[net-next-2.6.git] / tools / perf / builtin-stat.c

/*
 * builtin-stat.c
 *
 * Builtin stat command: Give a precise performance counters summary
 * overview about any workload, CPU or specific PID.
 *
 * Sample output:

   $ perf stat ~/hackbench 10
   Time: 0.104

    Performance counter stats for '/home/mingo/hackbench':

       1255.538611  task clock ticks     #      10.143 CPU utilization factor
             54011  context switches     #       0.043 M/sec
               385  CPU migrations       #       0.000 M/sec
             17755  pagefaults           #       0.014 M/sec
        3808323185  CPU cycles           #    3033.219 M/sec
        1575111190  instructions         #    1254.530 M/sec
          17367895  cache references     #      13.833 M/sec
           7674421  cache misses         #       6.112 M/sec

    Wall-clock time elapsed:   123.786620 msecs

 *
 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
 *
 * Improvements and fixes by:
 *
 *   Arjan van de Ven <arjan@linux.intel.com>
 *   Yanmin Zhang <yanmin.zhang@intel.com>
 *   Wu Fengguang <fengguang.wu@intel.com>
 *   Mike Galbraith <efault@gmx.de>
 *   Paul Mackerras <paulus@samba.org>
 *   Jaswinder Singh Rajput <jaswinder@kernel.org>
 *
 * Released under the GPL v2. (and only v2, not any later version)
 */

#include "perf.h"
#include "builtin.h"
#include "util/util.h"
#include "util/parse-options.h"
#include "util/parse-events.h"
#include "util/event.h"
#include "util/debug.h"
#include "util/header.h"
#include "util/cpumap.h"
#include "util/thread.h"

#include <sys/prctl.h>
#include <math.h>
#include <locale.h>

static struct perf_event_attr default_attrs[] = {

  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},

  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES	},
  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES		},

};

static bool			system_wide			=  false;
static unsigned int		nr_cpus				=  0;
static int			run_idx				=  0;

static int			run_count			=  1;
static bool			no_inherit			= false;
static bool			scale				=  true;
static pid_t			target_pid			= -1;
static pid_t			target_tid			= -1;
static pid_t			*all_tids			=  NULL;
static int			thread_num			=  0;
static pid_t			child_pid			= -1;
static bool			null_run			=  false;
static bool			big_num				=  false;


static int			*fd[MAX_NR_CPUS][MAX_COUNTERS];

static int			event_scaled[MAX_COUNTERS];

static volatile int done = 0;

struct stats
{
	double n, mean, M2;
};

static void update_stats(struct stats *stats, u64 val)
{
	double delta;

	stats->n++;
	delta = val - stats->mean;
	stats->mean += delta / stats->n;
	stats->M2 += delta*(val - stats->mean);
}

static double avg_stats(struct stats *stats)
{
	return stats->mean;
}

/*
 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
 *
 *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
 * s^2 = -------------------------------
 *                  n - 1
 *
 * http://en.wikipedia.org/wiki/Stddev
 *
 * The std dev of the mean is related to the std dev by:
 *
 *             s
 * s_mean = -------
 *          sqrt(n)
 *
 */
static double stddev_stats(struct stats *stats)
{
	double variance = stats->M2 / (stats->n - 1);
	double variance_mean = variance / stats->n;

	return sqrt(variance_mean);
}

struct stats			event_res_stats[MAX_COUNTERS][3];
struct stats			runtime_nsecs_stats;
struct stats			walltime_nsecs_stats;
struct stats			runtime_cycles_stats;
struct stats			runtime_branches_stats;

#define MATCH_EVENT(t, c, counter)			\
	(attrs[counter].type == PERF_TYPE_##t &&	\
	 attrs[counter].config == PERF_COUNT_##c)

#define ERR_PERF_OPEN \
"Error: counter %d, sys_perf_event_open() syscall returned with %d (%s)\n"

static int create_perf_stat_counter(int counter)
{
	struct perf_event_attr *attr = attrs + counter;
	int thread;
	int ncreated = 0;

	if (scale)
		attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
				    PERF_FORMAT_TOTAL_TIME_RUNNING;

	if (system_wide) {
		unsigned int cpu;

		for (cpu = 0; cpu < nr_cpus; cpu++) {
			fd[cpu][counter][0] = sys_perf_event_open(attr,
					-1, cpumap[cpu], -1, 0);
			if (fd[cpu][counter][0] < 0)
				pr_debug(ERR_PERF_OPEN, counter,
					 fd[cpu][counter][0], strerror(errno));
			else
				++ncreated;
		}
	} else {
		attr->inherit = !no_inherit;
		if (target_pid == -1 && target_tid == -1) {
			attr->disabled = 1;
			attr->enable_on_exec = 1;
		}
		for (thread = 0; thread < thread_num; thread++) {
			fd[0][counter][thread] = sys_perf_event_open(attr,
				all_tids[thread], -1, -1, 0);
			if (fd[0][counter][thread] < 0)
				pr_debug(ERR_PERF_OPEN, counter,
					 fd[0][counter][thread],
					 strerror(errno));
			else
				++ncreated;
		}
	}

	return ncreated;
}

/*
 * Does the counter have nsecs as a unit?
 */
static inline int nsec_counter(int counter)
{
	if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) ||
	    MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
		return 1;

	return 0;
}

/*
 * Read out the results of a single counter:
 */
static void read_counter(int counter)
{
	u64 count[3], single_count[3];
	unsigned int cpu;
	size_t res, nv;
	int scaled;
	int i, thread;

	count[0] = count[1] = count[2] = 0;

	nv = scale ? 3 : 1;
	for (cpu = 0; cpu < nr_cpus; cpu++) {
		for (thread = 0; thread < thread_num; thread++) {
			if (fd[cpu][counter][thread] < 0)
				continue;

			res = read(fd[cpu][counter][thread],
					single_count, nv * sizeof(u64));
			assert(res == nv * sizeof(u64));

			close(fd[cpu][counter][thread]);
			fd[cpu][counter][thread] = -1;

			count[0] += single_count[0];
			if (scale) {
				count[1] += single_count[1];
				count[2] += single_count[2];
			}
		}
	}

	scaled = 0;
	if (scale) {
		if (count[2] == 0) {
			event_scaled[counter] = -1;
			count[0] = 0;
			return;
		}

		if (count[2] < count[1]) {
			event_scaled[counter] = 1;
			count[0] = (unsigned long long)
				((double)count[0] * count[1] / count[2] + 0.5);
		}
	}

	for (i = 0; i < 3; i++)
		update_stats(&event_res_stats[counter][i], count[i]);

	if (verbose) {
		fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
				count[0], count[1], count[2]);
	}

	/*
	 * Save the full runtime - to allow normalization during printout:
	 */
	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
		update_stats(&runtime_nsecs_stats, count[0]);
	if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
		update_stats(&runtime_cycles_stats, count[0]);
	if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter))
		update_stats(&runtime_branches_stats, count[0]);
}

static int run_perf_stat(int argc __used, const char **argv)
{
	unsigned long long t0, t1;
	int status = 0;
	int counter, ncreated = 0;
	int child_ready_pipe[2], go_pipe[2];
	const bool forks = (argc > 0);
	char buf;

	if (!system_wide)
		nr_cpus = 1;

	if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
		perror("failed to create pipes");
		exit(1);
	}

	if (forks) {
		if ((child_pid = fork()) < 0)
			perror("failed to fork");

		if (!child_pid) {
			close(child_ready_pipe[0]);
			close(go_pipe[1]);
			fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);

			/*
			 * Do a dummy execvp to get the PLT entry resolved,
			 * so we avoid the resolver overhead on the real
			 * execvp call.
			 */
			execvp("", (char **)argv);

			/*
			 * Tell the parent we're ready to go
			 */
			close(child_ready_pipe[1]);

			/*
			 * Wait until the parent tells us to go.
			 */
			if (read(go_pipe[0], &buf, 1) == -1)
				perror("unable to read pipe");

			execvp(argv[0], (char **)argv);

			perror(argv[0]);
			exit(-1);
		}

		if (target_tid == -1 && target_pid == -1 && !system_wide)
			all_tids[0] = child_pid;

		/*
		 * Wait for the child to be ready to exec.
		 */
		close(child_ready_pipe[1]);
		close(go_pipe[0]);
		if (read(child_ready_pipe[0], &buf, 1) == -1)
			perror("unable to read pipe");
		close(child_ready_pipe[0]);
	}

	for (counter = 0; counter < nr_counters; counter++)
		ncreated += create_perf_stat_counter(counter);

	if (ncreated == 0) {
		pr_err("No permission to collect %sstats.\n"
		       "Consider tweaking /proc/sys/kernel/perf_event_paranoid.\n",
		       system_wide ? "system-wide " : "");
		if (child_pid != -1)
			kill(child_pid, SIGTERM);
		return -1;
	}

	/*
	 * Enable counters and exec the command:
	 */
	t0 = rdclock();

	if (forks) {
		close(go_pipe[1]);
		wait(&status);
	} else {
		while(!done) sleep(1);
	}

	t1 = rdclock();

	update_stats(&walltime_nsecs_stats, t1 - t0);

	for (counter = 0; counter < nr_counters; counter++)
		read_counter(counter);

	return WEXITSTATUS(status);
}

static void print_noise(int counter, double avg)
{
	if (run_count == 1)
		return;

	fprintf(stderr, "   ( +- %7.3f%% )",
			100 * stddev_stats(&event_res_stats[counter][0]) / avg);
}

static void nsec_printout(int counter, double avg)
{
	double msecs = avg / 1e6;

	fprintf(stderr, " %18.6f  %-24s", msecs, event_name(counter));

	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
		fprintf(stderr, " # %10.3f CPUs ",
				avg / avg_stats(&walltime_nsecs_stats));
	}
}

static void abs_printout(int counter, double avg)
{
	double total, ratio = 0.0;

	if (big_num)
		fprintf(stderr, " %'18.0f  %-24s", avg, event_name(counter));
	else
		fprintf(stderr, " %18.0f  %-24s", avg, event_name(counter));

	if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
		total = avg_stats(&runtime_cycles_stats);

		if (total)
			ratio = avg / total;

		fprintf(stderr, " # %10.3f IPC  ", ratio);
	} else if (MATCH_EVENT(HARDWARE, HW_BRANCH_MISSES, counter) &&
			runtime_branches_stats.n != 0) {
		total = avg_stats(&runtime_branches_stats);

		if (total)
			ratio = avg * 100 / total;

		fprintf(stderr, " # %10.3f %%    ", ratio);

	} else if (runtime_nsecs_stats.n != 0) {
		total = avg_stats(&runtime_nsecs_stats);

		if (total)
			ratio = 1000.0 * avg / total;

		fprintf(stderr, " # %10.3f M/sec", ratio);
	}
}

/*
 * Print out the results of a single counter:
 */
static void print_counter(int counter)
{
	double avg = avg_stats(&event_res_stats[counter][0]);
	int scaled = event_scaled[counter];

	if (scaled == -1) {
		fprintf(stderr, " %18s  %-24s\n",
			"<not counted>", event_name(counter));
		return;
	}

	if (nsec_counter(counter))
		nsec_printout(counter, avg);
	else
		abs_printout(counter, avg);

	print_noise(counter, avg);

	if (scaled) {
		double avg_enabled, avg_running;

		avg_enabled = avg_stats(&event_res_stats[counter][1]);
		avg_running = avg_stats(&event_res_stats[counter][2]);

		fprintf(stderr, "  (scaled from %.2f%%)",
				100 * avg_running / avg_enabled);
	}

	fprintf(stderr, "\n");
}

static void print_stat(int argc, const char **argv)
{
	int i, counter;

	fflush(stdout);

	fprintf(stderr, "\n");
	fprintf(stderr, " Performance counter stats for ");
	if(target_pid == -1 && target_tid == -1) {
		fprintf(stderr, "\'%s", argv[0]);
		for (i = 1; i < argc; i++)
			fprintf(stderr, " %s", argv[i]);
	} else if (target_pid != -1)
		fprintf(stderr, "process id \'%d", target_pid);
	else
		fprintf(stderr, "thread id \'%d", target_tid);

	fprintf(stderr, "\'");
	if (run_count > 1)
		fprintf(stderr, " (%d runs)", run_count);
	fprintf(stderr, ":\n\n");

	for (counter = 0; counter < nr_counters; counter++)
		print_counter(counter);

	fprintf(stderr, "\n");
	fprintf(stderr, " %18.9f  seconds time elapsed",
			avg_stats(&walltime_nsecs_stats)/1e9);
	if (run_count > 1) {
		fprintf(stderr, "   ( +- %7.3f%% )",
				100*stddev_stats(&walltime_nsecs_stats) /
				avg_stats(&walltime_nsecs_stats));
	}
	fprintf(stderr, "\n\n");
}

static volatile int signr = -1;

static void skip_signal(int signo)
{
	if(child_pid == -1)
		done = 1;

	signr = signo;
}

static void sig_atexit(void)
{
	if (child_pid != -1)
		kill(child_pid, SIGTERM);

	if (signr == -1)
		return;

	signal(signr, SIG_DFL);
	kill(getpid(), signr);
}

static const char * const stat_usage[] = {
	"perf stat [<options>] [<command>]",
	NULL
};

static const struct option options[] = {
	OPT_CALLBACK('e', "event", NULL, "event",
		     "event selector. use 'perf list' to list available events",
		     parse_events),
	OPT_BOOLEAN('i', "no-inherit", &no_inherit,
		    "child tasks do not inherit counters"),
	OPT_INTEGER('p', "pid", &target_pid,
		    "stat events on existing process id"),
	OPT_INTEGER('t', "tid", &target_tid,
		    "stat events on existing thread id"),
	OPT_BOOLEAN('a', "all-cpus", &system_wide,
		    "system-wide collection from all CPUs"),
	OPT_BOOLEAN('c', "scale", &scale,
		    "scale/normalize counters"),
	OPT_INCR('v', "verbose", &verbose,
		    "be more verbose (show counter open errors, etc)"),
	OPT_INTEGER('r', "repeat", &run_count,
		    "repeat command and print average + stddev (max: 100)"),
	OPT_BOOLEAN('n', "null", &null_run,
		    "null run - dont start any counters"),
	OPT_BOOLEAN('B', "big-num", &big_num,
		    "print large numbers with thousands\' separators"),
	OPT_END()
};

int cmd_stat(int argc, const char **argv, const char *prefix __used)
{
	int status;
	int i,j;

	setlocale(LC_ALL, "");

	argc = parse_options(argc, argv, options, stat_usage,
		PARSE_OPT_STOP_AT_NON_OPTION);
	if (!argc && target_pid == -1 && target_tid == -1)
		usage_with_options(stat_usage, options);
	if (run_count <= 0)
		usage_with_options(stat_usage, options);

	/* Set attrs and nr_counters if no event is selected and !null_run */
	if (!null_run && !nr_counters) {
		memcpy(attrs, default_attrs, sizeof(default_attrs));
		nr_counters = ARRAY_SIZE(default_attrs);
	}

	if (system_wide)
		nr_cpus = read_cpu_map();
	else
		nr_cpus = 1;

	if (target_pid != -1) {
		target_tid = target_pid;
		thread_num = find_all_tid(target_pid, &all_tids);
		if (thread_num <= 0) {
			fprintf(stderr, "Can't find all threads of pid %d\n",
					target_pid);
			usage_with_options(stat_usage, options);
		}
	} else {
		all_tids=malloc(sizeof(pid_t));
		if (!all_tids)
			return -ENOMEM;

		all_tids[0] = target_tid;
		thread_num = 1;
	}

	for (i = 0; i < MAX_NR_CPUS; i++) {
		for (j = 0; j < MAX_COUNTERS; j++) {
			fd[i][j] = malloc(sizeof(int)*thread_num);
			if (!fd[i][j])
				return -ENOMEM;
		}
	}

	/*
	 * We dont want to block the signals - that would cause
	 * child tasks to inherit that and Ctrl-C would not work.
	 * What we want is for Ctrl-C to work in the exec()-ed
	 * task, but being ignored by perf stat itself:
	 */
	atexit(sig_atexit);
	signal(SIGINT,  skip_signal);
	signal(SIGALRM, skip_signal);
	signal(SIGABRT, skip_signal);

	status = 0;
	for (run_idx = 0; run_idx < run_count; run_idx++) {
		if (run_count != 1 && verbose)
			fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
		status = run_perf_stat(argc, argv);
	}

	if (status != -1)
		print_stat(argc, argv);

	return status;
}
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
bf9e1876 IM	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>
6e750a8f	35	* Jaswinder Singh Rajput <jaswinder@kernel.org>
5242519b IM	36	*
5242519b IM	37	* Released under the GPL v2. (and only v2, not any later version)
ddcacfa0 IM	38	*/
ddcacfa0 IM	39
1a482f38	40	#include "perf.h"
16f762a2	41	#include "builtin.h"
148be2c1	42	#include "util/util.h"
5242519b IM	43	#include "util/parse-options.h"
5242519b IM	44	#include "util/parse-events.h"
8f28827a FW	45	#include "util/event.h"
8f28827a FW	46	#include "util/debug.h"
60666c63	47	#include "util/header.h"
a12b51c4	48	#include "util/cpumap.h"
d6d901c2	49	#include "util/thread.h"
ddcacfa0	50
ddcacfa0	51	#include <sys/prctl.h>
42202dd5	52	#include <math.h>
5af52b51	53	#include <locale.h>
16c8a109	54
cdd6c482	55	static struct perf_event_attr default_attrs[] = {
ddcacfa0	56
56aab464 IM	57	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
	58	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
	59	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
	60	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
f4dbfa8f	61
56aab464 IM	62	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
56aab464 IM	63	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
56aab464 IM	64	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
56aab464 IM	65	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
dd86e72a IM	66	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES },
dd86e72a IM	67	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES },
f4dbfa8f	68
ddcacfa0	69	};
5242519b	70
c0555642	71	static bool system_wide = false;
f37a291c	72	static unsigned int nr_cpus = 0;
3d632595	73	static int run_idx = 0;
ddcacfa0	74
3d632595	75	static int run_count = 1;
2e6cdf99	76	static bool no_inherit = false;
c0555642	77	static bool scale = true;
933da83a	78	static pid_t target_pid = -1;
d6d901c2 ZY	79	static pid_t target_tid = -1;
	80	static pid_t *all_tids = NULL;
	81	static int thread_num = 0;
933da83a	82	static pid_t child_pid = -1;
c0555642	83	static bool null_run = false;
5af52b51 SE	84	static bool big_num = false;
5af52b51 SE	85
ddcacfa0	86
d6d901c2	87	static int *fd[MAX_NR_CPUS][MAX_COUNTERS];
42202dd5	88
63d40deb	89	static int event_scaled[MAX_COUNTERS];
3d632595	90
60666c63 LW	91	static volatile int done = 0;
60666c63 LW	92
506d4bc8 PZ	93	struct stats
506d4bc8 PZ	94	{
8a02631a	95	double n, mean, M2;
506d4bc8	96	};
42202dd5	97
9e9772c4 PZ	98	static void update_stats(struct stats *stats, u64 val)
9e9772c4 PZ	99	{
8a02631a	100	double delta;
9e9772c4	101
8a02631a PZ	102	stats->n++;
	103	delta = val - stats->mean;
	104	stats->mean += delta / stats->n;
	105	stats->M2 += delta*(val - stats->mean);
9e9772c4 PZ	106	}
9e9772c4 PZ	107
506d4bc8 PZ	108	static double avg_stats(struct stats *stats)
506d4bc8 PZ	109	{
8a02631a	110	return stats->mean;
506d4bc8	111	}
42202dd5	112
506d4bc8	113	/*
63d40deb PZ	114	* http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
63d40deb PZ	115	*
8a02631a PZ	116	* (\Sum n_i^2) - ((\Sum n_i)^2)/n
	117	* s^2 = -------------------------------
	118	* n - 1
63d40deb PZ	119	*
	120	* http://en.wikipedia.org/wiki/Stddev
	121	*
	122	* The std dev of the mean is related to the std dev by:
	123	*
	124	* s
	125	* s_mean = -------
	126	* sqrt(n)
	127	*
506d4bc8 PZ	128	*/
	129	static double stddev_stats(struct stats *stats)
	130	{
8a02631a PZ	131	double variance = stats->M2 / (stats->n - 1);
8a02631a PZ	132	double variance_mean = variance / stats->n;
42202dd5	133
63d40deb	134	return sqrt(variance_mean);
506d4bc8	135	}
42202dd5	136
506d4bc8	137	struct stats event_res_stats[MAX_COUNTERS][3];
506d4bc8 PZ	138	struct stats runtime_nsecs_stats;
	139	struct stats walltime_nsecs_stats;
	140	struct stats runtime_cycles_stats;
11018201	141	struct stats runtime_branches_stats;
be1ac0d8	142
b9ebdcc0 JSR	143	#define MATCH_EVENT(t, c, counter) \
	144	(attrs[counter].type == PERF_TYPE_##t && \
	145	attrs[counter].config == PERF_COUNT_##c)
	146
cca03c0a	147	#define ERR_PERF_OPEN \
cdd6c482	148	"Error: counter %d, sys_perf_event_open() syscall returned with %d (%s)\n"
cca03c0a	149
084ab9f8	150	static int create_perf_stat_counter(int counter)
ddcacfa0	151	{
cdd6c482	152	struct perf_event_attr *attr = attrs + counter;
d6d901c2	153	int thread;
084ab9f8	154	int ncreated = 0;
16c8a109	155
ddcacfa0	156	if (scale)
a21ca2ca IM	157	attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED \|
a21ca2ca IM	158	PERF_FORMAT_TOTAL_TIME_RUNNING;
ddcacfa0 IM	159
ddcacfa0 IM	160	if (system_wide) {
f37a291c IM	161	unsigned int cpu;
f37a291c IM	162
cca03c0a	163	for (cpu = 0; cpu < nr_cpus; cpu++) {
d6d901c2 ZY	164	fd[cpu][counter][0] = sys_perf_event_open(attr,
d6d901c2 ZY	165	-1, cpumap[cpu], -1, 0);
084ab9f8 ACM	166	if (fd[cpu][counter][0] < 0)
	167	pr_debug(ERR_PERF_OPEN, counter,
	168	fd[cpu][counter][0], strerror(errno));
	169	else
	170	++ncreated;
ddcacfa0 IM	171	}
ddcacfa0 IM	172	} else {
2e6cdf99 SE	173	attr->inherit = !no_inherit;
2e6cdf99 SE	174	if (target_pid == -1 && target_tid == -1) {
6be2850e ZY	175	attr->disabled = 1;
	176	attr->enable_on_exec = 1;
	177	}
d6d901c2 ZY	178	for (thread = 0; thread < thread_num; thread++) {
	179	fd[0][counter][thread] = sys_perf_event_open(attr,
	180	all_tids[thread], -1, -1, 0);
084ab9f8 ACM	181	if (fd[0][counter][thread] < 0)
	182	pr_debug(ERR_PERF_OPEN, counter,
	183	fd[0][counter][thread],
	184	strerror(errno));
	185	else
	186	++ncreated;
d6d901c2	187	}
ddcacfa0	188	}
084ab9f8 ACM	189
084ab9f8 ACM	190	return ncreated;
ddcacfa0 IM	191	}
ddcacfa0 IM	192
c04f5e5d IM	193	/*
	194	* Does the counter have nsecs as a unit?
	195	*/
	196	static inline int nsec_counter(int counter)
	197	{
b9ebdcc0 JSR	198	if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) \|\|
b9ebdcc0 JSR	199	MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
c04f5e5d IM	200	return 1;
	201
	202	return 0;
	203	}
	204
	205	/*
2996f5dd	206	* Read out the results of a single counter:
c04f5e5d	207	*/
2996f5dd	208	static void read_counter(int counter)
c04f5e5d	209	{
849abde9	210	u64 count[3], single_count[3];
f37a291c IM	211	unsigned int cpu;
f37a291c IM	212	size_t res, nv;
c04f5e5d	213	int scaled;
d6d901c2	214	int i, thread;
c04f5e5d IM	215
c04f5e5d IM	216	count[0] = count[1] = count[2] = 0;
2996f5dd	217
c04f5e5d	218	nv = scale ? 3 : 1;
cca03c0a	219	for (cpu = 0; cpu < nr_cpus; cpu++) {
d6d901c2 ZY	220	for (thread = 0; thread < thread_num; thread++) {
	221	if (fd[cpu][counter][thread] < 0)
	222	continue;
	223
	224	res = read(fd[cpu][counter][thread],
	225	single_count, nv * sizeof(u64));
	226	assert(res == nv * sizeof(u64));
	227
	228	close(fd[cpu][counter][thread]);
	229	fd[cpu][counter][thread] = -1;
	230
	231	count[0] += single_count[0];
	232	if (scale) {
	233	count[1] += single_count[1];
	234	count[2] += single_count[2];
	235	}
c04f5e5d IM	236	}
	237	}
	238
	239	scaled = 0;
	240	if (scale) {
	241	if (count[2] == 0) {
9e9772c4	242	event_scaled[counter] = -1;
2996f5dd	243	count[0] = 0;
c04f5e5d IM	244	return;
c04f5e5d IM	245	}
2996f5dd	246
c04f5e5d	247	if (count[2] < count[1]) {
9e9772c4	248	event_scaled[counter] = 1;
c04f5e5d IM	249	count[0] = (unsigned long long)
	250	((double)count[0] * count[1] / count[2] + 0.5);
	251	}
	252	}
9e9772c4 PZ	253
	254	for (i = 0; i < 3; i++)
	255	update_stats(&event_res_stats[counter][i], count[i]);
	256
	257	if (verbose) {
	258	fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
	259	count[0], count[1], count[2]);
	260	}
	261
be1ac0d8 IM	262	/*
	263	* Save the full runtime - to allow normalization during printout:
	264	*/
b9ebdcc0	265	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
9e9772c4	266	update_stats(&runtime_nsecs_stats, count[0]);
b9ebdcc0	267	if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
9e9772c4	268	update_stats(&runtime_cycles_stats, count[0]);
11018201 AB	269	if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter))
11018201 AB	270	update_stats(&runtime_branches_stats, count[0]);
2996f5dd IM	271	}
2996f5dd IM	272
f37a291c	273	static int run_perf_stat(int argc __used, const char **argv)
42202dd5 IM	274	{
	275	unsigned long long t0, t1;
	276	int status = 0;
084ab9f8	277	int counter, ncreated = 0;
051ae7f7	278	int child_ready_pipe[2], go_pipe[2];
6be2850e	279	const bool forks = (argc > 0);
051ae7f7	280	char buf;
42202dd5 IM	281
	282	if (!system_wide)
	283	nr_cpus = 1;
	284
60666c63	285	if (forks && (pipe(child_ready_pipe) < 0 \|\| pipe(go_pipe) < 0)) {
051ae7f7 PM	286	perror("failed to create pipes");
	287	exit(1);
	288	}
	289
60666c63	290	if (forks) {
6be2850e	291	if ((child_pid = fork()) < 0)
60666c63 LW	292	perror("failed to fork");
60666c63 LW	293
6be2850e	294	if (!child_pid) {
60666c63 LW	295	close(child_ready_pipe[0]);
	296	close(go_pipe[1]);
	297	fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
	298
	299	/*
	300	* Do a dummy execvp to get the PLT entry resolved,
	301	* so we avoid the resolver overhead on the real
	302	* execvp call.
	303	*/
	304	execvp("", (char **)argv);
	305
	306	/*
	307	* Tell the parent we're ready to go
	308	*/
	309	close(child_ready_pipe[1]);
	310
	311	/*
	312	* Wait until the parent tells us to go.
	313	*/
	314	if (read(go_pipe[0], &buf, 1) == -1)
	315	perror("unable to read pipe");
	316
	317	execvp(argv[0], (char **)argv);
	318
	319	perror(argv[0]);
	320	exit(-1);
	321	}
051ae7f7	322
d6d901c2 ZY	323	if (target_tid == -1 && target_pid == -1 && !system_wide)
	324	all_tids[0] = child_pid;
	325
051ae7f7	326	/*
60666c63	327	* Wait for the child to be ready to exec.
051ae7f7 PM	328	*/
051ae7f7 PM	329	close(child_ready_pipe[1]);
60666c63 LW	330	close(go_pipe[0]);
60666c63 LW	331	if (read(child_ready_pipe[0], &buf, 1) == -1)
a92bef0f	332	perror("unable to read pipe");
60666c63	333	close(child_ready_pipe[0]);
051ae7f7 PM	334	}
051ae7f7 PM	335
42202dd5	336	for (counter = 0; counter < nr_counters; counter++)
084ab9f8 ACM	337	ncreated += create_perf_stat_counter(counter);
	338
	339	if (ncreated == 0) {
	340	pr_err("No permission to collect %sstats.\n"
	341	"Consider tweaking /proc/sys/kernel/perf_event_paranoid.\n",
	342	system_wide ? "system-wide " : "");
	343	if (child_pid != -1)
	344	kill(child_pid, SIGTERM);
	345	return -1;
	346	}
42202dd5 IM	347
	348	/*
	349	* Enable counters and exec the command:
	350	*/
	351	t0 = rdclock();
42202dd5	352
60666c63 LW	353	if (forks) {
	354	close(go_pipe[1]);
	355	wait(&status);
	356	} else {
6be2850e	357	while(!done) sleep(1);
60666c63	358	}
42202dd5	359
42202dd5 IM	360	t1 = rdclock();
42202dd5 IM	361
9e9772c4	362	update_stats(&walltime_nsecs_stats, t1 - t0);
42202dd5 IM	363
	364	for (counter = 0; counter < nr_counters; counter++)
	365	read_counter(counter);
	366
	367	return WEXITSTATUS(status);
	368	}
	369
849abde9	370	static void print_noise(int counter, double avg)
42202dd5	371	{
849abde9 PZ	372	if (run_count == 1)
	373	return;
	374
	375	fprintf(stderr, " ( +- %7.3f%% )",
	376	100 * stddev_stats(&event_res_stats[counter][0]) / avg);
42202dd5 IM	377	}
42202dd5 IM	378
849abde9	379	static void nsec_printout(int counter, double avg)
44175b6f	380	{
506d4bc8	381	double msecs = avg / 1e6;
44175b6f	382
5af52b51	383	fprintf(stderr, " %18.6f %-24s", msecs, event_name(counter));
44175b6f	384
b9ebdcc0	385	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
506d4bc8 PZ	386	fprintf(stderr, " # %10.3f CPUs ",
506d4bc8 PZ	387	avg / avg_stats(&walltime_nsecs_stats));
44175b6f IM	388	}
	389	}
	390
849abde9	391	static void abs_printout(int counter, double avg)
44175b6f	392	{
c7f7fea3 IM	393	double total, ratio = 0.0;
c7f7fea3 IM	394
5af52b51 SE	395	if (big_num)
	396	fprintf(stderr, " %'18.0f %-24s", avg, event_name(counter));
	397	else
	398	fprintf(stderr, " %18.0f %-24s", avg, event_name(counter));
44175b6f	399
506d4bc8	400	if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
c7f7fea3 IM	401	total = avg_stats(&runtime_cycles_stats);
	402
	403	if (total)
	404	ratio = avg / total;
	405
	406	fprintf(stderr, " # %10.3f IPC ", ratio);
7255fe2a LDM	407	} else if (MATCH_EVENT(HARDWARE, HW_BRANCH_MISSES, counter) &&
7255fe2a LDM	408	runtime_branches_stats.n != 0) {
11018201 AB	409	total = avg_stats(&runtime_branches_stats);
	410
	411	if (total)
	412	ratio = avg * 100 / total;
	413
dd86e72a	414	fprintf(stderr, " # %10.3f %% ", ratio);
11018201	415
7255fe2a	416	} else if (runtime_nsecs_stats.n != 0) {
c7f7fea3 IM	417	total = avg_stats(&runtime_nsecs_stats);
	418
	419	if (total)
	420	ratio = 1000.0 * avg / total;
	421
	422	fprintf(stderr, " # %10.3f M/sec", ratio);
44175b6f	423	}
44175b6f IM	424	}
44175b6f IM	425
2996f5dd IM	426	/*
	427	* Print out the results of a single counter:
	428	*/
	429	static void print_counter(int counter)
	430	{
849abde9	431	double avg = avg_stats(&event_res_stats[counter][0]);
63d40deb	432	int scaled = event_scaled[counter];
2996f5dd	433
2996f5dd	434	if (scaled == -1) {
5af52b51	435	fprintf(stderr, " %18s %-24s\n",
2996f5dd IM	436	"<not counted>", event_name(counter));
	437	return;
	438	}
c04f5e5d	439
44175b6f	440	if (nsec_counter(counter))
849abde9	441	nsec_printout(counter, avg);
44175b6f	442	else
849abde9 PZ	443	abs_printout(counter, avg);
	444
	445	print_noise(counter, avg);
506d4bc8 PZ	446
	447	if (scaled) {
	448	double avg_enabled, avg_running;
	449
	450	avg_enabled = avg_stats(&event_res_stats[counter][1]);
	451	avg_running = avg_stats(&event_res_stats[counter][2]);
d7c29318	452
210ad39f	453	fprintf(stderr, " (scaled from %.2f%%)",
506d4bc8 PZ	454	100 * avg_running / avg_enabled);
506d4bc8 PZ	455	}
44175b6f	456
c04f5e5d IM	457	fprintf(stderr, "\n");
	458	}
	459
42202dd5 IM	460	static void print_stat(int argc, const char **argv)
	461	{
	462	int i, counter;
	463
ddcacfa0 IM	464	fflush(stdout);
	465
	466	fprintf(stderr, "\n");
60666c63	467	fprintf(stderr, " Performance counter stats for ");
d6d901c2	468	if(target_pid == -1 && target_tid == -1) {
60666c63 LW	469	fprintf(stderr, "\'%s", argv[0]);
	470	for (i = 1; i < argc; i++)
	471	fprintf(stderr, " %s", argv[i]);
d6d901c2 ZY	472	} else if (target_pid != -1)
	473	fprintf(stderr, "process id \'%d", target_pid);
	474	else
	475	fprintf(stderr, "thread id \'%d", target_tid);
44db76c8	476
42202dd5 IM	477	fprintf(stderr, "\'");
	478	if (run_count > 1)
	479	fprintf(stderr, " (%d runs)", run_count);
	480	fprintf(stderr, ":\n\n");
2996f5dd	481
c04f5e5d IM	482	for (counter = 0; counter < nr_counters; counter++)
c04f5e5d IM	483	print_counter(counter);
ddcacfa0	484
ddcacfa0	485	fprintf(stderr, "\n");
5af52b51	486	fprintf(stderr, " %18.9f seconds time elapsed",
506d4bc8	487	avg_stats(&walltime_nsecs_stats)/1e9);
566747e6 IM	488	if (run_count > 1) {
566747e6 IM	489	fprintf(stderr, " ( +- %7.3f%% )",
506d4bc8 PZ	490	100*stddev_stats(&walltime_nsecs_stats) /
506d4bc8 PZ	491	avg_stats(&walltime_nsecs_stats));
566747e6 IM	492	}
566747e6 IM	493	fprintf(stderr, "\n\n");
ddcacfa0 IM	494	}
ddcacfa0 IM	495
f7b7c26e PZ	496	static volatile int signr = -1;
f7b7c26e PZ	497
5242519b	498	static void skip_signal(int signo)
ddcacfa0	499	{
6be2850e	500	if(child_pid == -1)
60666c63 LW	501	done = 1;
60666c63 LW	502
f7b7c26e PZ	503	signr = signo;
	504	}
	505
	506	static void sig_atexit(void)
	507	{
933da83a CW	508	if (child_pid != -1)
	509	kill(child_pid, SIGTERM);
	510
f7b7c26e PZ	511	if (signr == -1)
	512	return;
	513
	514	signal(signr, SIG_DFL);
	515	kill(getpid(), signr);
5242519b IM	516	}
	517
	518	static const char * const stat_usage[] = {
60666c63	519	"perf stat [<options>] [<command>]",
5242519b IM	520	NULL
	521	};
	522
5242519b IM	523	static const struct option options[] = {
5242519b IM	524	OPT_CALLBACK('e', "event", NULL, "event",
86847b62 TG	525	"event selector. use 'perf list' to list available events",
86847b62 TG	526	parse_events),
2e6cdf99 SE	527	OPT_BOOLEAN('i', "no-inherit", &no_inherit,
2e6cdf99 SE	528	"child tasks do not inherit counters"),
5242519b	529	OPT_INTEGER('p', "pid", &target_pid,
d6d901c2 ZY	530	"stat events on existing process id"),
	531	OPT_INTEGER('t', "tid", &target_tid,
	532	"stat events on existing thread id"),
5242519b	533	OPT_BOOLEAN('a', "all-cpus", &system_wide,
3d632595	534	"system-wide collection from all CPUs"),
b26bc5a7	535	OPT_BOOLEAN('c', "scale", &scale,
3d632595	536	"scale/normalize counters"),
c0555642	537	OPT_INCR('v', "verbose", &verbose,
743ee1f8	538	"be more verbose (show counter open errors, etc)"),
42202dd5 IM	539	OPT_INTEGER('r', "repeat", &run_count,
42202dd5 IM	540	"repeat command and print average + stddev (max: 100)"),
0cfb7a13 IM	541	OPT_BOOLEAN('n', "null", &null_run,
0cfb7a13 IM	542	"null run - dont start any counters"),
5af52b51 SE	543	OPT_BOOLEAN('B', "big-num", &big_num,
5af52b51 SE	544	"print large numbers with thousands\' separators"),
5242519b IM	545	OPT_END()
	546	};
	547
f37a291c	548	int cmd_stat(int argc, const char *argv, const char prefix __used)
5242519b	549	{
42202dd5	550	int status;
d6d901c2	551	int i,j;
42202dd5	552
5af52b51 SE	553	setlocale(LC_ALL, "");
5af52b51 SE	554
a0541234 AB	555	argc = parse_options(argc, argv, options, stat_usage,
a0541234 AB	556	PARSE_OPT_STOP_AT_NON_OPTION);
d6d901c2	557	if (!argc && target_pid == -1 && target_tid == -1)
5242519b	558	usage_with_options(stat_usage, options);
9e9772c4	559	if (run_count <= 0)
42202dd5	560	usage_with_options(stat_usage, options);
ddcacfa0	561
c3043569 JSR	562	/* Set attrs and nr_counters if no event is selected and !null_run */
	563	if (!null_run && !nr_counters) {
	564	memcpy(attrs, default_attrs, sizeof(default_attrs));
	565	nr_counters = ARRAY_SIZE(default_attrs);
	566	}
ddcacfa0	567
a12b51c4 PM	568	if (system_wide)
	569	nr_cpus = read_cpu_map();
	570	else
	571	nr_cpus = 1;
ddcacfa0	572
d6d901c2 ZY	573	if (target_pid != -1) {
	574	target_tid = target_pid;
	575	thread_num = find_all_tid(target_pid, &all_tids);
	576	if (thread_num <= 0) {
	577	fprintf(stderr, "Can't find all threads of pid %d\n",
	578	target_pid);
	579	usage_with_options(stat_usage, options);
	580	}
	581	} else {
	582	all_tids=malloc(sizeof(pid_t));
	583	if (!all_tids)
	584	return -ENOMEM;
	585
	586	all_tids[0] = target_tid;
	587	thread_num = 1;
	588	}
	589
	590	for (i = 0; i < MAX_NR_CPUS; i++) {
	591	for (j = 0; j < MAX_COUNTERS; j++) {
	592	fd[i][j] = malloc(sizeof(int)*thread_num);
	593	if (!fd[i][j])
	594	return -ENOMEM;
	595	}
	596	}
	597
58d7e993 IM	598	/*
	599	* We dont want to block the signals - that would cause
	600	* child tasks to inherit that and Ctrl-C would not work.
	601	* What we want is for Ctrl-C to work in the exec()-ed
	602	* task, but being ignored by perf stat itself:
	603	*/
f7b7c26e	604	atexit(sig_atexit);
58d7e993 IM	605	signal(SIGINT, skip_signal);
	606	signal(SIGALRM, skip_signal);
	607	signal(SIGABRT, skip_signal);
	608
42202dd5 IM	609	status = 0;
	610	for (run_idx = 0; run_idx < run_count; run_idx++) {
	611	if (run_count != 1 && verbose)
3d632595	612	fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
42202dd5 IM	613	status = run_perf_stat(argc, argv);
	614	}
	615
084ab9f8 ACM	616	if (status != -1)
084ab9f8 ACM	617	print_stat(argc, argv);
42202dd5 IM	618
42202dd5 IM	619	return status;
ddcacfa0	620	}