[mirror_ubuntu-hirsute-kernel.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	},

};

#define MAX_RUN			100

static int			system_wide			=  0;
static int			verbose				=  0;
static int			nr_cpus				=  0;
static int			run_idx				=  0;

static int			run_count			=  1;
static int			inherit				=  1;
static int			scale				=  1;
static int			target_pid			= -1;

static int			fd[MAX_NR_CPUS][MAX_COUNTERS];

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

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

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;

	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
3d632595 JSR	62	#define MAX_RUN 100
3d632595 JSR	63
a21ca2ca	64	static int system_wide = 0;
743ee1f8	65	static int verbose = 0;
ddcacfa0	66	static int nr_cpus = 0;
3d632595	67	static int run_idx = 0;
ddcacfa0	68
3d632595 JSR	69	static int run_count = 1;
3d632595 JSR	70	static int inherit = 1;
66cf7829	71	static int scale = 1;
3d632595	72	static int target_pid = -1;
ddcacfa0	73
3d632595	74	static int fd[MAX_NR_CPUS][MAX_COUNTERS];
42202dd5	75
9cffa8d5 PM	76	static u64 runtime_nsecs[MAX_RUN];
	77	static u64 walltime_nsecs[MAX_RUN];
	78	static u64 runtime_cycles[MAX_RUN];
42202dd5	79
3d632595 JSR	80	static u64 event_res[MAX_RUN][MAX_COUNTERS][3];
	81	static u64 event_scaled[MAX_RUN][MAX_COUNTERS];
	82
9cffa8d5 PM	83	static u64 event_res_avg[MAX_COUNTERS][3];
9cffa8d5 PM	84	static u64 event_res_noise[MAX_COUNTERS][3];
42202dd5	85
9cffa8d5	86	static u64 event_scaled_avg[MAX_COUNTERS];
42202dd5	87
9cffa8d5 PM	88	static u64 runtime_nsecs_avg;
9cffa8d5 PM	89	static u64 runtime_nsecs_noise;
42202dd5	90
9cffa8d5 PM	91	static u64 walltime_nsecs_avg;
9cffa8d5 PM	92	static u64 walltime_nsecs_noise;
42202dd5	93
9cffa8d5 PM	94	static u64 runtime_cycles_avg;
9cffa8d5 PM	95	static u64 runtime_cycles_noise;
be1ac0d8	96
cca03c0a JSR	97	#define ERR_PERF_OPEN \
	98	"Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
	99
743ee1f8	100	static void create_perf_stat_counter(int counter)
ddcacfa0	101	{
a21ca2ca	102	struct perf_counter_attr *attr = attrs + counter;
16c8a109	103
ddcacfa0	104	if (scale)
a21ca2ca IM	105	attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED \|
a21ca2ca IM	106	PERF_FORMAT_TOTAL_TIME_RUNNING;
ddcacfa0 IM	107
	108	if (system_wide) {
	109	int cpu;
cca03c0a	110	for (cpu = 0; cpu < nr_cpus; cpu++) {
a21ca2ca	111	fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
cca03c0a JSR	112	if (fd[cpu][counter] < 0 && verbose)
	113	fprintf(stderr, ERR_PERF_OPEN, counter,
	114	fd[cpu][counter], strerror(errno));
ddcacfa0 IM	115	}
ddcacfa0 IM	116	} else {
a21ca2ca IM	117	attr->inherit = inherit;
a21ca2ca IM	118	attr->disabled = 1;
ddcacfa0	119
a21ca2ca	120	fd[0][counter] = sys_perf_counter_open(attr, 0, -1, -1, 0);
cca03c0a JSR	121	if (fd[0][counter] < 0 && verbose)
	122	fprintf(stderr, ERR_PERF_OPEN, counter,
	123	fd[0][counter], strerror(errno));
ddcacfa0 IM	124	}
	125	}
	126
c04f5e5d IM	127	/*
	128	* Does the counter have nsecs as a unit?
	129	*/
	130	static inline int nsec_counter(int counter)
	131	{
a21ca2ca IM	132	if (attrs[counter].type != PERF_TYPE_SOFTWARE)
	133	return 0;
	134
f4dbfa8f	135	if (attrs[counter].config == PERF_COUNT_SW_CPU_CLOCK)
c04f5e5d	136	return 1;
a21ca2ca	137
f4dbfa8f	138	if (attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
c04f5e5d IM	139	return 1;
	140
	141	return 0;
	142	}
	143
	144	/*
2996f5dd	145	* Read out the results of a single counter:
c04f5e5d	146	*/
2996f5dd	147	static void read_counter(int counter)
c04f5e5d	148	{
9cffa8d5	149	u64 *count, single_count[3];
c04f5e5d IM	150	ssize_t res;
	151	int cpu, nv;
	152	int scaled;
	153
42202dd5	154	count = event_res[run_idx][counter];
2996f5dd	155
c04f5e5d	156	count[0] = count[1] = count[2] = 0;
2996f5dd	157
c04f5e5d	158	nv = scale ? 3 : 1;
cca03c0a	159	for (cpu = 0; cpu < nr_cpus; cpu++) {
743ee1f8 IM	160	if (fd[cpu][counter] < 0)
	161	continue;
	162
9cffa8d5 PM	163	res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
9cffa8d5 PM	164	assert(res == nv * sizeof(u64));
42202dd5 IM	165	close(fd[cpu][counter]);
42202dd5 IM	166	fd[cpu][counter] = -1;
c04f5e5d IM	167
	168	count[0] += single_count[0];
	169	if (scale) {
	170	count[1] += single_count[1];
	171	count[2] += single_count[2];
	172	}
	173	}
	174
	175	scaled = 0;
	176	if (scale) {
	177	if (count[2] == 0) {
42202dd5	178	event_scaled[run_idx][counter] = -1;
2996f5dd	179	count[0] = 0;
c04f5e5d IM	180	return;
c04f5e5d IM	181	}
2996f5dd	182
c04f5e5d	183	if (count[2] < count[1]) {
42202dd5	184	event_scaled[run_idx][counter] = 1;
c04f5e5d IM	185	count[0] = (unsigned long long)
	186	((double)count[0] * count[1] / count[2] + 0.5);
	187	}
	188	}
be1ac0d8 IM	189	/*
	190	* Save the full runtime - to allow normalization during printout:
	191	*/
a21ca2ca	192	if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
f4dbfa8f	193	attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
42202dd5	194	runtime_nsecs[run_idx] = count[0];
e779898a	195	if (attrs[counter].type == PERF_TYPE_HARDWARE &&
f4dbfa8f	196	attrs[counter].config == PERF_COUNT_HW_CPU_CYCLES)
42202dd5	197	runtime_cycles[run_idx] = count[0];
2996f5dd IM	198	}
2996f5dd IM	199
42202dd5 IM	200	static int run_perf_stat(int argc, const char **argv)
	201	{
	202	unsigned long long t0, t1;
	203	int status = 0;
	204	int counter;
	205	int pid;
	206
	207	if (!system_wide)
	208	nr_cpus = 1;
	209
	210	for (counter = 0; counter < nr_counters; counter++)
	211	create_perf_stat_counter(counter);
	212
	213	/*
	214	* Enable counters and exec the command:
	215	*/
	216	t0 = rdclock();
	217	prctl(PR_TASK_PERF_COUNTERS_ENABLE);
	218
	219	if ((pid = fork()) < 0)
	220	perror("failed to fork");
	221
	222	if (!pid) {
	223	if (execvp(argv[0], (char **)argv)) {
	224	perror(argv[0]);
	225	exit(-1);
	226	}
	227	}
	228
	229	wait(&status);
	230
	231	prctl(PR_TASK_PERF_COUNTERS_DISABLE);
	232	t1 = rdclock();
	233
	234	walltime_nsecs[run_idx] = t1 - t0;
	235
	236	for (counter = 0; counter < nr_counters; counter++)
	237	read_counter(counter);
	238
	239	return WEXITSTATUS(status);
	240	}
	241
9cffa8d5	242	static void print_noise(u64 count, u64 noise)
42202dd5 IM	243	{
	244	if (run_count > 1)
	245	fprintf(stderr, " ( +- %7.3f%% )",
	246	(double)noise[0]/(count[0]+1)*100.0);
	247	}
	248
9cffa8d5	249	static void nsec_printout(int counter, u64 count, u64 noise)
44175b6f IM	250	{
	251	double msecs = (double)count[0] / 1000000;
	252
	253	fprintf(stderr, " %14.6f %-20s", msecs, event_name(counter));
	254
	255	if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
	256	attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) {
	257
42202dd5 IM	258	if (walltime_nsecs_avg)
	259	fprintf(stderr, " # %10.3f CPUs ",
	260	(double)count[0] / (double)walltime_nsecs_avg);
44175b6f	261	}
42202dd5	262	print_noise(count, noise);
44175b6f IM	263	}
44175b6f IM	264
9cffa8d5	265	static void abs_printout(int counter, u64 count, u64 noise)
44175b6f IM	266	{
	267	fprintf(stderr, " %14Ld %-20s", count[0], event_name(counter));
	268
42202dd5	269	if (runtime_cycles_avg &&
44175b6f IM	270	attrs[counter].type == PERF_TYPE_HARDWARE &&
	271	attrs[counter].config == PERF_COUNT_HW_INSTRUCTIONS) {
	272
42202dd5 IM	273	fprintf(stderr, " # %10.3f IPC ",
	274	(double)count[0] / (double)runtime_cycles_avg);
	275	} else {
	276	if (runtime_nsecs_avg) {
	277	fprintf(stderr, " # %10.3f M/sec",
	278	(double)count[0]/runtime_nsecs_avg*1000.0);
	279	}
44175b6f	280	}
42202dd5	281	print_noise(count, noise);
44175b6f IM	282	}
44175b6f IM	283
2996f5dd IM	284	/*
	285	* Print out the results of a single counter:
	286	*/
	287	static void print_counter(int counter)
	288	{
9cffa8d5	289	u64 count, noise;
2996f5dd IM	290	int scaled;
2996f5dd IM	291
42202dd5 IM	292	count = event_res_avg[counter];
	293	noise = event_res_noise[counter];
	294	scaled = event_scaled_avg[counter];
2996f5dd IM	295
	296	if (scaled == -1) {
	297	fprintf(stderr, " %14s %-20s\n",
	298	"<not counted>", event_name(counter));
	299	return;
	300	}
c04f5e5d	301
44175b6f	302	if (nsec_counter(counter))
42202dd5	303	nsec_printout(counter, count, noise);
44175b6f	304	else
42202dd5	305	abs_printout(counter, count, noise);
d7c29318	306
c04f5e5d IM	307	if (scaled)
	308	fprintf(stderr, " (scaled from %.2f%%)",
	309	(double) count[2] / count[1] * 100);
44175b6f	310
c04f5e5d IM	311	fprintf(stderr, "\n");
	312	}
	313
42202dd5	314	/*
ef281a19	315	* normalize_noise noise values down to stddev:
42202dd5	316	*/
9cffa8d5	317	static void normalize_noise(u64 *val)
ddcacfa0	318	{
42202dd5	319	double res;
ddcacfa0	320
42202dd5	321	res = (double)val / (run_count sqrt((double)run_count));
ddcacfa0	322
9cffa8d5	323	*val = (u64)res;
42202dd5	324	}
ddcacfa0	325
9cffa8d5	326	static void update_avg(const char name, int idx, u64 avg, u64 *val)
ef281a19 IM	327	{
	328	avg += val;
	329
	330	if (verbose > 1)
	331	fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val);
	332	}
42202dd5 IM	333	/*
	334	* Calculate the averages and noises:
	335	*/
	336	static void calc_avg(void)
	337	{
	338	int i, j;
	339
ef281a19 IM	340	if (verbose > 1)
	341	fprintf(stderr, "\n");
	342
42202dd5	343	for (i = 0; i < run_count; i++) {
ef281a19 IM	344	update_avg("runtime", 0, &runtime_nsecs_avg, runtime_nsecs + i);
	345	update_avg("walltime", 0, &walltime_nsecs_avg, walltime_nsecs + i);
	346	update_avg("runtime_cycles", 0, &runtime_cycles_avg, runtime_cycles + i);
42202dd5 IM	347
42202dd5 IM	348	for (j = 0; j < nr_counters; j++) {
ef281a19 IM	349	update_avg("counter/0", j,
	350	event_res_avg[j]+0, event_res[i][j]+0);
	351	update_avg("counter/1", j,
	352	event_res_avg[j]+1, event_res[i][j]+1);
	353	update_avg("counter/2", j,
	354	event_res_avg[j]+2, event_res[i][j]+2);
	355	update_avg("scaled", j,
	356	event_scaled_avg + j, event_scaled[i]+j);
42202dd5 IM	357	}
	358	}
	359	runtime_nsecs_avg /= run_count;
	360	walltime_nsecs_avg /= run_count;
	361	runtime_cycles_avg /= run_count;
	362
	363	for (j = 0; j < nr_counters; j++) {
	364	event_res_avg[j][0] /= run_count;
	365	event_res_avg[j][1] /= run_count;
	366	event_res_avg[j][2] /= run_count;
	367	}
44db76c8	368
42202dd5 IM	369	for (i = 0; i < run_count; i++) {
42202dd5 IM	370	runtime_nsecs_noise +=
9cffa8d5	371	abs((s64)(runtime_nsecs[i] - runtime_nsecs_avg));
42202dd5	372	walltime_nsecs_noise +=
9cffa8d5	373	abs((s64)(walltime_nsecs[i] - walltime_nsecs_avg));
42202dd5	374	runtime_cycles_noise +=
9cffa8d5	375	abs((s64)(runtime_cycles[i] - runtime_cycles_avg));
42202dd5 IM	376
	377	for (j = 0; j < nr_counters; j++) {
	378	event_res_noise[j][0] +=
9cffa8d5	379	abs((s64)(event_res[i][j][0] - event_res_avg[j][0]));
42202dd5	380	event_res_noise[j][1] +=
9cffa8d5	381	abs((s64)(event_res[i][j][1] - event_res_avg[j][1]));
42202dd5	382	event_res_noise[j][2] +=
9cffa8d5	383	abs((s64)(event_res[i][j][2] - event_res_avg[j][2]));
ddcacfa0 IM	384	}
ddcacfa0 IM	385	}
44db76c8	386
ef281a19 IM	387	normalize_noise(&runtime_nsecs_noise);
	388	normalize_noise(&walltime_nsecs_noise);
	389	normalize_noise(&runtime_cycles_noise);
44db76c8	390
42202dd5	391	for (j = 0; j < nr_counters; j++) {
ef281a19 IM	392	normalize_noise(&event_res_noise[j][0]);
	393	normalize_noise(&event_res_noise[j][1]);
	394	normalize_noise(&event_res_noise[j][2]);
42202dd5 IM	395	}
	396	}
	397
	398	static void print_stat(int argc, const char **argv)
	399	{
	400	int i, counter;
	401
	402	calc_avg();
ddcacfa0 IM	403
	404	fflush(stdout);
	405
	406	fprintf(stderr, "\n");
44db76c8 IM	407	fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
	408
	409	for (i = 1; i < argc; i++)
	410	fprintf(stderr, " %s", argv[i]);
	411
42202dd5 IM	412	fprintf(stderr, "\'");
	413	if (run_count > 1)
	414	fprintf(stderr, " (%d runs)", run_count);
	415	fprintf(stderr, ":\n\n");
2996f5dd	416
c04f5e5d IM	417	for (counter = 0; counter < nr_counters; counter++)
c04f5e5d IM	418	print_counter(counter);
ddcacfa0	419
ddcacfa0	420
ddcacfa0	421	fprintf(stderr, "\n");
42202dd5 IM	422	fprintf(stderr, " %14.9f seconds time elapsed.\n",
42202dd5 IM	423	(double)walltime_nsecs_avg/1e9);
ddcacfa0	424	fprintf(stderr, "\n");
ddcacfa0 IM	425	}
ddcacfa0 IM	426
f7b7c26e PZ	427	static volatile int signr = -1;
f7b7c26e PZ	428
5242519b	429	static void skip_signal(int signo)
ddcacfa0	430	{
f7b7c26e PZ	431	signr = signo;
	432	}
	433
	434	static void sig_atexit(void)
	435	{
	436	if (signr == -1)
	437	return;
	438
	439	signal(signr, SIG_DFL);
	440	kill(getpid(), signr);
5242519b IM	441	}
	442
	443	static const char * const stat_usage[] = {
	444	"perf stat [<options>] <command>",
	445	NULL
	446	};
	447
5242519b IM	448	static const struct option options[] = {
5242519b IM	449	OPT_CALLBACK('e', "event", NULL, "event",
86847b62 TG	450	"event selector. use 'perf list' to list available events",
86847b62 TG	451	parse_events),
5242519b IM	452	OPT_BOOLEAN('i', "inherit", &inherit,
	453	"child tasks inherit counters"),
	454	OPT_INTEGER('p', "pid", &target_pid,
	455	"stat events on existing pid"),
	456	OPT_BOOLEAN('a', "all-cpus", &system_wide,
3d632595	457	"system-wide collection from all CPUs"),
86847b62	458	OPT_BOOLEAN('S', "scale", &scale,
3d632595	459	"scale/normalize counters"),
743ee1f8 IM	460	OPT_BOOLEAN('v', "verbose", &verbose,
743ee1f8 IM	461	"be more verbose (show counter open errors, etc)"),
42202dd5 IM	462	OPT_INTEGER('r', "repeat", &run_count,
42202dd5 IM	463	"repeat command and print average + stddev (max: 100)"),
5242519b IM	464	OPT_END()
	465	};
	466
	467	int cmd_stat(int argc, const char *argv, const char prefix)
	468	{
42202dd5 IM	469	int status;
42202dd5 IM	470
a21ca2ca	471	memcpy(attrs, default_attrs, sizeof(attrs));
5242519b IM	472
	473	argc = parse_options(argc, argv, options, stat_usage, 0);
	474	if (!argc)
	475	usage_with_options(stat_usage, options);
42202dd5 IM	476	if (run_count <= 0 \|\| run_count > MAX_RUN)
42202dd5 IM	477	usage_with_options(stat_usage, options);
ddcacfa0	478
a21ca2ca	479	if (!nr_counters)
ddcacfa0	480	nr_counters = 8;
ddcacfa0	481
ddcacfa0 IM	482	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
	483	assert(nr_cpus <= MAX_NR_CPUS);
	484	assert(nr_cpus >= 0);
	485
58d7e993 IM	486	/*
	487	* We dont want to block the signals - that would cause
	488	* child tasks to inherit that and Ctrl-C would not work.
	489	* What we want is for Ctrl-C to work in the exec()-ed
	490	* task, but being ignored by perf stat itself:
	491	*/
f7b7c26e	492	atexit(sig_atexit);
58d7e993 IM	493	signal(SIGINT, skip_signal);
	494	signal(SIGALRM, skip_signal);
	495	signal(SIGABRT, skip_signal);
	496
42202dd5 IM	497	status = 0;
	498	for (run_idx = 0; run_idx < run_count; run_idx++) {
	499	if (run_count != 1 && verbose)
3d632595	500	fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
42202dd5 IM	501	status = run_perf_stat(argc, argv);
	502	}
	503
	504	print_stat(argc, argv);
	505
	506	return status;
ddcacfa0	507	}