489 | if (create_perf_stat_counter(counter, first) < 0) { |
979987a5 DA |
490 | /* |
491 | * PPC returns ENXIO for HW counters until 2.6.37 | |
492 | * (behavior changed with commit b0a873e). | |
493 | */ | |
38f6ae1e | 494 | if (errno == EINVAL || errno == ENOSYS || |
979987a5 DA |
495 | errno == ENOENT || errno == EOPNOTSUPP || |
496 | errno == ENXIO) { | |
c63ca0c0 DA |
497 | if (verbose) |
498 | ui__warning("%s event is not supported by the kernel.\n", | |
499 | event_name(counter)); | |
2cee77c4 | 500 | counter->supported = false; |
ede70290 | 501 | continue; |
c63ca0c0 | 502 | } |
ede70290 IM |
503 | |
504 | if (errno == EPERM || errno == EACCES) { | |
48290609 ACM |
505 | error("You may not have permission to collect %sstats.\n" |
506 | "\t Consider tweaking" | |
507 | " /proc/sys/kernel/perf_event_paranoid or running as root.", | |
20f946b4 | 508 | target.system_wide ? "system-wide " : ""); |
48290609 ACM |
509 | } else { |
510 | error("open_counter returned with %d (%s). " | |
511 | "/bin/dmesg may provide additional information.\n", | |
512 | errno, strerror(errno)); | |
513 | } | |
514 | if (child_pid != -1) | |
515 | kill(child_pid, SIGTERM); | |
516 | die("Not all events could be opened.\n"); | |
517 | return -1; | |
518 | } | |
2cee77c4 | 519 | counter->supported = true; |
084ab9f8 | 520 | } |
42202dd5 | 521 | |
cfd748ae FW |
522 | if (perf_evlist__set_filters(evsel_list)) { |
523 | error("failed to set filter with %d (%s)\n", errno, | |
524 | strerror(errno)); | |
525 | return -1; | |
526 | } | |
527 | ||
42202dd5 IM |
528 | /* |
529 | * Enable counters and exec the command: | |
530 | */ | |
531 | t0 = rdclock(); | |
42202dd5 | 532 | |
60666c63 LW |
533 | if (forks) { |
534 | close(go_pipe[1]); | |
535 | wait(&status); | |
33e49ea7 AK |
536 | if (WIFSIGNALED(status)) |
537 | psignal(WTERMSIG(status), argv[0]); | |
60666c63 | 538 | } else { |
6be2850e | 539 | while(!done) sleep(1); |
60666c63 | 540 | } |
42202dd5 | 541 | |
42202dd5 IM |
542 | t1 = rdclock(); |
543 | ||
9e9772c4 | 544 | update_stats(&walltime_nsecs_stats, t1 - t0); |
42202dd5 | 545 | |
f5b4a9c3 | 546 | if (no_aggr) { |
361c99a6 | 547 | list_for_each_entry(counter, &evsel_list->entries, node) { |
f5b4a9c3 | 548 | read_counter(counter); |
7e2ed097 | 549 | perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1); |
c52b12ed | 550 | } |
f5b4a9c3 | 551 | } else { |
361c99a6 | 552 | list_for_each_entry(counter, &evsel_list->entries, node) { |
f5b4a9c3 | 553 | read_counter_aggr(counter); |
7e2ed097 ACM |
554 | perf_evsel__close_fd(counter, evsel_list->cpus->nr, |
555 | evsel_list->threads->nr); | |
c52b12ed | 556 | } |
f5b4a9c3 | 557 | } |
c52b12ed | 558 | |
42202dd5 IM |
559 | return WEXITSTATUS(status); |
560 | } | |
561 | ||
f99844cb IM |
562 | static void print_noise_pct(double total, double avg) |
563 | { | |
564 | double pct = 0.0; | |
565 | ||
566 | if (avg) | |
567 | pct = 100.0*total/avg; | |
568 | ||
3ae9a34d | 569 | if (csv_output) |
4aa9015f | 570 | fprintf(output, "%s%.2f%%", csv_sep, pct); |
a1bca6cc | 571 | else if (pct) |
4aa9015f | 572 | fprintf(output, " ( +-%6.2f%% )", pct); |
f99844cb IM |
573 | } |
574 | ||
69aad6f1 | 575 | static void print_noise(struct perf_evsel *evsel, double avg) |
42202dd5 | 576 | { |
69aad6f1 ACM |
577 | struct perf_stat *ps; |
578 | ||
849abde9 PZ |
579 | if (run_count == 1) |
580 | return; | |
581 | ||
69aad6f1 | 582 | ps = evsel->priv; |
f99844cb | 583 | print_noise_pct(stddev_stats(&ps->res_stats[0]), avg); |
42202dd5 IM |
584 | } |
585 | ||
daec78a0 | 586 | static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg) |
44175b6f | 587 | { |
506d4bc8 | 588 | double msecs = avg / 1e6; |
d7470b6a | 589 | char cpustr[16] = { '\0', }; |
2cba3ffb | 590 | const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s"; |
44175b6f | 591 | |
f5b4a9c3 | 592 | if (no_aggr) |
d7470b6a SE |
593 | sprintf(cpustr, "CPU%*d%s", |
594 | csv_output ? 0 : -4, | |
7e2ed097 | 595 | evsel_list->cpus->map[cpu], csv_sep); |
d7470b6a | 596 | |
4aa9015f | 597 | fprintf(output, fmt, cpustr, msecs, csv_sep, event_name(evsel)); |
d7470b6a | 598 | |
023695d9 | 599 | if (evsel->cgrp) |
4aa9015f | 600 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); |
023695d9 | 601 | |
d7470b6a SE |
602 | if (csv_output) |
603 | return; | |
44175b6f | 604 | |
daec78a0 | 605 | if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) |
4aa9015f SE |
606 | fprintf(output, " # %8.3f CPUs utilized ", |
607 | avg / avg_stats(&walltime_nsecs_stats)); | |
9dac6a29 NK |
608 | else |
609 | fprintf(output, " "); | |
44175b6f IM |
610 | } |
611 | ||
15e6392f NK |
612 | /* used for get_ratio_color() */ |
613 | enum grc_type { | |
614 | GRC_STALLED_CYCLES_FE, | |
615 | GRC_STALLED_CYCLES_BE, | |
616 | GRC_CACHE_MISSES, | |
617 | GRC_MAX_NR | |
618 | }; | |
619 | ||
620 | static const char *get_ratio_color(enum grc_type type, double ratio) | |
621 | { | |
622 | static const double grc_table[GRC_MAX_NR][3] = { | |
623 | [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 }, | |
624 | [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 }, | |
625 | [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 }, | |
626 | }; | |
627 | const char *color = PERF_COLOR_NORMAL; | |
628 | ||
629 | if (ratio > grc_table[type][0]) | |
630 | color = PERF_COLOR_RED; | |
631 | else if (ratio > grc_table[type][1]) | |
632 | color = PERF_COLOR_MAGENTA; | |
633 | else if (ratio > grc_table[type][2]) | |
634 | color = PERF_COLOR_YELLOW; | |
635 | ||
636 | return color; | |
637 | } | |
638 | ||
d3d1e86d IM |
639 | static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg) |
640 | { | |
641 | double total, ratio = 0.0; | |
642 | const char *color; | |
643 | ||
644 | total = avg_stats(&runtime_cycles_stats[cpu]); | |
645 | ||
646 | if (total) | |
647 | ratio = avg / total * 100.0; | |
648 | ||
15e6392f | 649 | color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio); |
d3d1e86d | 650 | |
4aa9015f SE |
651 | fprintf(output, " # "); |
652 | color_fprintf(output, color, "%6.2f%%", ratio); | |
653 | fprintf(output, " frontend cycles idle "); | |
d3d1e86d IM |
654 | } |
655 | ||
656 | static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg) | |
a5d243d0 IM |
657 | { |
658 | double total, ratio = 0.0; | |
659 | const char *color; | |
660 | ||
661 | total = avg_stats(&runtime_cycles_stats[cpu]); | |
662 | ||
663 | if (total) | |
664 | ratio = avg / total * 100.0; | |
665 | ||
15e6392f | 666 | color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio); |
a5d243d0 | 667 | |
4aa9015f SE |
668 | fprintf(output, " # "); |
669 | color_fprintf(output, color, "%6.2f%%", ratio); | |
670 | fprintf(output, " backend cycles idle "); | |
a5d243d0 IM |
671 | } |
672 | ||
c78df6c1 IM |
673 | static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg) |
674 | { | |
675 | double total, ratio = 0.0; | |
676 | const char *color; | |
677 | ||
678 | total = avg_stats(&runtime_branches_stats[cpu]); | |
679 | ||
680 | if (total) | |
681 | ratio = avg / total * 100.0; | |
682 | ||
15e6392f | 683 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c78df6c1 | 684 | |
4aa9015f SE |
685 | fprintf(output, " # "); |
686 | color_fprintf(output, color, "%6.2f%%", ratio); | |
687 | fprintf(output, " of all branches "); | |
c78df6c1 IM |
688 | } |
689 | ||
8bb6c79f IM |
690 | static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg) |
691 | { | |
692 | double total, ratio = 0.0; | |
693 | const char *color; | |
694 | ||
695 | total = avg_stats(&runtime_l1_dcache_stats[cpu]); | |
696 | ||
697 | if (total) | |
698 | ratio = avg / total * 100.0; | |
699 | ||
15e6392f | 700 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
8bb6c79f | 701 | |
4aa9015f SE |
702 | fprintf(output, " # "); |
703 | color_fprintf(output, color, "%6.2f%%", ratio); | |
704 | fprintf(output, " of all L1-dcache hits "); | |
8bb6c79f IM |
705 | } |
706 | ||
c3305257 IM |
707 | static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg) |
708 | { | |
709 | double total, ratio = 0.0; | |
710 | const char *color; | |
711 | ||
712 | total = avg_stats(&runtime_l1_icache_stats[cpu]); | |
713 | ||
714 | if (total) | |
715 | ratio = avg / total * 100.0; | |
716 | ||
15e6392f | 717 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 718 | |
4aa9015f SE |
719 | fprintf(output, " # "); |
720 | color_fprintf(output, color, "%6.2f%%", ratio); | |
721 | fprintf(output, " of all L1-icache hits "); | |
c3305257 IM |
722 | } |
723 | ||
724 | static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg) | |
725 | { | |
726 | double total, ratio = 0.0; | |
727 | const char *color; | |
728 | ||
729 | total = avg_stats(&runtime_dtlb_cache_stats[cpu]); | |
730 | ||
731 | if (total) | |
732 | ratio = avg / total * 100.0; | |
733 | ||
15e6392f | 734 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 735 | |
4aa9015f SE |
736 | fprintf(output, " # "); |
737 | color_fprintf(output, color, "%6.2f%%", ratio); | |
738 | fprintf(output, " of all dTLB cache hits "); | |
c3305257 IM |
739 | } |
740 | ||
741 | static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg) | |
742 | { | |
743 | double total, ratio = 0.0; | |
744 | const char *color; | |
745 | ||
746 | total = avg_stats(&runtime_itlb_cache_stats[cpu]); | |
747 | ||
748 | if (total) | |
749 | ratio = avg / total * 100.0; | |
750 | ||
15e6392f | 751 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 752 | |
4aa9015f SE |
753 | fprintf(output, " # "); |
754 | color_fprintf(output, color, "%6.2f%%", ratio); | |
755 | fprintf(output, " of all iTLB cache hits "); | |
c3305257 IM |
756 | } |
757 | ||
758 | static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg) | |
759 | { | |
760 | double total, ratio = 0.0; | |
761 | const char *color; | |
762 | ||
763 | total = avg_stats(&runtime_ll_cache_stats[cpu]); | |
764 | ||
765 | if (total) | |
766 | ratio = avg / total * 100.0; | |
767 | ||
15e6392f | 768 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 769 | |
4aa9015f SE |
770 | fprintf(output, " # "); |
771 | color_fprintf(output, color, "%6.2f%%", ratio); | |
772 | fprintf(output, " of all LL-cache hits "); | |
c3305257 IM |
773 | } |
774 | ||
daec78a0 | 775 | static void abs_printout(int cpu, struct perf_evsel *evsel, double avg) |
44175b6f | 776 | { |
c7f7fea3 | 777 | double total, ratio = 0.0; |
f5b4a9c3 | 778 | char cpustr[16] = { '\0', }; |
d7470b6a SE |
779 | const char *fmt; |
780 | ||
781 | if (csv_output) | |
782 | fmt = "%s%.0f%s%s"; | |
783 | else if (big_num) | |
2cba3ffb | 784 | fmt = "%s%'18.0f%s%-25s"; |
d7470b6a | 785 | else |
2cba3ffb | 786 | fmt = "%s%18.0f%s%-25s"; |
f5b4a9c3 SE |
787 | |
788 | if (no_aggr) | |
d7470b6a SE |
789 | sprintf(cpustr, "CPU%*d%s", |
790 | csv_output ? 0 : -4, | |
7e2ed097 | 791 | evsel_list->cpus->map[cpu], csv_sep); |
f5b4a9c3 SE |
792 | else |
793 | cpu = 0; | |
c7f7fea3 | 794 | |
4aa9015f | 795 | fprintf(output, fmt, cpustr, avg, csv_sep, event_name(evsel)); |
d7470b6a | 796 | |
023695d9 | 797 | if (evsel->cgrp) |
4aa9015f | 798 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); |
023695d9 | 799 | |
d7470b6a SE |
800 | if (csv_output) |
801 | return; | |
44175b6f | 802 | |
daec78a0 | 803 | if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) { |
f5b4a9c3 | 804 | total = avg_stats(&runtime_cycles_stats[cpu]); |
c7f7fea3 IM |
805 | |
806 | if (total) | |
807 | ratio = avg / total; | |
808 | ||
4aa9015f | 809 | fprintf(output, " # %5.2f insns per cycle ", ratio); |
481f988a | 810 | |
d3d1e86d IM |
811 | total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]); |
812 | total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu])); | |
481f988a IM |
813 | |
814 | if (total && avg) { | |
815 | ratio = total / avg; | |
4aa9015f | 816 | fprintf(output, "\n # %5.2f stalled cycles per insn", ratio); |
481f988a IM |
817 | } |
818 | ||
daec78a0 | 819 | } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) && |
f5b4a9c3 | 820 | runtime_branches_stats[cpu].n != 0) { |
c78df6c1 | 821 | print_branch_misses(cpu, evsel, avg); |
8bb6c79f IM |
822 | } else if ( |
823 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
824 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D | | |
825 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
826 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
c6264def | 827 | runtime_l1_dcache_stats[cpu].n != 0) { |
8bb6c79f | 828 | print_l1_dcache_misses(cpu, evsel, avg); |
c3305257 IM |
829 | } else if ( |
830 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
831 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I | | |
832 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
833 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
834 | runtime_l1_icache_stats[cpu].n != 0) { | |
835 | print_l1_icache_misses(cpu, evsel, avg); | |
836 | } else if ( | |
837 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
838 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB | | |
839 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
840 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
841 | runtime_dtlb_cache_stats[cpu].n != 0) { | |
842 | print_dtlb_cache_misses(cpu, evsel, avg); | |
843 | } else if ( | |
844 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
845 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB | | |
846 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
847 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
848 | runtime_itlb_cache_stats[cpu].n != 0) { | |
849 | print_itlb_cache_misses(cpu, evsel, avg); | |
850 | } else if ( | |
851 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
852 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL | | |
853 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
854 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
855 | runtime_ll_cache_stats[cpu].n != 0) { | |
856 | print_ll_cache_misses(cpu, evsel, avg); | |
d58f4c82 IM |
857 | } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) && |
858 | runtime_cacherefs_stats[cpu].n != 0) { | |
859 | total = avg_stats(&runtime_cacherefs_stats[cpu]); | |
860 | ||
861 | if (total) | |
862 | ratio = avg * 100 / total; | |
863 | ||
4aa9015f | 864 | fprintf(output, " # %8.3f %% of all cache refs ", ratio); |
d58f4c82 | 865 | |
d3d1e86d IM |
866 | } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) { |
867 | print_stalled_cycles_frontend(cpu, evsel, avg); | |
129c04cb | 868 | } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) { |
d3d1e86d | 869 | print_stalled_cycles_backend(cpu, evsel, avg); |
481f988a | 870 | } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) { |
f5b4a9c3 | 871 | total = avg_stats(&runtime_nsecs_stats[cpu]); |
c7f7fea3 IM |
872 | |
873 | if (total) | |
481f988a | 874 | ratio = 1.0 * avg / total; |
c7f7fea3 | 875 | |
4aa9015f | 876 | fprintf(output, " # %8.3f GHz ", ratio); |
481f988a | 877 | } else if (runtime_nsecs_stats[cpu].n != 0) { |
5fde2523 NK |
878 | char unit = 'M'; |
879 | ||
481f988a | 880 | total = avg_stats(&runtime_nsecs_stats[cpu]); |
11ba2b85 IM |
881 | |
882 | if (total) | |
481f988a | 883 | ratio = 1000.0 * avg / total; |
5fde2523 NK |
884 | if (ratio < 0.001) { |
885 | ratio *= 1000; | |
886 | unit = 'K'; | |
887 | } | |
11ba2b85 | 888 | |
5fde2523 | 889 | fprintf(output, " # %8.3f %c/sec ", ratio, unit); |
a5d243d0 | 890 | } else { |
4aa9015f | 891 | fprintf(output, " "); |
44175b6f | 892 | } |
44175b6f IM |
893 | } |
894 | ||
2996f5dd IM |
895 | /* |
896 | * Print out the results of a single counter: | |
f5b4a9c3 | 897 | * aggregated counts in system-wide mode |
2996f5dd | 898 | */ |
69aad6f1 | 899 | static void print_counter_aggr(struct perf_evsel *counter) |
2996f5dd | 900 | { |
69aad6f1 ACM |
901 | struct perf_stat *ps = counter->priv; |
902 | double avg = avg_stats(&ps->res_stats[0]); | |
c52b12ed | 903 | int scaled = counter->counts->scaled; |
2996f5dd | 904 | |
2996f5dd | 905 | if (scaled == -1) { |
4aa9015f | 906 | fprintf(output, "%*s%s%*s", |
d7470b6a | 907 | csv_output ? 0 : 18, |
2cee77c4 | 908 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, |
023695d9 SE |
909 | csv_sep, |
910 | csv_output ? 0 : -24, | |
911 | event_name(counter)); | |
912 | ||
913 | if (counter->cgrp) | |
4aa9015f | 914 | fprintf(output, "%s%s", csv_sep, counter->cgrp->name); |
023695d9 | 915 | |
4aa9015f | 916 | fputc('\n', output); |
2996f5dd IM |
917 | return; |
918 | } | |
c04f5e5d | 919 | |
44175b6f | 920 | if (nsec_counter(counter)) |
f5b4a9c3 | 921 | nsec_printout(-1, counter, avg); |
44175b6f | 922 | else |
f5b4a9c3 | 923 | abs_printout(-1, counter, avg); |
849abde9 | 924 | |
3ae9a34d ZH |
925 | print_noise(counter, avg); |
926 | ||
d7470b6a | 927 | if (csv_output) { |
4aa9015f | 928 | fputc('\n', output); |
d7470b6a SE |
929 | return; |
930 | } | |
931 | ||
506d4bc8 PZ |
932 | if (scaled) { |
933 | double avg_enabled, avg_running; | |
934 | ||
69aad6f1 ACM |
935 | avg_enabled = avg_stats(&ps->res_stats[1]); |
936 | avg_running = avg_stats(&ps->res_stats[2]); | |
d7c29318 | 937 | |
4aa9015f | 938 | fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled); |
506d4bc8 | 939 | } |
4aa9015f | 940 | fprintf(output, "\n"); |
c04f5e5d IM |
941 | } |
942 | ||
f5b4a9c3 SE |
943 | /* |
944 | * Print out the results of a single counter: | |
945 | * does not use aggregated count in system-wide | |
946 | */ | |
69aad6f1 | 947 | static void print_counter(struct perf_evsel *counter) |
f5b4a9c3 SE |
948 | { |
949 | u64 ena, run, val; | |
950 | int cpu; | |
951 | ||
7e2ed097 | 952 | for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) { |
c52b12ed ACM |
953 | val = counter->counts->cpu[cpu].val; |
954 | ena = counter->counts->cpu[cpu].ena; | |
955 | run = counter->counts->cpu[cpu].run; | |
f5b4a9c3 | 956 | if (run == 0 || ena == 0) { |
4aa9015f | 957 | fprintf(output, "CPU%*d%s%*s%s%*s", |
d7470b6a | 958 | csv_output ? 0 : -4, |
7e2ed097 | 959 | evsel_list->cpus->map[cpu], csv_sep, |
d7470b6a | 960 | csv_output ? 0 : 18, |
2cee77c4 DA |
961 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, |
962 | csv_sep, | |
023695d9 | 963 | csv_output ? 0 : -24, |
d7470b6a | 964 | event_name(counter)); |
f5b4a9c3 | 965 | |
023695d9 | 966 | if (counter->cgrp) |
4aa9015f SE |
967 | fprintf(output, "%s%s", |
968 | csv_sep, counter->cgrp->name); | |
023695d9 | 969 | |
4aa9015f | 970 | fputc('\n', output); |
f5b4a9c3 SE |
971 | continue; |
972 | } | |
973 | ||
974 | if (nsec_counter(counter)) | |
975 | nsec_printout(cpu, counter, val); | |
976 | else | |
977 | abs_printout(cpu, counter, val); | |
978 | ||
d7470b6a SE |
979 | if (!csv_output) { |
980 | print_noise(counter, 1.0); | |
f5b4a9c3 | 981 | |
c6264def | 982 | if (run != ena) |
4aa9015f SE |
983 | fprintf(output, " (%.2f%%)", |
984 | 100.0 * run / ena); | |
f5b4a9c3 | 985 | } |
4aa9015f | 986 | fputc('\n', output); |
f5b4a9c3 SE |
987 | } |
988 | } | |
989 | ||
42202dd5 IM |
990 | static void print_stat(int argc, const char **argv) |
991 | { | |
69aad6f1 ACM |
992 | struct perf_evsel *counter; |
993 | int i; | |
42202dd5 | 994 | |
ddcacfa0 IM |
995 | fflush(stdout); |
996 | ||
d7470b6a | 997 | if (!csv_output) { |
4aa9015f SE |
998 | fprintf(output, "\n"); |
999 | fprintf(output, " Performance counter stats for "); | |
aa22dd49 | 1000 | if (!perf_target__has_task(&target)) { |
4aa9015f | 1001 | fprintf(output, "\'%s", argv[0]); |
d7470b6a | 1002 | for (i = 1; i < argc; i++) |
4aa9015f | 1003 | fprintf(output, " %s", argv[i]); |
20f946b4 NK |
1004 | } else if (target.pid) |
1005 | fprintf(output, "process id \'%s", target.pid); | |
d7470b6a | 1006 | else |
20f946b4 | 1007 | fprintf(output, "thread id \'%s", target.tid); |
44db76c8 | 1008 | |
4aa9015f | 1009 | fprintf(output, "\'"); |
d7470b6a | 1010 | if (run_count > 1) |
4aa9015f SE |
1011 | fprintf(output, " (%d runs)", run_count); |
1012 | fprintf(output, ":\n\n"); | |
d7470b6a | 1013 | } |
2996f5dd | 1014 | |
f5b4a9c3 | 1015 | if (no_aggr) { |
361c99a6 | 1016 | list_for_each_entry(counter, &evsel_list->entries, node) |
f5b4a9c3 SE |
1017 | print_counter(counter); |
1018 | } else { | |
361c99a6 | 1019 | list_for_each_entry(counter, &evsel_list->entries, node) |
f5b4a9c3 SE |
1020 | print_counter_aggr(counter); |
1021 | } | |
ddcacfa0 | 1022 | |
d7470b6a | 1023 | if (!csv_output) { |
c3305257 | 1024 | if (!null_run) |
4aa9015f SE |
1025 | fprintf(output, "\n"); |
1026 | fprintf(output, " %17.9f seconds time elapsed", | |
d7470b6a SE |
1027 | avg_stats(&walltime_nsecs_stats)/1e9); |
1028 | if (run_count > 1) { | |
4aa9015f | 1029 | fprintf(output, " "); |
f99844cb IM |
1030 | print_noise_pct(stddev_stats(&walltime_nsecs_stats), |
1031 | avg_stats(&walltime_nsecs_stats)); | |
d7470b6a | 1032 | } |
4aa9015f | 1033 | fprintf(output, "\n\n"); |
566747e6 | 1034 | } |
ddcacfa0 IM |
1035 | } |
1036 | ||
f7b7c26e PZ |
1037 | static volatile int signr = -1; |
1038 | ||
5242519b | 1039 | static void skip_signal(int signo) |
ddcacfa0 | 1040 | { |
6be2850e | 1041 | if(child_pid == -1) |
60666c63 LW |
1042 | done = 1; |
1043 | ||
f7b7c26e PZ |
1044 | signr = signo; |
1045 | } | |
1046 | ||
1047 | static void sig_atexit(void) | |
1048 | { | |
933da83a CW |
1049 | if (child_pid != -1) |
1050 | kill(child_pid, SIGTERM); | |
1051 | ||
f7b7c26e PZ |
1052 | if (signr == -1) |
1053 | return; | |
1054 | ||
1055 | signal(signr, SIG_DFL); | |
1056 | kill(getpid(), signr); | |
5242519b IM |
1057 | } |
1058 | ||
1059 | static const char * const stat_usage[] = { | |
60666c63 | 1060 | "perf stat [<options>] [<command>]", |
5242519b IM |
1061 | NULL |
1062 | }; | |
1063 | ||
d7470b6a SE |
1064 | static int stat__set_big_num(const struct option *opt __used, |
1065 | const char *s __used, int unset) | |
1066 | { | |
1067 | big_num_opt = unset ? 0 : 1; | |
1068 | return 0; | |
1069 | } | |
1070 | ||
4aa9015f SE |
1071 | static bool append_file; |
1072 | ||
5242519b | 1073 | static const struct option options[] = { |
361c99a6 | 1074 | OPT_CALLBACK('e', "event", &evsel_list, "event", |
86847b62 | 1075 | "event selector. use 'perf list' to list available events", |
f120f9d5 | 1076 | parse_events_option), |
cfd748ae FW |
1077 | OPT_CALLBACK(0, "filter", &evsel_list, "filter", |
1078 | "event filter", parse_filter), | |
2e6cdf99 SE |
1079 | OPT_BOOLEAN('i', "no-inherit", &no_inherit, |
1080 | "child tasks do not inherit counters"), | |
20f946b4 | 1081 | OPT_STRING('p', "pid", &target.pid, "pid", |
b52956c9 | 1082 | "stat events on existing process id"), |
20f946b4 | 1083 | OPT_STRING('t', "tid", &target.tid, "tid", |
b52956c9 | 1084 | "stat events on existing thread id"), |
20f946b4 | 1085 | OPT_BOOLEAN('a', "all-cpus", &target.system_wide, |
3d632595 | 1086 | "system-wide collection from all CPUs"), |
43bece79 LM |
1087 | OPT_BOOLEAN('g', "group", &group, |
1088 | "put the counters into a counter group"), | |
b26bc5a7 | 1089 | OPT_BOOLEAN('c', "scale", &scale, |
3d632595 | 1090 | "scale/normalize counters"), |
c0555642 | 1091 | OPT_INCR('v', "verbose", &verbose, |
743ee1f8 | 1092 | "be more verbose (show counter open errors, etc)"), |
42202dd5 IM |
1093 | OPT_INTEGER('r', "repeat", &run_count, |
1094 | "repeat command and print average + stddev (max: 100)"), | |
0cfb7a13 IM |
1095 | OPT_BOOLEAN('n', "null", &null_run, |
1096 | "null run - dont start any counters"), | |
2cba3ffb | 1097 | OPT_INCR('d', "detailed", &detailed_run, |
c6264def | 1098 | "detailed run - start a lot of events"), |
f9cef0a9 IM |
1099 | OPT_BOOLEAN('S', "sync", &sync_run, |
1100 | "call sync() before starting a run"), | |
d7470b6a SE |
1101 | OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, |
1102 | "print large numbers with thousands\' separators", | |
1103 | stat__set_big_num), | |
20f946b4 | 1104 | OPT_STRING('C', "cpu", &target.cpu_list, "cpu", |
c45c6ea2 | 1105 | "list of cpus to monitor in system-wide"), |
f5b4a9c3 SE |
1106 | OPT_BOOLEAN('A', "no-aggr", &no_aggr, |
1107 | "disable CPU count aggregation"), | |
d7470b6a SE |
1108 | OPT_STRING('x', "field-separator", &csv_sep, "separator", |
1109 | "print counts with custom separator"), | |
023695d9 SE |
1110 | OPT_CALLBACK('G', "cgroup", &evsel_list, "name", |
1111 | "monitor event in cgroup name only", | |
1112 | parse_cgroups), | |
4aa9015f SE |
1113 | OPT_STRING('o', "output", &output_name, "file", |
1114 | "output file name"), | |
1115 | OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), | |
56f3bae7 JC |
1116 | OPT_INTEGER(0, "log-fd", &output_fd, |
1117 | "log output to fd, instead of stderr"), | |
5242519b IM |
1118 | OPT_END() |
1119 | }; | |
1120 | ||
2cba3ffb IM |
1121 | /* |
1122 | * Add default attributes, if there were no attributes specified or | |
1123 | * if -d/--detailed, -d -d or -d -d -d is used: | |
1124 | */ | |
1125 | static int add_default_attributes(void) | |
1126 | { | |
2cba3ffb IM |
1127 | /* Set attrs if no event is selected and !null_run: */ |
1128 | if (null_run) | |
1129 | return 0; | |
1130 | ||
1131 | if (!evsel_list->nr_entries) { | |
79695e1b | 1132 | if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0) |
50d08e47 | 1133 | return -1; |
2cba3ffb IM |
1134 | } |
1135 | ||
1136 | /* Detailed events get appended to the event list: */ | |
1137 | ||
1138 | if (detailed_run < 1) | |
1139 | return 0; | |
1140 | ||
1141 | /* Append detailed run extra attributes: */ | |
79695e1b | 1142 | if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) |
50d08e47 | 1143 | return -1; |
2cba3ffb IM |
1144 | |
1145 | if (detailed_run < 2) | |
1146 | return 0; | |
1147 | ||
1148 | /* Append very detailed run extra attributes: */ | |
79695e1b | 1149 | if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) |
50d08e47 | 1150 | return -1; |
2cba3ffb IM |
1151 | |
1152 | if (detailed_run < 3) | |
1153 | return 0; | |
1154 | ||
1155 | /* Append very, very detailed run extra attributes: */ | |
79695e1b | 1156 | return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); |
2cba3ffb IM |
1157 | } |
1158 | ||
f37a291c | 1159 | int cmd_stat(int argc, const char **argv, const char *prefix __used) |
5242519b | 1160 | { |
69aad6f1 ACM |
1161 | struct perf_evsel *pos; |
1162 | int status = -ENOMEM; | |
4aa9015f | 1163 | const char *mode; |
42202dd5 | 1164 | |
5af52b51 SE |
1165 | setlocale(LC_ALL, ""); |
1166 | ||
7e2ed097 | 1167 | evsel_list = perf_evlist__new(NULL, NULL); |
361c99a6 ACM |
1168 | if (evsel_list == NULL) |
1169 | return -ENOMEM; | |
1170 | ||
a0541234 AB |
1171 | argc = parse_options(argc, argv, options, stat_usage, |
1172 | PARSE_OPT_STOP_AT_NON_OPTION); | |
d7470b6a | 1173 | |
4aa9015f SE |
1174 | output = stderr; |
1175 | if (output_name && strcmp(output_name, "-")) | |
1176 | output = NULL; | |
1177 | ||
56f3bae7 JC |
1178 | if (output_name && output_fd) { |
1179 | fprintf(stderr, "cannot use both --output and --log-fd\n"); | |
1180 | usage_with_options(stat_usage, options); | |
1181 | } | |
4aa9015f SE |
1182 | if (!output) { |
1183 | struct timespec tm; | |
1184 | mode = append_file ? "a" : "w"; | |
1185 | ||
1186 | output = fopen(output_name, mode); | |
1187 | if (!output) { | |
1188 | perror("failed to create output file"); | |
1189 | exit(-1); | |
1190 | } | |
1191 | clock_gettime(CLOCK_REALTIME, &tm); | |
1192 | fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); | |
56f3bae7 JC |
1193 | } else if (output_fd != 2) { |
1194 | mode = append_file ? "a" : "w"; | |
1195 | output = fdopen(output_fd, mode); | |
1196 | if (!output) { | |
1197 | perror("Failed opening logfd"); | |
1198 | return -errno; | |
1199 | } | |
4aa9015f SE |
1200 | } |
1201 | ||
d4ffd04d | 1202 | if (csv_sep) { |
d7470b6a | 1203 | csv_output = true; |
d4ffd04d JC |
1204 | if (!strcmp(csv_sep, "\\t")) |
1205 | csv_sep = "\t"; | |
1206 | } else | |
d7470b6a SE |
1207 | csv_sep = DEFAULT_SEPARATOR; |
1208 | ||
1209 | /* | |
1210 | * let the spreadsheet do the pretty-printing | |
1211 | */ | |
1212 | if (csv_output) { | |
61a9f324 | 1213 | /* User explicitly passed -B? */ |
d7470b6a SE |
1214 | if (big_num_opt == 1) { |
1215 | fprintf(stderr, "-B option not supported with -x\n"); | |
1216 | usage_with_options(stat_usage, options); | |
1217 | } else /* Nope, so disable big number formatting */ | |
1218 | big_num = false; | |
1219 | } else if (big_num_opt == 0) /* User passed --no-big-num */ | |
1220 | big_num = false; | |
1221 | ||
aa22dd49 | 1222 | if (!argc && !perf_target__has_task(&target)) |
5242519b | 1223 | usage_with_options(stat_usage, options); |
9e9772c4 | 1224 | if (run_count <= 0) |
42202dd5 | 1225 | usage_with_options(stat_usage, options); |
ddcacfa0 | 1226 | |
023695d9 | 1227 | /* no_aggr, cgroup are for system-wide only */ |
aa22dd49 | 1228 | if ((no_aggr || nr_cgroups) && !perf_target__has_cpu(&target)) { |
023695d9 SE |
1229 | fprintf(stderr, "both cgroup and no-aggregation " |
1230 | "modes only available in system-wide mode\n"); | |
1231 | ||
f5b4a9c3 | 1232 | usage_with_options(stat_usage, options); |
023695d9 | 1233 | } |
f5b4a9c3 | 1234 | |
2cba3ffb IM |
1235 | if (add_default_attributes()) |
1236 | goto out; | |
ddcacfa0 | 1237 | |
4bd0f2d2 | 1238 | perf_target__validate(&target); |
5c98d466 | 1239 | |
77a6f014 | 1240 | if (perf_evlist__create_maps(evsel_list, &target) < 0) { |
aa22dd49 | 1241 | if (perf_target__has_task(&target)) |
77a6f014 | 1242 | pr_err("Problems finding threads of monitor\n"); |
aa22dd49 | 1243 | if (perf_target__has_cpu(&target)) |
77a6f014 | 1244 | perror("failed to parse CPUs map"); |
ddcacfa0 | 1245 | |
c45c6ea2 | 1246 | usage_with_options(stat_usage, options); |
60d567e2 ACM |
1247 | return -1; |
1248 | } | |
c45c6ea2 | 1249 | |
361c99a6 | 1250 | list_for_each_entry(pos, &evsel_list->entries, node) { |
c52b12ed | 1251 | if (perf_evsel__alloc_stat_priv(pos) < 0 || |
806fb630 | 1252 | perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0) |
69aad6f1 | 1253 | goto out_free_fd; |
d6d901c2 ZY |
1254 | } |
1255 | ||
58d7e993 IM |
1256 | /* |
1257 | * We dont want to block the signals - that would cause | |
1258 | * child tasks to inherit that and Ctrl-C would not work. | |
1259 | * What we want is for Ctrl-C to work in the exec()-ed | |
1260 | * task, but being ignored by perf stat itself: | |
1261 | */ | |
f7b7c26e | 1262 | atexit(sig_atexit); |
58d7e993 IM |
1263 | signal(SIGINT, skip_signal); |
1264 | signal(SIGALRM, skip_signal); | |
1265 | signal(SIGABRT, skip_signal); | |
1266 | ||
42202dd5 IM |
1267 | status = 0; |
1268 | for (run_idx = 0; run_idx < run_count; run_idx++) { | |
1269 | if (run_count != 1 && verbose) | |
4aa9015f SE |
1270 | fprintf(output, "[ perf stat: executing run #%d ... ]\n", |
1271 | run_idx + 1); | |
f9cef0a9 IM |
1272 | |
1273 | if (sync_run) | |
1274 | sync(); | |
1275 | ||
42202dd5 IM |
1276 | status = run_perf_stat(argc, argv); |
1277 | } | |
1278 | ||
084ab9f8 ACM |
1279 | if (status != -1) |
1280 | print_stat(argc, argv); | |
69aad6f1 | 1281 | out_free_fd: |
361c99a6 | 1282 | list_for_each_entry(pos, &evsel_list->entries, node) |
69aad6f1 | 1283 | perf_evsel__free_stat_priv(pos); |
7e2ed097 | 1284 | perf_evlist__delete_maps(evsel_list); |
0015e2e1 ACM |
1285 | out: |
1286 | perf_evlist__delete(evsel_list); | |
42202dd5 | 1287 | return status; |
ddcacfa0 | 1288 | } |