]>
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 | |
bf9e1876 IM |
9 | $ perf stat ~/hackbench 10 |
10 | Time: 0.104 | |
ddcacfa0 | 11 | |
bf9e1876 | 12 | Performance counter stats for '/home/mingo/hackbench': |
ddcacfa0 | 13 | |
bf9e1876 IM |
14 | 1255.538611 task clock ticks # 10.143 CPU utilization factor |
15 | 54011 context switches # 0.043 M/sec | |
16 | 385 CPU migrations # 0.000 M/sec | |
17 | 17755 pagefaults # 0.014 M/sec | |
18 | 3808323185 CPU cycles # 3033.219 M/sec | |
19 | 1575111190 instructions # 1254.530 M/sec | |
20 | 17367895 cache references # 13.833 M/sec | |
21 | 7674421 cache misses # 6.112 M/sec | |
ddcacfa0 | 22 | |
bf9e1876 | 23 | Wall-clock time elapsed: 123.786620 msecs |
ddcacfa0 | 24 | |
5242519b IM |
25 | * |
26 | * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com> | |
27 | * | |
28 | * Improvements and fixes by: | |
29 | * | |
30 | * Arjan van de Ven <arjan@linux.intel.com> | |
31 | * Yanmin Zhang <yanmin.zhang@intel.com> | |
32 | * Wu Fengguang <fengguang.wu@intel.com> | |
33 | * Mike Galbraith <efault@gmx.de> | |
34 | * Paul Mackerras <paulus@samba.org> | |
35 | * | |
36 | * Released under the GPL v2. (and only v2, not any later version) | |
ddcacfa0 IM |
37 | */ |
38 | ||
1a482f38 | 39 | #include "perf.h" |
16f762a2 | 40 | #include "builtin.h" |
148be2c1 | 41 | #include "util/util.h" |
5242519b IM |
42 | #include "util/parse-options.h" |
43 | #include "util/parse-events.h" | |
ddcacfa0 | 44 | |
ddcacfa0 | 45 | #include <sys/prctl.h> |
16c8a109 | 46 | |
ddcacfa0 | 47 | static int system_wide = 0; |
5242519b | 48 | static int inherit = 1; |
ddcacfa0 | 49 | |
5242519b | 50 | static __u64 default_event_id[MAX_COUNTERS] = { |
ddcacfa0 IM |
51 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK), |
52 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CONTEXT_SWITCHES), | |
53 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_MIGRATIONS), | |
54 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS), | |
55 | ||
56 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_CPU_CYCLES), | |
57 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_INSTRUCTIONS), | |
58 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_REFERENCES), | |
59 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_MISSES), | |
60 | }; | |
5242519b | 61 | |
ddcacfa0 IM |
62 | static int default_interval = 100000; |
63 | static int event_count[MAX_COUNTERS]; | |
64 | static int fd[MAX_NR_CPUS][MAX_COUNTERS]; | |
65 | ||
5242519b | 66 | static int target_pid = -1; |
ddcacfa0 | 67 | static int nr_cpus = 0; |
ddcacfa0 IM |
68 | static unsigned int page_size; |
69 | ||
66cf7829 | 70 | static int scale = 1; |
ddcacfa0 IM |
71 | |
72 | static const unsigned int default_count[] = { | |
73 | 1000000, | |
74 | 1000000, | |
75 | 10000, | |
76 | 10000, | |
77 | 1000000, | |
78 | 10000, | |
79 | }; | |
80 | ||
2996f5dd IM |
81 | static __u64 event_res[MAX_COUNTERS][3]; |
82 | static __u64 event_scaled[MAX_COUNTERS]; | |
83 | ||
be1ac0d8 | 84 | static __u64 runtime_nsecs; |
d7c29318 | 85 | static __u64 walltime_nsecs; |
be1ac0d8 | 86 | |
ddcacfa0 IM |
87 | static void create_perfstat_counter(int counter) |
88 | { | |
c70975bc | 89 | struct perf_counter_attr attr; |
ddcacfa0 | 90 | |
c70975bc PZ |
91 | memset(&attr, 0, sizeof(attr)); |
92 | attr.config = event_id[counter]; | |
93 | attr.sample_type = 0; | |
94 | attr.exclude_kernel = event_mask[counter] & EVENT_MASK_KERNEL; | |
95 | attr.exclude_user = event_mask[counter] & EVENT_MASK_USER; | |
16c8a109 | 96 | |
ddcacfa0 | 97 | if (scale) |
c70975bc | 98 | attr.read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | |
ddcacfa0 IM |
99 | PERF_FORMAT_TOTAL_TIME_RUNNING; |
100 | ||
101 | if (system_wide) { | |
102 | int cpu; | |
103 | for (cpu = 0; cpu < nr_cpus; cpu ++) { | |
c70975bc | 104 | fd[cpu][counter] = sys_perf_counter_open(&attr, -1, cpu, -1, 0); |
ddcacfa0 IM |
105 | if (fd[cpu][counter] < 0) { |
106 | printf("perfstat error: syscall returned with %d (%s)\n", | |
107 | fd[cpu][counter], strerror(errno)); | |
108 | exit(-1); | |
109 | } | |
110 | } | |
111 | } else { | |
c70975bc PZ |
112 | attr.inherit = inherit; |
113 | attr.disabled = 1; | |
ddcacfa0 | 114 | |
c70975bc | 115 | fd[0][counter] = sys_perf_counter_open(&attr, 0, -1, -1, 0); |
ddcacfa0 IM |
116 | if (fd[0][counter] < 0) { |
117 | printf("perfstat error: syscall returned with %d (%s)\n", | |
118 | fd[0][counter], strerror(errno)); | |
119 | exit(-1); | |
120 | } | |
121 | } | |
122 | } | |
123 | ||
c04f5e5d IM |
124 | /* |
125 | * Does the counter have nsecs as a unit? | |
126 | */ | |
127 | static inline int nsec_counter(int counter) | |
128 | { | |
129 | if (event_id[counter] == EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_CLOCK)) | |
130 | return 1; | |
131 | if (event_id[counter] == EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK)) | |
132 | return 1; | |
133 | ||
134 | return 0; | |
135 | } | |
136 | ||
137 | /* | |
2996f5dd | 138 | * Read out the results of a single counter: |
c04f5e5d | 139 | */ |
2996f5dd | 140 | static void read_counter(int counter) |
c04f5e5d | 141 | { |
2996f5dd | 142 | __u64 *count, single_count[3]; |
c04f5e5d IM |
143 | ssize_t res; |
144 | int cpu, nv; | |
145 | int scaled; | |
146 | ||
2996f5dd IM |
147 | count = event_res[counter]; |
148 | ||
c04f5e5d | 149 | count[0] = count[1] = count[2] = 0; |
2996f5dd | 150 | |
c04f5e5d IM |
151 | nv = scale ? 3 : 1; |
152 | for (cpu = 0; cpu < nr_cpus; cpu ++) { | |
153 | res = read(fd[cpu][counter], single_count, nv * sizeof(__u64)); | |
154 | assert(res == nv * sizeof(__u64)); | |
155 | ||
156 | count[0] += single_count[0]; | |
157 | if (scale) { | |
158 | count[1] += single_count[1]; | |
159 | count[2] += single_count[2]; | |
160 | } | |
161 | } | |
162 | ||
163 | scaled = 0; | |
164 | if (scale) { | |
165 | if (count[2] == 0) { | |
2996f5dd IM |
166 | event_scaled[counter] = -1; |
167 | count[0] = 0; | |
c04f5e5d IM |
168 | return; |
169 | } | |
2996f5dd | 170 | |
c04f5e5d | 171 | if (count[2] < count[1]) { |
2996f5dd | 172 | event_scaled[counter] = 1; |
c04f5e5d IM |
173 | count[0] = (unsigned long long) |
174 | ((double)count[0] * count[1] / count[2] + 0.5); | |
175 | } | |
176 | } | |
be1ac0d8 IM |
177 | /* |
178 | * Save the full runtime - to allow normalization during printout: | |
179 | */ | |
180 | if (event_id[counter] == EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK)) | |
181 | runtime_nsecs = count[0]; | |
2996f5dd IM |
182 | } |
183 | ||
184 | /* | |
185 | * Print out the results of a single counter: | |
186 | */ | |
187 | static void print_counter(int counter) | |
188 | { | |
189 | __u64 *count; | |
190 | int scaled; | |
191 | ||
192 | count = event_res[counter]; | |
193 | scaled = event_scaled[counter]; | |
194 | ||
195 | if (scaled == -1) { | |
196 | fprintf(stderr, " %14s %-20s\n", | |
197 | "<not counted>", event_name(counter)); | |
198 | return; | |
199 | } | |
c04f5e5d IM |
200 | |
201 | if (nsec_counter(counter)) { | |
202 | double msecs = (double)count[0] / 1000000; | |
203 | ||
d7c29318 | 204 | fprintf(stderr, " %14.6f %-20s", |
c04f5e5d | 205 | msecs, event_name(counter)); |
d7c29318 IM |
206 | if (event_id[counter] == |
207 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK)) { | |
208 | ||
209 | fprintf(stderr, " # %11.3f CPU utilization factor", | |
210 | (double)count[0] / (double)walltime_nsecs); | |
211 | } | |
c04f5e5d | 212 | } else { |
be1ac0d8 | 213 | fprintf(stderr, " %14Ld %-20s", |
c04f5e5d | 214 | count[0], event_name(counter)); |
be1ac0d8 | 215 | if (runtime_nsecs) |
d7c29318 | 216 | fprintf(stderr, " # %11.3f M/sec", |
be1ac0d8 | 217 | (double)count[0]/runtime_nsecs*1000.0); |
c04f5e5d IM |
218 | } |
219 | if (scaled) | |
220 | fprintf(stderr, " (scaled from %.2f%%)", | |
221 | (double) count[2] / count[1] * 100); | |
222 | fprintf(stderr, "\n"); | |
223 | } | |
224 | ||
16f762a2 | 225 | static int do_perfstat(int argc, const char **argv) |
ddcacfa0 IM |
226 | { |
227 | unsigned long long t0, t1; | |
228 | int counter; | |
ddcacfa0 IM |
229 | int status; |
230 | int pid; | |
44db76c8 | 231 | int i; |
ddcacfa0 IM |
232 | |
233 | if (!system_wide) | |
234 | nr_cpus = 1; | |
235 | ||
236 | for (counter = 0; counter < nr_counters; counter++) | |
237 | create_perfstat_counter(counter); | |
238 | ||
ddcacfa0 IM |
239 | /* |
240 | * Enable counters and exec the command: | |
241 | */ | |
242 | t0 = rdclock(); | |
243 | prctl(PR_TASK_PERF_COUNTERS_ENABLE); | |
244 | ||
245 | if ((pid = fork()) < 0) | |
246 | perror("failed to fork"); | |
44db76c8 | 247 | |
ddcacfa0 | 248 | if (!pid) { |
5242519b | 249 | if (execvp(argv[0], (char **)argv)) { |
ddcacfa0 IM |
250 | perror(argv[0]); |
251 | exit(-1); | |
252 | } | |
253 | } | |
44db76c8 | 254 | |
ddcacfa0 IM |
255 | while (wait(&status) >= 0) |
256 | ; | |
44db76c8 | 257 | |
ddcacfa0 IM |
258 | prctl(PR_TASK_PERF_COUNTERS_DISABLE); |
259 | t1 = rdclock(); | |
260 | ||
d7c29318 IM |
261 | walltime_nsecs = t1 - t0; |
262 | ||
ddcacfa0 IM |
263 | fflush(stdout); |
264 | ||
265 | fprintf(stderr, "\n"); | |
44db76c8 IM |
266 | fprintf(stderr, " Performance counter stats for \'%s", argv[0]); |
267 | ||
268 | for (i = 1; i < argc; i++) | |
269 | fprintf(stderr, " %s", argv[i]); | |
270 | ||
271 | fprintf(stderr, "\':\n"); | |
ddcacfa0 IM |
272 | fprintf(stderr, "\n"); |
273 | ||
2996f5dd IM |
274 | for (counter = 0; counter < nr_counters; counter++) |
275 | read_counter(counter); | |
276 | ||
c04f5e5d IM |
277 | for (counter = 0; counter < nr_counters; counter++) |
278 | print_counter(counter); | |
ddcacfa0 | 279 | |
ddcacfa0 | 280 | |
ddcacfa0 IM |
281 | fprintf(stderr, "\n"); |
282 | fprintf(stderr, " Wall-clock time elapsed: %12.6f msecs\n", | |
283 | (double)(t1-t0)/1e6); | |
284 | fprintf(stderr, "\n"); | |
285 | ||
286 | return 0; | |
287 | } | |
288 | ||
5242519b | 289 | static void skip_signal(int signo) |
ddcacfa0 | 290 | { |
5242519b IM |
291 | } |
292 | ||
293 | static const char * const stat_usage[] = { | |
294 | "perf stat [<options>] <command>", | |
295 | NULL | |
296 | }; | |
297 | ||
298 | static char events_help_msg[EVENTS_HELP_MAX]; | |
299 | ||
300 | static const struct option options[] = { | |
301 | OPT_CALLBACK('e', "event", NULL, "event", | |
302 | events_help_msg, parse_events), | |
303 | OPT_INTEGER('c', "count", &default_interval, | |
304 | "event period to sample"), | |
305 | OPT_BOOLEAN('i', "inherit", &inherit, | |
306 | "child tasks inherit counters"), | |
307 | OPT_INTEGER('p', "pid", &target_pid, | |
308 | "stat events on existing pid"), | |
309 | OPT_BOOLEAN('a', "all-cpus", &system_wide, | |
310 | "system-wide collection from all CPUs"), | |
311 | OPT_BOOLEAN('l', "scale", &scale, | |
312 | "scale/normalize counters"), | |
313 | OPT_END() | |
314 | }; | |
315 | ||
316 | int cmd_stat(int argc, const char **argv, const char *prefix) | |
317 | { | |
318 | int counter; | |
319 | ||
320 | page_size = sysconf(_SC_PAGE_SIZE); | |
321 | ||
322 | create_events_help(events_help_msg); | |
323 | memcpy(event_id, default_event_id, sizeof(default_event_id)); | |
324 | ||
325 | argc = parse_options(argc, argv, options, stat_usage, 0); | |
326 | if (!argc) | |
327 | usage_with_options(stat_usage, options); | |
ddcacfa0 IM |
328 | |
329 | if (!nr_counters) { | |
330 | nr_counters = 8; | |
331 | } | |
332 | ||
333 | for (counter = 0; counter < nr_counters; counter++) { | |
334 | if (event_count[counter]) | |
335 | continue; | |
336 | ||
337 | event_count[counter] = default_interval; | |
338 | } | |
ddcacfa0 IM |
339 | nr_cpus = sysconf(_SC_NPROCESSORS_ONLN); |
340 | assert(nr_cpus <= MAX_NR_CPUS); | |
341 | assert(nr_cpus >= 0); | |
342 | ||
58d7e993 IM |
343 | /* |
344 | * We dont want to block the signals - that would cause | |
345 | * child tasks to inherit that and Ctrl-C would not work. | |
346 | * What we want is for Ctrl-C to work in the exec()-ed | |
347 | * task, but being ignored by perf stat itself: | |
348 | */ | |
349 | signal(SIGINT, skip_signal); | |
350 | signal(SIGALRM, skip_signal); | |
351 | signal(SIGABRT, skip_signal); | |
352 | ||
ddcacfa0 IM |
353 | return do_perfstat(argc, argv); |
354 | } |