Commit | Line | Data |
---|---|---|
ddcacfa0 | 1 | /* |
5242519b | 2 | * perf stat: /usr/bin/time -alike performance counter statistics utility |
ddcacfa0 IM |
3 | |
4 | It summarizes the counter events of all tasks (and child tasks), | |
5 | covering all CPUs that the command (or workload) executes on. | |
6 | It only counts the per-task events of the workload started, | |
7 | independent of how many other tasks run on those CPUs. | |
8 | ||
9 | Sample output: | |
10 | ||
5242519b | 11 | $ perf stat -e 1 -e 3 -e 5 ls -lR /usr/include/ >/dev/null |
ddcacfa0 IM |
12 | |
13 | Performance counter stats for 'ls': | |
14 | ||
15 | 163516953 instructions | |
16 | 2295 cache-misses | |
17 | 2855182 branch-misses | |
5242519b IM |
18 | * |
19 | * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com> | |
20 | * | |
21 | * Improvements and fixes by: | |
22 | * | |
23 | * Arjan van de Ven <arjan@linux.intel.com> | |
24 | * Yanmin Zhang <yanmin.zhang@intel.com> | |
25 | * Wu Fengguang <fengguang.wu@intel.com> | |
26 | * Mike Galbraith <efault@gmx.de> | |
27 | * Paul Mackerras <paulus@samba.org> | |
28 | * | |
29 | * Released under the GPL v2. (and only v2, not any later version) | |
ddcacfa0 IM |
30 | */ |
31 | ||
1a482f38 | 32 | #include "perf.h" |
16f762a2 | 33 | #include "builtin.h" |
148be2c1 | 34 | #include "util/util.h" |
5242519b IM |
35 | #include "util/parse-options.h" |
36 | #include "util/parse-events.h" | |
ddcacfa0 | 37 | |
ddcacfa0 | 38 | #include <sys/prctl.h> |
16c8a109 | 39 | |
ddcacfa0 | 40 | static int system_wide = 0; |
5242519b | 41 | static int inherit = 1; |
ddcacfa0 | 42 | |
5242519b | 43 | static __u64 default_event_id[MAX_COUNTERS] = { |
ddcacfa0 IM |
44 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK), |
45 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CONTEXT_SWITCHES), | |
46 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_MIGRATIONS), | |
47 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS), | |
48 | ||
49 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_CPU_CYCLES), | |
50 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_INSTRUCTIONS), | |
51 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_REFERENCES), | |
52 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_MISSES), | |
53 | }; | |
5242519b | 54 | |
ddcacfa0 IM |
55 | static int default_interval = 100000; |
56 | static int event_count[MAX_COUNTERS]; | |
57 | static int fd[MAX_NR_CPUS][MAX_COUNTERS]; | |
58 | ||
5242519b | 59 | static int target_pid = -1; |
ddcacfa0 | 60 | static int nr_cpus = 0; |
ddcacfa0 IM |
61 | static unsigned int page_size; |
62 | ||
66cf7829 | 63 | static int scale = 1; |
ddcacfa0 IM |
64 | |
65 | static const unsigned int default_count[] = { | |
66 | 1000000, | |
67 | 1000000, | |
68 | 10000, | |
69 | 10000, | |
70 | 1000000, | |
71 | 10000, | |
72 | }; | |
73 | ||
2996f5dd IM |
74 | static __u64 event_res[MAX_COUNTERS][3]; |
75 | static __u64 event_scaled[MAX_COUNTERS]; | |
76 | ||
be1ac0d8 IM |
77 | static __u64 runtime_nsecs; |
78 | ||
ddcacfa0 IM |
79 | static void create_perfstat_counter(int counter) |
80 | { | |
81 | struct perf_counter_hw_event hw_event; | |
82 | ||
83 | memset(&hw_event, 0, sizeof(hw_event)); | |
84 | hw_event.config = event_id[counter]; | |
85 | hw_event.record_type = 0; | |
5242519b | 86 | hw_event.nmi = 1; |
16c8a109 PZ |
87 | hw_event.exclude_kernel = event_mask[counter] & EVENT_MASK_KERNEL; |
88 | hw_event.exclude_user = event_mask[counter] & EVENT_MASK_USER; | |
89 | ||
ddcacfa0 IM |
90 | if (scale) |
91 | hw_event.read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | | |
92 | PERF_FORMAT_TOTAL_TIME_RUNNING; | |
93 | ||
94 | if (system_wide) { | |
95 | int cpu; | |
96 | for (cpu = 0; cpu < nr_cpus; cpu ++) { | |
97 | fd[cpu][counter] = sys_perf_counter_open(&hw_event, -1, cpu, -1, 0); | |
98 | if (fd[cpu][counter] < 0) { | |
99 | printf("perfstat error: syscall returned with %d (%s)\n", | |
100 | fd[cpu][counter], strerror(errno)); | |
101 | exit(-1); | |
102 | } | |
103 | } | |
104 | } else { | |
5242519b | 105 | hw_event.inherit = inherit; |
ddcacfa0 IM |
106 | hw_event.disabled = 1; |
107 | ||
108 | fd[0][counter] = sys_perf_counter_open(&hw_event, 0, -1, -1, 0); | |
109 | if (fd[0][counter] < 0) { | |
110 | printf("perfstat error: syscall returned with %d (%s)\n", | |
111 | fd[0][counter], strerror(errno)); | |
112 | exit(-1); | |
113 | } | |
114 | } | |
115 | } | |
116 | ||
c04f5e5d IM |
117 | /* |
118 | * Does the counter have nsecs as a unit? | |
119 | */ | |
120 | static inline int nsec_counter(int counter) | |
121 | { | |
122 | if (event_id[counter] == EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_CLOCK)) | |
123 | return 1; | |
124 | if (event_id[counter] == EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK)) | |
125 | return 1; | |
126 | ||
127 | return 0; | |
128 | } | |
129 | ||
130 | /* | |
2996f5dd | 131 | * Read out the results of a single counter: |
c04f5e5d | 132 | */ |
2996f5dd | 133 | static void read_counter(int counter) |
c04f5e5d | 134 | { |
2996f5dd | 135 | __u64 *count, single_count[3]; |
c04f5e5d IM |
136 | ssize_t res; |
137 | int cpu, nv; | |
138 | int scaled; | |
139 | ||
2996f5dd IM |
140 | count = event_res[counter]; |
141 | ||
c04f5e5d | 142 | count[0] = count[1] = count[2] = 0; |
2996f5dd | 143 | |
c04f5e5d IM |
144 | nv = scale ? 3 : 1; |
145 | for (cpu = 0; cpu < nr_cpus; cpu ++) { | |
146 | res = read(fd[cpu][counter], single_count, nv * sizeof(__u64)); | |
147 | assert(res == nv * sizeof(__u64)); | |
148 | ||
149 | count[0] += single_count[0]; | |
150 | if (scale) { | |
151 | count[1] += single_count[1]; | |
152 | count[2] += single_count[2]; | |
153 | } | |
154 | } | |
155 | ||
156 | scaled = 0; | |
157 | if (scale) { | |
158 | if (count[2] == 0) { | |
2996f5dd IM |
159 | event_scaled[counter] = -1; |
160 | count[0] = 0; | |
c04f5e5d IM |
161 | return; |
162 | } | |
2996f5dd | 163 | |
c04f5e5d | 164 | if (count[2] < count[1]) { |
2996f5dd | 165 | event_scaled[counter] = 1; |
c04f5e5d IM |
166 | count[0] = (unsigned long long) |
167 | ((double)count[0] * count[1] / count[2] + 0.5); | |
168 | } | |
169 | } | |
be1ac0d8 IM |
170 | /* |
171 | * Save the full runtime - to allow normalization during printout: | |
172 | */ | |
173 | if (event_id[counter] == EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK)) | |
174 | runtime_nsecs = count[0]; | |
2996f5dd IM |
175 | } |
176 | ||
177 | /* | |
178 | * Print out the results of a single counter: | |
179 | */ | |
180 | static void print_counter(int counter) | |
181 | { | |
182 | __u64 *count; | |
183 | int scaled; | |
184 | ||
185 | count = event_res[counter]; | |
186 | scaled = event_scaled[counter]; | |
187 | ||
188 | if (scaled == -1) { | |
189 | fprintf(stderr, " %14s %-20s\n", | |
190 | "<not counted>", event_name(counter)); | |
191 | return; | |
192 | } | |
c04f5e5d IM |
193 | |
194 | if (nsec_counter(counter)) { | |
195 | double msecs = (double)count[0] / 1000000; | |
196 | ||
197 | fprintf(stderr, " %14.6f %-20s (msecs)", | |
198 | msecs, event_name(counter)); | |
199 | } else { | |
be1ac0d8 | 200 | fprintf(stderr, " %14Ld %-20s", |
c04f5e5d | 201 | count[0], event_name(counter)); |
be1ac0d8 IM |
202 | if (runtime_nsecs) |
203 | fprintf(stderr, " # %12.3f M/sec", | |
204 | (double)count[0]/runtime_nsecs*1000.0); | |
c04f5e5d IM |
205 | } |
206 | if (scaled) | |
207 | fprintf(stderr, " (scaled from %.2f%%)", | |
208 | (double) count[2] / count[1] * 100); | |
209 | fprintf(stderr, "\n"); | |
210 | } | |
211 | ||
16f762a2 | 212 | static int do_perfstat(int argc, const char **argv) |
ddcacfa0 IM |
213 | { |
214 | unsigned long long t0, t1; | |
215 | int counter; | |
ddcacfa0 IM |
216 | int status; |
217 | int pid; | |
218 | ||
219 | if (!system_wide) | |
220 | nr_cpus = 1; | |
221 | ||
222 | for (counter = 0; counter < nr_counters; counter++) | |
223 | create_perfstat_counter(counter); | |
224 | ||
ddcacfa0 IM |
225 | /* |
226 | * Enable counters and exec the command: | |
227 | */ | |
228 | t0 = rdclock(); | |
229 | prctl(PR_TASK_PERF_COUNTERS_ENABLE); | |
230 | ||
231 | if ((pid = fork()) < 0) | |
232 | perror("failed to fork"); | |
233 | if (!pid) { | |
5242519b | 234 | if (execvp(argv[0], (char **)argv)) { |
ddcacfa0 IM |
235 | perror(argv[0]); |
236 | exit(-1); | |
237 | } | |
238 | } | |
239 | while (wait(&status) >= 0) | |
240 | ; | |
241 | prctl(PR_TASK_PERF_COUNTERS_DISABLE); | |
242 | t1 = rdclock(); | |
243 | ||
244 | fflush(stdout); | |
245 | ||
246 | fprintf(stderr, "\n"); | |
247 | fprintf(stderr, " Performance counter stats for \'%s\':\n", | |
248 | argv[0]); | |
249 | fprintf(stderr, "\n"); | |
250 | ||
2996f5dd IM |
251 | for (counter = 0; counter < nr_counters; counter++) |
252 | read_counter(counter); | |
253 | ||
c04f5e5d IM |
254 | for (counter = 0; counter < nr_counters; counter++) |
255 | print_counter(counter); | |
ddcacfa0 | 256 | |
ddcacfa0 | 257 | |
ddcacfa0 IM |
258 | fprintf(stderr, "\n"); |
259 | fprintf(stderr, " Wall-clock time elapsed: %12.6f msecs\n", | |
260 | (double)(t1-t0)/1e6); | |
261 | fprintf(stderr, "\n"); | |
262 | ||
263 | return 0; | |
264 | } | |
265 | ||
5242519b | 266 | static void skip_signal(int signo) |
ddcacfa0 | 267 | { |
5242519b IM |
268 | } |
269 | ||
270 | static const char * const stat_usage[] = { | |
271 | "perf stat [<options>] <command>", | |
272 | NULL | |
273 | }; | |
274 | ||
275 | static char events_help_msg[EVENTS_HELP_MAX]; | |
276 | ||
277 | static const struct option options[] = { | |
278 | OPT_CALLBACK('e', "event", NULL, "event", | |
279 | events_help_msg, parse_events), | |
280 | OPT_INTEGER('c', "count", &default_interval, | |
281 | "event period to sample"), | |
282 | OPT_BOOLEAN('i', "inherit", &inherit, | |
283 | "child tasks inherit counters"), | |
284 | OPT_INTEGER('p', "pid", &target_pid, | |
285 | "stat events on existing pid"), | |
286 | OPT_BOOLEAN('a', "all-cpus", &system_wide, | |
287 | "system-wide collection from all CPUs"), | |
288 | OPT_BOOLEAN('l', "scale", &scale, | |
289 | "scale/normalize counters"), | |
290 | OPT_END() | |
291 | }; | |
292 | ||
293 | int cmd_stat(int argc, const char **argv, const char *prefix) | |
294 | { | |
295 | int counter; | |
296 | ||
297 | page_size = sysconf(_SC_PAGE_SIZE); | |
298 | ||
299 | create_events_help(events_help_msg); | |
300 | memcpy(event_id, default_event_id, sizeof(default_event_id)); | |
301 | ||
302 | argc = parse_options(argc, argv, options, stat_usage, 0); | |
303 | if (!argc) | |
304 | usage_with_options(stat_usage, options); | |
ddcacfa0 IM |
305 | |
306 | if (!nr_counters) { | |
307 | nr_counters = 8; | |
308 | } | |
309 | ||
310 | for (counter = 0; counter < nr_counters; counter++) { | |
311 | if (event_count[counter]) | |
312 | continue; | |
313 | ||
314 | event_count[counter] = default_interval; | |
315 | } | |
ddcacfa0 IM |
316 | nr_cpus = sysconf(_SC_NPROCESSORS_ONLN); |
317 | assert(nr_cpus <= MAX_NR_CPUS); | |
318 | assert(nr_cpus >= 0); | |
319 | ||
58d7e993 IM |
320 | /* |
321 | * We dont want to block the signals - that would cause | |
322 | * child tasks to inherit that and Ctrl-C would not work. | |
323 | * What we want is for Ctrl-C to work in the exec()-ed | |
324 | * task, but being ignored by perf stat itself: | |
325 | */ | |
326 | signal(SIGINT, skip_signal); | |
327 | signal(SIGALRM, skip_signal); | |
328 | signal(SIGABRT, skip_signal); | |
329 | ||
ddcacfa0 IM |
330 | return do_perfstat(argc, argv); |
331 | } |