4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ~/hackbench 10
12 Performance counter stats for '/home/mingo/hackbench':
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
23 Wall-clock time elapsed: 123.786620 msecs
26 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
28 * Improvements and fixes by:
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>
36 * Released under the GPL v2. (and only v2, not any later version)
41 #include "util/util.h"
42 #include "util/parse-options.h"
43 #include "util/parse-events.h"
45 #include <sys/prctl.h>
47 static struct perf_counter_attr default_attrs[MAX_COUNTERS] = {
49 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_TASK_CLOCK },
50 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_CONTEXT_SWITCHES },
51 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_CPU_MIGRATIONS },
52 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_PAGE_FAULTS },
54 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_CPU_CYCLES },
55 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_INSTRUCTIONS },
56 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_CACHE_REFERENCES },
57 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_CACHE_MISSES },
60 static int system_wide = 0;
61 static int inherit = 1;
63 static int fd[MAX_NR_CPUS][MAX_COUNTERS];
65 static int target_pid = -1;
66 static int nr_cpus = 0;
67 static unsigned int page_size;
71 static const unsigned int default_count[] = {
80 static __u64 event_res[MAX_COUNTERS][3];
81 static __u64 event_scaled[MAX_COUNTERS];
83 static __u64 runtime_nsecs;
84 static __u64 walltime_nsecs;
86 static void create_perfstat_counter(int counter)
88 struct perf_counter_attr *attr = attrs + counter;
91 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
92 PERF_FORMAT_TOTAL_TIME_RUNNING;
96 for (cpu = 0; cpu < nr_cpus; cpu ++) {
97 fd[cpu][counter] = sys_perf_counter_open(attr, -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));
105 attr->inherit = inherit;
108 fd[0][counter] = sys_perf_counter_open(attr, 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));
118 * Does the counter have nsecs as a unit?
120 static inline int nsec_counter(int counter)
122 if (attrs[counter].type != PERF_TYPE_SOFTWARE)
125 if (attrs[counter].config == PERF_COUNT_CPU_CLOCK)
128 if (attrs[counter].config == PERF_COUNT_TASK_CLOCK)
135 * Read out the results of a single counter:
137 static void read_counter(int counter)
139 __u64 *count, single_count[3];
144 count = event_res[counter];
146 count[0] = count[1] = count[2] = 0;
149 for (cpu = 0; cpu < nr_cpus; cpu ++) {
150 res = read(fd[cpu][counter], single_count, nv * sizeof(__u64));
151 assert(res == nv * sizeof(__u64));
153 count[0] += single_count[0];
155 count[1] += single_count[1];
156 count[2] += single_count[2];
163 event_scaled[counter] = -1;
168 if (count[2] < count[1]) {
169 event_scaled[counter] = 1;
170 count[0] = (unsigned long long)
171 ((double)count[0] * count[1] / count[2] + 0.5);
175 * Save the full runtime - to allow normalization during printout:
177 if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
178 attrs[counter].config == PERF_COUNT_TASK_CLOCK)
179 runtime_nsecs = count[0];
183 * Print out the results of a single counter:
185 static void print_counter(int counter)
190 count = event_res[counter];
191 scaled = event_scaled[counter];
194 fprintf(stderr, " %14s %-20s\n",
195 "<not counted>", event_name(counter));
199 if (nsec_counter(counter)) {
200 double msecs = (double)count[0] / 1000000;
202 fprintf(stderr, " %14.6f %-20s",
203 msecs, event_name(counter));
204 if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
205 attrs[counter].config == PERF_COUNT_TASK_CLOCK) {
207 fprintf(stderr, " # %11.3f CPU utilization factor",
208 (double)count[0] / (double)walltime_nsecs);
211 fprintf(stderr, " %14Ld %-20s",
212 count[0], event_name(counter));
214 fprintf(stderr, " # %11.3f M/sec",
215 (double)count[0]/runtime_nsecs*1000.0);
218 fprintf(stderr, " (scaled from %.2f%%)",
219 (double) count[2] / count[1] * 100);
220 fprintf(stderr, "\n");
223 static int do_perfstat(int argc, const char **argv)
225 unsigned long long t0, t1;
234 for (counter = 0; counter < nr_counters; counter++)
235 create_perfstat_counter(counter);
238 * Enable counters and exec the command:
241 prctl(PR_TASK_PERF_COUNTERS_ENABLE);
243 if ((pid = fork()) < 0)
244 perror("failed to fork");
247 if (execvp(argv[0], (char **)argv)) {
253 while (wait(&status) >= 0)
256 prctl(PR_TASK_PERF_COUNTERS_DISABLE);
259 walltime_nsecs = t1 - t0;
263 fprintf(stderr, "\n");
264 fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
266 for (i = 1; i < argc; i++)
267 fprintf(stderr, " %s", argv[i]);
269 fprintf(stderr, "\':\n");
270 fprintf(stderr, "\n");
272 for (counter = 0; counter < nr_counters; counter++)
273 read_counter(counter);
275 for (counter = 0; counter < nr_counters; counter++)
276 print_counter(counter);
279 fprintf(stderr, "\n");
280 fprintf(stderr, " Wall-clock time elapsed: %12.6f msecs\n",
281 (double)(t1-t0)/1e6);
282 fprintf(stderr, "\n");
287 static void skip_signal(int signo)
291 static const char * const stat_usage[] = {
292 "perf stat [<options>] <command>",
296 static const struct option options[] = {
297 OPT_CALLBACK('e', "event", NULL, "event",
298 "event selector. use 'perf list' to list available events",
300 OPT_BOOLEAN('i', "inherit", &inherit,
301 "child tasks inherit counters"),
302 OPT_INTEGER('p', "pid", &target_pid,
303 "stat events on existing pid"),
304 OPT_BOOLEAN('a', "all-cpus", &system_wide,
305 "system-wide collection from all CPUs"),
306 OPT_BOOLEAN('S', "scale", &scale,
307 "scale/normalize counters"),
311 int cmd_stat(int argc, const char **argv, const char *prefix)
313 page_size = sysconf(_SC_PAGE_SIZE);
315 memcpy(attrs, default_attrs, sizeof(attrs));
317 argc = parse_options(argc, argv, options, stat_usage, 0);
319 usage_with_options(stat_usage, options);
324 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
325 assert(nr_cpus <= MAX_NR_CPUS);
326 assert(nr_cpus >= 0);
329 * We dont want to block the signals - that would cause
330 * child tasks to inherit that and Ctrl-C would not work.
331 * What we want is for Ctrl-C to work in the exec()-ed
332 * task, but being ignored by perf stat itself:
334 signal(SIGINT, skip_signal);
335 signal(SIGALRM, skip_signal);
336 signal(SIGABRT, skip_signal);
338 return do_perfstat(argc, argv);