hwmon: (lm95241) Rewrite to avoid using macros
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / tools / perf / builtin-stat.c
blob02b2d8013a61e537bf80e5d110d6c2f2c3f6b150
1 /*
2 * builtin-stat.c
4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
7 * Sample output:
9 $ perf stat ~/hackbench 10
10 Time: 0.104
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)
40 #include "perf.h"
41 #include "builtin.h"
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/evsel.h"
47 #include "util/debug.h"
48 #include "util/header.h"
49 #include "util/cpumap.h"
50 #include "util/thread.h"
52 #include <sys/prctl.h>
53 #include <math.h>
54 #include <locale.h>
56 #define DEFAULT_SEPARATOR " "
58 static struct perf_event_attr default_attrs[] = {
60 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
61 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
62 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
63 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
65 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
66 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
67 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
68 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
69 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES },
70 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES },
74 static bool system_wide = false;
75 static struct cpu_map *cpus;
76 static int run_idx = 0;
78 static int run_count = 1;
79 static bool no_inherit = false;
80 static bool scale = true;
81 static bool no_aggr = false;
82 static pid_t target_pid = -1;
83 static pid_t target_tid = -1;
84 static struct thread_map *threads;
85 static pid_t child_pid = -1;
86 static bool null_run = false;
87 static bool big_num = true;
88 static int big_num_opt = -1;
89 static const char *cpu_list;
90 static const char *csv_sep = NULL;
91 static bool csv_output = false;
93 static volatile int done = 0;
95 struct stats
97 double n, mean, M2;
100 struct perf_stat {
101 struct stats res_stats[3];
104 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
106 evsel->priv = zalloc(sizeof(struct perf_stat));
107 return evsel->priv == NULL ? -ENOMEM : 0;
110 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
112 free(evsel->priv);
113 evsel->priv = NULL;
116 static void update_stats(struct stats *stats, u64 val)
118 double delta;
120 stats->n++;
121 delta = val - stats->mean;
122 stats->mean += delta / stats->n;
123 stats->M2 += delta*(val - stats->mean);
126 static double avg_stats(struct stats *stats)
128 return stats->mean;
132 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
134 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
135 * s^2 = -------------------------------
136 * n - 1
138 * http://en.wikipedia.org/wiki/Stddev
140 * The std dev of the mean is related to the std dev by:
143 * s_mean = -------
144 * sqrt(n)
147 static double stddev_stats(struct stats *stats)
149 double variance = stats->M2 / (stats->n - 1);
150 double variance_mean = variance / stats->n;
152 return sqrt(variance_mean);
155 struct stats runtime_nsecs_stats[MAX_NR_CPUS];
156 struct stats runtime_cycles_stats[MAX_NR_CPUS];
157 struct stats runtime_branches_stats[MAX_NR_CPUS];
158 struct stats walltime_nsecs_stats;
160 static int create_perf_stat_counter(struct perf_evsel *evsel)
162 struct perf_event_attr *attr = &evsel->attr;
164 if (scale)
165 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
166 PERF_FORMAT_TOTAL_TIME_RUNNING;
168 if (system_wide)
169 return perf_evsel__open_per_cpu(evsel, cpus);
171 attr->inherit = !no_inherit;
172 if (target_pid == -1 && target_tid == -1) {
173 attr->disabled = 1;
174 attr->enable_on_exec = 1;
177 return perf_evsel__open_per_thread(evsel, threads);
181 * Does the counter have nsecs as a unit?
183 static inline int nsec_counter(struct perf_evsel *evsel)
185 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
186 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
187 return 1;
189 return 0;
193 * Read out the results of a single counter:
194 * aggregate counts across CPUs in system-wide mode
196 static int read_counter_aggr(struct perf_evsel *counter)
198 struct perf_stat *ps = counter->priv;
199 u64 *count = counter->counts->aggr.values;
200 int i;
202 if (__perf_evsel__read(counter, cpus->nr, threads->nr, scale) < 0)
203 return -1;
205 for (i = 0; i < 3; i++)
206 update_stats(&ps->res_stats[i], count[i]);
208 if (verbose) {
209 fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
210 count[0], count[1], count[2]);
214 * Save the full runtime - to allow normalization during printout:
216 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
217 update_stats(&runtime_nsecs_stats[0], count[0]);
218 if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
219 update_stats(&runtime_cycles_stats[0], count[0]);
220 if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
221 update_stats(&runtime_branches_stats[0], count[0]);
223 return 0;
227 * Read out the results of a single counter:
228 * do not aggregate counts across CPUs in system-wide mode
230 static int read_counter(struct perf_evsel *counter)
232 u64 *count;
233 int cpu;
235 for (cpu = 0; cpu < cpus->nr; cpu++) {
236 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
237 return -1;
239 count = counter->counts->cpu[cpu].values;
241 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
242 update_stats(&runtime_nsecs_stats[cpu], count[0]);
243 if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
244 update_stats(&runtime_cycles_stats[cpu], count[0]);
245 if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
246 update_stats(&runtime_branches_stats[cpu], count[0]);
249 return 0;
252 static int run_perf_stat(int argc __used, const char **argv)
254 unsigned long long t0, t1;
255 struct perf_evsel *counter;
256 int status = 0;
257 int child_ready_pipe[2], go_pipe[2];
258 const bool forks = (argc > 0);
259 char buf;
261 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
262 perror("failed to create pipes");
263 exit(1);
266 if (forks) {
267 if ((child_pid = fork()) < 0)
268 perror("failed to fork");
270 if (!child_pid) {
271 close(child_ready_pipe[0]);
272 close(go_pipe[1]);
273 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
276 * Do a dummy execvp to get the PLT entry resolved,
277 * so we avoid the resolver overhead on the real
278 * execvp call.
280 execvp("", (char **)argv);
283 * Tell the parent we're ready to go
285 close(child_ready_pipe[1]);
288 * Wait until the parent tells us to go.
290 if (read(go_pipe[0], &buf, 1) == -1)
291 perror("unable to read pipe");
293 execvp(argv[0], (char **)argv);
295 perror(argv[0]);
296 exit(-1);
299 if (target_tid == -1 && target_pid == -1 && !system_wide)
300 threads->map[0] = child_pid;
303 * Wait for the child to be ready to exec.
305 close(child_ready_pipe[1]);
306 close(go_pipe[0]);
307 if (read(child_ready_pipe[0], &buf, 1) == -1)
308 perror("unable to read pipe");
309 close(child_ready_pipe[0]);
312 list_for_each_entry(counter, &evsel_list, node) {
313 if (create_perf_stat_counter(counter) < 0) {
314 if (errno == -EPERM || errno == -EACCES) {
315 error("You may not have permission to collect %sstats.\n"
316 "\t Consider tweaking"
317 " /proc/sys/kernel/perf_event_paranoid or running as root.",
318 system_wide ? "system-wide " : "");
319 } else {
320 error("open_counter returned with %d (%s). "
321 "/bin/dmesg may provide additional information.\n",
322 errno, strerror(errno));
324 if (child_pid != -1)
325 kill(child_pid, SIGTERM);
326 die("Not all events could be opened.\n");
327 return -1;
332 * Enable counters and exec the command:
334 t0 = rdclock();
336 if (forks) {
337 close(go_pipe[1]);
338 wait(&status);
339 } else {
340 while(!done) sleep(1);
343 t1 = rdclock();
345 update_stats(&walltime_nsecs_stats, t1 - t0);
347 if (no_aggr) {
348 list_for_each_entry(counter, &evsel_list, node) {
349 read_counter(counter);
350 perf_evsel__close_fd(counter, cpus->nr, 1);
352 } else {
353 list_for_each_entry(counter, &evsel_list, node) {
354 read_counter_aggr(counter);
355 perf_evsel__close_fd(counter, cpus->nr, threads->nr);
359 return WEXITSTATUS(status);
362 static void print_noise(struct perf_evsel *evsel, double avg)
364 struct perf_stat *ps;
366 if (run_count == 1)
367 return;
369 ps = evsel->priv;
370 fprintf(stderr, " ( +- %7.3f%% )",
371 100 * stddev_stats(&ps->res_stats[0]) / avg);
374 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
376 double msecs = avg / 1e6;
377 char cpustr[16] = { '\0', };
378 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
380 if (no_aggr)
381 sprintf(cpustr, "CPU%*d%s",
382 csv_output ? 0 : -4,
383 cpus->map[cpu], csv_sep);
385 fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel));
387 if (csv_output)
388 return;
390 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
391 fprintf(stderr, " # %10.3f CPUs ",
392 avg / avg_stats(&walltime_nsecs_stats));
395 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
397 double total, ratio = 0.0;
398 char cpustr[16] = { '\0', };
399 const char *fmt;
401 if (csv_output)
402 fmt = "%s%.0f%s%s";
403 else if (big_num)
404 fmt = "%s%'18.0f%s%-24s";
405 else
406 fmt = "%s%18.0f%s%-24s";
408 if (no_aggr)
409 sprintf(cpustr, "CPU%*d%s",
410 csv_output ? 0 : -4,
411 cpus->map[cpu], csv_sep);
412 else
413 cpu = 0;
415 fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel));
417 if (csv_output)
418 return;
420 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
421 total = avg_stats(&runtime_cycles_stats[cpu]);
423 if (total)
424 ratio = avg / total;
426 fprintf(stderr, " # %10.3f IPC ", ratio);
427 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
428 runtime_branches_stats[cpu].n != 0) {
429 total = avg_stats(&runtime_branches_stats[cpu]);
431 if (total)
432 ratio = avg * 100 / total;
434 fprintf(stderr, " # %10.3f %% ", ratio);
436 } else if (runtime_nsecs_stats[cpu].n != 0) {
437 total = avg_stats(&runtime_nsecs_stats[cpu]);
439 if (total)
440 ratio = 1000.0 * avg / total;
442 fprintf(stderr, " # %10.3f M/sec", ratio);
447 * Print out the results of a single counter:
448 * aggregated counts in system-wide mode
450 static void print_counter_aggr(struct perf_evsel *counter)
452 struct perf_stat *ps = counter->priv;
453 double avg = avg_stats(&ps->res_stats[0]);
454 int scaled = counter->counts->scaled;
456 if (scaled == -1) {
457 fprintf(stderr, "%*s%s%-24s\n",
458 csv_output ? 0 : 18,
459 "<not counted>", csv_sep, event_name(counter));
460 return;
463 if (nsec_counter(counter))
464 nsec_printout(-1, counter, avg);
465 else
466 abs_printout(-1, counter, avg);
468 if (csv_output) {
469 fputc('\n', stderr);
470 return;
473 print_noise(counter, avg);
475 if (scaled) {
476 double avg_enabled, avg_running;
478 avg_enabled = avg_stats(&ps->res_stats[1]);
479 avg_running = avg_stats(&ps->res_stats[2]);
481 fprintf(stderr, " (scaled from %.2f%%)",
482 100 * avg_running / avg_enabled);
485 fprintf(stderr, "\n");
489 * Print out the results of a single counter:
490 * does not use aggregated count in system-wide
492 static void print_counter(struct perf_evsel *counter)
494 u64 ena, run, val;
495 int cpu;
497 for (cpu = 0; cpu < cpus->nr; cpu++) {
498 val = counter->counts->cpu[cpu].val;
499 ena = counter->counts->cpu[cpu].ena;
500 run = counter->counts->cpu[cpu].run;
501 if (run == 0 || ena == 0) {
502 fprintf(stderr, "CPU%*d%s%*s%s%-24s",
503 csv_output ? 0 : -4,
504 cpus->map[cpu], csv_sep,
505 csv_output ? 0 : 18,
506 "<not counted>", csv_sep,
507 event_name(counter));
509 fprintf(stderr, "\n");
510 continue;
513 if (nsec_counter(counter))
514 nsec_printout(cpu, counter, val);
515 else
516 abs_printout(cpu, counter, val);
518 if (!csv_output) {
519 print_noise(counter, 1.0);
521 if (run != ena) {
522 fprintf(stderr, " (scaled from %.2f%%)",
523 100.0 * run / ena);
526 fprintf(stderr, "\n");
530 static void print_stat(int argc, const char **argv)
532 struct perf_evsel *counter;
533 int i;
535 fflush(stdout);
537 if (!csv_output) {
538 fprintf(stderr, "\n");
539 fprintf(stderr, " Performance counter stats for ");
540 if(target_pid == -1 && target_tid == -1) {
541 fprintf(stderr, "\'%s", argv[0]);
542 for (i = 1; i < argc; i++)
543 fprintf(stderr, " %s", argv[i]);
544 } else if (target_pid != -1)
545 fprintf(stderr, "process id \'%d", target_pid);
546 else
547 fprintf(stderr, "thread id \'%d", target_tid);
549 fprintf(stderr, "\'");
550 if (run_count > 1)
551 fprintf(stderr, " (%d runs)", run_count);
552 fprintf(stderr, ":\n\n");
555 if (no_aggr) {
556 list_for_each_entry(counter, &evsel_list, node)
557 print_counter(counter);
558 } else {
559 list_for_each_entry(counter, &evsel_list, node)
560 print_counter_aggr(counter);
563 if (!csv_output) {
564 fprintf(stderr, "\n");
565 fprintf(stderr, " %18.9f seconds time elapsed",
566 avg_stats(&walltime_nsecs_stats)/1e9);
567 if (run_count > 1) {
568 fprintf(stderr, " ( +- %7.3f%% )",
569 100*stddev_stats(&walltime_nsecs_stats) /
570 avg_stats(&walltime_nsecs_stats));
572 fprintf(stderr, "\n\n");
576 static volatile int signr = -1;
578 static void skip_signal(int signo)
580 if(child_pid == -1)
581 done = 1;
583 signr = signo;
586 static void sig_atexit(void)
588 if (child_pid != -1)
589 kill(child_pid, SIGTERM);
591 if (signr == -1)
592 return;
594 signal(signr, SIG_DFL);
595 kill(getpid(), signr);
598 static const char * const stat_usage[] = {
599 "perf stat [<options>] [<command>]",
600 NULL
603 static int stat__set_big_num(const struct option *opt __used,
604 const char *s __used, int unset)
606 big_num_opt = unset ? 0 : 1;
607 return 0;
610 static const struct option options[] = {
611 OPT_CALLBACK('e', "event", NULL, "event",
612 "event selector. use 'perf list' to list available events",
613 parse_events),
614 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
615 "child tasks do not inherit counters"),
616 OPT_INTEGER('p', "pid", &target_pid,
617 "stat events on existing process id"),
618 OPT_INTEGER('t', "tid", &target_tid,
619 "stat events on existing thread id"),
620 OPT_BOOLEAN('a', "all-cpus", &system_wide,
621 "system-wide collection from all CPUs"),
622 OPT_BOOLEAN('c', "scale", &scale,
623 "scale/normalize counters"),
624 OPT_INCR('v', "verbose", &verbose,
625 "be more verbose (show counter open errors, etc)"),
626 OPT_INTEGER('r', "repeat", &run_count,
627 "repeat command and print average + stddev (max: 100)"),
628 OPT_BOOLEAN('n', "null", &null_run,
629 "null run - dont start any counters"),
630 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
631 "print large numbers with thousands\' separators",
632 stat__set_big_num),
633 OPT_STRING('C', "cpu", &cpu_list, "cpu",
634 "list of cpus to monitor in system-wide"),
635 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
636 "disable CPU count aggregation"),
637 OPT_STRING('x', "field-separator", &csv_sep, "separator",
638 "print counts with custom separator"),
639 OPT_END()
642 int cmd_stat(int argc, const char **argv, const char *prefix __used)
644 struct perf_evsel *pos;
645 int status = -ENOMEM;
647 setlocale(LC_ALL, "");
649 argc = parse_options(argc, argv, options, stat_usage,
650 PARSE_OPT_STOP_AT_NON_OPTION);
652 if (csv_sep)
653 csv_output = true;
654 else
655 csv_sep = DEFAULT_SEPARATOR;
658 * let the spreadsheet do the pretty-printing
660 if (csv_output) {
661 /* User explicitely passed -B? */
662 if (big_num_opt == 1) {
663 fprintf(stderr, "-B option not supported with -x\n");
664 usage_with_options(stat_usage, options);
665 } else /* Nope, so disable big number formatting */
666 big_num = false;
667 } else if (big_num_opt == 0) /* User passed --no-big-num */
668 big_num = false;
670 if (!argc && target_pid == -1 && target_tid == -1)
671 usage_with_options(stat_usage, options);
672 if (run_count <= 0)
673 usage_with_options(stat_usage, options);
675 /* no_aggr is for system-wide only */
676 if (no_aggr && !system_wide)
677 usage_with_options(stat_usage, options);
679 /* Set attrs and nr_counters if no event is selected and !null_run */
680 if (!null_run && !nr_counters) {
681 size_t c;
683 nr_counters = ARRAY_SIZE(default_attrs);
685 for (c = 0; c < ARRAY_SIZE(default_attrs); ++c) {
686 pos = perf_evsel__new(default_attrs[c].type,
687 default_attrs[c].config,
688 nr_counters);
689 if (pos == NULL)
690 goto out;
691 list_add(&pos->node, &evsel_list);
695 if (target_pid != -1)
696 target_tid = target_pid;
698 threads = thread_map__new(target_pid, target_tid);
699 if (threads == NULL) {
700 pr_err("Problems finding threads of monitor\n");
701 usage_with_options(stat_usage, options);
704 if (system_wide)
705 cpus = cpu_map__new(cpu_list);
706 else
707 cpus = cpu_map__dummy_new();
709 if (cpus == NULL) {
710 perror("failed to parse CPUs map");
711 usage_with_options(stat_usage, options);
712 return -1;
715 list_for_each_entry(pos, &evsel_list, node) {
716 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
717 perf_evsel__alloc_counts(pos, cpus->nr) < 0 ||
718 perf_evsel__alloc_fd(pos, cpus->nr, threads->nr) < 0)
719 goto out_free_fd;
723 * We dont want to block the signals - that would cause
724 * child tasks to inherit that and Ctrl-C would not work.
725 * What we want is for Ctrl-C to work in the exec()-ed
726 * task, but being ignored by perf stat itself:
728 atexit(sig_atexit);
729 signal(SIGINT, skip_signal);
730 signal(SIGALRM, skip_signal);
731 signal(SIGABRT, skip_signal);
733 status = 0;
734 for (run_idx = 0; run_idx < run_count; run_idx++) {
735 if (run_count != 1 && verbose)
736 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
737 status = run_perf_stat(argc, argv);
740 if (status != -1)
741 print_stat(argc, argv);
742 out_free_fd:
743 list_for_each_entry(pos, &evsel_list, node)
744 perf_evsel__free_stat_priv(pos);
745 out:
746 thread_map__delete(threads);
747 threads = NULL;
748 return status;