Commit | Line | Data |
---|---|---|
ddcacfa0 | 1 | /* |
bf9e1876 IM |
2 | * builtin-stat.c |
3 | * | |
4 | * Builtin stat command: Give a precise performance counters summary | |
5 | * overview about any workload, CPU or specific PID. | |
6 | * | |
7 | * Sample output: | |
ddcacfa0 | 8 | |
2cba3ffb | 9 | $ perf stat ./hackbench 10 |
ddcacfa0 | 10 | |
2cba3ffb | 11 | Time: 0.118 |
ddcacfa0 | 12 | |
2cba3ffb | 13 | Performance counter stats for './hackbench 10': |
ddcacfa0 | 14 | |
2cba3ffb IM |
15 | 1708.761321 task-clock # 11.037 CPUs utilized |
16 | 41,190 context-switches # 0.024 M/sec | |
17 | 6,735 CPU-migrations # 0.004 M/sec | |
18 | 17,318 page-faults # 0.010 M/sec | |
19 | 5,205,202,243 cycles # 3.046 GHz | |
20 | 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle | |
21 | 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle | |
22 | 2,603,501,247 instructions # 0.50 insns per cycle | |
23 | # 1.48 stalled cycles per insn | |
24 | 484,357,498 branches # 283.455 M/sec | |
25 | 6,388,934 branch-misses # 1.32% of all branches | |
26 | ||
27 | 0.154822978 seconds time elapsed | |
ddcacfa0 | 28 | |
5242519b | 29 | * |
2cba3ffb | 30 | * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com> |
5242519b IM |
31 | * |
32 | * Improvements and fixes by: | |
33 | * | |
34 | * Arjan van de Ven <arjan@linux.intel.com> | |
35 | * Yanmin Zhang <yanmin.zhang@intel.com> | |
36 | * Wu Fengguang <fengguang.wu@intel.com> | |
37 | * Mike Galbraith <efault@gmx.de> | |
38 | * Paul Mackerras <paulus@samba.org> | |
6e750a8f | 39 | * Jaswinder Singh Rajput <jaswinder@kernel.org> |
5242519b IM |
40 | * |
41 | * Released under the GPL v2. (and only v2, not any later version) | |
ddcacfa0 IM |
42 | */ |
43 | ||
1a482f38 | 44 | #include "perf.h" |
16f762a2 | 45 | #include "builtin.h" |
f14d5707 | 46 | #include "util/cgroup.h" |
148be2c1 | 47 | #include "util/util.h" |
5242519b IM |
48 | #include "util/parse-options.h" |
49 | #include "util/parse-events.h" | |
4cabc3d1 | 50 | #include "util/pmu.h" |
8f28827a | 51 | #include "util/event.h" |
361c99a6 | 52 | #include "util/evlist.h" |
69aad6f1 | 53 | #include "util/evsel.h" |
8f28827a | 54 | #include "util/debug.h" |
a5d243d0 | 55 | #include "util/color.h" |
0007ecea | 56 | #include "util/stat.h" |
60666c63 | 57 | #include "util/header.h" |
a12b51c4 | 58 | #include "util/cpumap.h" |
d6d901c2 | 59 | #include "util/thread.h" |
fd78260b | 60 | #include "util/thread_map.h" |
ddcacfa0 | 61 | |
1f16c575 | 62 | #include <stdlib.h> |
ddcacfa0 | 63 | #include <sys/prctl.h> |
5af52b51 | 64 | #include <locale.h> |
16c8a109 | 65 | |
d7470b6a | 66 | #define DEFAULT_SEPARATOR " " |
2cee77c4 DA |
67 | #define CNTR_NOT_SUPPORTED "<not supported>" |
68 | #define CNTR_NOT_COUNTED "<not counted>" | |
d7470b6a | 69 | |
13370a9b SE |
70 | static void print_stat(int argc, const char **argv); |
71 | static void print_counter_aggr(struct perf_evsel *counter, char *prefix); | |
72 | static void print_counter(struct perf_evsel *counter, char *prefix); | |
86ee6e18 | 73 | static void print_aggr(char *prefix); |
13370a9b | 74 | |
4cabc3d1 AK |
75 | /* Default events used for perf stat -T */ |
76 | static const char * const transaction_attrs[] = { | |
77 | "task-clock", | |
78 | "{" | |
79 | "instructions," | |
80 | "cycles," | |
81 | "cpu/cycles-t/," | |
82 | "cpu/tx-start/," | |
83 | "cpu/el-start/," | |
84 | "cpu/cycles-ct/" | |
85 | "}" | |
86 | }; | |
87 | ||
88 | /* More limited version when the CPU does not have all events. */ | |
89 | static const char * const transaction_limited_attrs[] = { | |
90 | "task-clock", | |
91 | "{" | |
92 | "instructions," | |
93 | "cycles," | |
94 | "cpu/cycles-t/," | |
95 | "cpu/tx-start/" | |
96 | "}" | |
97 | }; | |
98 | ||
99 | /* must match transaction_attrs and the beginning limited_attrs */ | |
100 | enum { | |
101 | T_TASK_CLOCK, | |
102 | T_INSTRUCTIONS, | |
103 | T_CYCLES, | |
104 | T_CYCLES_IN_TX, | |
105 | T_TRANSACTION_START, | |
106 | T_ELISION_START, | |
107 | T_CYCLES_IN_TX_CP, | |
108 | }; | |
109 | ||
666e6d48 | 110 | static struct perf_evlist *evsel_list; |
361c99a6 | 111 | |
602ad878 | 112 | static struct target target = { |
77a6f014 NK |
113 | .uid = UINT_MAX, |
114 | }; | |
ddcacfa0 | 115 | |
86ee6e18 SE |
116 | enum aggr_mode { |
117 | AGGR_NONE, | |
118 | AGGR_GLOBAL, | |
119 | AGGR_SOCKET, | |
12c08a9f | 120 | AGGR_CORE, |
86ee6e18 SE |
121 | }; |
122 | ||
3d632595 | 123 | static int run_count = 1; |
2e6cdf99 | 124 | static bool no_inherit = false; |
c0555642 | 125 | static bool scale = true; |
86ee6e18 | 126 | static enum aggr_mode aggr_mode = AGGR_GLOBAL; |
d07f0b12 | 127 | static volatile pid_t child_pid = -1; |
c0555642 | 128 | static bool null_run = false; |
2cba3ffb | 129 | static int detailed_run = 0; |
4cabc3d1 | 130 | static bool transaction_run; |
201e0b06 | 131 | static bool big_num = true; |
d7470b6a | 132 | static int big_num_opt = -1; |
d7470b6a SE |
133 | static const char *csv_sep = NULL; |
134 | static bool csv_output = false; | |
43bece79 | 135 | static bool group = false; |
4aa9015f | 136 | static FILE *output = NULL; |
1f16c575 PZ |
137 | static const char *pre_cmd = NULL; |
138 | static const char *post_cmd = NULL; | |
139 | static bool sync_run = false; | |
13370a9b | 140 | static unsigned int interval = 0; |
41191688 | 141 | static unsigned int initial_delay = 0; |
410136f5 | 142 | static unsigned int unit_width = 4; /* strlen("unit") */ |
a7e191c3 | 143 | static bool forever = false; |
13370a9b | 144 | static struct timespec ref_time; |
86ee6e18 SE |
145 | static struct cpu_map *aggr_map; |
146 | static int (*aggr_get_id)(struct cpu_map *m, int cpu); | |
5af52b51 | 147 | |
60666c63 LW |
148 | static volatile int done = 0; |
149 | ||
69aad6f1 ACM |
150 | struct perf_stat { |
151 | struct stats res_stats[3]; | |
69aad6f1 ACM |
152 | }; |
153 | ||
13370a9b SE |
154 | static inline void diff_timespec(struct timespec *r, struct timespec *a, |
155 | struct timespec *b) | |
156 | { | |
157 | r->tv_sec = a->tv_sec - b->tv_sec; | |
158 | if (a->tv_nsec < b->tv_nsec) { | |
159 | r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec; | |
160 | r->tv_sec--; | |
161 | } else { | |
162 | r->tv_nsec = a->tv_nsec - b->tv_nsec ; | |
163 | } | |
164 | } | |
165 | ||
166 | static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel) | |
167 | { | |
168 | return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus; | |
169 | } | |
170 | ||
171 | static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel) | |
172 | { | |
173 | return perf_evsel__cpus(evsel)->nr; | |
174 | } | |
175 | ||
a7e191c3 FD |
176 | static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel) |
177 | { | |
90f6bb6c AK |
178 | int i; |
179 | struct perf_stat *ps = evsel->priv; | |
180 | ||
181 | for (i = 0; i < 3; i++) | |
182 | init_stats(&ps->res_stats[i]); | |
a7e191c3 FD |
183 | } |
184 | ||
c52b12ed | 185 | static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel) |
69aad6f1 | 186 | { |
c52b12ed | 187 | evsel->priv = zalloc(sizeof(struct perf_stat)); |
d180ac14 | 188 | if (evsel->priv == NULL) |
90f6bb6c AK |
189 | return -ENOMEM; |
190 | perf_evsel__reset_stat_priv(evsel); | |
191 | return 0; | |
69aad6f1 ACM |
192 | } |
193 | ||
194 | static void perf_evsel__free_stat_priv(struct perf_evsel *evsel) | |
195 | { | |
04662523 | 196 | zfree(&evsel->priv); |
69aad6f1 ACM |
197 | } |
198 | ||
13370a9b | 199 | static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel) |
7ae92e74 | 200 | { |
13370a9b SE |
201 | void *addr; |
202 | size_t sz; | |
203 | ||
204 | sz = sizeof(*evsel->counts) + | |
205 | (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values)); | |
206 | ||
207 | addr = zalloc(sz); | |
208 | if (!addr) | |
209 | return -ENOMEM; | |
210 | ||
211 | evsel->prev_raw_counts = addr; | |
212 | ||
213 | return 0; | |
7ae92e74 YZ |
214 | } |
215 | ||
13370a9b | 216 | static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel) |
7ae92e74 | 217 | { |
04662523 | 218 | zfree(&evsel->prev_raw_counts); |
7ae92e74 YZ |
219 | } |
220 | ||
d134ffb9 ACM |
221 | static void perf_evlist__free_stats(struct perf_evlist *evlist) |
222 | { | |
223 | struct perf_evsel *evsel; | |
224 | ||
0050f7aa | 225 | evlist__for_each(evlist, evsel) { |
d134ffb9 ACM |
226 | perf_evsel__free_stat_priv(evsel); |
227 | perf_evsel__free_counts(evsel); | |
228 | perf_evsel__free_prev_raw_counts(evsel); | |
229 | } | |
230 | } | |
231 | ||
232 | static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw) | |
233 | { | |
234 | struct perf_evsel *evsel; | |
235 | ||
0050f7aa | 236 | evlist__for_each(evlist, evsel) { |
d134ffb9 ACM |
237 | if (perf_evsel__alloc_stat_priv(evsel) < 0 || |
238 | perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 || | |
239 | (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0)) | |
240 | goto out_free; | |
241 | } | |
242 | ||
243 | return 0; | |
244 | ||
245 | out_free: | |
246 | perf_evlist__free_stats(evlist); | |
247 | return -1; | |
248 | } | |
249 | ||
666e6d48 RR |
250 | static struct stats runtime_nsecs_stats[MAX_NR_CPUS]; |
251 | static struct stats runtime_cycles_stats[MAX_NR_CPUS]; | |
252 | static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS]; | |
253 | static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS]; | |
254 | static struct stats runtime_branches_stats[MAX_NR_CPUS]; | |
255 | static struct stats runtime_cacherefs_stats[MAX_NR_CPUS]; | |
256 | static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS]; | |
257 | static struct stats runtime_l1_icache_stats[MAX_NR_CPUS]; | |
258 | static struct stats runtime_ll_cache_stats[MAX_NR_CPUS]; | |
259 | static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS]; | |
260 | static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS]; | |
4cabc3d1 | 261 | static struct stats runtime_cycles_in_tx_stats[MAX_NR_CPUS]; |
666e6d48 | 262 | static struct stats walltime_nsecs_stats; |
4cabc3d1 AK |
263 | static struct stats runtime_transaction_stats[MAX_NR_CPUS]; |
264 | static struct stats runtime_elision_stats[MAX_NR_CPUS]; | |
be1ac0d8 | 265 | |
d134ffb9 | 266 | static void perf_stat__reset_stats(struct perf_evlist *evlist) |
a7e191c3 | 267 | { |
d134ffb9 ACM |
268 | struct perf_evsel *evsel; |
269 | ||
0050f7aa | 270 | evlist__for_each(evlist, evsel) { |
d134ffb9 ACM |
271 | perf_evsel__reset_stat_priv(evsel); |
272 | perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel)); | |
273 | } | |
274 | ||
a7e191c3 FD |
275 | memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats)); |
276 | memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats)); | |
277 | memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats)); | |
278 | memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats)); | |
279 | memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats)); | |
280 | memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats)); | |
281 | memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats)); | |
282 | memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats)); | |
283 | memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats)); | |
284 | memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats)); | |
285 | memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats)); | |
4cabc3d1 AK |
286 | memset(runtime_cycles_in_tx_stats, 0, |
287 | sizeof(runtime_cycles_in_tx_stats)); | |
288 | memset(runtime_transaction_stats, 0, | |
289 | sizeof(runtime_transaction_stats)); | |
290 | memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats)); | |
a7e191c3 FD |
291 | memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats)); |
292 | } | |
293 | ||
cac21425 | 294 | static int create_perf_stat_counter(struct perf_evsel *evsel) |
ddcacfa0 | 295 | { |
69aad6f1 | 296 | struct perf_event_attr *attr = &evsel->attr; |
727ab04e | 297 | |
ddcacfa0 | 298 | if (scale) |
a21ca2ca IM |
299 | attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | |
300 | PERF_FORMAT_TOTAL_TIME_RUNNING; | |
ddcacfa0 | 301 | |
5d2cd909 ACM |
302 | attr->inherit = !no_inherit; |
303 | ||
602ad878 | 304 | if (target__has_cpu(&target)) |
594ac61a | 305 | return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel)); |
5622c07b | 306 | |
602ad878 | 307 | if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) { |
48290609 | 308 | attr->disabled = 1; |
41191688 AK |
309 | if (!initial_delay) |
310 | attr->enable_on_exec = 1; | |
ddcacfa0 | 311 | } |
084ab9f8 | 312 | |
594ac61a | 313 | return perf_evsel__open_per_thread(evsel, evsel_list->threads); |
ddcacfa0 IM |
314 | } |
315 | ||
c04f5e5d IM |
316 | /* |
317 | * Does the counter have nsecs as a unit? | |
318 | */ | |
daec78a0 | 319 | static inline int nsec_counter(struct perf_evsel *evsel) |
c04f5e5d | 320 | { |
daec78a0 ACM |
321 | if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || |
322 | perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) | |
c04f5e5d IM |
323 | return 1; |
324 | ||
325 | return 0; | |
326 | } | |
327 | ||
4cabc3d1 AK |
328 | static struct perf_evsel *nth_evsel(int n) |
329 | { | |
330 | static struct perf_evsel **array; | |
331 | static int array_len; | |
332 | struct perf_evsel *ev; | |
333 | int j; | |
334 | ||
335 | /* Assumes this only called when evsel_list does not change anymore. */ | |
336 | if (!array) { | |
0050f7aa | 337 | evlist__for_each(evsel_list, ev) |
4cabc3d1 AK |
338 | array_len++; |
339 | array = malloc(array_len * sizeof(void *)); | |
340 | if (!array) | |
341 | exit(ENOMEM); | |
342 | j = 0; | |
0050f7aa | 343 | evlist__for_each(evsel_list, ev) |
4cabc3d1 AK |
344 | array[j++] = ev; |
345 | } | |
346 | if (n < array_len) | |
347 | return array[n]; | |
348 | return NULL; | |
349 | } | |
350 | ||
dcd9936a IM |
351 | /* |
352 | * Update various tracking values we maintain to print | |
353 | * more semantic information such as miss/hit ratios, | |
354 | * instruction rates, etc: | |
355 | */ | |
356 | static void update_shadow_stats(struct perf_evsel *counter, u64 *count) | |
357 | { | |
358 | if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK)) | |
359 | update_stats(&runtime_nsecs_stats[0], count[0]); | |
360 | else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES)) | |
361 | update_stats(&runtime_cycles_stats[0], count[0]); | |
4cabc3d1 AK |
362 | else if (transaction_run && |
363 | perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX))) | |
364 | update_stats(&runtime_cycles_in_tx_stats[0], count[0]); | |
365 | else if (transaction_run && | |
366 | perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START))) | |
367 | update_stats(&runtime_transaction_stats[0], count[0]); | |
368 | else if (transaction_run && | |
369 | perf_evsel__cmp(counter, nth_evsel(T_ELISION_START))) | |
370 | update_stats(&runtime_elision_stats[0], count[0]); | |
d3d1e86d IM |
371 | else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) |
372 | update_stats(&runtime_stalled_cycles_front_stats[0], count[0]); | |
129c04cb | 373 | else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND)) |
d3d1e86d | 374 | update_stats(&runtime_stalled_cycles_back_stats[0], count[0]); |
dcd9936a IM |
375 | else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS)) |
376 | update_stats(&runtime_branches_stats[0], count[0]); | |
377 | else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES)) | |
378 | update_stats(&runtime_cacherefs_stats[0], count[0]); | |
8bb6c79f IM |
379 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D)) |
380 | update_stats(&runtime_l1_dcache_stats[0], count[0]); | |
c3305257 IM |
381 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I)) |
382 | update_stats(&runtime_l1_icache_stats[0], count[0]); | |
383 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL)) | |
384 | update_stats(&runtime_ll_cache_stats[0], count[0]); | |
385 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB)) | |
386 | update_stats(&runtime_dtlb_cache_stats[0], count[0]); | |
387 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB)) | |
388 | update_stats(&runtime_itlb_cache_stats[0], count[0]); | |
dcd9936a IM |
389 | } |
390 | ||
779d0b99 JO |
391 | static void zero_per_pkg(struct perf_evsel *counter) |
392 | { | |
393 | if (counter->per_pkg_mask) | |
394 | memset(counter->per_pkg_mask, 0, MAX_NR_CPUS); | |
395 | } | |
396 | ||
397 | static int check_per_pkg(struct perf_evsel *counter, int cpu, bool *skip) | |
398 | { | |
399 | unsigned long *mask = counter->per_pkg_mask; | |
400 | struct cpu_map *cpus = perf_evsel__cpus(counter); | |
401 | int s; | |
402 | ||
403 | *skip = false; | |
404 | ||
405 | if (!counter->per_pkg) | |
406 | return 0; | |
407 | ||
408 | if (cpu_map__empty(cpus)) | |
409 | return 0; | |
410 | ||
411 | if (!mask) { | |
412 | mask = zalloc(MAX_NR_CPUS); | |
413 | if (!mask) | |
414 | return -ENOMEM; | |
415 | ||
416 | counter->per_pkg_mask = mask; | |
417 | } | |
418 | ||
419 | s = cpu_map__get_socket(cpus, cpu); | |
420 | if (s < 0) | |
421 | return -1; | |
422 | ||
423 | *skip = test_and_set_bit(s, mask) == 1; | |
424 | return 0; | |
425 | } | |
426 | ||
060c4f9c JO |
427 | static int read_cb(struct perf_evsel *evsel, int cpu, int thread __maybe_unused, |
428 | struct perf_counts_values *count) | |
429 | { | |
1971f59f | 430 | struct perf_counts_values *aggr = &evsel->counts->aggr; |
779d0b99 JO |
431 | static struct perf_counts_values zero; |
432 | bool skip = false; | |
433 | ||
434 | if (check_per_pkg(evsel, cpu, &skip)) { | |
435 | pr_err("failed to read per-pkg counter\n"); | |
436 | return -1; | |
437 | } | |
438 | ||
439 | if (skip) | |
440 | count = &zero; | |
1971f59f | 441 | |
060c4f9c JO |
442 | switch (aggr_mode) { |
443 | case AGGR_CORE: | |
444 | case AGGR_SOCKET: | |
445 | case AGGR_NONE: | |
6c0345b7 JO |
446 | if (!evsel->snapshot) |
447 | perf_evsel__compute_deltas(evsel, cpu, count); | |
060c4f9c JO |
448 | perf_counts_values__scale(count, scale, NULL); |
449 | evsel->counts->cpu[cpu] = *count; | |
450 | update_shadow_stats(evsel, count->values); | |
451 | break; | |
452 | case AGGR_GLOBAL: | |
1971f59f JO |
453 | aggr->val += count->val; |
454 | if (scale) { | |
455 | aggr->ena += count->ena; | |
456 | aggr->run += count->run; | |
457 | } | |
060c4f9c JO |
458 | default: |
459 | break; | |
460 | } | |
461 | ||
462 | return 0; | |
463 | } | |
464 | ||
1971f59f JO |
465 | static int read_counter(struct perf_evsel *counter); |
466 | ||
c04f5e5d | 467 | /* |
2996f5dd | 468 | * Read out the results of a single counter: |
f5b4a9c3 | 469 | * aggregate counts across CPUs in system-wide mode |
c04f5e5d | 470 | */ |
c52b12ed | 471 | static int read_counter_aggr(struct perf_evsel *counter) |
c04f5e5d | 472 | { |
1971f59f | 473 | struct perf_counts_values *aggr = &counter->counts->aggr; |
69aad6f1 | 474 | struct perf_stat *ps = counter->priv; |
c52b12ed ACM |
475 | u64 *count = counter->counts->aggr.values; |
476 | int i; | |
2996f5dd | 477 | |
1971f59f JO |
478 | aggr->val = aggr->ena = aggr->run = 0; |
479 | ||
480 | if (read_counter(counter)) | |
c52b12ed | 481 | return -1; |
9e9772c4 | 482 | |
6c0345b7 JO |
483 | if (!counter->snapshot) |
484 | perf_evsel__compute_deltas(counter, -1, aggr); | |
1971f59f JO |
485 | perf_counts_values__scale(aggr, scale, &counter->counts->scaled); |
486 | ||
9e9772c4 | 487 | for (i = 0; i < 3; i++) |
69aad6f1 | 488 | update_stats(&ps->res_stats[i], count[i]); |
9e9772c4 PZ |
489 | |
490 | if (verbose) { | |
4aa9015f | 491 | fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", |
7289f83c | 492 | perf_evsel__name(counter), count[0], count[1], count[2]); |
9e9772c4 PZ |
493 | } |
494 | ||
be1ac0d8 IM |
495 | /* |
496 | * Save the full runtime - to allow normalization during printout: | |
497 | */ | |
dcd9936a | 498 | update_shadow_stats(counter, count); |
c52b12ed ACM |
499 | |
500 | return 0; | |
f5b4a9c3 SE |
501 | } |
502 | ||
503 | /* | |
504 | * Read out the results of a single counter: | |
505 | * do not aggregate counts across CPUs in system-wide mode | |
506 | */ | |
c52b12ed | 507 | static int read_counter(struct perf_evsel *counter) |
f5b4a9c3 | 508 | { |
9bf1a529 JO |
509 | int nthreads = thread_map__nr(evsel_list->threads); |
510 | int ncpus = perf_evsel__nr_cpus(counter); | |
511 | int cpu, thread; | |
f5b4a9c3 | 512 | |
9bf1a529 JO |
513 | if (counter->system_wide) |
514 | nthreads = 1; | |
515 | ||
779d0b99 JO |
516 | if (counter->per_pkg) |
517 | zero_per_pkg(counter); | |
518 | ||
9bf1a529 JO |
519 | for (thread = 0; thread < nthreads; thread++) { |
520 | for (cpu = 0; cpu < ncpus; cpu++) { | |
521 | if (perf_evsel__read_cb(counter, cpu, thread, read_cb)) | |
522 | return -1; | |
523 | } | |
f5b4a9c3 | 524 | } |
c52b12ed ACM |
525 | |
526 | return 0; | |
2996f5dd IM |
527 | } |
528 | ||
13370a9b SE |
529 | static void print_interval(void) |
530 | { | |
531 | static int num_print_interval; | |
532 | struct perf_evsel *counter; | |
533 | struct perf_stat *ps; | |
534 | struct timespec ts, rs; | |
535 | char prefix[64]; | |
536 | ||
86ee6e18 | 537 | if (aggr_mode == AGGR_GLOBAL) { |
0050f7aa | 538 | evlist__for_each(evsel_list, counter) { |
13370a9b SE |
539 | ps = counter->priv; |
540 | memset(ps->res_stats, 0, sizeof(ps->res_stats)); | |
86ee6e18 | 541 | read_counter_aggr(counter); |
13370a9b | 542 | } |
86ee6e18 | 543 | } else { |
0050f7aa | 544 | evlist__for_each(evsel_list, counter) { |
13370a9b SE |
545 | ps = counter->priv; |
546 | memset(ps->res_stats, 0, sizeof(ps->res_stats)); | |
86ee6e18 | 547 | read_counter(counter); |
13370a9b SE |
548 | } |
549 | } | |
86ee6e18 | 550 | |
13370a9b SE |
551 | clock_gettime(CLOCK_MONOTONIC, &ts); |
552 | diff_timespec(&rs, &ts, &ref_time); | |
553 | sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep); | |
554 | ||
555 | if (num_print_interval == 0 && !csv_output) { | |
86ee6e18 SE |
556 | switch (aggr_mode) { |
557 | case AGGR_SOCKET: | |
410136f5 | 558 | fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit"); |
86ee6e18 | 559 | break; |
12c08a9f | 560 | case AGGR_CORE: |
410136f5 | 561 | fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit"); |
12c08a9f | 562 | break; |
86ee6e18 | 563 | case AGGR_NONE: |
410136f5 | 564 | fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit"); |
86ee6e18 SE |
565 | break; |
566 | case AGGR_GLOBAL: | |
567 | default: | |
410136f5 | 568 | fprintf(output, "# time counts %*s events\n", unit_width, "unit"); |
86ee6e18 | 569 | } |
13370a9b SE |
570 | } |
571 | ||
572 | if (++num_print_interval == 25) | |
573 | num_print_interval = 0; | |
574 | ||
86ee6e18 | 575 | switch (aggr_mode) { |
12c08a9f | 576 | case AGGR_CORE: |
86ee6e18 SE |
577 | case AGGR_SOCKET: |
578 | print_aggr(prefix); | |
579 | break; | |
580 | case AGGR_NONE: | |
0050f7aa | 581 | evlist__for_each(evsel_list, counter) |
13370a9b | 582 | print_counter(counter, prefix); |
86ee6e18 SE |
583 | break; |
584 | case AGGR_GLOBAL: | |
585 | default: | |
0050f7aa | 586 | evlist__for_each(evsel_list, counter) |
13370a9b SE |
587 | print_counter_aggr(counter, prefix); |
588 | } | |
2bbf03f1 AK |
589 | |
590 | fflush(output); | |
13370a9b SE |
591 | } |
592 | ||
41191688 AK |
593 | static void handle_initial_delay(void) |
594 | { | |
595 | struct perf_evsel *counter; | |
596 | ||
597 | if (initial_delay) { | |
598 | const int ncpus = cpu_map__nr(evsel_list->cpus), | |
599 | nthreads = thread_map__nr(evsel_list->threads); | |
600 | ||
601 | usleep(initial_delay * 1000); | |
0050f7aa | 602 | evlist__for_each(evsel_list, counter) |
41191688 AK |
603 | perf_evsel__enable(counter, ncpus, nthreads); |
604 | } | |
605 | } | |
606 | ||
f33cbe72 | 607 | static volatile int workload_exec_errno; |
6af206fd ACM |
608 | |
609 | /* | |
610 | * perf_evlist__prepare_workload will send a SIGUSR1 | |
611 | * if the fork fails, since we asked by setting its | |
612 | * want_signal to true. | |
613 | */ | |
f33cbe72 ACM |
614 | static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, |
615 | void *ucontext __maybe_unused) | |
6af206fd | 616 | { |
f33cbe72 | 617 | workload_exec_errno = info->si_value.sival_int; |
6af206fd ACM |
618 | } |
619 | ||
acf28922 | 620 | static int __run_perf_stat(int argc, const char **argv) |
42202dd5 | 621 | { |
56e52e85 | 622 | char msg[512]; |
42202dd5 | 623 | unsigned long long t0, t1; |
cac21425 | 624 | struct perf_evsel *counter; |
13370a9b | 625 | struct timespec ts; |
410136f5 | 626 | size_t l; |
42202dd5 | 627 | int status = 0; |
6be2850e | 628 | const bool forks = (argc > 0); |
42202dd5 | 629 | |
13370a9b SE |
630 | if (interval) { |
631 | ts.tv_sec = interval / 1000; | |
632 | ts.tv_nsec = (interval % 1000) * 1000000; | |
633 | } else { | |
634 | ts.tv_sec = 1; | |
635 | ts.tv_nsec = 0; | |
636 | } | |
637 | ||
60666c63 | 638 | if (forks) { |
735f7e0b ACM |
639 | if (perf_evlist__prepare_workload(evsel_list, &target, argv, false, |
640 | workload_exec_failed_signal) < 0) { | |
acf28922 NK |
641 | perror("failed to prepare workload"); |
642 | return -1; | |
60666c63 | 643 | } |
d20a47e7 | 644 | child_pid = evsel_list->workload.pid; |
051ae7f7 PM |
645 | } |
646 | ||
6a4bb04c | 647 | if (group) |
63dab225 | 648 | perf_evlist__set_leader(evsel_list); |
6a4bb04c | 649 | |
0050f7aa | 650 | evlist__for_each(evsel_list, counter) { |
cac21425 | 651 | if (create_perf_stat_counter(counter) < 0) { |
979987a5 DA |
652 | /* |
653 | * PPC returns ENXIO for HW counters until 2.6.37 | |
654 | * (behavior changed with commit b0a873e). | |
655 | */ | |
38f6ae1e | 656 | if (errno == EINVAL || errno == ENOSYS || |
979987a5 DA |
657 | errno == ENOENT || errno == EOPNOTSUPP || |
658 | errno == ENXIO) { | |
c63ca0c0 DA |
659 | if (verbose) |
660 | ui__warning("%s event is not supported by the kernel.\n", | |
7289f83c | 661 | perf_evsel__name(counter)); |
2cee77c4 | 662 | counter->supported = false; |
ede70290 | 663 | continue; |
c63ca0c0 | 664 | } |
ede70290 | 665 | |
56e52e85 ACM |
666 | perf_evsel__open_strerror(counter, &target, |
667 | errno, msg, sizeof(msg)); | |
668 | ui__error("%s\n", msg); | |
669 | ||
48290609 ACM |
670 | if (child_pid != -1) |
671 | kill(child_pid, SIGTERM); | |
fceda7fe | 672 | |
48290609 ACM |
673 | return -1; |
674 | } | |
2cee77c4 | 675 | counter->supported = true; |
410136f5 SE |
676 | |
677 | l = strlen(counter->unit); | |
678 | if (l > unit_width) | |
679 | unit_width = l; | |
084ab9f8 | 680 | } |
42202dd5 | 681 | |
1491a632 | 682 | if (perf_evlist__apply_filters(evsel_list)) { |
cfd748ae | 683 | error("failed to set filter with %d (%s)\n", errno, |
759e612b | 684 | strerror_r(errno, msg, sizeof(msg))); |
cfd748ae FW |
685 | return -1; |
686 | } | |
687 | ||
42202dd5 IM |
688 | /* |
689 | * Enable counters and exec the command: | |
690 | */ | |
691 | t0 = rdclock(); | |
13370a9b | 692 | clock_gettime(CLOCK_MONOTONIC, &ref_time); |
42202dd5 | 693 | |
60666c63 | 694 | if (forks) { |
acf28922 | 695 | perf_evlist__start_workload(evsel_list); |
41191688 | 696 | handle_initial_delay(); |
acf28922 | 697 | |
13370a9b SE |
698 | if (interval) { |
699 | while (!waitpid(child_pid, &status, WNOHANG)) { | |
700 | nanosleep(&ts, NULL); | |
701 | print_interval(); | |
702 | } | |
703 | } | |
60666c63 | 704 | wait(&status); |
6af206fd | 705 | |
f33cbe72 ACM |
706 | if (workload_exec_errno) { |
707 | const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg)); | |
708 | pr_err("Workload failed: %s\n", emsg); | |
6af206fd | 709 | return -1; |
f33cbe72 | 710 | } |
6af206fd | 711 | |
33e49ea7 AK |
712 | if (WIFSIGNALED(status)) |
713 | psignal(WTERMSIG(status), argv[0]); | |
60666c63 | 714 | } else { |
41191688 | 715 | handle_initial_delay(); |
13370a9b SE |
716 | while (!done) { |
717 | nanosleep(&ts, NULL); | |
718 | if (interval) | |
719 | print_interval(); | |
720 | } | |
60666c63 | 721 | } |
42202dd5 | 722 | |
42202dd5 IM |
723 | t1 = rdclock(); |
724 | ||
9e9772c4 | 725 | update_stats(&walltime_nsecs_stats, t1 - t0); |
42202dd5 | 726 | |
86ee6e18 | 727 | if (aggr_mode == AGGR_GLOBAL) { |
0050f7aa | 728 | evlist__for_each(evsel_list, counter) { |
f5b4a9c3 | 729 | read_counter_aggr(counter); |
7ae92e74 | 730 | perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), |
b3a319d5 | 731 | thread_map__nr(evsel_list->threads)); |
c52b12ed | 732 | } |
86ee6e18 | 733 | } else { |
0050f7aa | 734 | evlist__for_each(evsel_list, counter) { |
86ee6e18 SE |
735 | read_counter(counter); |
736 | perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1); | |
737 | } | |
f5b4a9c3 | 738 | } |
c52b12ed | 739 | |
42202dd5 IM |
740 | return WEXITSTATUS(status); |
741 | } | |
742 | ||
41cde476 | 743 | static int run_perf_stat(int argc, const char **argv) |
1f16c575 PZ |
744 | { |
745 | int ret; | |
746 | ||
747 | if (pre_cmd) { | |
748 | ret = system(pre_cmd); | |
749 | if (ret) | |
750 | return ret; | |
751 | } | |
752 | ||
753 | if (sync_run) | |
754 | sync(); | |
755 | ||
756 | ret = __run_perf_stat(argc, argv); | |
757 | if (ret) | |
758 | return ret; | |
759 | ||
760 | if (post_cmd) { | |
761 | ret = system(post_cmd); | |
762 | if (ret) | |
763 | return ret; | |
764 | } | |
765 | ||
766 | return ret; | |
767 | } | |
768 | ||
f99844cb IM |
769 | static void print_noise_pct(double total, double avg) |
770 | { | |
0007ecea | 771 | double pct = rel_stddev_stats(total, avg); |
f99844cb | 772 | |
3ae9a34d | 773 | if (csv_output) |
4aa9015f | 774 | fprintf(output, "%s%.2f%%", csv_sep, pct); |
a1bca6cc | 775 | else if (pct) |
4aa9015f | 776 | fprintf(output, " ( +-%6.2f%% )", pct); |
f99844cb IM |
777 | } |
778 | ||
69aad6f1 | 779 | static void print_noise(struct perf_evsel *evsel, double avg) |
42202dd5 | 780 | { |
69aad6f1 ACM |
781 | struct perf_stat *ps; |
782 | ||
849abde9 PZ |
783 | if (run_count == 1) |
784 | return; | |
785 | ||
69aad6f1 | 786 | ps = evsel->priv; |
f99844cb | 787 | print_noise_pct(stddev_stats(&ps->res_stats[0]), avg); |
42202dd5 IM |
788 | } |
789 | ||
12c08a9f | 790 | static void aggr_printout(struct perf_evsel *evsel, int id, int nr) |
44175b6f | 791 | { |
86ee6e18 | 792 | switch (aggr_mode) { |
12c08a9f SE |
793 | case AGGR_CORE: |
794 | fprintf(output, "S%d-C%*d%s%*d%s", | |
795 | cpu_map__id_to_socket(id), | |
796 | csv_output ? 0 : -8, | |
797 | cpu_map__id_to_cpu(id), | |
798 | csv_sep, | |
799 | csv_output ? 0 : 4, | |
800 | nr, | |
801 | csv_sep); | |
802 | break; | |
86ee6e18 SE |
803 | case AGGR_SOCKET: |
804 | fprintf(output, "S%*d%s%*d%s", | |
d7e7a451 | 805 | csv_output ? 0 : -5, |
12c08a9f | 806 | id, |
d7e7a451 SE |
807 | csv_sep, |
808 | csv_output ? 0 : 4, | |
809 | nr, | |
810 | csv_sep); | |
86ee6e18 SE |
811 | break; |
812 | case AGGR_NONE: | |
813 | fprintf(output, "CPU%*d%s", | |
d7470b6a | 814 | csv_output ? 0 : -4, |
12c08a9f | 815 | perf_evsel__cpus(evsel)->map[id], csv_sep); |
86ee6e18 SE |
816 | break; |
817 | case AGGR_GLOBAL: | |
818 | default: | |
819 | break; | |
820 | } | |
821 | } | |
822 | ||
da88c7f7 | 823 | static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg) |
86ee6e18 SE |
824 | { |
825 | double msecs = avg / 1e6; | |
410136f5 | 826 | const char *fmt_v, *fmt_n; |
4bbe5a61 | 827 | char name[25]; |
86ee6e18 | 828 | |
410136f5 SE |
829 | fmt_v = csv_output ? "%.6f%s" : "%18.6f%s"; |
830 | fmt_n = csv_output ? "%s" : "%-25s"; | |
831 | ||
da88c7f7 | 832 | aggr_printout(evsel, id, nr); |
d7470b6a | 833 | |
4bbe5a61 DA |
834 | scnprintf(name, sizeof(name), "%s%s", |
835 | perf_evsel__name(evsel), csv_output ? "" : " (msec)"); | |
410136f5 SE |
836 | |
837 | fprintf(output, fmt_v, msecs, csv_sep); | |
838 | ||
839 | if (csv_output) | |
840 | fprintf(output, "%s%s", evsel->unit, csv_sep); | |
841 | else | |
842 | fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep); | |
843 | ||
844 | fprintf(output, fmt_n, name); | |
d7470b6a | 845 | |
023695d9 | 846 | if (evsel->cgrp) |
4aa9015f | 847 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); |
023695d9 | 848 | |
13370a9b | 849 | if (csv_output || interval) |
d7470b6a | 850 | return; |
44175b6f | 851 | |
daec78a0 | 852 | if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) |
4aa9015f SE |
853 | fprintf(output, " # %8.3f CPUs utilized ", |
854 | avg / avg_stats(&walltime_nsecs_stats)); | |
9dac6a29 NK |
855 | else |
856 | fprintf(output, " "); | |
44175b6f IM |
857 | } |
858 | ||
15e6392f NK |
859 | /* used for get_ratio_color() */ |
860 | enum grc_type { | |
861 | GRC_STALLED_CYCLES_FE, | |
862 | GRC_STALLED_CYCLES_BE, | |
863 | GRC_CACHE_MISSES, | |
864 | GRC_MAX_NR | |
865 | }; | |
866 | ||
867 | static const char *get_ratio_color(enum grc_type type, double ratio) | |
868 | { | |
869 | static const double grc_table[GRC_MAX_NR][3] = { | |
870 | [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 }, | |
871 | [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 }, | |
872 | [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 }, | |
873 | }; | |
874 | const char *color = PERF_COLOR_NORMAL; | |
875 | ||
876 | if (ratio > grc_table[type][0]) | |
877 | color = PERF_COLOR_RED; | |
878 | else if (ratio > grc_table[type][1]) | |
879 | color = PERF_COLOR_MAGENTA; | |
880 | else if (ratio > grc_table[type][2]) | |
881 | color = PERF_COLOR_YELLOW; | |
882 | ||
883 | return color; | |
884 | } | |
885 | ||
1d037ca1 IT |
886 | static void print_stalled_cycles_frontend(int cpu, |
887 | struct perf_evsel *evsel | |
888 | __maybe_unused, double avg) | |
d3d1e86d IM |
889 | { |
890 | double total, ratio = 0.0; | |
891 | const char *color; | |
892 | ||
893 | total = avg_stats(&runtime_cycles_stats[cpu]); | |
894 | ||
895 | if (total) | |
896 | ratio = avg / total * 100.0; | |
897 | ||
15e6392f | 898 | color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio); |
d3d1e86d | 899 | |
4aa9015f SE |
900 | fprintf(output, " # "); |
901 | color_fprintf(output, color, "%6.2f%%", ratio); | |
902 | fprintf(output, " frontend cycles idle "); | |
d3d1e86d IM |
903 | } |
904 | ||
1d037ca1 IT |
905 | static void print_stalled_cycles_backend(int cpu, |
906 | struct perf_evsel *evsel | |
907 | __maybe_unused, double avg) | |
a5d243d0 IM |
908 | { |
909 | double total, ratio = 0.0; | |
910 | const char *color; | |
911 | ||
912 | total = avg_stats(&runtime_cycles_stats[cpu]); | |
913 | ||
914 | if (total) | |
915 | ratio = avg / total * 100.0; | |
916 | ||
15e6392f | 917 | color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio); |
a5d243d0 | 918 | |
4aa9015f SE |
919 | fprintf(output, " # "); |
920 | color_fprintf(output, color, "%6.2f%%", ratio); | |
921 | fprintf(output, " backend cycles idle "); | |
a5d243d0 IM |
922 | } |
923 | ||
1d037ca1 IT |
924 | static void print_branch_misses(int cpu, |
925 | struct perf_evsel *evsel __maybe_unused, | |
926 | double avg) | |
c78df6c1 IM |
927 | { |
928 | double total, ratio = 0.0; | |
929 | const char *color; | |
930 | ||
931 | total = avg_stats(&runtime_branches_stats[cpu]); | |
932 | ||
933 | if (total) | |
934 | ratio = avg / total * 100.0; | |
935 | ||
15e6392f | 936 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c78df6c1 | 937 | |
4aa9015f SE |
938 | fprintf(output, " # "); |
939 | color_fprintf(output, color, "%6.2f%%", ratio); | |
940 | fprintf(output, " of all branches "); | |
c78df6c1 IM |
941 | } |
942 | ||
1d037ca1 IT |
943 | static void print_l1_dcache_misses(int cpu, |
944 | struct perf_evsel *evsel __maybe_unused, | |
945 | double avg) | |
8bb6c79f IM |
946 | { |
947 | double total, ratio = 0.0; | |
948 | const char *color; | |
949 | ||
950 | total = avg_stats(&runtime_l1_dcache_stats[cpu]); | |
951 | ||
952 | if (total) | |
953 | ratio = avg / total * 100.0; | |
954 | ||
15e6392f | 955 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
8bb6c79f | 956 | |
4aa9015f SE |
957 | fprintf(output, " # "); |
958 | color_fprintf(output, color, "%6.2f%%", ratio); | |
959 | fprintf(output, " of all L1-dcache hits "); | |
8bb6c79f IM |
960 | } |
961 | ||
1d037ca1 IT |
962 | static void print_l1_icache_misses(int cpu, |
963 | struct perf_evsel *evsel __maybe_unused, | |
964 | double avg) | |
c3305257 IM |
965 | { |
966 | double total, ratio = 0.0; | |
967 | const char *color; | |
968 | ||
969 | total = avg_stats(&runtime_l1_icache_stats[cpu]); | |
970 | ||
971 | if (total) | |
972 | ratio = avg / total * 100.0; | |
973 | ||
15e6392f | 974 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 975 | |
4aa9015f SE |
976 | fprintf(output, " # "); |
977 | color_fprintf(output, color, "%6.2f%%", ratio); | |
978 | fprintf(output, " of all L1-icache hits "); | |
c3305257 IM |
979 | } |
980 | ||
1d037ca1 IT |
981 | static void print_dtlb_cache_misses(int cpu, |
982 | struct perf_evsel *evsel __maybe_unused, | |
983 | double avg) | |
c3305257 IM |
984 | { |
985 | double total, ratio = 0.0; | |
986 | const char *color; | |
987 | ||
988 | total = avg_stats(&runtime_dtlb_cache_stats[cpu]); | |
989 | ||
990 | if (total) | |
991 | ratio = avg / total * 100.0; | |
992 | ||
15e6392f | 993 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 994 | |
4aa9015f SE |
995 | fprintf(output, " # "); |
996 | color_fprintf(output, color, "%6.2f%%", ratio); | |
997 | fprintf(output, " of all dTLB cache hits "); | |
c3305257 IM |
998 | } |
999 | ||
1d037ca1 IT |
1000 | static void print_itlb_cache_misses(int cpu, |
1001 | struct perf_evsel *evsel __maybe_unused, | |
1002 | double avg) | |
c3305257 IM |
1003 | { |
1004 | double total, ratio = 0.0; | |
1005 | const char *color; | |
1006 | ||
1007 | total = avg_stats(&runtime_itlb_cache_stats[cpu]); | |
1008 | ||
1009 | if (total) | |
1010 | ratio = avg / total * 100.0; | |
1011 | ||
15e6392f | 1012 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 1013 | |
4aa9015f SE |
1014 | fprintf(output, " # "); |
1015 | color_fprintf(output, color, "%6.2f%%", ratio); | |
1016 | fprintf(output, " of all iTLB cache hits "); | |
c3305257 IM |
1017 | } |
1018 | ||
1d037ca1 IT |
1019 | static void print_ll_cache_misses(int cpu, |
1020 | struct perf_evsel *evsel __maybe_unused, | |
1021 | double avg) | |
c3305257 IM |
1022 | { |
1023 | double total, ratio = 0.0; | |
1024 | const char *color; | |
1025 | ||
1026 | total = avg_stats(&runtime_ll_cache_stats[cpu]); | |
1027 | ||
1028 | if (total) | |
1029 | ratio = avg / total * 100.0; | |
1030 | ||
15e6392f | 1031 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); |
c3305257 | 1032 | |
4aa9015f SE |
1033 | fprintf(output, " # "); |
1034 | color_fprintf(output, color, "%6.2f%%", ratio); | |
1035 | fprintf(output, " of all LL-cache hits "); | |
c3305257 IM |
1036 | } |
1037 | ||
da88c7f7 | 1038 | static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg) |
44175b6f | 1039 | { |
4cabc3d1 | 1040 | double total, ratio = 0.0, total2; |
410136f5 | 1041 | double sc = evsel->scale; |
d7470b6a | 1042 | const char *fmt; |
da88c7f7 | 1043 | int cpu = cpu_map__id_to_cpu(id); |
d7470b6a | 1044 | |
410136f5 SE |
1045 | if (csv_output) { |
1046 | fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s"; | |
1047 | } else { | |
1048 | if (big_num) | |
1049 | fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s"; | |
1050 | else | |
1051 | fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s"; | |
1052 | } | |
f5b4a9c3 | 1053 | |
da88c7f7 | 1054 | aggr_printout(evsel, id, nr); |
86ee6e18 SE |
1055 | |
1056 | if (aggr_mode == AGGR_GLOBAL) | |
f5b4a9c3 | 1057 | cpu = 0; |
c7f7fea3 | 1058 | |
410136f5 SE |
1059 | fprintf(output, fmt, avg, csv_sep); |
1060 | ||
1061 | if (evsel->unit) | |
1062 | fprintf(output, "%-*s%s", | |
1063 | csv_output ? 0 : unit_width, | |
1064 | evsel->unit, csv_sep); | |
1065 | ||
1066 | fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel)); | |
d7470b6a | 1067 | |
023695d9 | 1068 | if (evsel->cgrp) |
4aa9015f | 1069 | fprintf(output, "%s%s", csv_sep, evsel->cgrp->name); |
023695d9 | 1070 | |
13370a9b | 1071 | if (csv_output || interval) |
d7470b6a | 1072 | return; |
44175b6f | 1073 | |
daec78a0 | 1074 | if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) { |
f5b4a9c3 | 1075 | total = avg_stats(&runtime_cycles_stats[cpu]); |
3e7a0817 | 1076 | if (total) { |
c7f7fea3 | 1077 | ratio = avg / total; |
3e7a0817 RR |
1078 | fprintf(output, " # %5.2f insns per cycle ", ratio); |
1079 | } | |
d3d1e86d IM |
1080 | total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]); |
1081 | total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu])); | |
481f988a IM |
1082 | |
1083 | if (total && avg) { | |
1084 | ratio = total / avg; | |
410136f5 SE |
1085 | fprintf(output, "\n"); |
1086 | if (aggr_mode == AGGR_NONE) | |
1087 | fprintf(output, " "); | |
1088 | fprintf(output, " # %5.2f stalled cycles per insn", ratio); | |
481f988a IM |
1089 | } |
1090 | ||
daec78a0 | 1091 | } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) && |
f5b4a9c3 | 1092 | runtime_branches_stats[cpu].n != 0) { |
c78df6c1 | 1093 | print_branch_misses(cpu, evsel, avg); |
8bb6c79f IM |
1094 | } else if ( |
1095 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
1096 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D | | |
1097 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
1098 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
c6264def | 1099 | runtime_l1_dcache_stats[cpu].n != 0) { |
8bb6c79f | 1100 | print_l1_dcache_misses(cpu, evsel, avg); |
c3305257 IM |
1101 | } else if ( |
1102 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
1103 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I | | |
1104 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
1105 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
1106 | runtime_l1_icache_stats[cpu].n != 0) { | |
1107 | print_l1_icache_misses(cpu, evsel, avg); | |
1108 | } else if ( | |
1109 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
1110 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB | | |
1111 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
1112 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
1113 | runtime_dtlb_cache_stats[cpu].n != 0) { | |
1114 | print_dtlb_cache_misses(cpu, evsel, avg); | |
1115 | } else if ( | |
1116 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
1117 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB | | |
1118 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
1119 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
1120 | runtime_itlb_cache_stats[cpu].n != 0) { | |
1121 | print_itlb_cache_misses(cpu, evsel, avg); | |
1122 | } else if ( | |
1123 | evsel->attr.type == PERF_TYPE_HW_CACHE && | |
1124 | evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL | | |
1125 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | | |
1126 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) && | |
1127 | runtime_ll_cache_stats[cpu].n != 0) { | |
1128 | print_ll_cache_misses(cpu, evsel, avg); | |
d58f4c82 IM |
1129 | } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) && |
1130 | runtime_cacherefs_stats[cpu].n != 0) { | |
1131 | total = avg_stats(&runtime_cacherefs_stats[cpu]); | |
1132 | ||
1133 | if (total) | |
1134 | ratio = avg * 100 / total; | |
1135 | ||
4aa9015f | 1136 | fprintf(output, " # %8.3f %% of all cache refs ", ratio); |
d58f4c82 | 1137 | |
d3d1e86d IM |
1138 | } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) { |
1139 | print_stalled_cycles_frontend(cpu, evsel, avg); | |
129c04cb | 1140 | } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) { |
d3d1e86d | 1141 | print_stalled_cycles_backend(cpu, evsel, avg); |
481f988a | 1142 | } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) { |
f5b4a9c3 | 1143 | total = avg_stats(&runtime_nsecs_stats[cpu]); |
c7f7fea3 | 1144 | |
c458fe62 RR |
1145 | if (total) { |
1146 | ratio = avg / total; | |
1147 | fprintf(output, " # %8.3f GHz ", ratio); | |
1148 | } | |
4cabc3d1 AK |
1149 | } else if (transaction_run && |
1150 | perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) { | |
1151 | total = avg_stats(&runtime_cycles_stats[cpu]); | |
1152 | if (total) | |
1153 | fprintf(output, | |
1154 | " # %5.2f%% transactional cycles ", | |
1155 | 100.0 * (avg / total)); | |
1156 | } else if (transaction_run && | |
1157 | perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) { | |
1158 | total = avg_stats(&runtime_cycles_stats[cpu]); | |
1159 | total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]); | |
1160 | if (total2 < avg) | |
1161 | total2 = avg; | |
1162 | if (total) | |
1163 | fprintf(output, | |
1164 | " # %5.2f%% aborted cycles ", | |
1165 | 100.0 * ((total2-avg) / total)); | |
1166 | } else if (transaction_run && | |
1167 | perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) && | |
1168 | avg > 0 && | |
1169 | runtime_cycles_in_tx_stats[cpu].n != 0) { | |
1170 | total = avg_stats(&runtime_cycles_in_tx_stats[cpu]); | |
1171 | ||
1172 | if (total) | |
1173 | ratio = total / avg; | |
1174 | ||
1175 | fprintf(output, " # %8.0f cycles / transaction ", ratio); | |
1176 | } else if (transaction_run && | |
1177 | perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) && | |
1178 | avg > 0 && | |
1179 | runtime_cycles_in_tx_stats[cpu].n != 0) { | |
1180 | total = avg_stats(&runtime_cycles_in_tx_stats[cpu]); | |
1181 | ||
1182 | if (total) | |
1183 | ratio = total / avg; | |
1184 | ||
1185 | fprintf(output, " # %8.0f cycles / elision ", ratio); | |
481f988a | 1186 | } else if (runtime_nsecs_stats[cpu].n != 0) { |
5fde2523 NK |
1187 | char unit = 'M'; |
1188 | ||
481f988a | 1189 | total = avg_stats(&runtime_nsecs_stats[cpu]); |
11ba2b85 IM |
1190 | |
1191 | if (total) | |
481f988a | 1192 | ratio = 1000.0 * avg / total; |
5fde2523 NK |
1193 | if (ratio < 0.001) { |
1194 | ratio *= 1000; | |
1195 | unit = 'K'; | |
1196 | } | |
11ba2b85 | 1197 | |
5fde2523 | 1198 | fprintf(output, " # %8.3f %c/sec ", ratio, unit); |
a5d243d0 | 1199 | } else { |
4aa9015f | 1200 | fprintf(output, " "); |
44175b6f | 1201 | } |
44175b6f IM |
1202 | } |
1203 | ||
86ee6e18 | 1204 | static void print_aggr(char *prefix) |
d7e7a451 SE |
1205 | { |
1206 | struct perf_evsel *counter; | |
582ec082 | 1207 | int cpu, cpu2, s, s2, id, nr; |
410136f5 | 1208 | double uval; |
d7e7a451 | 1209 | u64 ena, run, val; |
d7e7a451 | 1210 | |
86ee6e18 | 1211 | if (!(aggr_map || aggr_get_id)) |
d7e7a451 SE |
1212 | return; |
1213 | ||
86ee6e18 SE |
1214 | for (s = 0; s < aggr_map->nr; s++) { |
1215 | id = aggr_map->map[s]; | |
0050f7aa | 1216 | evlist__for_each(evsel_list, counter) { |
d7e7a451 SE |
1217 | val = ena = run = 0; |
1218 | nr = 0; | |
1219 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { | |
582ec082 SE |
1220 | cpu2 = perf_evsel__cpus(counter)->map[cpu]; |
1221 | s2 = aggr_get_id(evsel_list->cpus, cpu2); | |
86ee6e18 | 1222 | if (s2 != id) |
d7e7a451 SE |
1223 | continue; |
1224 | val += counter->counts->cpu[cpu].val; | |
1225 | ena += counter->counts->cpu[cpu].ena; | |
1226 | run += counter->counts->cpu[cpu].run; | |
1227 | nr++; | |
1228 | } | |
1229 | if (prefix) | |
1230 | fprintf(output, "%s", prefix); | |
1231 | ||
1232 | if (run == 0 || ena == 0) { | |
582ec082 | 1233 | aggr_printout(counter, id, nr); |
86ee6e18 | 1234 | |
410136f5 | 1235 | fprintf(output, "%*s%s", |
d7e7a451 SE |
1236 | csv_output ? 0 : 18, |
1237 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, | |
410136f5 SE |
1238 | csv_sep); |
1239 | ||
1240 | fprintf(output, "%-*s%s", | |
1241 | csv_output ? 0 : unit_width, | |
1242 | counter->unit, csv_sep); | |
1243 | ||
1244 | fprintf(output, "%*s", | |
1245 | csv_output ? 0 : -25, | |
d7e7a451 | 1246 | perf_evsel__name(counter)); |
86ee6e18 | 1247 | |
d7e7a451 SE |
1248 | if (counter->cgrp) |
1249 | fprintf(output, "%s%s", | |
1250 | csv_sep, counter->cgrp->name); | |
1251 | ||
1252 | fputc('\n', output); | |
1253 | continue; | |
1254 | } | |
410136f5 | 1255 | uval = val * counter->scale; |
d7e7a451 SE |
1256 | |
1257 | if (nsec_counter(counter)) | |
410136f5 | 1258 | nsec_printout(id, nr, counter, uval); |
d7e7a451 | 1259 | else |
410136f5 | 1260 | abs_printout(id, nr, counter, uval); |
d7e7a451 SE |
1261 | |
1262 | if (!csv_output) { | |
1263 | print_noise(counter, 1.0); | |
1264 | ||
1265 | if (run != ena) | |
1266 | fprintf(output, " (%.2f%%)", | |
1267 | 100.0 * run / ena); | |
1268 | } | |
1269 | fputc('\n', output); | |
1270 | } | |
1271 | } | |
1272 | } | |
1273 | ||
2996f5dd IM |
1274 | /* |
1275 | * Print out the results of a single counter: | |
f5b4a9c3 | 1276 | * aggregated counts in system-wide mode |
2996f5dd | 1277 | */ |
13370a9b | 1278 | static void print_counter_aggr(struct perf_evsel *counter, char *prefix) |
2996f5dd | 1279 | { |
69aad6f1 ACM |
1280 | struct perf_stat *ps = counter->priv; |
1281 | double avg = avg_stats(&ps->res_stats[0]); | |
c52b12ed | 1282 | int scaled = counter->counts->scaled; |
410136f5 | 1283 | double uval; |
2996f5dd | 1284 | |
13370a9b SE |
1285 | if (prefix) |
1286 | fprintf(output, "%s", prefix); | |
1287 | ||
2996f5dd | 1288 | if (scaled == -1) { |
410136f5 | 1289 | fprintf(output, "%*s%s", |
d7470b6a | 1290 | csv_output ? 0 : 18, |
2cee77c4 | 1291 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, |
410136f5 SE |
1292 | csv_sep); |
1293 | fprintf(output, "%-*s%s", | |
1294 | csv_output ? 0 : unit_width, | |
1295 | counter->unit, csv_sep); | |
1296 | fprintf(output, "%*s", | |
1297 | csv_output ? 0 : -25, | |
7289f83c | 1298 | perf_evsel__name(counter)); |
023695d9 SE |
1299 | |
1300 | if (counter->cgrp) | |
4aa9015f | 1301 | fprintf(output, "%s%s", csv_sep, counter->cgrp->name); |
023695d9 | 1302 | |
4aa9015f | 1303 | fputc('\n', output); |
2996f5dd IM |
1304 | return; |
1305 | } | |
c04f5e5d | 1306 | |
410136f5 SE |
1307 | uval = avg * counter->scale; |
1308 | ||
44175b6f | 1309 | if (nsec_counter(counter)) |
410136f5 | 1310 | nsec_printout(-1, 0, counter, uval); |
44175b6f | 1311 | else |
410136f5 | 1312 | abs_printout(-1, 0, counter, uval); |
849abde9 | 1313 | |
3ae9a34d ZH |
1314 | print_noise(counter, avg); |
1315 | ||
d7470b6a | 1316 | if (csv_output) { |
4aa9015f | 1317 | fputc('\n', output); |
d7470b6a SE |
1318 | return; |
1319 | } | |
1320 | ||
506d4bc8 PZ |
1321 | if (scaled) { |
1322 | double avg_enabled, avg_running; | |
1323 | ||
69aad6f1 ACM |
1324 | avg_enabled = avg_stats(&ps->res_stats[1]); |
1325 | avg_running = avg_stats(&ps->res_stats[2]); | |
d7c29318 | 1326 | |
4aa9015f | 1327 | fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled); |
506d4bc8 | 1328 | } |
4aa9015f | 1329 | fprintf(output, "\n"); |
c04f5e5d IM |
1330 | } |
1331 | ||
f5b4a9c3 SE |
1332 | /* |
1333 | * Print out the results of a single counter: | |
1334 | * does not use aggregated count in system-wide | |
1335 | */ | |
13370a9b | 1336 | static void print_counter(struct perf_evsel *counter, char *prefix) |
f5b4a9c3 SE |
1337 | { |
1338 | u64 ena, run, val; | |
410136f5 | 1339 | double uval; |
f5b4a9c3 SE |
1340 | int cpu; |
1341 | ||
7ae92e74 | 1342 | for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { |
c52b12ed ACM |
1343 | val = counter->counts->cpu[cpu].val; |
1344 | ena = counter->counts->cpu[cpu].ena; | |
1345 | run = counter->counts->cpu[cpu].run; | |
13370a9b SE |
1346 | |
1347 | if (prefix) | |
1348 | fprintf(output, "%s", prefix); | |
1349 | ||
f5b4a9c3 | 1350 | if (run == 0 || ena == 0) { |
410136f5 | 1351 | fprintf(output, "CPU%*d%s%*s%s", |
d7470b6a | 1352 | csv_output ? 0 : -4, |
7ae92e74 | 1353 | perf_evsel__cpus(counter)->map[cpu], csv_sep, |
d7470b6a | 1354 | csv_output ? 0 : 18, |
2cee77c4 | 1355 | counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, |
410136f5 SE |
1356 | csv_sep); |
1357 | ||
1358 | fprintf(output, "%-*s%s", | |
1359 | csv_output ? 0 : unit_width, | |
1360 | counter->unit, csv_sep); | |
1361 | ||
1362 | fprintf(output, "%*s", | |
1363 | csv_output ? 0 : -25, | |
1364 | perf_evsel__name(counter)); | |
f5b4a9c3 | 1365 | |
023695d9 | 1366 | if (counter->cgrp) |
4aa9015f SE |
1367 | fprintf(output, "%s%s", |
1368 | csv_sep, counter->cgrp->name); | |
023695d9 | 1369 | |
4aa9015f | 1370 | fputc('\n', output); |
f5b4a9c3 SE |
1371 | continue; |
1372 | } | |
1373 | ||
410136f5 SE |
1374 | uval = val * counter->scale; |
1375 | ||
f5b4a9c3 | 1376 | if (nsec_counter(counter)) |
410136f5 | 1377 | nsec_printout(cpu, 0, counter, uval); |
f5b4a9c3 | 1378 | else |
410136f5 | 1379 | abs_printout(cpu, 0, counter, uval); |
f5b4a9c3 | 1380 | |
d7470b6a SE |
1381 | if (!csv_output) { |
1382 | print_noise(counter, 1.0); | |
f5b4a9c3 | 1383 | |
c6264def | 1384 | if (run != ena) |
4aa9015f SE |
1385 | fprintf(output, " (%.2f%%)", |
1386 | 100.0 * run / ena); | |
f5b4a9c3 | 1387 | } |
4aa9015f | 1388 | fputc('\n', output); |
f5b4a9c3 SE |
1389 | } |
1390 | } | |
1391 | ||
42202dd5 IM |
1392 | static void print_stat(int argc, const char **argv) |
1393 | { | |
69aad6f1 ACM |
1394 | struct perf_evsel *counter; |
1395 | int i; | |
42202dd5 | 1396 | |
ddcacfa0 IM |
1397 | fflush(stdout); |
1398 | ||
d7470b6a | 1399 | if (!csv_output) { |
4aa9015f SE |
1400 | fprintf(output, "\n"); |
1401 | fprintf(output, " Performance counter stats for "); | |
62d3b617 DA |
1402 | if (target.system_wide) |
1403 | fprintf(output, "\'system wide"); | |
1404 | else if (target.cpu_list) | |
1405 | fprintf(output, "\'CPU(s) %s", target.cpu_list); | |
602ad878 | 1406 | else if (!target__has_task(&target)) { |
4aa9015f | 1407 | fprintf(output, "\'%s", argv[0]); |
d7470b6a | 1408 | for (i = 1; i < argc; i++) |
4aa9015f | 1409 | fprintf(output, " %s", argv[i]); |
20f946b4 NK |
1410 | } else if (target.pid) |
1411 | fprintf(output, "process id \'%s", target.pid); | |
d7470b6a | 1412 | else |
20f946b4 | 1413 | fprintf(output, "thread id \'%s", target.tid); |
44db76c8 | 1414 | |
4aa9015f | 1415 | fprintf(output, "\'"); |
d7470b6a | 1416 | if (run_count > 1) |
4aa9015f SE |
1417 | fprintf(output, " (%d runs)", run_count); |
1418 | fprintf(output, ":\n\n"); | |
d7470b6a | 1419 | } |
2996f5dd | 1420 | |
86ee6e18 | 1421 | switch (aggr_mode) { |
12c08a9f | 1422 | case AGGR_CORE: |
86ee6e18 SE |
1423 | case AGGR_SOCKET: |
1424 | print_aggr(NULL); | |
1425 | break; | |
1426 | case AGGR_GLOBAL: | |
0050f7aa | 1427 | evlist__for_each(evsel_list, counter) |
13370a9b | 1428 | print_counter_aggr(counter, NULL); |
86ee6e18 SE |
1429 | break; |
1430 | case AGGR_NONE: | |
0050f7aa | 1431 | evlist__for_each(evsel_list, counter) |
86ee6e18 SE |
1432 | print_counter(counter, NULL); |
1433 | break; | |
1434 | default: | |
1435 | break; | |
f5b4a9c3 | 1436 | } |
ddcacfa0 | 1437 | |
d7470b6a | 1438 | if (!csv_output) { |
c3305257 | 1439 | if (!null_run) |
4aa9015f SE |
1440 | fprintf(output, "\n"); |
1441 | fprintf(output, " %17.9f seconds time elapsed", | |
d7470b6a SE |
1442 | avg_stats(&walltime_nsecs_stats)/1e9); |
1443 | if (run_count > 1) { | |
4aa9015f | 1444 | fprintf(output, " "); |
f99844cb IM |
1445 | print_noise_pct(stddev_stats(&walltime_nsecs_stats), |
1446 | avg_stats(&walltime_nsecs_stats)); | |
d7470b6a | 1447 | } |
4aa9015f | 1448 | fprintf(output, "\n\n"); |
566747e6 | 1449 | } |
ddcacfa0 IM |
1450 | } |
1451 | ||
f7b7c26e PZ |
1452 | static volatile int signr = -1; |
1453 | ||
5242519b | 1454 | static void skip_signal(int signo) |
ddcacfa0 | 1455 | { |
13370a9b | 1456 | if ((child_pid == -1) || interval) |
60666c63 LW |
1457 | done = 1; |
1458 | ||
f7b7c26e | 1459 | signr = signo; |
d07f0b12 SE |
1460 | /* |
1461 | * render child_pid harmless | |
1462 | * won't send SIGTERM to a random | |
1463 | * process in case of race condition | |
1464 | * and fast PID recycling | |
1465 | */ | |
1466 | child_pid = -1; | |
f7b7c26e PZ |
1467 | } |
1468 | ||
1469 | static void sig_atexit(void) | |
1470 | { | |
d07f0b12 SE |
1471 | sigset_t set, oset; |
1472 | ||
1473 | /* | |
1474 | * avoid race condition with SIGCHLD handler | |
1475 | * in skip_signal() which is modifying child_pid | |
1476 | * goal is to avoid send SIGTERM to a random | |
1477 | * process | |
1478 | */ | |
1479 | sigemptyset(&set); | |
1480 | sigaddset(&set, SIGCHLD); | |
1481 | sigprocmask(SIG_BLOCK, &set, &oset); | |
1482 | ||
933da83a CW |
1483 | if (child_pid != -1) |
1484 | kill(child_pid, SIGTERM); | |
1485 | ||
d07f0b12 SE |
1486 | sigprocmask(SIG_SETMASK, &oset, NULL); |
1487 | ||
f7b7c26e PZ |
1488 | if (signr == -1) |
1489 | return; | |
1490 | ||
1491 | signal(signr, SIG_DFL); | |
1492 | kill(getpid(), signr); | |
5242519b IM |
1493 | } |
1494 | ||
1d037ca1 IT |
1495 | static int stat__set_big_num(const struct option *opt __maybe_unused, |
1496 | const char *s __maybe_unused, int unset) | |
d7470b6a SE |
1497 | { |
1498 | big_num_opt = unset ? 0 : 1; | |
1499 | return 0; | |
1500 | } | |
1501 | ||
86ee6e18 SE |
1502 | static int perf_stat_init_aggr_mode(void) |
1503 | { | |
1504 | switch (aggr_mode) { | |
1505 | case AGGR_SOCKET: | |
1506 | if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) { | |
1507 | perror("cannot build socket map"); | |
1508 | return -1; | |
1509 | } | |
1510 | aggr_get_id = cpu_map__get_socket; | |
1511 | break; | |
12c08a9f SE |
1512 | case AGGR_CORE: |
1513 | if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) { | |
1514 | perror("cannot build core map"); | |
1515 | return -1; | |
1516 | } | |
1517 | aggr_get_id = cpu_map__get_core; | |
1518 | break; | |
86ee6e18 SE |
1519 | case AGGR_NONE: |
1520 | case AGGR_GLOBAL: | |
1521 | default: | |
1522 | break; | |
1523 | } | |
1524 | return 0; | |
1525 | } | |
1526 | ||
4cabc3d1 AK |
1527 | static int setup_events(const char * const *attrs, unsigned len) |
1528 | { | |
1529 | unsigned i; | |
1530 | ||
1531 | for (i = 0; i < len; i++) { | |
1532 | if (parse_events(evsel_list, attrs[i])) | |
1533 | return -1; | |
1534 | } | |
1535 | return 0; | |
1536 | } | |
86ee6e18 | 1537 | |
2cba3ffb IM |
1538 | /* |
1539 | * Add default attributes, if there were no attributes specified or | |
1540 | * if -d/--detailed, -d -d or -d -d -d is used: | |
1541 | */ | |
1542 | static int add_default_attributes(void) | |
1543 | { | |
b070a547 ACM |
1544 | struct perf_event_attr default_attrs[] = { |
1545 | ||
1546 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, | |
1547 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, | |
1548 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, | |
1549 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, | |
1550 | ||
1551 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, | |
1552 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, | |
1553 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, | |
1554 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, | |
1555 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, | |
1556 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, | |
1557 | ||
1558 | }; | |
1559 | ||
1560 | /* | |
1561 | * Detailed stats (-d), covering the L1 and last level data caches: | |
1562 | */ | |
1563 | struct perf_event_attr detailed_attrs[] = { | |
1564 | ||
1565 | { .type = PERF_TYPE_HW_CACHE, | |
1566 | .config = | |
1567 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1568 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1569 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1570 | ||
1571 | { .type = PERF_TYPE_HW_CACHE, | |
1572 | .config = | |
1573 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1574 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1575 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1576 | ||
1577 | { .type = PERF_TYPE_HW_CACHE, | |
1578 | .config = | |
1579 | PERF_COUNT_HW_CACHE_LL << 0 | | |
1580 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1581 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1582 | ||
1583 | { .type = PERF_TYPE_HW_CACHE, | |
1584 | .config = | |
1585 | PERF_COUNT_HW_CACHE_LL << 0 | | |
1586 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1587 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1588 | }; | |
1589 | ||
1590 | /* | |
1591 | * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: | |
1592 | */ | |
1593 | struct perf_event_attr very_detailed_attrs[] = { | |
1594 | ||
1595 | { .type = PERF_TYPE_HW_CACHE, | |
1596 | .config = | |
1597 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
1598 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1599 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1600 | ||
1601 | { .type = PERF_TYPE_HW_CACHE, | |
1602 | .config = | |
1603 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
1604 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1605 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1606 | ||
1607 | { .type = PERF_TYPE_HW_CACHE, | |
1608 | .config = | |
1609 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1610 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1611 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1612 | ||
1613 | { .type = PERF_TYPE_HW_CACHE, | |
1614 | .config = | |
1615 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1616 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1617 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1618 | ||
1619 | { .type = PERF_TYPE_HW_CACHE, | |
1620 | .config = | |
1621 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
1622 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1623 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1624 | ||
1625 | { .type = PERF_TYPE_HW_CACHE, | |
1626 | .config = | |
1627 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
1628 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1629 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1630 | ||
1631 | }; | |
1632 | ||
1633 | /* | |
1634 | * Very, very detailed stats (-d -d -d), adding prefetch events: | |
1635 | */ | |
1636 | struct perf_event_attr very_very_detailed_attrs[] = { | |
1637 | ||
1638 | { .type = PERF_TYPE_HW_CACHE, | |
1639 | .config = | |
1640 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1641 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
1642 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1643 | ||
1644 | { .type = PERF_TYPE_HW_CACHE, | |
1645 | .config = | |
1646 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1647 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
1648 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1649 | }; | |
1650 | ||
2cba3ffb IM |
1651 | /* Set attrs if no event is selected and !null_run: */ |
1652 | if (null_run) | |
1653 | return 0; | |
1654 | ||
4cabc3d1 AK |
1655 | if (transaction_run) { |
1656 | int err; | |
1657 | if (pmu_have_event("cpu", "cycles-ct") && | |
1658 | pmu_have_event("cpu", "el-start")) | |
1659 | err = setup_events(transaction_attrs, | |
1660 | ARRAY_SIZE(transaction_attrs)); | |
1661 | else | |
1662 | err = setup_events(transaction_limited_attrs, | |
1663 | ARRAY_SIZE(transaction_limited_attrs)); | |
1664 | if (err < 0) { | |
1665 | fprintf(stderr, "Cannot set up transaction events\n"); | |
1666 | return -1; | |
1667 | } | |
1668 | return 0; | |
1669 | } | |
1670 | ||
2cba3ffb | 1671 | if (!evsel_list->nr_entries) { |
79695e1b | 1672 | if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0) |
50d08e47 | 1673 | return -1; |
2cba3ffb IM |
1674 | } |
1675 | ||
1676 | /* Detailed events get appended to the event list: */ | |
1677 | ||
1678 | if (detailed_run < 1) | |
1679 | return 0; | |
1680 | ||
1681 | /* Append detailed run extra attributes: */ | |
79695e1b | 1682 | if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) |
50d08e47 | 1683 | return -1; |
2cba3ffb IM |
1684 | |
1685 | if (detailed_run < 2) | |
1686 | return 0; | |
1687 | ||
1688 | /* Append very detailed run extra attributes: */ | |
79695e1b | 1689 | if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) |
50d08e47 | 1690 | return -1; |
2cba3ffb IM |
1691 | |
1692 | if (detailed_run < 3) | |
1693 | return 0; | |
1694 | ||
1695 | /* Append very, very detailed run extra attributes: */ | |
79695e1b | 1696 | return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); |
2cba3ffb IM |
1697 | } |
1698 | ||
1d037ca1 | 1699 | int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused) |
5242519b | 1700 | { |
1f16c575 | 1701 | bool append_file = false; |
b070a547 ACM |
1702 | int output_fd = 0; |
1703 | const char *output_name = NULL; | |
1704 | const struct option options[] = { | |
4cabc3d1 AK |
1705 | OPT_BOOLEAN('T', "transaction", &transaction_run, |
1706 | "hardware transaction statistics"), | |
b070a547 ACM |
1707 | OPT_CALLBACK('e', "event", &evsel_list, "event", |
1708 | "event selector. use 'perf list' to list available events", | |
1709 | parse_events_option), | |
1710 | OPT_CALLBACK(0, "filter", &evsel_list, "filter", | |
1711 | "event filter", parse_filter), | |
1712 | OPT_BOOLEAN('i', "no-inherit", &no_inherit, | |
1713 | "child tasks do not inherit counters"), | |
1714 | OPT_STRING('p', "pid", &target.pid, "pid", | |
1715 | "stat events on existing process id"), | |
1716 | OPT_STRING('t', "tid", &target.tid, "tid", | |
1717 | "stat events on existing thread id"), | |
1718 | OPT_BOOLEAN('a', "all-cpus", &target.system_wide, | |
1719 | "system-wide collection from all CPUs"), | |
1720 | OPT_BOOLEAN('g', "group", &group, | |
1721 | "put the counters into a counter group"), | |
1722 | OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"), | |
1723 | OPT_INCR('v', "verbose", &verbose, | |
1724 | "be more verbose (show counter open errors, etc)"), | |
1725 | OPT_INTEGER('r', "repeat", &run_count, | |
a7e191c3 | 1726 | "repeat command and print average + stddev (max: 100, forever: 0)"), |
b070a547 ACM |
1727 | OPT_BOOLEAN('n', "null", &null_run, |
1728 | "null run - dont start any counters"), | |
1729 | OPT_INCR('d', "detailed", &detailed_run, | |
1730 | "detailed run - start a lot of events"), | |
1731 | OPT_BOOLEAN('S', "sync", &sync_run, | |
1732 | "call sync() before starting a run"), | |
48000a1a | 1733 | OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, |
b070a547 ACM |
1734 | "print large numbers with thousands\' separators", |
1735 | stat__set_big_num), | |
1736 | OPT_STRING('C', "cpu", &target.cpu_list, "cpu", | |
1737 | "list of cpus to monitor in system-wide"), | |
86ee6e18 SE |
1738 | OPT_SET_UINT('A', "no-aggr", &aggr_mode, |
1739 | "disable CPU count aggregation", AGGR_NONE), | |
b070a547 ACM |
1740 | OPT_STRING('x', "field-separator", &csv_sep, "separator", |
1741 | "print counts with custom separator"), | |
1742 | OPT_CALLBACK('G', "cgroup", &evsel_list, "name", | |
1743 | "monitor event in cgroup name only", parse_cgroups), | |
1744 | OPT_STRING('o', "output", &output_name, "file", "output file name"), | |
1745 | OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), | |
1746 | OPT_INTEGER(0, "log-fd", &output_fd, | |
1747 | "log output to fd, instead of stderr"), | |
1f16c575 PZ |
1748 | OPT_STRING(0, "pre", &pre_cmd, "command", |
1749 | "command to run prior to the measured command"), | |
1750 | OPT_STRING(0, "post", &post_cmd, "command", | |
1751 | "command to run after to the measured command"), | |
13370a9b SE |
1752 | OPT_UINTEGER('I', "interval-print", &interval, |
1753 | "print counts at regular interval in ms (>= 100)"), | |
d4304958 | 1754 | OPT_SET_UINT(0, "per-socket", &aggr_mode, |
86ee6e18 | 1755 | "aggregate counts per processor socket", AGGR_SOCKET), |
12c08a9f SE |
1756 | OPT_SET_UINT(0, "per-core", &aggr_mode, |
1757 | "aggregate counts per physical processor core", AGGR_CORE), | |
41191688 AK |
1758 | OPT_UINTEGER('D', "delay", &initial_delay, |
1759 | "ms to wait before starting measurement after program start"), | |
b070a547 ACM |
1760 | OPT_END() |
1761 | }; | |
1762 | const char * const stat_usage[] = { | |
1763 | "perf stat [<options>] [<command>]", | |
1764 | NULL | |
1765 | }; | |
cc03c542 | 1766 | int status = -EINVAL, run_idx; |
4aa9015f | 1767 | const char *mode; |
42202dd5 | 1768 | |
5af52b51 SE |
1769 | setlocale(LC_ALL, ""); |
1770 | ||
334fe7a3 | 1771 | evsel_list = perf_evlist__new(); |
361c99a6 ACM |
1772 | if (evsel_list == NULL) |
1773 | return -ENOMEM; | |
1774 | ||
a0541234 AB |
1775 | argc = parse_options(argc, argv, options, stat_usage, |
1776 | PARSE_OPT_STOP_AT_NON_OPTION); | |
d7470b6a | 1777 | |
4aa9015f SE |
1778 | output = stderr; |
1779 | if (output_name && strcmp(output_name, "-")) | |
1780 | output = NULL; | |
1781 | ||
56f3bae7 JC |
1782 | if (output_name && output_fd) { |
1783 | fprintf(stderr, "cannot use both --output and --log-fd\n"); | |
cc03c542 NK |
1784 | parse_options_usage(stat_usage, options, "o", 1); |
1785 | parse_options_usage(NULL, options, "log-fd", 0); | |
1786 | goto out; | |
56f3bae7 | 1787 | } |
fc3e4d07 SE |
1788 | |
1789 | if (output_fd < 0) { | |
1790 | fprintf(stderr, "argument to --log-fd must be a > 0\n"); | |
cc03c542 NK |
1791 | parse_options_usage(stat_usage, options, "log-fd", 0); |
1792 | goto out; | |
fc3e4d07 SE |
1793 | } |
1794 | ||
4aa9015f SE |
1795 | if (!output) { |
1796 | struct timespec tm; | |
1797 | mode = append_file ? "a" : "w"; | |
1798 | ||
1799 | output = fopen(output_name, mode); | |
1800 | if (!output) { | |
1801 | perror("failed to create output file"); | |
fceda7fe | 1802 | return -1; |
4aa9015f SE |
1803 | } |
1804 | clock_gettime(CLOCK_REALTIME, &tm); | |
1805 | fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); | |
fc3e4d07 | 1806 | } else if (output_fd > 0) { |
56f3bae7 JC |
1807 | mode = append_file ? "a" : "w"; |
1808 | output = fdopen(output_fd, mode); | |
1809 | if (!output) { | |
1810 | perror("Failed opening logfd"); | |
1811 | return -errno; | |
1812 | } | |
4aa9015f SE |
1813 | } |
1814 | ||
d4ffd04d | 1815 | if (csv_sep) { |
d7470b6a | 1816 | csv_output = true; |
d4ffd04d JC |
1817 | if (!strcmp(csv_sep, "\\t")) |
1818 | csv_sep = "\t"; | |
1819 | } else | |
d7470b6a SE |
1820 | csv_sep = DEFAULT_SEPARATOR; |
1821 | ||
1822 | /* | |
1823 | * let the spreadsheet do the pretty-printing | |
1824 | */ | |
1825 | if (csv_output) { | |
61a9f324 | 1826 | /* User explicitly passed -B? */ |
d7470b6a SE |
1827 | if (big_num_opt == 1) { |
1828 | fprintf(stderr, "-B option not supported with -x\n"); | |
cc03c542 NK |
1829 | parse_options_usage(stat_usage, options, "B", 1); |
1830 | parse_options_usage(NULL, options, "x", 1); | |
1831 | goto out; | |
d7470b6a SE |
1832 | } else /* Nope, so disable big number formatting */ |
1833 | big_num = false; | |
1834 | } else if (big_num_opt == 0) /* User passed --no-big-num */ | |
1835 | big_num = false; | |
1836 | ||
602ad878 | 1837 | if (!argc && target__none(&target)) |
5242519b | 1838 | usage_with_options(stat_usage, options); |
ac3063bd | 1839 | |
a7e191c3 | 1840 | if (run_count < 0) { |
cc03c542 NK |
1841 | pr_err("Run count must be a positive number\n"); |
1842 | parse_options_usage(stat_usage, options, "r", 1); | |
1843 | goto out; | |
a7e191c3 FD |
1844 | } else if (run_count == 0) { |
1845 | forever = true; | |
1846 | run_count = 1; | |
1847 | } | |
ddcacfa0 | 1848 | |
023695d9 | 1849 | /* no_aggr, cgroup are for system-wide only */ |
602ad878 ACM |
1850 | if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) && |
1851 | !target__has_cpu(&target)) { | |
023695d9 SE |
1852 | fprintf(stderr, "both cgroup and no-aggregation " |
1853 | "modes only available in system-wide mode\n"); | |
1854 | ||
cc03c542 NK |
1855 | parse_options_usage(stat_usage, options, "G", 1); |
1856 | parse_options_usage(NULL, options, "A", 1); | |
1857 | parse_options_usage(NULL, options, "a", 1); | |
1858 | goto out; | |
d7e7a451 SE |
1859 | } |
1860 | ||
2cba3ffb IM |
1861 | if (add_default_attributes()) |
1862 | goto out; | |
ddcacfa0 | 1863 | |
602ad878 | 1864 | target__validate(&target); |
5c98d466 | 1865 | |
77a6f014 | 1866 | if (perf_evlist__create_maps(evsel_list, &target) < 0) { |
602ad878 | 1867 | if (target__has_task(&target)) { |
77a6f014 | 1868 | pr_err("Problems finding threads of monitor\n"); |
cc03c542 NK |
1869 | parse_options_usage(stat_usage, options, "p", 1); |
1870 | parse_options_usage(NULL, options, "t", 1); | |
602ad878 | 1871 | } else if (target__has_cpu(&target)) { |
77a6f014 | 1872 | perror("failed to parse CPUs map"); |
cc03c542 NK |
1873 | parse_options_usage(stat_usage, options, "C", 1); |
1874 | parse_options_usage(NULL, options, "a", 1); | |
1875 | } | |
1876 | goto out; | |
60d567e2 | 1877 | } |
13370a9b SE |
1878 | if (interval && interval < 100) { |
1879 | pr_err("print interval must be >= 100ms\n"); | |
cc03c542 | 1880 | parse_options_usage(stat_usage, options, "I", 1); |
03ad9747 | 1881 | goto out; |
13370a9b | 1882 | } |
c45c6ea2 | 1883 | |
d134ffb9 | 1884 | if (perf_evlist__alloc_stats(evsel_list, interval)) |
03ad9747 | 1885 | goto out; |
d6d901c2 | 1886 | |
86ee6e18 | 1887 | if (perf_stat_init_aggr_mode()) |
03ad9747 | 1888 | goto out; |
86ee6e18 | 1889 | |
58d7e993 IM |
1890 | /* |
1891 | * We dont want to block the signals - that would cause | |
1892 | * child tasks to inherit that and Ctrl-C would not work. | |
1893 | * What we want is for Ctrl-C to work in the exec()-ed | |
1894 | * task, but being ignored by perf stat itself: | |
1895 | */ | |
f7b7c26e | 1896 | atexit(sig_atexit); |
a7e191c3 FD |
1897 | if (!forever) |
1898 | signal(SIGINT, skip_signal); | |
13370a9b | 1899 | signal(SIGCHLD, skip_signal); |
58d7e993 IM |
1900 | signal(SIGALRM, skip_signal); |
1901 | signal(SIGABRT, skip_signal); | |
1902 | ||
42202dd5 | 1903 | status = 0; |
a7e191c3 | 1904 | for (run_idx = 0; forever || run_idx < run_count; run_idx++) { |
42202dd5 | 1905 | if (run_count != 1 && verbose) |
4aa9015f SE |
1906 | fprintf(output, "[ perf stat: executing run #%d ... ]\n", |
1907 | run_idx + 1); | |
f9cef0a9 | 1908 | |
42202dd5 | 1909 | status = run_perf_stat(argc, argv); |
a7e191c3 FD |
1910 | if (forever && status != -1) { |
1911 | print_stat(argc, argv); | |
d134ffb9 | 1912 | perf_stat__reset_stats(evsel_list); |
a7e191c3 | 1913 | } |
42202dd5 IM |
1914 | } |
1915 | ||
a7e191c3 | 1916 | if (!forever && status != -1 && !interval) |
084ab9f8 | 1917 | print_stat(argc, argv); |
d134ffb9 ACM |
1918 | |
1919 | perf_evlist__free_stats(evsel_list); | |
0015e2e1 ACM |
1920 | out: |
1921 | perf_evlist__delete(evsel_list); | |
42202dd5 | 1922 | return status; |
ddcacfa0 | 1923 | } |