[mirror_ubuntu-focal-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>
 *   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 <sys/prctl.h>
#include <math.h>

static struct perf_counter_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_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			null_run			=  0;

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  %-24s", 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  %-24s", 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  %-24s\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);
			if (event_scaled[i][j] != -1)
				update_avg("scaled", j,
					event_scaled_avg + j, event_scaled[i]+j);
			else
				event_scaled_avg[j] = -1;
		}
	}
	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",
			(double)walltime_nsecs_avg/1e9);
	if (run_count > 1) {
		fprintf(stderr, "   ( +- %7.3f%% )",
			100.0*(double)walltime_nsecs_noise/(double)walltime_nsecs_avg);
	}
	fprintf(stderr, "\n\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_BOOLEAN('n', "null", &null_run,
		    "null run - dont start any counters"),
	OPT_END()
};

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

	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);

	/* 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);
	}

	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>
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"
ddcacfa0	45
ddcacfa0	46	#include <sys/prctl.h>
42202dd5	47	#include <math.h>
16c8a109	48
c3043569	49	static struct perf_counter_attr default_attrs[] = {
ddcacfa0	50
f4dbfa8f PZ	51	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
	52	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
	53	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
	54	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
	55
	56	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
	57	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
	58	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
	59	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES },
	60
ddcacfa0	61	};
5242519b	62
3d632595 JSR	63	#define MAX_RUN 100
3d632595 JSR	64
a21ca2ca	65	static int system_wide = 0;
743ee1f8	66	static int verbose = 0;
ddcacfa0	67	static int nr_cpus = 0;
3d632595	68	static int run_idx = 0;
ddcacfa0	69
3d632595 JSR	70	static int run_count = 1;
3d632595 JSR	71	static int inherit = 1;
66cf7829	72	static int scale = 1;
3d632595	73	static int target_pid = -1;
0cfb7a13	74	static int null_run = 0;
ddcacfa0	75
3d632595	76	static int fd[MAX_NR_CPUS][MAX_COUNTERS];
42202dd5	77
9cffa8d5 PM	78	static u64 runtime_nsecs[MAX_RUN];
	79	static u64 walltime_nsecs[MAX_RUN];
	80	static u64 runtime_cycles[MAX_RUN];
42202dd5	81
3d632595 JSR	82	static u64 event_res[MAX_RUN][MAX_COUNTERS][3];
	83	static u64 event_scaled[MAX_RUN][MAX_COUNTERS];
	84
9cffa8d5 PM	85	static u64 event_res_avg[MAX_COUNTERS][3];
9cffa8d5 PM	86	static u64 event_res_noise[MAX_COUNTERS][3];
42202dd5	87
9cffa8d5	88	static u64 event_scaled_avg[MAX_COUNTERS];
42202dd5	89
9cffa8d5 PM	90	static u64 runtime_nsecs_avg;
9cffa8d5 PM	91	static u64 runtime_nsecs_noise;
42202dd5	92
9cffa8d5 PM	93	static u64 walltime_nsecs_avg;
9cffa8d5 PM	94	static u64 walltime_nsecs_noise;
42202dd5	95
9cffa8d5 PM	96	static u64 runtime_cycles_avg;
9cffa8d5 PM	97	static u64 runtime_cycles_noise;
be1ac0d8	98
cca03c0a JSR	99	#define ERR_PERF_OPEN \
	100	"Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
	101
743ee1f8	102	static void create_perf_stat_counter(int counter)
ddcacfa0	103	{
a21ca2ca	104	struct perf_counter_attr *attr = attrs + counter;
16c8a109	105
ddcacfa0	106	if (scale)
a21ca2ca IM	107	attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED \|
a21ca2ca IM	108	PERF_FORMAT_TOTAL_TIME_RUNNING;
ddcacfa0 IM	109
	110	if (system_wide) {
	111	int cpu;
cca03c0a	112	for (cpu = 0; cpu < nr_cpus; cpu++) {
a21ca2ca	113	fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
cca03c0a JSR	114	if (fd[cpu][counter] < 0 && verbose)
	115	fprintf(stderr, ERR_PERF_OPEN, counter,
	116	fd[cpu][counter], strerror(errno));
ddcacfa0 IM	117	}
ddcacfa0 IM	118	} else {
a21ca2ca IM	119	attr->inherit = inherit;
a21ca2ca IM	120	attr->disabled = 1;
ddcacfa0	121
a21ca2ca	122	fd[0][counter] = sys_perf_counter_open(attr, 0, -1, -1, 0);
cca03c0a JSR	123	if (fd[0][counter] < 0 && verbose)
	124	fprintf(stderr, ERR_PERF_OPEN, counter,
	125	fd[0][counter], strerror(errno));
ddcacfa0 IM	126	}
	127	}
	128
c04f5e5d IM	129	/*
	130	* Does the counter have nsecs as a unit?
	131	*/
	132	static inline int nsec_counter(int counter)
	133	{
a21ca2ca IM	134	if (attrs[counter].type != PERF_TYPE_SOFTWARE)
	135	return 0;
	136
f4dbfa8f	137	if (attrs[counter].config == PERF_COUNT_SW_CPU_CLOCK)
c04f5e5d	138	return 1;
a21ca2ca	139
f4dbfa8f	140	if (attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
c04f5e5d IM	141	return 1;
	142
	143	return 0;
	144	}
	145
	146	/*
2996f5dd	147	* Read out the results of a single counter:
c04f5e5d	148	*/
2996f5dd	149	static void read_counter(int counter)
c04f5e5d	150	{
9cffa8d5	151	u64 *count, single_count[3];
c04f5e5d IM	152	ssize_t res;
	153	int cpu, nv;
	154	int scaled;
	155
42202dd5	156	count = event_res[run_idx][counter];
2996f5dd	157
c04f5e5d	158	count[0] = count[1] = count[2] = 0;
2996f5dd	159
c04f5e5d	160	nv = scale ? 3 : 1;
cca03c0a	161	for (cpu = 0; cpu < nr_cpus; cpu++) {
743ee1f8 IM	162	if (fd[cpu][counter] < 0)
	163	continue;
	164
9cffa8d5 PM	165	res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
9cffa8d5 PM	166	assert(res == nv * sizeof(u64));
42202dd5 IM	167	close(fd[cpu][counter]);
42202dd5 IM	168	fd[cpu][counter] = -1;
c04f5e5d IM	169
	170	count[0] += single_count[0];
	171	if (scale) {
	172	count[1] += single_count[1];
	173	count[2] += single_count[2];
	174	}
	175	}
	176
	177	scaled = 0;
	178	if (scale) {
	179	if (count[2] == 0) {
42202dd5	180	event_scaled[run_idx][counter] = -1;
2996f5dd	181	count[0] = 0;
c04f5e5d IM	182	return;
c04f5e5d IM	183	}
2996f5dd	184
c04f5e5d	185	if (count[2] < count[1]) {
42202dd5	186	event_scaled[run_idx][counter] = 1;
c04f5e5d IM	187	count[0] = (unsigned long long)
	188	((double)count[0] * count[1] / count[2] + 0.5);
	189	}
	190	}
be1ac0d8 IM	191	/*
	192	* Save the full runtime - to allow normalization during printout:
	193	*/
a21ca2ca	194	if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
f4dbfa8f	195	attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
42202dd5	196	runtime_nsecs[run_idx] = count[0];
e779898a	197	if (attrs[counter].type == PERF_TYPE_HARDWARE &&
f4dbfa8f	198	attrs[counter].config == PERF_COUNT_HW_CPU_CYCLES)
42202dd5	199	runtime_cycles[run_idx] = count[0];
2996f5dd IM	200	}
2996f5dd IM	201
42202dd5 IM	202	static int run_perf_stat(int argc, const char **argv)
	203	{
	204	unsigned long long t0, t1;
	205	int status = 0;
	206	int counter;
	207	int pid;
	208
	209	if (!system_wide)
	210	nr_cpus = 1;
	211
	212	for (counter = 0; counter < nr_counters; counter++)
	213	create_perf_stat_counter(counter);
	214
	215	/*
	216	* Enable counters and exec the command:
	217	*/
	218	t0 = rdclock();
	219	prctl(PR_TASK_PERF_COUNTERS_ENABLE);
	220
	221	if ((pid = fork()) < 0)
	222	perror("failed to fork");
	223
	224	if (!pid) {
	225	if (execvp(argv[0], (char **)argv)) {
	226	perror(argv[0]);
	227	exit(-1);
	228	}
	229	}
	230
	231	wait(&status);
	232
	233	prctl(PR_TASK_PERF_COUNTERS_DISABLE);
	234	t1 = rdclock();
	235
	236	walltime_nsecs[run_idx] = t1 - t0;
	237
	238	for (counter = 0; counter < nr_counters; counter++)
	239	read_counter(counter);
	240
	241	return WEXITSTATUS(status);
	242	}
	243
9cffa8d5	244	static void print_noise(u64 count, u64 noise)
42202dd5 IM	245	{
	246	if (run_count > 1)
	247	fprintf(stderr, " ( +- %7.3f%% )",
	248	(double)noise[0]/(count[0]+1)*100.0);
	249	}
	250
9cffa8d5	251	static void nsec_printout(int counter, u64 count, u64 noise)
44175b6f IM	252	{
	253	double msecs = (double)count[0] / 1000000;
	254
6e750a8f	255	fprintf(stderr, " %14.6f %-24s", msecs, event_name(counter));
44175b6f IM	256
	257	if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
	258	attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) {
	259
42202dd5 IM	260	if (walltime_nsecs_avg)
	261	fprintf(stderr, " # %10.3f CPUs ",
	262	(double)count[0] / (double)walltime_nsecs_avg);
44175b6f	263	}
42202dd5	264	print_noise(count, noise);
44175b6f IM	265	}
44175b6f IM	266
9cffa8d5	267	static void abs_printout(int counter, u64 count, u64 noise)
44175b6f	268	{
6e750a8f	269	fprintf(stderr, " %14Ld %-24s", count[0], event_name(counter));
44175b6f	270
42202dd5	271	if (runtime_cycles_avg &&
44175b6f IM	272	attrs[counter].type == PERF_TYPE_HARDWARE &&
	273	attrs[counter].config == PERF_COUNT_HW_INSTRUCTIONS) {
	274
42202dd5 IM	275	fprintf(stderr, " # %10.3f IPC ",
	276	(double)count[0] / (double)runtime_cycles_avg);
	277	} else {
	278	if (runtime_nsecs_avg) {
	279	fprintf(stderr, " # %10.3f M/sec",
	280	(double)count[0]/runtime_nsecs_avg*1000.0);
	281	}
44175b6f	282	}
42202dd5	283	print_noise(count, noise);
44175b6f IM	284	}
44175b6f IM	285
2996f5dd IM	286	/*
	287	* Print out the results of a single counter:
	288	*/
	289	static void print_counter(int counter)
	290	{
9cffa8d5	291	u64 count, noise;
2996f5dd IM	292	int scaled;
2996f5dd IM	293
42202dd5 IM	294	count = event_res_avg[counter];
	295	noise = event_res_noise[counter];
	296	scaled = event_scaled_avg[counter];
2996f5dd IM	297
2996f5dd IM	298	if (scaled == -1) {
6e750a8f	299	fprintf(stderr, " %14s %-24s\n",
2996f5dd IM	300	"<not counted>", event_name(counter));
	301	return;
	302	}
c04f5e5d	303
44175b6f	304	if (nsec_counter(counter))
42202dd5	305	nsec_printout(counter, count, noise);
44175b6f	306	else
42202dd5	307	abs_printout(counter, count, noise);
d7c29318	308
c04f5e5d	309	if (scaled)
210ad39f IM	310	fprintf(stderr, " (scaled from %.2f%%)",
210ad39f IM	311	(double) count[2] / count[1] * 100);
44175b6f	312
c04f5e5d IM	313	fprintf(stderr, "\n");
	314	}
	315
42202dd5	316	/*
ef281a19	317	* normalize_noise noise values down to stddev:
42202dd5	318	*/
9cffa8d5	319	static void normalize_noise(u64 *val)
ddcacfa0	320	{
42202dd5	321	double res;
ddcacfa0	322
42202dd5	323	res = (double)val / (run_count sqrt((double)run_count));
ddcacfa0	324
9cffa8d5	325	*val = (u64)res;
42202dd5	326	}
ddcacfa0	327
9cffa8d5	328	static void update_avg(const char name, int idx, u64 avg, u64 *val)
ef281a19 IM	329	{
	330	avg += val;
	331
	332	if (verbose > 1)
	333	fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val);
	334	}
42202dd5 IM	335	/*
	336	* Calculate the averages and noises:
	337	*/
	338	static void calc_avg(void)
	339	{
	340	int i, j;
	341
ef281a19 IM	342	if (verbose > 1)
	343	fprintf(stderr, "\n");
	344
42202dd5	345	for (i = 0; i < run_count; i++) {
ef281a19 IM	346	update_avg("runtime", 0, &runtime_nsecs_avg, runtime_nsecs + i);
	347	update_avg("walltime", 0, &walltime_nsecs_avg, walltime_nsecs + i);
	348	update_avg("runtime_cycles", 0, &runtime_cycles_avg, runtime_cycles + i);
42202dd5 IM	349
42202dd5 IM	350	for (j = 0; j < nr_counters; j++) {
ef281a19 IM	351	update_avg("counter/0", j,
	352	event_res_avg[j]+0, event_res[i][j]+0);
	353	update_avg("counter/1", j,
	354	event_res_avg[j]+1, event_res[i][j]+1);
	355	update_avg("counter/2", j,
	356	event_res_avg[j]+2, event_res[i][j]+2);
566747e6 IM	357	if (event_scaled[i][j] != -1)
	358	update_avg("scaled", j,
	359	event_scaled_avg + j, event_scaled[i]+j);
	360	else
	361	event_scaled_avg[j] = -1;
42202dd5 IM	362	}
	363	}
	364	runtime_nsecs_avg /= run_count;
	365	walltime_nsecs_avg /= run_count;
	366	runtime_cycles_avg /= run_count;
	367
	368	for (j = 0; j < nr_counters; j++) {
	369	event_res_avg[j][0] /= run_count;
	370	event_res_avg[j][1] /= run_count;
	371	event_res_avg[j][2] /= run_count;
	372	}
44db76c8	373
42202dd5 IM	374	for (i = 0; i < run_count; i++) {
42202dd5 IM	375	runtime_nsecs_noise +=
9cffa8d5	376	abs((s64)(runtime_nsecs[i] - runtime_nsecs_avg));
42202dd5	377	walltime_nsecs_noise +=
9cffa8d5	378	abs((s64)(walltime_nsecs[i] - walltime_nsecs_avg));
42202dd5	379	runtime_cycles_noise +=
9cffa8d5	380	abs((s64)(runtime_cycles[i] - runtime_cycles_avg));
42202dd5 IM	381
	382	for (j = 0; j < nr_counters; j++) {
	383	event_res_noise[j][0] +=
9cffa8d5	384	abs((s64)(event_res[i][j][0] - event_res_avg[j][0]));
42202dd5	385	event_res_noise[j][1] +=
9cffa8d5	386	abs((s64)(event_res[i][j][1] - event_res_avg[j][1]));
42202dd5	387	event_res_noise[j][2] +=
9cffa8d5	388	abs((s64)(event_res[i][j][2] - event_res_avg[j][2]));
ddcacfa0 IM	389	}
ddcacfa0 IM	390	}
44db76c8	391
ef281a19 IM	392	normalize_noise(&runtime_nsecs_noise);
	393	normalize_noise(&walltime_nsecs_noise);
	394	normalize_noise(&runtime_cycles_noise);
44db76c8	395
42202dd5	396	for (j = 0; j < nr_counters; j++) {
ef281a19 IM	397	normalize_noise(&event_res_noise[j][0]);
	398	normalize_noise(&event_res_noise[j][1]);
	399	normalize_noise(&event_res_noise[j][2]);
42202dd5 IM	400	}
	401	}
	402
	403	static void print_stat(int argc, const char **argv)
	404	{
	405	int i, counter;
	406
	407	calc_avg();
ddcacfa0 IM	408
	409	fflush(stdout);
	410
	411	fprintf(stderr, "\n");
44db76c8 IM	412	fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
	413
	414	for (i = 1; i < argc; i++)
	415	fprintf(stderr, " %s", argv[i]);
	416
42202dd5 IM	417	fprintf(stderr, "\'");
	418	if (run_count > 1)
	419	fprintf(stderr, " (%d runs)", run_count);
	420	fprintf(stderr, ":\n\n");
2996f5dd	421
c04f5e5d IM	422	for (counter = 0; counter < nr_counters; counter++)
c04f5e5d IM	423	print_counter(counter);
ddcacfa0	424
ddcacfa0	425	fprintf(stderr, "\n");
566747e6	426	fprintf(stderr, " %14.9f seconds time elapsed",
42202dd5	427	(double)walltime_nsecs_avg/1e9);
566747e6 IM	428	if (run_count > 1) {
	429	fprintf(stderr, " ( +- %7.3f%% )",
	430	100.0*(double)walltime_nsecs_noise/(double)walltime_nsecs_avg);
	431	}
	432	fprintf(stderr, "\n\n");
ddcacfa0 IM	433	}
ddcacfa0 IM	434
f7b7c26e PZ	435	static volatile int signr = -1;
f7b7c26e PZ	436
5242519b	437	static void skip_signal(int signo)
ddcacfa0	438	{
f7b7c26e PZ	439	signr = signo;
	440	}
	441
	442	static void sig_atexit(void)
	443	{
	444	if (signr == -1)
	445	return;
	446
	447	signal(signr, SIG_DFL);
	448	kill(getpid(), signr);
5242519b IM	449	}
	450
	451	static const char * const stat_usage[] = {
	452	"perf stat [<options>] <command>",
	453	NULL
	454	};
	455
5242519b IM	456	static const struct option options[] = {
5242519b IM	457	OPT_CALLBACK('e', "event", NULL, "event",
86847b62 TG	458	"event selector. use 'perf list' to list available events",
86847b62 TG	459	parse_events),
5242519b IM	460	OPT_BOOLEAN('i', "inherit", &inherit,
	461	"child tasks inherit counters"),
	462	OPT_INTEGER('p', "pid", &target_pid,
	463	"stat events on existing pid"),
	464	OPT_BOOLEAN('a', "all-cpus", &system_wide,
3d632595	465	"system-wide collection from all CPUs"),
86847b62	466	OPT_BOOLEAN('S', "scale", &scale,
3d632595	467	"scale/normalize counters"),
743ee1f8 IM	468	OPT_BOOLEAN('v', "verbose", &verbose,
743ee1f8 IM	469	"be more verbose (show counter open errors, etc)"),
42202dd5 IM	470	OPT_INTEGER('r', "repeat", &run_count,
42202dd5 IM	471	"repeat command and print average + stddev (max: 100)"),
0cfb7a13 IM	472	OPT_BOOLEAN('n', "null", &null_run,
0cfb7a13 IM	473	"null run - dont start any counters"),
5242519b IM	474	OPT_END()
	475	};
	476
	477	int cmd_stat(int argc, const char *argv, const char prefix)
	478	{
42202dd5 IM	479	int status;
42202dd5 IM	480
5242519b IM	481	argc = parse_options(argc, argv, options, stat_usage, 0);
	482	if (!argc)
	483	usage_with_options(stat_usage, options);
42202dd5 IM	484	if (run_count <= 0 \|\| run_count > MAX_RUN)
42202dd5 IM	485	usage_with_options(stat_usage, options);
ddcacfa0	486
c3043569 JSR	487	/* Set attrs and nr_counters if no event is selected and !null_run */
	488	if (!null_run && !nr_counters) {
	489	memcpy(attrs, default_attrs, sizeof(default_attrs));
	490	nr_counters = ARRAY_SIZE(default_attrs);
	491	}
ddcacfa0	492
ddcacfa0 IM	493	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
	494	assert(nr_cpus <= MAX_NR_CPUS);
	495	assert(nr_cpus >= 0);
	496
58d7e993 IM	497	/*
	498	* We dont want to block the signals - that would cause
	499	* child tasks to inherit that and Ctrl-C would not work.
	500	* What we want is for Ctrl-C to work in the exec()-ed
	501	* task, but being ignored by perf stat itself:
	502	*/
f7b7c26e	503	atexit(sig_atexit);
58d7e993 IM	504	signal(SIGINT, skip_signal);
	505	signal(SIGALRM, skip_signal);
	506	signal(SIGABRT, skip_signal);
	507
42202dd5 IM	508	status = 0;
	509	for (run_idx = 0; run_idx < run_count; run_idx++) {
	510	if (run_count != 1 && verbose)
3d632595	511	fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
42202dd5 IM	512	status = run_perf_stat(argc, argv);
	513	}
	514
	515	print_stat(argc, argv);
	516
	517	return status;
ddcacfa0	518	}