projects / deliverable / linux.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
perf build-id: Fix filename size calculation
[deliverable/linux.git] / tools / perf / builtin-stat.c
CommitLineData
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 67static 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 */
86static 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 */
116static 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 */
159static 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
361c99a6
ACM		 176struct perf_evlist *evsel_list;
177
c0555642 178static bool system_wide = false;
3d632595 179static int run_idx = 0;
ddcacfa0 180
3d632595 181static int run_count = 1;
2e6cdf99 182static bool no_inherit = false;
c0555642 183static bool scale = true;
f5b4a9c3 184static bool no_aggr = false;
b52956c9
DA		 185static const char *target_pid;
186static const char *target_tid;
933da83a 187static pid_t child_pid = -1;
c0555642 188static bool null_run = false;
2cba3ffb 189static int detailed_run = 0;
f9cef0a9 190static bool sync_run = false;
201e0b06 191static bool big_num = true;
d7470b6a 192static int big_num_opt = -1;
c45c6ea2 193static const char *cpu_list;
d7470b6a
SE		 194static const char *csv_sep = NULL;
195static bool csv_output = false;
43bece79 196static bool group = false;
4aa9015f
SE		 197static const char *output_name = NULL;
198static FILE *output = NULL;
56f3bae7 199static int output_fd;
5af52b51 200
60666c63
LW		 201static volatile int done = 0;
202
506d4bc8
PZ		 203struct stats
204{
8a02631a 205 double n, mean, M2;
506d4bc8 206};
42202dd5 207
69aad6f1
ACM		 208struct perf_stat {
209 struct stats res_stats[3];
69aad6f1
ACM		 210};
211
c52b12ed 212static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
69aad6f1 213{
c52b12ed 214 evsel->priv = zalloc(sizeof(struct perf_stat));
69aad6f1
ACM		 215 return evsel->priv == NULL ? -ENOMEM : 0;
216}
217
218static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
219{
220 free(evsel->priv);
221 evsel->priv = NULL;
222}
223
9e9772c4
PZ		 224static void update_stats(struct stats *stats, u64 val)
225{
8a02631a 226 double delta;
9e9772c4 227
8a02631a
PZ		 228 stats->n++;
229 delta = val - stats->mean;
230 stats->mean += delta / stats->n;
231 stats->M2 += delta*(val - stats->mean);
9e9772c4
PZ		 232}
233
506d4bc8
PZ		 234static double avg_stats(struct stats *stats)
235{
8a02631a 236 return stats->mean;
506d4bc8 237}
42202dd5 238
506d4bc8 239/*
63d40deb
PZ		 240 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
241 *
8a02631a
PZ		 242 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
243 * s^2 = -------------------------------
244 * n - 1
63d40deb
PZ		 245 *
246 * http://en.wikipedia.org/wiki/Stddev
247 *
248 * The std dev of the mean is related to the std dev by:
249 *
250 * s
251 * s_mean = -------
252 * sqrt(n)
253 *
506d4bc8
PZ		 254 */
255static double stddev_stats(struct stats *stats)
256{
19f47402
JC		 257 double variance, variance_mean;
258
259 if (!stats->n)
260 return 0.0;
261
262 variance = stats->M2 / (stats->n - 1);
263 variance_mean = variance / stats->n;
42202dd5 264
63d40deb 265 return sqrt(variance_mean);
506d4bc8 266}
42202dd5 267
f5b4a9c3
SE		 268struct stats runtime_nsecs_stats[MAX_NR_CPUS];
269struct stats runtime_cycles_stats[MAX_NR_CPUS];
d3d1e86d
IM		 270struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
271struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
f5b4a9c3 272struct stats runtime_branches_stats[MAX_NR_CPUS];
d58f4c82 273struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
8bb6c79f 274struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
c3305257
IM		 275struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
276struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
277struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
278struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
506d4bc8 279struct stats walltime_nsecs_stats;
be1ac0d8 280
727ab04e
ACM		 281static int create_perf_stat_counter(struct perf_evsel *evsel,
282 struct perf_evsel *first)
ddcacfa0 283{
69aad6f1 284 struct perf_event_attr *attr = &evsel->attr;
727ab04e
ACM		 285 struct xyarray *group_fd = NULL;
286
287 if (group && evsel != first)
288 group_fd = first->fd;
16c8a109 289
ddcacfa0 290 if (scale)
a21ca2ca
IM		 291 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
292 PERF_FORMAT_TOTAL_TIME_RUNNING;
ddcacfa0 293
5d2cd909
ACM		 294 attr->inherit = !no_inherit;
295
48290609 296 if (system_wide)
727ab04e
ACM		 297 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus,
298 group, group_fd);
4c19ea45 299 if (!target_pid && !target_tid && (!group || evsel == first)) {
48290609
ACM		 300 attr->disabled = 1;
301 attr->enable_on_exec = 1;
ddcacfa0 302 }
084ab9f8 303
727ab04e
ACM		 304 return perf_evsel__open_per_thread(evsel, evsel_list->threads,
305 group, group_fd);
ddcacfa0
IM		 306}
307
c04f5e5d
IM		 308/*
309 * Does the counter have nsecs as a unit?
310 */
daec78a0 311static inline int nsec_counter(struct perf_evsel *evsel)
c04f5e5d 312{
daec78a0
ACM		 313 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
314 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
c04f5e5d
IM		 315 return 1;
316
317 return 0;
318}
319
dcd9936a
IM		 320/*
321 * Update various tracking values we maintain to print
322 * more semantic information such as miss/hit ratios,
323 * instruction rates, etc:
324 */
325static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
326{
327 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
328 update_stats(&runtime_nsecs_stats[0], count[0]);
329 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
330 update_stats(&runtime_cycles_stats[0], count[0]);
d3d1e86d
IM		 331 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
332 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
129c04cb 333 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
d3d1e86d 334 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
dcd9936a
IM		 335 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
336 update_stats(&runtime_branches_stats[0], count[0]);
337 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
338 update_stats(&runtime_cacherefs_stats[0], count[0]);
8bb6c79f
IM		 339 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
340 update_stats(&runtime_l1_dcache_stats[0], count[0]);
c3305257
IM		 341 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
342 update_stats(&runtime_l1_icache_stats[0], count[0]);
343 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
344 update_stats(&runtime_ll_cache_stats[0], count[0]);
345 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
346 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
347 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
348 update_stats(&runtime_itlb_cache_stats[0], count[0]);
dcd9936a
IM		 349}
350
c04f5e5d 351/*
2996f5dd 352 * Read out the results of a single counter:
f5b4a9c3 353 * aggregate counts across CPUs in system-wide mode
c04f5e5d 354 */
c52b12ed 355static int read_counter_aggr(struct perf_evsel *counter)
c04f5e5d 356{
69aad6f1 357 struct perf_stat *ps = counter->priv;
c52b12ed
ACM		 358 u64 *count = counter->counts->aggr.values;
359 int i;
2996f5dd 360
7e2ed097
ACM		 361 if (__perf_evsel__read(counter, evsel_list->cpus->nr,
362 evsel_list->threads->nr, scale) < 0)
c52b12ed 363 return -1;
9e9772c4
PZ		 364
365 for (i = 0; i < 3; i++)
69aad6f1 366 update_stats(&ps->res_stats[i], count[i]);
9e9772c4
PZ		 367
368 if (verbose) {
4aa9015f 369 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
9486aa38 370 event_name(counter), count[0], count[1], count[2]);
9e9772c4
PZ		 371 }
372
be1ac0d8
IM		 373 /*
374 * Save the full runtime - to allow normalization during printout:
375 */
dcd9936a 376 update_shadow_stats(counter, count);
c52b12ed
ACM		 377
378 return 0;
f5b4a9c3
SE		 379}
380
381/*
382 * Read out the results of a single counter:
383 * do not aggregate counts across CPUs in system-wide mode
384 */
c52b12ed 385static int read_counter(struct perf_evsel *counter)
f5b4a9c3 386{
c52b12ed 387 u64 *count;
f5b4a9c3 388 int cpu;
f5b4a9c3 389
7e2ed097 390 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
c52b12ed
ACM		 391 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
392 return -1;
f5b4a9c3 393
c52b12ed 394 count = counter->counts->cpu[cpu].values;
f5b4a9c3 395
dcd9936a 396 update_shadow_stats(counter, count);
f5b4a9c3 397 }
c52b12ed
ACM		 398
399 return 0;
2996f5dd
IM		 400}
401
f37a291c 402static int run_perf_stat(int argc __used, const char **argv)
42202dd5
IM		 403{
404 unsigned long long t0, t1;
727ab04e 405 struct perf_evsel *counter, *first;
42202dd5 406 int status = 0;
051ae7f7 407 int child_ready_pipe[2], go_pipe[2];
6be2850e 408 const bool forks = (argc > 0);
051ae7f7 409 char buf;
42202dd5 410
60666c63 411 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
051ae7f7
PM		 412 perror("failed to create pipes");
413 exit(1);
414 }
415
60666c63 416 if (forks) {
6be2850e 417 if ((child_pid = fork()) < 0)
60666c63
LW		 418 perror("failed to fork");
419
6be2850e 420 if (!child_pid) {
60666c63
LW		 421 close(child_ready_pipe[0]);
422 close(go_pipe[1]);
423 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
424
425 /*
426 * Do a dummy execvp to get the PLT entry resolved,
427 * so we avoid the resolver overhead on the real
428 * execvp call.
429 */
430 execvp("", (char **)argv);
431
432 /*
433 * Tell the parent we're ready to go
434 */
435 close(child_ready_pipe[1]);
436
437 /*
438 * Wait until the parent tells us to go.
439 */
440 if (read(go_pipe[0], &buf, 1) == -1)
441 perror("unable to read pipe");
442
443 execvp(argv[0], (char **)argv);
444
445 perror(argv[0]);
446 exit(-1);
447 }
051ae7f7 448
b52956c9 449 if (!target_tid && !target_pid && !system_wide)
7e2ed097 450 evsel_list->threads->map[0] = child_pid;
d6d901c2 451
051ae7f7 452 /*
60666c63 453 * Wait for the child to be ready to exec.
051ae7f7
PM		 454 */
455 close(child_ready_pipe[1]);
60666c63
LW		 456 close(go_pipe[0]);
457 if (read(child_ready_pipe[0], &buf, 1) == -1)
a92bef0f 458 perror("unable to read pipe");
60666c63 459 close(child_ready_pipe[0]);
051ae7f7
PM		 460 }
461
727ab04e
ACM		 462 first = list_entry(evsel_list->entries.next, struct perf_evsel, node);
463
361c99a6 464 list_for_each_entry(counter, &evsel_list->entries, node) {
727ab04e 465 if (create_perf_stat_counter(counter, first) < 0) {
38f6ae1e
AB		 466 if (errno == EINVAL || errno == ENOSYS ||
467 errno == ENOENT || errno == EOPNOTSUPP) {
c63ca0c0
DA		 468 if (verbose)
469 ui__warning("%s event is not supported by the kernel.\n",
470 event_name(counter));
2cee77c4 471 counter->supported = false;
ede70290 472 continue;
c63ca0c0 473 }
ede70290
IM		 474
475 if (errno == EPERM || errno == EACCES) {
48290609
ACM		 476 error("You may not have permission to collect %sstats.\n"
477 "\t Consider tweaking"
478 " /proc/sys/kernel/perf_event_paranoid or running as root.",
479 system_wide ? "system-wide " : "");
480 } else {
481 error("open_counter returned with %d (%s). "
482 "/bin/dmesg may provide additional information.\n",
483 errno, strerror(errno));
484 }
485 if (child_pid != -1)
486 kill(child_pid, SIGTERM);
487 die("Not all events could be opened.\n");
488 return -1;
489 }
2cee77c4 490 counter->supported = true;
084ab9f8 491 }
42202dd5 492
cfd748ae
FW		 493 if (perf_evlist__set_filters(evsel_list)) {
494 error("failed to set filter with %d (%s)\n", errno,
495 strerror(errno));
496 return -1;
497 }
498
42202dd5
IM		 499 /*
500 * Enable counters and exec the command:
501 */
502 t0 = rdclock();
42202dd5 503
60666c63
LW		 504 if (forks) {
505 close(go_pipe[1]);
506 wait(&status);
33e49ea7
AK		 507 if (WIFSIGNALED(status))
508 psignal(WTERMSIG(status), argv[0]);
60666c63 509 } else {
6be2850e 510 while(!done) sleep(1);
60666c63 511 }
42202dd5 512
42202dd5
IM		 513 t1 = rdclock();
514
9e9772c4 515 update_stats(&walltime_nsecs_stats, t1 - t0);
42202dd5 516
f5b4a9c3 517 if (no_aggr) {
361c99a6 518 list_for_each_entry(counter, &evsel_list->entries, node) {
f5b4a9c3 519 read_counter(counter);
7e2ed097 520 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
c52b12ed 521 }
f5b4a9c3 522 } else {
361c99a6 523 list_for_each_entry(counter, &evsel_list->entries, node) {
f5b4a9c3 524 read_counter_aggr(counter);
7e2ed097
ACM		 525 perf_evsel__close_fd(counter, evsel_list->cpus->nr,
526 evsel_list->threads->nr);
c52b12ed 527 }
f5b4a9c3 528 }
c52b12ed 529
42202dd5
IM		 530 return WEXITSTATUS(status);
531}
532
f99844cb
IM		 533static void print_noise_pct(double total, double avg)
534{
535 double pct = 0.0;
536
537 if (avg)
538 pct = 100.0*total/avg;
539
3ae9a34d 540 if (csv_output)
4aa9015f 541 fprintf(output, "%s%.2f%%", csv_sep, pct);
a1bca6cc 542 else if (pct)
4aa9015f 543 fprintf(output, " ( +-%6.2f%% )", pct);
f99844cb
IM		 544}
545
69aad6f1 546static void print_noise(struct perf_evsel *evsel, double avg)
42202dd5 547{
69aad6f1
ACM		 548 struct perf_stat *ps;
549
849abde9
PZ		 550 if (run_count == 1)
551 return;
552
69aad6f1 553 ps = evsel->priv;
f99844cb 554 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
42202dd5
IM		 555}
556
daec78a0 557static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
44175b6f 558{
506d4bc8 559 double msecs = avg / 1e6;
d7470b6a 560 char cpustr[16] = { '\0', };
2cba3ffb 561 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
44175b6f 562
f5b4a9c3 563 if (no_aggr)
d7470b6a
SE		 564 sprintf(cpustr, "CPU%*d%s",
565 csv_output ? 0 : -4,
7e2ed097 566 evsel_list->cpus->map[cpu], csv_sep);
d7470b6a 567
4aa9015f 568 fprintf(output, fmt, cpustr, msecs, csv_sep, event_name(evsel));
d7470b6a 569
023695d9 570 if (evsel->cgrp)
4aa9015f 571 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
023695d9 572
d7470b6a
SE		 573 if (csv_output)
574 return;
44175b6f 575
daec78a0 576 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
4aa9015f
SE		 577 fprintf(output, " # %8.3f CPUs utilized ",
578 avg / avg_stats(&walltime_nsecs_stats));
9dac6a29
NK		 579 else
580 fprintf(output, " ");
44175b6f
IM		 581}
582
15e6392f
NK		 583/* used for get_ratio_color() */
584enum grc_type {
585 GRC_STALLED_CYCLES_FE,
586 GRC_STALLED_CYCLES_BE,
587 GRC_CACHE_MISSES,
588 GRC_MAX_NR
589};
590
591static const char *get_ratio_color(enum grc_type type, double ratio)
592{
593 static const double grc_table[GRC_MAX_NR][3] = {
594 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
595 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
596 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
597 };
598 const char *color = PERF_COLOR_NORMAL;
599
600 if (ratio > grc_table[type][0])
601 color = PERF_COLOR_RED;
602 else if (ratio > grc_table[type][1])
603 color = PERF_COLOR_MAGENTA;
604 else if (ratio > grc_table[type][2])
605 color = PERF_COLOR_YELLOW;
606
607 return color;
608}
609
d3d1e86d
IM		 610static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg)
611{
612 double total, ratio = 0.0;
613 const char *color;
614
615 total = avg_stats(&runtime_cycles_stats[cpu]);
616
617 if (total)
618 ratio = avg / total * 100.0;
619
15e6392f 620 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
d3d1e86d 621
4aa9015f
SE		 622 fprintf(output, " # ");
623 color_fprintf(output, color, "%6.2f%%", ratio);
624 fprintf(output, " frontend cycles idle ");
d3d1e86d
IM		 625}
626
627static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg)
a5d243d0
IM		 628{
629 double total, ratio = 0.0;
630 const char *color;
631
632 total = avg_stats(&runtime_cycles_stats[cpu]);
633
634 if (total)
635 ratio = avg / total * 100.0;
636
15e6392f 637 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
a5d243d0 638
4aa9015f
SE		 639 fprintf(output, " # ");
640 color_fprintf(output, color, "%6.2f%%", ratio);
641 fprintf(output, " backend cycles idle ");
a5d243d0
IM		 642}
643
c78df6c1
IM		 644static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
645{
646 double total, ratio = 0.0;
647 const char *color;
648
649 total = avg_stats(&runtime_branches_stats[cpu]);
650
651 if (total)
652 ratio = avg / total * 100.0;
653
15e6392f 654 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
c78df6c1 655
4aa9015f
SE		 656 fprintf(output, " # ");
657 color_fprintf(output, color, "%6.2f%%", ratio);
658 fprintf(output, " of all branches ");
c78df6c1
IM		 659}
660
8bb6c79f
IM		 661static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
662{
663 double total, ratio = 0.0;
664 const char *color;
665
666 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
667
668 if (total)
669 ratio = avg / total * 100.0;
670
15e6392f 671 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
8bb6c79f 672
4aa9015f
SE		 673 fprintf(output, " # ");
674 color_fprintf(output, color, "%6.2f%%", ratio);
675 fprintf(output, " of all L1-dcache hits ");
8bb6c79f
IM		 676}
677
c3305257
IM		 678static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
679{
680 double total, ratio = 0.0;
681 const char *color;
682
683 total = avg_stats(&runtime_l1_icache_stats[cpu]);
684
685 if (total)
686 ratio = avg / total * 100.0;
687
15e6392f 688 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
c3305257 689
4aa9015f
SE		 690 fprintf(output, " # ");
691 color_fprintf(output, color, "%6.2f%%", ratio);
692 fprintf(output, " of all L1-icache hits ");
c3305257
IM		 693}
694
695static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
696{
697 double total, ratio = 0.0;
698 const char *color;
699
700 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
701
702 if (total)
703 ratio = avg / total * 100.0;
704
15e6392f 705 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
c3305257 706
4aa9015f
SE		 707 fprintf(output, " # ");
708 color_fprintf(output, color, "%6.2f%%", ratio);
709 fprintf(output, " of all dTLB cache hits ");
c3305257
IM		 710}
711
712static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
713{
714 double total, ratio = 0.0;
715 const char *color;
716
717 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
718
719 if (total)
720 ratio = avg / total * 100.0;
721
15e6392f 722 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
c3305257 723
4aa9015f
SE		 724 fprintf(output, " # ");
725 color_fprintf(output, color, "%6.2f%%", ratio);
726 fprintf(output, " of all iTLB cache hits ");
c3305257
IM		 727}
728
729static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
730{
731 double total, ratio = 0.0;
732 const char *color;
733
734 total = avg_stats(&runtime_ll_cache_stats[cpu]);
735
736 if (total)
737 ratio = avg / total * 100.0;
738
15e6392f 739 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
c3305257 740
4aa9015f
SE		 741 fprintf(output, " # ");
742 color_fprintf(output, color, "%6.2f%%", ratio);
743 fprintf(output, " of all LL-cache hits ");
c3305257
IM		 744}
745
daec78a0 746static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
44175b6f 747{
c7f7fea3 748 double total, ratio = 0.0;
f5b4a9c3 749 char cpustr[16] = { '\0', };
d7470b6a
SE		 750 const char *fmt;
751
752 if (csv_output)
753 fmt = "%s%.0f%s%s";
754 else if (big_num)
2cba3ffb 755 fmt = "%s%'18.0f%s%-25s";
d7470b6a 756 else
2cba3ffb 757 fmt = "%s%18.0f%s%-25s";
f5b4a9c3
SE		 758
759 if (no_aggr)
d7470b6a
SE		 760 sprintf(cpustr, "CPU%*d%s",
761 csv_output ? 0 : -4,
7e2ed097 762 evsel_list->cpus->map[cpu], csv_sep);
f5b4a9c3
SE		 763 else
764 cpu = 0;
c7f7fea3 765
4aa9015f 766 fprintf(output, fmt, cpustr, avg, csv_sep, event_name(evsel));
d7470b6a 767
023695d9 768 if (evsel->cgrp)
4aa9015f 769 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
023695d9 770
d7470b6a
SE		 771 if (csv_output)
772 return;
44175b6f 773
daec78a0 774 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
f5b4a9c3 775 total = avg_stats(&runtime_cycles_stats[cpu]);
c7f7fea3
IM		 776
777 if (total)
778 ratio = avg / total;
779
4aa9015f 780 fprintf(output, " # %5.2f insns per cycle ", ratio);
481f988a 781
d3d1e86d
IM		 782 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
783 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
481f988a
IM		 784
785 if (total && avg) {
786 ratio = total / avg;
4aa9015f 787 fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
481f988a
IM		 788 }
789
daec78a0 790 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
f5b4a9c3 791 runtime_branches_stats[cpu].n != 0) {
c78df6c1 792 print_branch_misses(cpu, evsel, avg);
8bb6c79f
IM		 793 } else if (
794 evsel->attr.type == PERF_TYPE_HW_CACHE &&
795 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
796 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
797 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
c6264def 798 runtime_l1_dcache_stats[cpu].n != 0) {
8bb6c79f 799 print_l1_dcache_misses(cpu, evsel, avg);
c3305257
IM		 800 } else if (
801 evsel->attr.type == PERF_TYPE_HW_CACHE &&
802 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
803 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
804 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
805 runtime_l1_icache_stats[cpu].n != 0) {
806 print_l1_icache_misses(cpu, evsel, avg);
807 } else if (
808 evsel->attr.type == PERF_TYPE_HW_CACHE &&
809 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
810 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
811 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
812 runtime_dtlb_cache_stats[cpu].n != 0) {
813 print_dtlb_cache_misses(cpu, evsel, avg);
814 } else if (
815 evsel->attr.type == PERF_TYPE_HW_CACHE &&
816 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
817 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
818 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
819 runtime_itlb_cache_stats[cpu].n != 0) {
820 print_itlb_cache_misses(cpu, evsel, avg);
821 } else if (
822 evsel->attr.type == PERF_TYPE_HW_CACHE &&
823 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
824 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
825 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
826 runtime_ll_cache_stats[cpu].n != 0) {
827 print_ll_cache_misses(cpu, evsel, avg);
d58f4c82
IM		 828 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
829 runtime_cacherefs_stats[cpu].n != 0) {
830 total = avg_stats(&runtime_cacherefs_stats[cpu]);
831
832 if (total)
833 ratio = avg * 100 / total;
834
4aa9015f 835 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
d58f4c82 836
d3d1e86d
IM		 837 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
838 print_stalled_cycles_frontend(cpu, evsel, avg);
129c04cb 839 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
d3d1e86d 840 print_stalled_cycles_backend(cpu, evsel, avg);
481f988a 841 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
f5b4a9c3 842 total = avg_stats(&runtime_nsecs_stats[cpu]);
c7f7fea3
IM		 843
844 if (total)
481f988a 845 ratio = 1.0 * avg / total;
c7f7fea3 846
4aa9015f 847 fprintf(output, " # %8.3f GHz ", ratio);
481f988a 848 } else if (runtime_nsecs_stats[cpu].n != 0) {
5fde2523
NK		 849 char unit = 'M';
850
481f988a 851 total = avg_stats(&runtime_nsecs_stats[cpu]);
11ba2b85
IM		 852
853 if (total)
481f988a 854 ratio = 1000.0 * avg / total;
5fde2523
NK		 855 if (ratio < 0.001) {
856 ratio *= 1000;
857 unit = 'K';
858 }
11ba2b85 859
5fde2523 860 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
a5d243d0 861 } else {
4aa9015f 862 fprintf(output, " ");
44175b6f 863 }
44175b6f
IM		 864}
865
2996f5dd
IM		 866/*
867 * Print out the results of a single counter:
f5b4a9c3 868 * aggregated counts in system-wide mode
2996f5dd 869 */
69aad6f1 870static void print_counter_aggr(struct perf_evsel *counter)
2996f5dd 871{
69aad6f1
ACM		 872 struct perf_stat *ps = counter->priv;
873 double avg = avg_stats(&ps->res_stats[0]);
c52b12ed 874 int scaled = counter->counts->scaled;
2996f5dd 875
2996f5dd 876 if (scaled == -1) {
4aa9015f 877 fprintf(output, "%*s%s%*s",
d7470b6a 878 csv_output ? 0 : 18,
2cee77c4 879 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
023695d9
SE		 880 csv_sep,
881 csv_output ? 0 : -24,
882 event_name(counter));
883
884 if (counter->cgrp)
4aa9015f 885 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
023695d9 886
4aa9015f 887 fputc('\n', output);
2996f5dd
IM		 888 return;
889 }
c04f5e5d 890
44175b6f 891 if (nsec_counter(counter))
f5b4a9c3 892 nsec_printout(-1, counter, avg);
44175b6f 893 else
f5b4a9c3 894 abs_printout(-1, counter, avg);
849abde9 895
3ae9a34d
ZH		 896 print_noise(counter, avg);
897
d7470b6a 898 if (csv_output) {
4aa9015f 899 fputc('\n', output);
d7470b6a
SE		 900 return;
901 }
902
506d4bc8
PZ		 903 if (scaled) {
904 double avg_enabled, avg_running;
905
69aad6f1
ACM		 906 avg_enabled = avg_stats(&ps->res_stats[1]);
907 avg_running = avg_stats(&ps->res_stats[2]);
d7c29318 908
4aa9015f 909 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
506d4bc8 910 }
4aa9015f 911 fprintf(output, "\n");
c04f5e5d
IM		 912}
913
f5b4a9c3
SE		 914/*
915 * Print out the results of a single counter:
916 * does not use aggregated count in system-wide
917 */
69aad6f1 918static void print_counter(struct perf_evsel *counter)
f5b4a9c3
SE		 919{
920 u64 ena, run, val;
921 int cpu;
922
7e2ed097 923 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
c52b12ed
ACM		 924 val = counter->counts->cpu[cpu].val;
925 ena = counter->counts->cpu[cpu].ena;
926 run = counter->counts->cpu[cpu].run;
f5b4a9c3 927 if (run == 0 || ena == 0) {
4aa9015f 928 fprintf(output, "CPU%*d%s%*s%s%*s",
d7470b6a 929 csv_output ? 0 : -4,
7e2ed097 930 evsel_list->cpus->map[cpu], csv_sep,
d7470b6a 931 csv_output ? 0 : 18,
2cee77c4
DA		 932 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
933 csv_sep,
023695d9 934 csv_output ? 0 : -24,
d7470b6a 935 event_name(counter));
f5b4a9c3 936
023695d9 937 if (counter->cgrp)
4aa9015f
SE		 938 fprintf(output, "%s%s",
939 csv_sep, counter->cgrp->name);
023695d9 940
4aa9015f 941 fputc('\n', output);
f5b4a9c3
SE		 942 continue;
943 }
944
945 if (nsec_counter(counter))
946 nsec_printout(cpu, counter, val);
947 else
948 abs_printout(cpu, counter, val);
949
d7470b6a
SE		 950 if (!csv_output) {
951 print_noise(counter, 1.0);
f5b4a9c3 952
c6264def 953 if (run != ena)
4aa9015f
SE		 954 fprintf(output, " (%.2f%%)",
955 100.0 * run / ena);
f5b4a9c3 956 }
4aa9015f 957 fputc('\n', output);
f5b4a9c3
SE		 958 }
959}
960
42202dd5
IM		 961static void print_stat(int argc, const char **argv)
962{
69aad6f1
ACM		 963 struct perf_evsel *counter;
964 int i;
42202dd5 965
ddcacfa0
IM		 966 fflush(stdout);
967
d7470b6a 968 if (!csv_output) {
4aa9015f
SE		 969 fprintf(output, "\n");
970 fprintf(output, " Performance counter stats for ");
b52956c9 971 if (!target_pid && !target_tid) {
4aa9015f 972 fprintf(output, "\'%s", argv[0]);
d7470b6a 973 for (i = 1; i < argc; i++)
4aa9015f 974 fprintf(output, " %s", argv[i]);
b52956c9
DA		 975 } else if (target_pid)
976 fprintf(output, "process id \'%s", target_pid);
d7470b6a 977 else
b52956c9 978 fprintf(output, "thread id \'%s", target_tid);
44db76c8 979
4aa9015f 980 fprintf(output, "\'");
d7470b6a 981 if (run_count > 1)
4aa9015f
SE		 982 fprintf(output, " (%d runs)", run_count);
983 fprintf(output, ":\n\n");
d7470b6a 984 }
2996f5dd 985
f5b4a9c3 986 if (no_aggr) {
361c99a6 987 list_for_each_entry(counter, &evsel_list->entries, node)
f5b4a9c3
SE		 988 print_counter(counter);
989 } else {
361c99a6 990 list_for_each_entry(counter, &evsel_list->entries, node)
f5b4a9c3
SE		 991 print_counter_aggr(counter);
992 }
ddcacfa0 993
d7470b6a 994 if (!csv_output) {
c3305257 995 if (!null_run)
4aa9015f
SE		 996 fprintf(output, "\n");
997 fprintf(output, " %17.9f seconds time elapsed",
d7470b6a
SE		 998 avg_stats(&walltime_nsecs_stats)/1e9);
999 if (run_count > 1) {
4aa9015f 1000 fprintf(output, " ");
f99844cb
IM		 1001 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1002 avg_stats(&walltime_nsecs_stats));
d7470b6a 1003 }
4aa9015f 1004 fprintf(output, "\n\n");
566747e6 1005 }
ddcacfa0
IM		 1006}
1007
f7b7c26e
PZ		 1008static volatile int signr = -1;
1009
5242519b 1010static void skip_signal(int signo)
ddcacfa0 1011{
6be2850e 1012 if(child_pid == -1)
60666c63
LW		 1013 done = 1;
1014
f7b7c26e
PZ		 1015 signr = signo;
1016}
1017
1018static void sig_atexit(void)
1019{
933da83a
CW		 1020 if (child_pid != -1)
1021 kill(child_pid, SIGTERM);
1022
f7b7c26e
PZ		 1023 if (signr == -1)
1024 return;
1025
1026 signal(signr, SIG_DFL);
1027 kill(getpid(), signr);
5242519b
IM		 1028}
1029
1030static const char * const stat_usage[] = {
60666c63 1031 "perf stat [<options>] [<command>]",
5242519b
IM		 1032 NULL
1033};
1034
d7470b6a
SE		 1035static int stat__set_big_num(const struct option *opt __used,
1036 const char *s __used, int unset)
1037{
1038 big_num_opt = unset ? 0 : 1;
1039 return 0;
1040}
1041
4aa9015f
SE		 1042static bool append_file;
1043
5242519b 1044static const struct option options[] = {
361c99a6 1045 OPT_CALLBACK('e', "event", &evsel_list, "event",
86847b62 1046 "event selector. use 'perf list' to list available events",
f120f9d5 1047 parse_events_option),
cfd748ae
FW		 1048 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1049 "event filter", parse_filter),
2e6cdf99
SE		 1050 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1051 "child tasks do not inherit counters"),
b52956c9
DA		 1052 OPT_STRING('p', "pid", &target_pid, "pid",
1053 "stat events on existing process id"),
1054 OPT_STRING('t', "tid", &target_tid, "tid",
1055 "stat events on existing thread id"),
5242519b 1056 OPT_BOOLEAN('a', "all-cpus", &system_wide,
3d632595 1057 "system-wide collection from all CPUs"),
43bece79
LM		 1058 OPT_BOOLEAN('g', "group", &group,
1059 "put the counters into a counter group"),
b26bc5a7 1060 OPT_BOOLEAN('c', "scale", &scale,
3d632595 1061 "scale/normalize counters"),
c0555642 1062 OPT_INCR('v', "verbose", &verbose,
743ee1f8 1063 "be more verbose (show counter open errors, etc)"),
42202dd5
IM		 1064 OPT_INTEGER('r', "repeat", &run_count,
1065 "repeat command and print average + stddev (max: 100)"),
0cfb7a13
IM		 1066 OPT_BOOLEAN('n', "null", &null_run,
1067 "null run - dont start any counters"),
2cba3ffb 1068 OPT_INCR('d', "detailed", &detailed_run,
c6264def 1069 "detailed run - start a lot of events"),
f9cef0a9
IM		 1070 OPT_BOOLEAN('S', "sync", &sync_run,
1071 "call sync() before starting a run"),
d7470b6a
SE		 1072 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1073 "print large numbers with thousands\' separators",
1074 stat__set_big_num),
c45c6ea2
SE		 1075 OPT_STRING('C', "cpu", &cpu_list, "cpu",
1076 "list of cpus to monitor in system-wide"),
f5b4a9c3
SE		 1077 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1078 "disable CPU count aggregation"),
d7470b6a
SE		 1079 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1080 "print counts with custom separator"),
023695d9
SE		 1081 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1082 "monitor event in cgroup name only",
1083 parse_cgroups),
4aa9015f
SE		 1084 OPT_STRING('o', "output", &output_name, "file",
1085 "output file name"),
1086 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
56f3bae7
JC		 1087 OPT_INTEGER(0, "log-fd", &output_fd,
1088 "log output to fd, instead of stderr"),
5242519b
IM		 1089 OPT_END()
1090};
1091
2cba3ffb
IM		 1092/*
1093 * Add default attributes, if there were no attributes specified or
1094 * if -d/--detailed, -d -d or -d -d -d is used:
1095 */
1096static int add_default_attributes(void)
1097{
2cba3ffb
IM		 1098 /* Set attrs if no event is selected and !null_run: */
1099 if (null_run)
1100 return 0;
1101
1102 if (!evsel_list->nr_entries) {
50d08e47
ACM		 1103 if (perf_evlist__add_attrs_array(evsel_list, default_attrs) < 0)
1104 return -1;
2cba3ffb
IM		 1105 }
1106
1107 /* Detailed events get appended to the event list: */
1108
1109 if (detailed_run < 1)
1110 return 0;
1111
1112 /* Append detailed run extra attributes: */
50d08e47
ACM		 1113 if (perf_evlist__add_attrs_array(evsel_list, detailed_attrs) < 0)
1114 return -1;
2cba3ffb
IM		 1115
1116 if (detailed_run < 2)
1117 return 0;
1118
1119 /* Append very detailed run extra attributes: */
50d08e47
ACM		 1120 if (perf_evlist__add_attrs_array(evsel_list, very_detailed_attrs) < 0)
1121 return -1;
2cba3ffb
IM		 1122
1123 if (detailed_run < 3)
1124 return 0;
1125
1126 /* Append very, very detailed run extra attributes: */
50d08e47 1127 return perf_evlist__add_attrs_array(evsel_list, very_very_detailed_attrs);
2cba3ffb
IM		 1128}
1129
f37a291c 1130int cmd_stat(int argc, const char **argv, const char *prefix __used)
5242519b 1131{
69aad6f1
ACM		 1132 struct perf_evsel *pos;
1133 int status = -ENOMEM;
4aa9015f 1134 const char *mode;
42202dd5 1135
5af52b51
SE		 1136 setlocale(LC_ALL, "");
1137
7e2ed097 1138 evsel_list = perf_evlist__new(NULL, NULL);
361c99a6
ACM		 1139 if (evsel_list == NULL)
1140 return -ENOMEM;
1141
a0541234
AB		 1142 argc = parse_options(argc, argv, options, stat_usage,
1143 PARSE_OPT_STOP_AT_NON_OPTION);
d7470b6a 1144
4aa9015f
SE		 1145 output = stderr;
1146 if (output_name && strcmp(output_name, "-"))
1147 output = NULL;
1148
56f3bae7
JC		 1149 if (output_name && output_fd) {
1150 fprintf(stderr, "cannot use both --output and --log-fd\n");
1151 usage_with_options(stat_usage, options);
1152 }
4aa9015f
SE		 1153 if (!output) {
1154 struct timespec tm;
1155 mode = append_file ? "a" : "w";
1156
1157 output = fopen(output_name, mode);
1158 if (!output) {
1159 perror("failed to create output file");
1160 exit(-1);
1161 }
1162 clock_gettime(CLOCK_REALTIME, &tm);
1163 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
56f3bae7
JC		 1164 } else if (output_fd != 2) {
1165 mode = append_file ? "a" : "w";
1166 output = fdopen(output_fd, mode);
1167 if (!output) {
1168 perror("Failed opening logfd");
1169 return -errno;
1170 }
4aa9015f
SE		 1171 }
1172
d4ffd04d 1173 if (csv_sep) {
d7470b6a 1174 csv_output = true;
d4ffd04d
JC		 1175 if (!strcmp(csv_sep, "\\t"))
1176 csv_sep = "\t";
1177 } else
d7470b6a
SE		 1178 csv_sep = DEFAULT_SEPARATOR;
1179
1180 /*
1181 * let the spreadsheet do the pretty-printing
1182 */
1183 if (csv_output) {
61a9f324 1184 /* User explicitly passed -B? */
d7470b6a
SE		 1185 if (big_num_opt == 1) {
1186 fprintf(stderr, "-B option not supported with -x\n");
1187 usage_with_options(stat_usage, options);
1188 } else /* Nope, so disable big number formatting */
1189 big_num = false;
1190 } else if (big_num_opt == 0) /* User passed --no-big-num */
1191 big_num = false;
1192
b52956c9 1193 if (!argc && !target_pid && !target_tid)
5242519b 1194 usage_with_options(stat_usage, options);
9e9772c4 1195 if (run_count <= 0)
42202dd5 1196 usage_with_options(stat_usage, options);
ddcacfa0 1197
023695d9
SE		 1198 /* no_aggr, cgroup are for system-wide only */
1199 if ((no_aggr || nr_cgroups) && !system_wide) {
1200 fprintf(stderr, "both cgroup and no-aggregation "
1201 "modes only available in system-wide mode\n");
1202
f5b4a9c3 1203 usage_with_options(stat_usage, options);
023695d9 1204 }
f5b4a9c3 1205
2cba3ffb
IM		 1206 if (add_default_attributes())
1207 goto out;
ddcacfa0 1208
b52956c9 1209 if (target_pid)
5c98d466
ACM		 1210 target_tid = target_pid;
1211
b52956c9
DA		 1212 evsel_list->threads = thread_map__new_str(target_pid,
1213 target_tid, UINT_MAX);
7e2ed097 1214 if (evsel_list->threads == NULL) {
5c98d466
ACM		 1215 pr_err("Problems finding threads of monitor\n");
1216 usage_with_options(stat_usage, options);
1217 }
1218
a12b51c4 1219 if (system_wide)
7e2ed097 1220 evsel_list->cpus = cpu_map__new(cpu_list);
a12b51c4 1221 else
7e2ed097 1222 evsel_list->cpus = cpu_map__dummy_new();
ddcacfa0 1223
7e2ed097 1224 if (evsel_list->cpus == NULL) {
60d567e2 1225 perror("failed to parse CPUs map");
c45c6ea2 1226 usage_with_options(stat_usage, options);
60d567e2
ACM		 1227 return -1;
1228 }
c45c6ea2 1229
361c99a6 1230 list_for_each_entry(pos, &evsel_list->entries, node) {
c52b12ed 1231 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
806fb630 1232 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0)
69aad6f1 1233 goto out_free_fd;
d6d901c2
ZY		 1234 }
1235
58d7e993
IM		 1236 /*
1237 * We dont want to block the signals - that would cause
1238 * child tasks to inherit that and Ctrl-C would not work.
1239 * What we want is for Ctrl-C to work in the exec()-ed
1240 * task, but being ignored by perf stat itself:
1241 */
f7b7c26e 1242 atexit(sig_atexit);
58d7e993
IM		 1243 signal(SIGINT, skip_signal);
1244 signal(SIGALRM, skip_signal);
1245 signal(SIGABRT, skip_signal);
1246
42202dd5
IM		 1247 status = 0;
1248 for (run_idx = 0; run_idx < run_count; run_idx++) {
1249 if (run_count != 1 && verbose)
4aa9015f
SE		 1250 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1251 run_idx + 1);
f9cef0a9
IM		 1252
1253 if (sync_run)
1254 sync();
1255
42202dd5
IM		 1256 status = run_perf_stat(argc, argv);
1257 }
1258
084ab9f8
ACM		 1259 if (status != -1)
1260 print_stat(argc, argv);
69aad6f1 1261out_free_fd:
361c99a6 1262 list_for_each_entry(pos, &evsel_list->entries, node)
69aad6f1 1263 perf_evsel__free_stat_priv(pos);
7e2ed097 1264 perf_evlist__delete_maps(evsel_list);
0015e2e1
ACM		 1265out:
1266 perf_evlist__delete(evsel_list);
42202dd5 1267 return status;
ddcacfa0 1268}
Linux kernel - Deliverables
This page took 0.223988 seconds and 5 git commands to generate.