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>
35 * Jaswinder Singh Rajput <jaswinder@kernel.org>
37 * Released under the GPL v2. (and only v2, not any later version)
42 #include "util/util.h"
43 #include "util/parse-options.h"
44 #include "util/parse-events.h"
45 #include "util/event.h"
46 #include "util/debug.h"
48 #include <sys/prctl.h>
51 static struct perf_counter_attr default_attrs[] = {
53 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
54 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
55 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
56 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
58 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
59 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
60 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
61 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES },
65 static int system_wide = 0;
66 static unsigned int nr_cpus = 0;
67 static int run_idx = 0;
69 static int run_count = 1;
70 static int inherit = 1;
72 static int target_pid = -1;
73 static int null_run = 0;
75 static int fd[MAX_NR_CPUS][MAX_COUNTERS];
77 static u64 event_res[MAX_COUNTERS][3];
78 static int event_scaled[MAX_COUNTERS];
86 static void update_stats(struct stats *stats, u64 val)
91 stats->sum_sq += sq * sq;
94 static double avg_stats(struct stats *stats)
96 return stats->sum / run_count;
100 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
102 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
103 * s^2 -------------------------------
106 * http://en.wikipedia.org/wiki/Stddev
108 * The std dev of the mean is related to the std dev by:
115 static double stddev_stats(struct stats *stats)
117 double avg = stats->sum / run_count;
118 double variance = (stats->sum_sq - stats->sum*avg)/(run_count - 1);
119 double variance_mean = variance / run_count;
121 return sqrt(variance_mean);
124 struct stats event_res_stats[MAX_COUNTERS][3];
125 struct stats runtime_nsecs_stats;
126 struct stats walltime_nsecs_stats;
127 struct stats runtime_cycles_stats;
129 #define MATCH_EVENT(t, c, counter) \
130 (attrs[counter].type == PERF_TYPE_##t && \
131 attrs[counter].config == PERF_COUNT_##c)
133 #define ERR_PERF_OPEN \
134 "Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
136 static void create_perf_stat_counter(int counter, int pid)
138 struct perf_counter_attr *attr = attrs + counter;
141 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
142 PERF_FORMAT_TOTAL_TIME_RUNNING;
147 for (cpu = 0; cpu < nr_cpus; cpu++) {
148 fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
149 if (fd[cpu][counter] < 0 && verbose)
150 fprintf(stderr, ERR_PERF_OPEN, counter,
151 fd[cpu][counter], strerror(errno));
154 attr->inherit = inherit;
156 attr->enable_on_exec = 1;
158 fd[0][counter] = sys_perf_counter_open(attr, pid, -1, -1, 0);
159 if (fd[0][counter] < 0 && verbose)
160 fprintf(stderr, ERR_PERF_OPEN, counter,
161 fd[0][counter], strerror(errno));
166 * Does the counter have nsecs as a unit?
168 static inline int nsec_counter(int counter)
170 if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) ||
171 MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
178 * Read out the results of a single counter:
180 static void read_counter(int counter)
182 u64 *count, single_count[3];
188 count = event_res[counter];
190 count[0] = count[1] = count[2] = 0;
193 for (cpu = 0; cpu < nr_cpus; cpu++) {
194 if (fd[cpu][counter] < 0)
197 res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
198 assert(res == nv * sizeof(u64));
200 close(fd[cpu][counter]);
201 fd[cpu][counter] = -1;
203 count[0] += single_count[0];
205 count[1] += single_count[1];
206 count[2] += single_count[2];
213 event_scaled[counter] = -1;
218 if (count[2] < count[1]) {
219 event_scaled[counter] = 1;
220 count[0] = (unsigned long long)
221 ((double)count[0] * count[1] / count[2] + 0.5);
225 for (i = 0; i < 3; i++)
226 update_stats(&event_res_stats[counter][i], count[i]);
229 fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
230 count[0], count[1], count[2]);
234 * Save the full runtime - to allow normalization during printout:
236 if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
237 update_stats(&runtime_nsecs_stats, count[0]);
238 if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
239 update_stats(&runtime_cycles_stats, count[0]);
242 static int run_perf_stat(int argc __used, const char **argv)
244 unsigned long long t0, t1;
248 int child_ready_pipe[2], go_pipe[2];
254 if (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0) {
255 perror("failed to create pipes");
259 if ((pid = fork()) < 0)
260 perror("failed to fork");
263 close(child_ready_pipe[0]);
265 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
268 * Do a dummy execvp to get the PLT entry resolved,
269 * so we avoid the resolver overhead on the real
272 execvp("", (char **)argv);
275 * Tell the parent we're ready to go
277 close(child_ready_pipe[1]);
280 * Wait until the parent tells us to go.
282 if (read(go_pipe[0], &buf, 1) == -1)
283 perror("unable to read pipe");
285 execvp(argv[0], (char **)argv);
292 * Wait for the child to be ready to exec.
294 close(child_ready_pipe[1]);
296 if (read(child_ready_pipe[0], &buf, 1) == -1)
297 perror("unable to read pipe");
298 close(child_ready_pipe[0]);
300 for (counter = 0; counter < nr_counters; counter++)
301 create_perf_stat_counter(counter, pid);
304 * Enable counters and exec the command:
313 update_stats(&walltime_nsecs_stats, t1 - t0);
315 for (counter = 0; counter < nr_counters; counter++)
316 read_counter(counter);
318 return WEXITSTATUS(status);
321 static void print_noise(double avg, double stddev)
324 fprintf(stderr, " ( +- %7.3f%% )", 100*stddev / avg);
327 static void nsec_printout(int counter, double avg, double stddev)
329 double msecs = avg / 1e6;
331 fprintf(stderr, " %14.6f %-24s", msecs, event_name(counter));
333 if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
334 fprintf(stderr, " # %10.3f CPUs ",
335 avg / avg_stats(&walltime_nsecs_stats));
337 print_noise(avg, stddev);
340 static void abs_printout(int counter, double avg, double stddev)
342 fprintf(stderr, " %14.0f %-24s", avg, event_name(counter));
344 if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
345 fprintf(stderr, " # %10.3f IPC ",
346 avg / avg_stats(&runtime_cycles_stats));
348 fprintf(stderr, " # %10.3f M/sec",
349 1000.0 * avg / avg_stats(&runtime_nsecs_stats));
351 print_noise(avg, stddev);
355 * Print out the results of a single counter:
357 static void print_counter(int counter)
360 int scaled = event_scaled[counter];
362 avg = avg_stats(&event_res_stats[counter][0]);
363 stddev = stddev_stats(&event_res_stats[counter][0]);
366 fprintf(stderr, " %14s %-24s\n",
367 "<not counted>", event_name(counter));
371 if (nsec_counter(counter))
372 nsec_printout(counter, avg, stddev);
374 abs_printout(counter, avg, stddev);
377 double avg_enabled, avg_running;
379 avg_enabled = avg_stats(&event_res_stats[counter][1]);
380 avg_running = avg_stats(&event_res_stats[counter][2]);
382 fprintf(stderr, " (scaled from %.2f%%)",
383 100 * avg_running / avg_enabled);
386 fprintf(stderr, "\n");
389 static void print_stat(int argc, const char **argv)
395 fprintf(stderr, "\n");
396 fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
398 for (i = 1; i < argc; i++)
399 fprintf(stderr, " %s", argv[i]);
401 fprintf(stderr, "\'");
403 fprintf(stderr, " (%d runs)", run_count);
404 fprintf(stderr, ":\n\n");
406 for (counter = 0; counter < nr_counters; counter++)
407 print_counter(counter);
409 fprintf(stderr, "\n");
410 fprintf(stderr, " %14.9f seconds time elapsed",
411 avg_stats(&walltime_nsecs_stats)/1e9);
413 fprintf(stderr, " ( +- %7.3f%% )",
414 100*stddev_stats(&walltime_nsecs_stats) /
415 avg_stats(&walltime_nsecs_stats));
417 fprintf(stderr, "\n\n");
420 static volatile int signr = -1;
422 static void skip_signal(int signo)
427 static void sig_atexit(void)
432 signal(signr, SIG_DFL);
433 kill(getpid(), signr);
436 static const char * const stat_usage[] = {
437 "perf stat [<options>] <command>",
441 static const struct option options[] = {
442 OPT_CALLBACK('e', "event", NULL, "event",
443 "event selector. use 'perf list' to list available events",
445 OPT_BOOLEAN('i', "inherit", &inherit,
446 "child tasks inherit counters"),
447 OPT_INTEGER('p', "pid", &target_pid,
448 "stat events on existing pid"),
449 OPT_BOOLEAN('a', "all-cpus", &system_wide,
450 "system-wide collection from all CPUs"),
451 OPT_BOOLEAN('c', "scale", &scale,
452 "scale/normalize counters"),
453 OPT_BOOLEAN('v', "verbose", &verbose,
454 "be more verbose (show counter open errors, etc)"),
455 OPT_INTEGER('r', "repeat", &run_count,
456 "repeat command and print average + stddev (max: 100)"),
457 OPT_BOOLEAN('n', "null", &null_run,
458 "null run - dont start any counters"),
462 int cmd_stat(int argc, const char **argv, const char *prefix __used)
466 argc = parse_options(argc, argv, options, stat_usage,
467 PARSE_OPT_STOP_AT_NON_OPTION);
469 usage_with_options(stat_usage, options);
471 usage_with_options(stat_usage, options);
473 /* Set attrs and nr_counters if no event is selected and !null_run */
474 if (!null_run && !nr_counters) {
475 memcpy(attrs, default_attrs, sizeof(default_attrs));
476 nr_counters = ARRAY_SIZE(default_attrs);
479 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
480 assert(nr_cpus <= MAX_NR_CPUS);
481 assert((int)nr_cpus >= 0);
484 * We dont want to block the signals - that would cause
485 * child tasks to inherit that and Ctrl-C would not work.
486 * What we want is for Ctrl-C to work in the exec()-ed
487 * task, but being ignored by perf stat itself:
490 signal(SIGINT, skip_signal);
491 signal(SIGALRM, skip_signal);
492 signal(SIGABRT, skip_signal);
495 for (run_idx = 0; run_idx < run_count; run_idx++) {
496 if (run_count != 1 && verbose)
497 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
498 status = run_perf_stat(argc, argv);
501 print_stat(argc, argv);