[linux-2.6-block.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>
 *
 * 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 <sys/prctl.h>
#include <math.h>

static struct perf_counter_attr default_attrs[MAX_COUNTERS] = {

  { .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_CACHE_REFERENCES},
  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES	},

};

static int			system_wide			=  0;
static int			inherit				=  1;
static int			verbose				=  0;

static int			fd[MAX_NR_CPUS][MAX_COUNTERS];

static int			target_pid			= -1;
static int			nr_cpus				=  0;
static unsigned int		page_size;

static int			scale				=  1;

static const unsigned int default_count[] = {
	1000000,
	1000000,
	  10000,
	  10000,
	1000000,
	  10000,
};

#define MAX_RUN 100

static int			run_count		=  1;
static int			run_idx			=  0;

static u64			event_res[MAX_RUN][MAX_COUNTERS][3];
static u64			event_scaled[MAX_RUN][MAX_COUNTERS];

//static u64			event_hist[MAX_RUN][MAX_COUNTERS][3];


static u64			runtime_nsecs[MAX_RUN];
static u64			walltime_nsecs[MAX_RUN];
static u64			runtime_cycles[MAX_RUN];

static u64			event_res_avg[MAX_COUNTERS][3];
static u64			event_res_noise[MAX_COUNTERS][3];

static u64			event_scaled_avg[MAX_COUNTERS];

static u64			runtime_nsecs_avg;
static u64			runtime_nsecs_noise;

static u64			walltime_nsecs_avg;
static u64			walltime_nsecs_noise;

static u64			runtime_cycles_avg;
static u64			runtime_cycles_noise;


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

static void create_perf_stat_counter(int counter)
{
	struct perf_counter_attr *attr = attrs + counter;

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

	if (system_wide) {
		int cpu;
		for (cpu = 0; cpu < nr_cpus; cpu++) {
			fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
			if (fd[cpu][counter] < 0 && verbose)
				fprintf(stderr, ERR_PERF_OPEN, counter,
					fd[cpu][counter], strerror(errno));
		}
	} else {
		attr->inherit	= inherit;
		attr->disabled	= 1;

		fd[0][counter] = sys_perf_counter_open(attr, 0, -1, -1, 0);
		if (fd[0][counter] < 0 && verbose)
			fprintf(stderr, ERR_PERF_OPEN, counter,
				fd[0][counter], strerror(errno));
	}
}

/*
 * Does the counter have nsecs as a unit?
 */
static inline int nsec_counter(int counter)
{
	if (attrs[counter].type != PERF_TYPE_SOFTWARE)
		return 0;

	if (attrs[counter].config == PERF_COUNT_SW_CPU_CLOCK)
		return 1;

	if (attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
		return 1;

	return 0;
}

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

	count = event_res[run_idx][counter];

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

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

		res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
		assert(res == nv * sizeof(u64));
		close(fd[cpu][counter]);
		fd[cpu][counter] = -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[run_idx][counter] = -1;
			count[0] = 0;
			return;
		}

		if (count[2] < count[1]) {
			event_scaled[run_idx][counter] = 1;
			count[0] = (unsigned long long)
				((double)count[0] * count[1] / count[2] + 0.5);
		}
	}
	/*
	 * Save the full runtime - to allow normalization during printout:
	 */
	if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
		attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
		runtime_nsecs[run_idx] = count[0];
	if (attrs[counter].type == PERF_TYPE_HARDWARE &&
		attrs[counter].config == PERF_COUNT_HW_CPU_CYCLES)
		runtime_cycles[run_idx] = count[0];
}

static int run_perf_stat(int argc, const char **argv)
{
	unsigned long long t0, t1;
	int status = 0;
	int counter;
	int pid;

	if (!system_wide)
		nr_cpus = 1;

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

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

	if ((pid = fork()) < 0)
		perror("failed to fork");

	if (!pid) {
		if (execvp(argv[0], (char **)argv)) {
			perror(argv[0]);
			exit(-1);
		}
	}

	wait(&status);

	prctl(PR_TASK_PERF_COUNTERS_DISABLE);
	t1 = rdclock();

	walltime_nsecs[run_idx] = t1 - t0;

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

	return WEXITSTATUS(status);
}

static void print_noise(u64 *count, u64 *noise)
{
	if (run_count > 1)
		fprintf(stderr, "   ( +- %7.3f%% )",
			(double)noise[0]/(count[0]+1)*100.0);
}

static void nsec_printout(int counter, u64 *count, u64 *noise)
{
	double msecs = (double)count[0] / 1000000;

	fprintf(stderr, " %14.6f  %-20s", msecs, event_name(counter));

	if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
		attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) {

		if (walltime_nsecs_avg)
			fprintf(stderr, " # %10.3f CPUs ",
				(double)count[0] / (double)walltime_nsecs_avg);
	}
	print_noise(count, noise);
}

static void abs_printout(int counter, u64 *count, u64 *noise)
{
	fprintf(stderr, " %14Ld  %-20s", count[0], event_name(counter));

	if (runtime_cycles_avg &&
		attrs[counter].type == PERF_TYPE_HARDWARE &&
			attrs[counter].config == PERF_COUNT_HW_INSTRUCTIONS) {

		fprintf(stderr, " # %10.3f IPC  ",
			(double)count[0] / (double)runtime_cycles_avg);
	} else {
		if (runtime_nsecs_avg) {
			fprintf(stderr, " # %10.3f M/sec",
				(double)count[0]/runtime_nsecs_avg*1000.0);
		}
	}
	print_noise(count, noise);
}

/*
 * Print out the results of a single counter:
 */
static void print_counter(int counter)
{
	u64 *count, *noise;
	int scaled;

	count = event_res_avg[counter];
	noise = event_res_noise[counter];
	scaled = event_scaled_avg[counter];

	if (scaled == -1) {
		fprintf(stderr, " %14s  %-20s\n",
			"<not counted>", event_name(counter));
		return;
	}

	if (nsec_counter(counter))
		nsec_printout(counter, count, noise);
	else
		abs_printout(counter, count, noise);

	if (scaled)
		fprintf(stderr, "  (scaled from %.2f%%)",
			(double) count[2] / count[1] * 100);

	fprintf(stderr, "\n");
}

/*
 * normalize_noise noise values down to stddev:
 */
static void normalize_noise(u64 *val)
{
	double res;

	res = (double)*val / (run_count * sqrt((double)run_count));

	*val = (u64)res;
}

static void update_avg(const char *name, int idx, u64 *avg, u64 *val)
{
	*avg += *val;

	if (verbose > 1)
		fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val);
}
/*
 * Calculate the averages and noises:
 */
static void calc_avg(void)
{
	int i, j;

	if (verbose > 1)
		fprintf(stderr, "\n");

	for (i = 0; i < run_count; i++) {
		update_avg("runtime", 0, &runtime_nsecs_avg, runtime_nsecs + i);
		update_avg("walltime", 0, &walltime_nsecs_avg, walltime_nsecs + i);
		update_avg("runtime_cycles", 0, &runtime_cycles_avg, runtime_cycles + i);

		for (j = 0; j < nr_counters; j++) {
			update_avg("counter/0", j,
				event_res_avg[j]+0, event_res[i][j]+0);
			update_avg("counter/1", j,
				event_res_avg[j]+1, event_res[i][j]+1);
			update_avg("counter/2", j,
				event_res_avg[j]+2, event_res[i][j]+2);
			update_avg("scaled", j,
				event_scaled_avg + j, event_scaled[i]+j);
		}
	}
	runtime_nsecs_avg /= run_count;
	walltime_nsecs_avg /= run_count;
	runtime_cycles_avg /= run_count;

	for (j = 0; j < nr_counters; j++) {
		event_res_avg[j][0] /= run_count;
		event_res_avg[j][1] /= run_count;
		event_res_avg[j][2] /= run_count;
	}

	for (i = 0; i < run_count; i++) {
		runtime_nsecs_noise +=
			abs((s64)(runtime_nsecs[i] - runtime_nsecs_avg));
		walltime_nsecs_noise +=
			abs((s64)(walltime_nsecs[i] - walltime_nsecs_avg));
		runtime_cycles_noise +=
			abs((s64)(runtime_cycles[i] - runtime_cycles_avg));

		for (j = 0; j < nr_counters; j++) {
			event_res_noise[j][0] +=
				abs((s64)(event_res[i][j][0] - event_res_avg[j][0]));
			event_res_noise[j][1] +=
				abs((s64)(event_res[i][j][1] - event_res_avg[j][1]));
			event_res_noise[j][2] +=
				abs((s64)(event_res[i][j][2] - event_res_avg[j][2]));
		}
	}

	normalize_noise(&runtime_nsecs_noise);
	normalize_noise(&walltime_nsecs_noise);
	normalize_noise(&runtime_cycles_noise);

	for (j = 0; j < nr_counters; j++) {
		normalize_noise(&event_res_noise[j][0]);
		normalize_noise(&event_res_noise[j][1]);
		normalize_noise(&event_res_noise[j][2]);
	}
}

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

	calc_avg();

	fflush(stdout);

	fprintf(stderr, "\n");
	fprintf(stderr, " Performance counter stats for \'%s", argv[0]);

	for (i = 1; i < argc; i++)
		fprintf(stderr, " %s", argv[i]);

	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, " %14.9f  seconds time elapsed.\n",
			(double)walltime_nsecs_avg/1e9);
	fprintf(stderr, "\n");
}

static volatile int signr = -1;

static void skip_signal(int signo)
{
	signr = signo;
}

static void sig_atexit(void)
{
	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', "inherit", &inherit,
		    "child tasks inherit counters"),
	OPT_INTEGER('p', "pid", &target_pid,
		    "stat events on existing pid"),
	OPT_BOOLEAN('a', "all-cpus", &system_wide,
			    "system-wide collection from all CPUs"),
	OPT_BOOLEAN('S', "scale", &scale,
			    "scale/normalize counters"),
	OPT_BOOLEAN('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_END()
};

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

	page_size = sysconf(_SC_PAGE_SIZE);

	memcpy(attrs, default_attrs, sizeof(attrs));

	argc = parse_options(argc, argv, options, stat_usage, 0);
	if (!argc)
		usage_with_options(stat_usage, options);
	if (run_count <= 0 || run_count > MAX_RUN)
		usage_with_options(stat_usage, options);

	if (!nr_counters)
		nr_counters = 8;

	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
	assert(nr_cpus <= MAX_NR_CPUS);
	assert(nr_cpus >= 0);

	/*
	 * 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);
	}

	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>
	35	*
	36	* Released under the GPL v2. (and only v2, not any later version)
ddcacfa0 IM	37	*/
ddcacfa0 IM	38
1a482f38	39	#include "perf.h"
16f762a2	40	#include "builtin.h"
148be2c1	41	#include "util/util.h"
5242519b IM	42	#include "util/parse-options.h"
5242519b IM	43	#include "util/parse-events.h"
ddcacfa0	44
ddcacfa0	45	#include <sys/prctl.h>
42202dd5	46	#include <math.h>
16c8a109	47
a21ca2ca	48	static struct perf_counter_attr default_attrs[MAX_COUNTERS] = {
ddcacfa0	49
f4dbfa8f PZ	50	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
	51	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
	52	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
	53	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
	54
	55	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
	56	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
	57	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
	58	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES },
	59
ddcacfa0	60	};
5242519b	61
a21ca2ca IM	62	static int system_wide = 0;
a21ca2ca IM	63	static int inherit = 1;
743ee1f8	64	static int verbose = 0;
a21ca2ca	65
ddcacfa0 IM	66	static int fd[MAX_NR_CPUS][MAX_COUNTERS];
ddcacfa0 IM	67
5242519b	68	static int target_pid = -1;
ddcacfa0	69	static int nr_cpus = 0;
ddcacfa0 IM	70	static unsigned int page_size;
ddcacfa0 IM	71
66cf7829	72	static int scale = 1;
ddcacfa0 IM	73
	74	static const unsigned int default_count[] = {
	75	1000000,
	76	1000000,
	77	10000,
	78	10000,
	79	1000000,
	80	10000,
	81	};
	82
42202dd5	83	#define MAX_RUN 100
2996f5dd	84
42202dd5 IM	85	static int run_count = 1;
	86	static int run_idx = 0;
	87
9cffa8d5 PM	88	static u64 event_res[MAX_RUN][MAX_COUNTERS][3];
9cffa8d5 PM	89	static u64 event_scaled[MAX_RUN][MAX_COUNTERS];
42202dd5	90
9cffa8d5	91	//static u64 event_hist[MAX_RUN][MAX_COUNTERS][3];
42202dd5 IM	92
42202dd5 IM	93
9cffa8d5 PM	94	static u64 runtime_nsecs[MAX_RUN];
	95	static u64 walltime_nsecs[MAX_RUN];
	96	static u64 runtime_cycles[MAX_RUN];
42202dd5	97
9cffa8d5 PM	98	static u64 event_res_avg[MAX_COUNTERS][3];
9cffa8d5 PM	99	static u64 event_res_noise[MAX_COUNTERS][3];
42202dd5	100
9cffa8d5	101	static u64 event_scaled_avg[MAX_COUNTERS];
42202dd5	102
9cffa8d5 PM	103	static u64 runtime_nsecs_avg;
9cffa8d5 PM	104	static u64 runtime_nsecs_noise;
42202dd5	105
9cffa8d5 PM	106	static u64 walltime_nsecs_avg;
9cffa8d5 PM	107	static u64 walltime_nsecs_noise;
42202dd5	108
9cffa8d5 PM	109	static u64 runtime_cycles_avg;
9cffa8d5 PM	110	static u64 runtime_cycles_noise;
be1ac0d8	111
cca03c0a JSR	112
	113	#define ERR_PERF_OPEN \
	114	"Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
	115
743ee1f8	116	static void create_perf_stat_counter(int counter)
ddcacfa0	117	{
a21ca2ca	118	struct perf_counter_attr *attr = attrs + counter;
16c8a109	119
ddcacfa0	120	if (scale)
a21ca2ca IM	121	attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED \|
a21ca2ca IM	122	PERF_FORMAT_TOTAL_TIME_RUNNING;
ddcacfa0 IM	123
	124	if (system_wide) {
	125	int cpu;
cca03c0a	126	for (cpu = 0; cpu < nr_cpus; cpu++) {
a21ca2ca	127	fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
cca03c0a JSR	128	if (fd[cpu][counter] < 0 && verbose)
	129	fprintf(stderr, ERR_PERF_OPEN, counter,
	130	fd[cpu][counter], strerror(errno));
ddcacfa0 IM	131	}
ddcacfa0 IM	132	} else {
a21ca2ca IM	133	attr->inherit = inherit;
a21ca2ca IM	134	attr->disabled = 1;
ddcacfa0	135
a21ca2ca	136	fd[0][counter] = sys_perf_counter_open(attr, 0, -1, -1, 0);
cca03c0a JSR	137	if (fd[0][counter] < 0 && verbose)
	138	fprintf(stderr, ERR_PERF_OPEN, counter,
	139	fd[0][counter], strerror(errno));
ddcacfa0 IM	140	}
	141	}
	142
c04f5e5d IM	143	/*
	144	* Does the counter have nsecs as a unit?
	145	*/
	146	static inline int nsec_counter(int counter)
	147	{
a21ca2ca IM	148	if (attrs[counter].type != PERF_TYPE_SOFTWARE)
	149	return 0;
	150
f4dbfa8f	151	if (attrs[counter].config == PERF_COUNT_SW_CPU_CLOCK)
c04f5e5d	152	return 1;
a21ca2ca	153
f4dbfa8f	154	if (attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
c04f5e5d IM	155	return 1;
	156
	157	return 0;
	158	}
	159
	160	/*
2996f5dd	161	* Read out the results of a single counter:
c04f5e5d	162	*/
2996f5dd	163	static void read_counter(int counter)
c04f5e5d	164	{
9cffa8d5	165	u64 *count, single_count[3];
c04f5e5d IM	166	ssize_t res;
	167	int cpu, nv;
	168	int scaled;
	169
42202dd5	170	count = event_res[run_idx][counter];
2996f5dd	171
c04f5e5d	172	count[0] = count[1] = count[2] = 0;
2996f5dd	173
c04f5e5d	174	nv = scale ? 3 : 1;
cca03c0a	175	for (cpu = 0; cpu < nr_cpus; cpu++) {
743ee1f8 IM	176	if (fd[cpu][counter] < 0)
	177	continue;
	178
9cffa8d5 PM	179	res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
9cffa8d5 PM	180	assert(res == nv * sizeof(u64));
42202dd5 IM	181	close(fd[cpu][counter]);
42202dd5 IM	182	fd[cpu][counter] = -1;
c04f5e5d IM	183
	184	count[0] += single_count[0];
	185	if (scale) {
	186	count[1] += single_count[1];
	187	count[2] += single_count[2];
	188	}
	189	}
	190
	191	scaled = 0;
	192	if (scale) {
	193	if (count[2] == 0) {
42202dd5	194	event_scaled[run_idx][counter] = -1;
2996f5dd	195	count[0] = 0;
c04f5e5d IM	196	return;
c04f5e5d IM	197	}
2996f5dd	198
c04f5e5d	199	if (count[2] < count[1]) {
42202dd5	200	event_scaled[run_idx][counter] = 1;
c04f5e5d IM	201	count[0] = (unsigned long long)
	202	((double)count[0] * count[1] / count[2] + 0.5);
	203	}
	204	}
be1ac0d8 IM	205	/*
	206	* Save the full runtime - to allow normalization during printout:
	207	*/
a21ca2ca	208	if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
f4dbfa8f	209	attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
42202dd5	210	runtime_nsecs[run_idx] = count[0];
e779898a	211	if (attrs[counter].type == PERF_TYPE_HARDWARE &&
f4dbfa8f	212	attrs[counter].config == PERF_COUNT_HW_CPU_CYCLES)
42202dd5	213	runtime_cycles[run_idx] = count[0];
2996f5dd IM	214	}
2996f5dd IM	215
42202dd5 IM	216	static int run_perf_stat(int argc, const char **argv)
	217	{
	218	unsigned long long t0, t1;
	219	int status = 0;
	220	int counter;
	221	int pid;
	222
	223	if (!system_wide)
	224	nr_cpus = 1;
	225
	226	for (counter = 0; counter < nr_counters; counter++)
	227	create_perf_stat_counter(counter);
	228
	229	/*
	230	* Enable counters and exec the command:
	231	*/
	232	t0 = rdclock();
	233	prctl(PR_TASK_PERF_COUNTERS_ENABLE);
	234
	235	if ((pid = fork()) < 0)
	236	perror("failed to fork");
	237
	238	if (!pid) {
	239	if (execvp(argv[0], (char **)argv)) {
	240	perror(argv[0]);
	241	exit(-1);
	242	}
	243	}
	244
	245	wait(&status);
	246
	247	prctl(PR_TASK_PERF_COUNTERS_DISABLE);
	248	t1 = rdclock();
	249
	250	walltime_nsecs[run_idx] = t1 - t0;
	251
	252	for (counter = 0; counter < nr_counters; counter++)
	253	read_counter(counter);
	254
	255	return WEXITSTATUS(status);
	256	}
	257
9cffa8d5	258	static void print_noise(u64 count, u64 noise)
42202dd5 IM	259	{
	260	if (run_count > 1)
	261	fprintf(stderr, " ( +- %7.3f%% )",
	262	(double)noise[0]/(count[0]+1)*100.0);
	263	}
	264
9cffa8d5	265	static void nsec_printout(int counter, u64 count, u64 noise)
44175b6f IM	266	{
	267	double msecs = (double)count[0] / 1000000;
	268
	269	fprintf(stderr, " %14.6f %-20s", msecs, event_name(counter));
	270
	271	if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
	272	attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) {
	273
42202dd5 IM	274	if (walltime_nsecs_avg)
	275	fprintf(stderr, " # %10.3f CPUs ",
	276	(double)count[0] / (double)walltime_nsecs_avg);
44175b6f	277	}
42202dd5	278	print_noise(count, noise);
44175b6f IM	279	}
44175b6f IM	280
9cffa8d5	281	static void abs_printout(int counter, u64 count, u64 noise)
44175b6f IM	282	{
	283	fprintf(stderr, " %14Ld %-20s", count[0], event_name(counter));
	284
42202dd5	285	if (runtime_cycles_avg &&
44175b6f IM	286	attrs[counter].type == PERF_TYPE_HARDWARE &&
	287	attrs[counter].config == PERF_COUNT_HW_INSTRUCTIONS) {
	288
42202dd5 IM	289	fprintf(stderr, " # %10.3f IPC ",
	290	(double)count[0] / (double)runtime_cycles_avg);
	291	} else {
	292	if (runtime_nsecs_avg) {
	293	fprintf(stderr, " # %10.3f M/sec",
	294	(double)count[0]/runtime_nsecs_avg*1000.0);
	295	}
44175b6f	296	}
42202dd5	297	print_noise(count, noise);
44175b6f IM	298	}
44175b6f IM	299
2996f5dd IM	300	/*
	301	* Print out the results of a single counter:
	302	*/
	303	static void print_counter(int counter)
	304	{
9cffa8d5	305	u64 count, noise;
2996f5dd IM	306	int scaled;
2996f5dd IM	307
42202dd5 IM	308	count = event_res_avg[counter];
	309	noise = event_res_noise[counter];
	310	scaled = event_scaled_avg[counter];
2996f5dd IM	311
	312	if (scaled == -1) {
	313	fprintf(stderr, " %14s %-20s\n",
	314	"<not counted>", event_name(counter));
	315	return;
	316	}
c04f5e5d	317
44175b6f	318	if (nsec_counter(counter))
42202dd5	319	nsec_printout(counter, count, noise);
44175b6f	320	else
42202dd5	321	abs_printout(counter, count, noise);
d7c29318	322
c04f5e5d IM	323	if (scaled)
	324	fprintf(stderr, " (scaled from %.2f%%)",
	325	(double) count[2] / count[1] * 100);
44175b6f	326
c04f5e5d IM	327	fprintf(stderr, "\n");
	328	}
	329
42202dd5	330	/*
ef281a19	331	* normalize_noise noise values down to stddev:
42202dd5	332	*/
9cffa8d5	333	static void normalize_noise(u64 *val)
ddcacfa0	334	{
42202dd5	335	double res;
ddcacfa0	336
42202dd5	337	res = (double)val / (run_count sqrt((double)run_count));
ddcacfa0	338
9cffa8d5	339	*val = (u64)res;
42202dd5	340	}
ddcacfa0	341
9cffa8d5	342	static void update_avg(const char name, int idx, u64 avg, u64 *val)
ef281a19 IM	343	{
	344	avg += val;
	345
	346	if (verbose > 1)
	347	fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val);
	348	}
42202dd5 IM	349	/*
	350	* Calculate the averages and noises:
	351	*/
	352	static void calc_avg(void)
	353	{
	354	int i, j;
	355
ef281a19 IM	356	if (verbose > 1)
	357	fprintf(stderr, "\n");
	358
42202dd5	359	for (i = 0; i < run_count; i++) {
ef281a19 IM	360	update_avg("runtime", 0, &runtime_nsecs_avg, runtime_nsecs + i);
	361	update_avg("walltime", 0, &walltime_nsecs_avg, walltime_nsecs + i);
	362	update_avg("runtime_cycles", 0, &runtime_cycles_avg, runtime_cycles + i);
42202dd5 IM	363
42202dd5 IM	364	for (j = 0; j < nr_counters; j++) {
ef281a19 IM	365	update_avg("counter/0", j,
	366	event_res_avg[j]+0, event_res[i][j]+0);
	367	update_avg("counter/1", j,
	368	event_res_avg[j]+1, event_res[i][j]+1);
	369	update_avg("counter/2", j,
	370	event_res_avg[j]+2, event_res[i][j]+2);
	371	update_avg("scaled", j,
	372	event_scaled_avg + j, event_scaled[i]+j);
42202dd5 IM	373	}
	374	}
	375	runtime_nsecs_avg /= run_count;
	376	walltime_nsecs_avg /= run_count;
	377	runtime_cycles_avg /= run_count;
	378
	379	for (j = 0; j < nr_counters; j++) {
	380	event_res_avg[j][0] /= run_count;
	381	event_res_avg[j][1] /= run_count;
	382	event_res_avg[j][2] /= run_count;
	383	}
44db76c8	384
42202dd5 IM	385	for (i = 0; i < run_count; i++) {
42202dd5 IM	386	runtime_nsecs_noise +=
9cffa8d5	387	abs((s64)(runtime_nsecs[i] - runtime_nsecs_avg));
42202dd5	388	walltime_nsecs_noise +=
9cffa8d5	389	abs((s64)(walltime_nsecs[i] - walltime_nsecs_avg));
42202dd5	390	runtime_cycles_noise +=
9cffa8d5	391	abs((s64)(runtime_cycles[i] - runtime_cycles_avg));
42202dd5 IM	392
	393	for (j = 0; j < nr_counters; j++) {
	394	event_res_noise[j][0] +=
9cffa8d5	395	abs((s64)(event_res[i][j][0] - event_res_avg[j][0]));
42202dd5	396	event_res_noise[j][1] +=
9cffa8d5	397	abs((s64)(event_res[i][j][1] - event_res_avg[j][1]));
42202dd5	398	event_res_noise[j][2] +=
9cffa8d5	399	abs((s64)(event_res[i][j][2] - event_res_avg[j][2]));
ddcacfa0 IM	400	}
ddcacfa0 IM	401	}
44db76c8	402
ef281a19 IM	403	normalize_noise(&runtime_nsecs_noise);
	404	normalize_noise(&walltime_nsecs_noise);
	405	normalize_noise(&runtime_cycles_noise);
44db76c8	406
42202dd5	407	for (j = 0; j < nr_counters; j++) {
ef281a19 IM	408	normalize_noise(&event_res_noise[j][0]);
	409	normalize_noise(&event_res_noise[j][1]);
	410	normalize_noise(&event_res_noise[j][2]);
42202dd5 IM	411	}
	412	}
	413
	414	static void print_stat(int argc, const char **argv)
	415	{
	416	int i, counter;
	417
	418	calc_avg();
ddcacfa0 IM	419
	420	fflush(stdout);
	421
	422	fprintf(stderr, "\n");
44db76c8 IM	423	fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
	424
	425	for (i = 1; i < argc; i++)
	426	fprintf(stderr, " %s", argv[i]);
	427
42202dd5 IM	428	fprintf(stderr, "\'");
	429	if (run_count > 1)
	430	fprintf(stderr, " (%d runs)", run_count);
	431	fprintf(stderr, ":\n\n");
2996f5dd	432
c04f5e5d IM	433	for (counter = 0; counter < nr_counters; counter++)
c04f5e5d IM	434	print_counter(counter);
ddcacfa0	435
ddcacfa0	436
ddcacfa0	437	fprintf(stderr, "\n");
42202dd5 IM	438	fprintf(stderr, " %14.9f seconds time elapsed.\n",
42202dd5 IM	439	(double)walltime_nsecs_avg/1e9);
ddcacfa0	440	fprintf(stderr, "\n");
ddcacfa0 IM	441	}
ddcacfa0 IM	442
f7b7c26e PZ	443	static volatile int signr = -1;
f7b7c26e PZ	444
5242519b	445	static void skip_signal(int signo)
ddcacfa0	446	{
f7b7c26e PZ	447	signr = signo;
	448	}
	449
	450	static void sig_atexit(void)
	451	{
	452	if (signr == -1)
	453	return;
	454
	455	signal(signr, SIG_DFL);
	456	kill(getpid(), signr);
5242519b IM	457	}
	458
	459	static const char * const stat_usage[] = {
	460	"perf stat [<options>] <command>",
	461	NULL
	462	};
	463
5242519b IM	464	static const struct option options[] = {
5242519b IM	465	OPT_CALLBACK('e', "event", NULL, "event",
86847b62 TG	466	"event selector. use 'perf list' to list available events",
86847b62 TG	467	parse_events),
5242519b IM	468	OPT_BOOLEAN('i', "inherit", &inherit,
	469	"child tasks inherit counters"),
	470	OPT_INTEGER('p', "pid", &target_pid,
	471	"stat events on existing pid"),
	472	OPT_BOOLEAN('a', "all-cpus", &system_wide,
	473	"system-wide collection from all CPUs"),
86847b62	474	OPT_BOOLEAN('S', "scale", &scale,
5242519b	475	"scale/normalize counters"),
743ee1f8 IM	476	OPT_BOOLEAN('v', "verbose", &verbose,
743ee1f8 IM	477	"be more verbose (show counter open errors, etc)"),
42202dd5 IM	478	OPT_INTEGER('r', "repeat", &run_count,
42202dd5 IM	479	"repeat command and print average + stddev (max: 100)"),
5242519b IM	480	OPT_END()
	481	};
	482
	483	int cmd_stat(int argc, const char *argv, const char prefix)
	484	{
42202dd5 IM	485	int status;
42202dd5 IM	486
5242519b IM	487	page_size = sysconf(_SC_PAGE_SIZE);
5242519b IM	488
a21ca2ca	489	memcpy(attrs, default_attrs, sizeof(attrs));
5242519b IM	490
	491	argc = parse_options(argc, argv, options, stat_usage, 0);
	492	if (!argc)
	493	usage_with_options(stat_usage, options);
42202dd5 IM	494	if (run_count <= 0 \|\| run_count > MAX_RUN)
42202dd5 IM	495	usage_with_options(stat_usage, options);
ddcacfa0	496
a21ca2ca	497	if (!nr_counters)
ddcacfa0	498	nr_counters = 8;
ddcacfa0	499
ddcacfa0 IM	500	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
	501	assert(nr_cpus <= MAX_NR_CPUS);
	502	assert(nr_cpus >= 0);
	503
58d7e993 IM	504	/*
	505	* We dont want to block the signals - that would cause
	506	* child tasks to inherit that and Ctrl-C would not work.
	507	* What we want is for Ctrl-C to work in the exec()-ed
	508	* task, but being ignored by perf stat itself:
	509	*/
f7b7c26e	510	atexit(sig_atexit);
58d7e993 IM	511	signal(SIGINT, skip_signal);
	512	signal(SIGALRM, skip_signal);
	513	signal(SIGABRT, skip_signal);
	514
42202dd5 IM	515	status = 0;
	516	for (run_idx = 0; run_idx < run_count; run_idx++) {
	517	if (run_count != 1 && verbose)
	518	fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx+1);
	519	status = run_perf_stat(argc, argv);
	520	}
	521
	522	print_stat(argc, argv);
	523
	524	return status;
ddcacfa0	525	}