[GitHub/mt8127/android_kernel_alcatel_ttab.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 unsigned 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 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_counter_open() syscall returned with %d (%s)\n"

static void create_perf_stat_counter(int counter, int pid)
{
	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) {
		unsigned 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;
		attr->enable_on_exec = 1;

		fd[0][counter] = sys_perf_counter_open(attr, pid, -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 (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, single_count[3];
	unsigned int cpu;
	size_t res, 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 (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
		runtime_nsecs[run_idx] = count[0];
	if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
		runtime_cycles[run_idx] = count[0];
}

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

	if (!system_wide)
		nr_cpus = 1;

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

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

	if (!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);
	}

	/*
	 * 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++)
		create_perf_stat_counter(counter, pid);

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

	close(go_pipe[1]);
	wait(&status);

	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 (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
		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 &&
	    MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
		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] != (u64)-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('c', "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 __used)
{
	int status;

	argc = parse_options(argc, argv, options, stat_usage,
		PARSE_OPT_STOP_AT_NON_OPTION);
	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((int)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;
f37a291c	66	static unsigned 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;
0cfb7a13	73	static int null_run = 0;
ddcacfa0	74
3d632595	75	static int fd[MAX_NR_CPUS][MAX_COUNTERS];
42202dd5	76
9cffa8d5 PM	77	static u64 runtime_nsecs[MAX_RUN];
	78	static u64 walltime_nsecs[MAX_RUN];
	79	static u64 runtime_cycles[MAX_RUN];
42202dd5	80
3d632595 JSR	81	static u64 event_res[MAX_RUN][MAX_COUNTERS][3];
	82	static u64 event_scaled[MAX_RUN][MAX_COUNTERS];
	83
9cffa8d5 PM	84	static u64 event_res_avg[MAX_COUNTERS][3];
9cffa8d5 PM	85	static u64 event_res_noise[MAX_COUNTERS][3];
42202dd5	86
9cffa8d5	87	static u64 event_scaled_avg[MAX_COUNTERS];
42202dd5	88
9cffa8d5 PM	89	static u64 runtime_nsecs_avg;
9cffa8d5 PM	90	static u64 runtime_nsecs_noise;
42202dd5	91
9cffa8d5 PM	92	static u64 walltime_nsecs_avg;
9cffa8d5 PM	93	static u64 walltime_nsecs_noise;
42202dd5	94
9cffa8d5 PM	95	static u64 runtime_cycles_avg;
9cffa8d5 PM	96	static u64 runtime_cycles_noise;
be1ac0d8	97
b9ebdcc0 JSR	98	#define MATCH_EVENT(t, c, counter) \
	99	(attrs[counter].type == PERF_TYPE_##t && \
	100	attrs[counter].config == PERF_COUNT_##c)
	101
cca03c0a JSR	102	#define ERR_PERF_OPEN \
	103	"Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
	104
051ae7f7	105	static void create_perf_stat_counter(int counter, int pid)
ddcacfa0	106	{
a21ca2ca	107	struct perf_counter_attr *attr = attrs + counter;
16c8a109	108
ddcacfa0	109	if (scale)
a21ca2ca IM	110	attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED \|
a21ca2ca IM	111	PERF_FORMAT_TOTAL_TIME_RUNNING;
ddcacfa0 IM	112
ddcacfa0 IM	113	if (system_wide) {
f37a291c IM	114	unsigned int cpu;
f37a291c IM	115
cca03c0a	116	for (cpu = 0; cpu < nr_cpus; cpu++) {
a21ca2ca	117	fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
cca03c0a JSR	118	if (fd[cpu][counter] < 0 && verbose)
	119	fprintf(stderr, ERR_PERF_OPEN, counter,
	120	fd[cpu][counter], strerror(errno));
ddcacfa0 IM	121	}
ddcacfa0 IM	122	} else {
57e7986e PM	123	attr->inherit = inherit;
	124	attr->disabled = 1;
	125	attr->enable_on_exec = 1;
ddcacfa0	126
051ae7f7	127	fd[0][counter] = sys_perf_counter_open(attr, pid, -1, -1, 0);
cca03c0a JSR	128	if (fd[0][counter] < 0 && verbose)
	129	fprintf(stderr, ERR_PERF_OPEN, counter,
	130	fd[0][counter], strerror(errno));
ddcacfa0 IM	131	}
	132	}
	133
c04f5e5d IM	134	/*
	135	* Does the counter have nsecs as a unit?
	136	*/
	137	static inline int nsec_counter(int counter)
	138	{
b9ebdcc0 JSR	139	if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) \|\|
b9ebdcc0 JSR	140	MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
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];
f37a291c IM	152	unsigned int cpu;
f37a291c IM	153	size_t res, nv;
c04f5e5d IM	154	int scaled;
c04f5e5d IM	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));
f37a291c	167
42202dd5 IM	168	close(fd[cpu][counter]);
42202dd5 IM	169	fd[cpu][counter] = -1;
c04f5e5d IM	170
	171	count[0] += single_count[0];
	172	if (scale) {
	173	count[1] += single_count[1];
	174	count[2] += single_count[2];
	175	}
	176	}
	177
	178	scaled = 0;
	179	if (scale) {
	180	if (count[2] == 0) {
42202dd5	181	event_scaled[run_idx][counter] = -1;
2996f5dd	182	count[0] = 0;
c04f5e5d IM	183	return;
c04f5e5d IM	184	}
2996f5dd	185
c04f5e5d	186	if (count[2] < count[1]) {
42202dd5	187	event_scaled[run_idx][counter] = 1;
c04f5e5d IM	188	count[0] = (unsigned long long)
	189	((double)count[0] * count[1] / count[2] + 0.5);
	190	}
	191	}
be1ac0d8 IM	192	/*
	193	* Save the full runtime - to allow normalization during printout:
	194	*/
b9ebdcc0	195	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
42202dd5	196	runtime_nsecs[run_idx] = count[0];
b9ebdcc0	197	if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
42202dd5	198	runtime_cycles[run_idx] = count[0];
2996f5dd IM	199	}
2996f5dd IM	200
f37a291c	201	static int run_perf_stat(int argc __used, const char **argv)
42202dd5 IM	202	{
	203	unsigned long long t0, t1;
	204	int status = 0;
	205	int counter;
	206	int pid;
051ae7f7 PM	207	int child_ready_pipe[2], go_pipe[2];
051ae7f7 PM	208	char buf;
42202dd5 IM	209
	210	if (!system_wide)
	211	nr_cpus = 1;
	212
051ae7f7 PM	213	if (pipe(child_ready_pipe) < 0 \|\| pipe(go_pipe) < 0) {
	214	perror("failed to create pipes");
	215	exit(1);
	216	}
	217
	218	if ((pid = fork()) < 0)
	219	perror("failed to fork");
	220
	221	if (!pid) {
	222	close(child_ready_pipe[0]);
	223	close(go_pipe[1]);
	224	fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
	225
	226	/*
	227	* Do a dummy execvp to get the PLT entry resolved,
	228	* so we avoid the resolver overhead on the real
	229	* execvp call.
	230	*/
	231	execvp("", (char **)argv);
	232
	233	/*
	234	* Tell the parent we're ready to go
	235	*/
	236	close(child_ready_pipe[1]);
	237
	238	/*
	239	* Wait until the parent tells us to go.
	240	*/
a92bef0f FW	241	if (read(go_pipe[0], &buf, 1) == -1)
a92bef0f FW	242	perror("unable to read pipe");
051ae7f7 PM	243
	244	execvp(argv[0], (char **)argv);
	245
	246	perror(argv[0]);
	247	exit(-1);
	248	}
	249
	250	/*
	251	* Wait for the child to be ready to exec.
	252	*/
	253	close(child_ready_pipe[1]);
	254	close(go_pipe[0]);
a92bef0f FW	255	if (read(child_ready_pipe[0], &buf, 1) == -1)
a92bef0f FW	256	perror("unable to read pipe");
051ae7f7 PM	257	close(child_ready_pipe[0]);
051ae7f7 PM	258
42202dd5	259	for (counter = 0; counter < nr_counters; counter++)
051ae7f7	260	create_perf_stat_counter(counter, pid);
42202dd5 IM	261
	262	/*
	263	* Enable counters and exec the command:
	264	*/
	265	t0 = rdclock();
42202dd5	266
051ae7f7	267	close(go_pipe[1]);
42202dd5 IM	268	wait(&status);
42202dd5 IM	269
42202dd5 IM	270	t1 = rdclock();
	271
	272	walltime_nsecs[run_idx] = t1 - t0;
	273
	274	for (counter = 0; counter < nr_counters; counter++)
	275	read_counter(counter);
	276
	277	return WEXITSTATUS(status);
	278	}
	279
9cffa8d5	280	static void print_noise(u64 count, u64 noise)
42202dd5 IM	281	{
	282	if (run_count > 1)
	283	fprintf(stderr, " ( +- %7.3f%% )",
	284	(double)noise[0]/(count[0]+1)*100.0);
	285	}
	286
9cffa8d5	287	static void nsec_printout(int counter, u64 count, u64 noise)
44175b6f IM	288	{
	289	double msecs = (double)count[0] / 1000000;
	290
6e750a8f	291	fprintf(stderr, " %14.6f %-24s", msecs, event_name(counter));
44175b6f	292
b9ebdcc0	293	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
42202dd5 IM	294	if (walltime_nsecs_avg)
	295	fprintf(stderr, " # %10.3f CPUs ",
	296	(double)count[0] / (double)walltime_nsecs_avg);
44175b6f	297	}
42202dd5	298	print_noise(count, noise);
44175b6f IM	299	}
44175b6f IM	300
9cffa8d5	301	static void abs_printout(int counter, u64 count, u64 noise)
44175b6f	302	{
6e750a8f	303	fprintf(stderr, " %14Ld %-24s", count[0], event_name(counter));
44175b6f	304
42202dd5	305	if (runtime_cycles_avg &&
b9ebdcc0	306	MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
42202dd5 IM	307	fprintf(stderr, " # %10.3f IPC ",
	308	(double)count[0] / (double)runtime_cycles_avg);
	309	} else {
	310	if (runtime_nsecs_avg) {
	311	fprintf(stderr, " # %10.3f M/sec",
	312	(double)count[0]/runtime_nsecs_avg*1000.0);
	313	}
44175b6f	314	}
42202dd5	315	print_noise(count, noise);
44175b6f IM	316	}
44175b6f IM	317
2996f5dd IM	318	/*
	319	* Print out the results of a single counter:
	320	*/
	321	static void print_counter(int counter)
	322	{
9cffa8d5	323	u64 count, noise;
2996f5dd IM	324	int scaled;
2996f5dd IM	325
42202dd5 IM	326	count = event_res_avg[counter];
	327	noise = event_res_noise[counter];
	328	scaled = event_scaled_avg[counter];
2996f5dd IM	329
2996f5dd IM	330	if (scaled == -1) {
6e750a8f	331	fprintf(stderr, " %14s %-24s\n",
2996f5dd IM	332	"<not counted>", event_name(counter));
	333	return;
	334	}
c04f5e5d	335
44175b6f	336	if (nsec_counter(counter))
42202dd5	337	nsec_printout(counter, count, noise);
44175b6f	338	else
42202dd5	339	abs_printout(counter, count, noise);
d7c29318	340
c04f5e5d	341	if (scaled)
210ad39f IM	342	fprintf(stderr, " (scaled from %.2f%%)",
210ad39f IM	343	(double) count[2] / count[1] * 100);
44175b6f	344
c04f5e5d IM	345	fprintf(stderr, "\n");
	346	}
	347
42202dd5	348	/*
ef281a19	349	* normalize_noise noise values down to stddev:
42202dd5	350	*/
9cffa8d5	351	static void normalize_noise(u64 *val)
ddcacfa0	352	{
42202dd5	353	double res;
ddcacfa0	354
42202dd5	355	res = (double)val / (run_count sqrt((double)run_count));
ddcacfa0	356
9cffa8d5	357	*val = (u64)res;
42202dd5	358	}
ddcacfa0	359
9cffa8d5	360	static void update_avg(const char name, int idx, u64 avg, u64 *val)
ef281a19 IM	361	{
	362	avg += val;
	363
	364	if (verbose > 1)
	365	fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val);
	366	}
42202dd5 IM	367	/*
	368	* Calculate the averages and noises:
	369	*/
	370	static void calc_avg(void)
	371	{
	372	int i, j;
	373
ef281a19 IM	374	if (verbose > 1)
	375	fprintf(stderr, "\n");
	376
42202dd5	377	for (i = 0; i < run_count; i++) {
ef281a19 IM	378	update_avg("runtime", 0, &runtime_nsecs_avg, runtime_nsecs + i);
	379	update_avg("walltime", 0, &walltime_nsecs_avg, walltime_nsecs + i);
	380	update_avg("runtime_cycles", 0, &runtime_cycles_avg, runtime_cycles + i);
42202dd5 IM	381
42202dd5 IM	382	for (j = 0; j < nr_counters; j++) {
ef281a19 IM	383	update_avg("counter/0", j,
	384	event_res_avg[j]+0, event_res[i][j]+0);
	385	update_avg("counter/1", j,
	386	event_res_avg[j]+1, event_res[i][j]+1);
	387	update_avg("counter/2", j,
	388	event_res_avg[j]+2, event_res[i][j]+2);
f37a291c	389	if (event_scaled[i][j] != (u64)-1)
566747e6 IM	390	update_avg("scaled", j,
	391	event_scaled_avg + j, event_scaled[i]+j);
	392	else
	393	event_scaled_avg[j] = -1;
42202dd5 IM	394	}
	395	}
	396	runtime_nsecs_avg /= run_count;
	397	walltime_nsecs_avg /= run_count;
	398	runtime_cycles_avg /= run_count;
	399
	400	for (j = 0; j < nr_counters; j++) {
	401	event_res_avg[j][0] /= run_count;
	402	event_res_avg[j][1] /= run_count;
	403	event_res_avg[j][2] /= run_count;
	404	}
44db76c8	405
42202dd5 IM	406	for (i = 0; i < run_count; i++) {
42202dd5 IM	407	runtime_nsecs_noise +=
9cffa8d5	408	abs((s64)(runtime_nsecs[i] - runtime_nsecs_avg));
42202dd5	409	walltime_nsecs_noise +=
9cffa8d5	410	abs((s64)(walltime_nsecs[i] - walltime_nsecs_avg));
42202dd5	411	runtime_cycles_noise +=
9cffa8d5	412	abs((s64)(runtime_cycles[i] - runtime_cycles_avg));
42202dd5 IM	413
	414	for (j = 0; j < nr_counters; j++) {
	415	event_res_noise[j][0] +=
9cffa8d5	416	abs((s64)(event_res[i][j][0] - event_res_avg[j][0]));
42202dd5	417	event_res_noise[j][1] +=
9cffa8d5	418	abs((s64)(event_res[i][j][1] - event_res_avg[j][1]));
42202dd5	419	event_res_noise[j][2] +=
9cffa8d5	420	abs((s64)(event_res[i][j][2] - event_res_avg[j][2]));
ddcacfa0 IM	421	}
ddcacfa0 IM	422	}
44db76c8	423
ef281a19 IM	424	normalize_noise(&runtime_nsecs_noise);
	425	normalize_noise(&walltime_nsecs_noise);
	426	normalize_noise(&runtime_cycles_noise);
44db76c8	427
42202dd5	428	for (j = 0; j < nr_counters; j++) {
ef281a19 IM	429	normalize_noise(&event_res_noise[j][0]);
	430	normalize_noise(&event_res_noise[j][1]);
	431	normalize_noise(&event_res_noise[j][2]);
42202dd5 IM	432	}
	433	}
	434
	435	static void print_stat(int argc, const char **argv)
	436	{
	437	int i, counter;
	438
	439	calc_avg();
ddcacfa0 IM	440
	441	fflush(stdout);
	442
	443	fprintf(stderr, "\n");
44db76c8 IM	444	fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
	445
	446	for (i = 1; i < argc; i++)
	447	fprintf(stderr, " %s", argv[i]);
	448
42202dd5 IM	449	fprintf(stderr, "\'");
	450	if (run_count > 1)
	451	fprintf(stderr, " (%d runs)", run_count);
	452	fprintf(stderr, ":\n\n");
2996f5dd	453
c04f5e5d IM	454	for (counter = 0; counter < nr_counters; counter++)
c04f5e5d IM	455	print_counter(counter);
ddcacfa0	456
ddcacfa0	457	fprintf(stderr, "\n");
566747e6	458	fprintf(stderr, " %14.9f seconds time elapsed",
42202dd5	459	(double)walltime_nsecs_avg/1e9);
566747e6 IM	460	if (run_count > 1) {
	461	fprintf(stderr, " ( +- %7.3f%% )",
	462	100.0*(double)walltime_nsecs_noise/(double)walltime_nsecs_avg);
	463	}
	464	fprintf(stderr, "\n\n");
ddcacfa0 IM	465	}
ddcacfa0 IM	466
f7b7c26e PZ	467	static volatile int signr = -1;
f7b7c26e PZ	468
5242519b	469	static void skip_signal(int signo)
ddcacfa0	470	{
f7b7c26e PZ	471	signr = signo;
	472	}
	473
	474	static void sig_atexit(void)
	475	{
	476	if (signr == -1)
	477	return;
	478
	479	signal(signr, SIG_DFL);
	480	kill(getpid(), signr);
5242519b IM	481	}
	482
	483	static const char * const stat_usage[] = {
	484	"perf stat [<options>] <command>",
	485	NULL
	486	};
	487
5242519b IM	488	static const struct option options[] = {
5242519b IM	489	OPT_CALLBACK('e', "event", NULL, "event",
86847b62 TG	490	"event selector. use 'perf list' to list available events",
86847b62 TG	491	parse_events),
5242519b IM	492	OPT_BOOLEAN('i', "inherit", &inherit,
	493	"child tasks inherit counters"),
	494	OPT_INTEGER('p', "pid", &target_pid,
	495	"stat events on existing pid"),
	496	OPT_BOOLEAN('a', "all-cpus", &system_wide,
3d632595	497	"system-wide collection from all CPUs"),
b26bc5a7	498	OPT_BOOLEAN('c', "scale", &scale,
3d632595	499	"scale/normalize counters"),
743ee1f8 IM	500	OPT_BOOLEAN('v', "verbose", &verbose,
743ee1f8 IM	501	"be more verbose (show counter open errors, etc)"),
42202dd5 IM	502	OPT_INTEGER('r', "repeat", &run_count,
42202dd5 IM	503	"repeat command and print average + stddev (max: 100)"),
0cfb7a13 IM	504	OPT_BOOLEAN('n', "null", &null_run,
0cfb7a13 IM	505	"null run - dont start any counters"),
5242519b IM	506	OPT_END()
	507	};
	508
f37a291c	509	int cmd_stat(int argc, const char *argv, const char prefix __used)
5242519b	510	{
42202dd5 IM	511	int status;
42202dd5 IM	512
a0541234 AB	513	argc = parse_options(argc, argv, options, stat_usage,
a0541234 AB	514	PARSE_OPT_STOP_AT_NON_OPTION);
5242519b IM	515	if (!argc)
5242519b IM	516	usage_with_options(stat_usage, options);
42202dd5 IM	517	if (run_count <= 0 \|\| run_count > MAX_RUN)
42202dd5 IM	518	usage_with_options(stat_usage, options);
ddcacfa0	519
c3043569 JSR	520	/* Set attrs and nr_counters if no event is selected and !null_run */
	521	if (!null_run && !nr_counters) {
	522	memcpy(attrs, default_attrs, sizeof(default_attrs));
	523	nr_counters = ARRAY_SIZE(default_attrs);
	524	}
ddcacfa0	525
ddcacfa0 IM	526	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
ddcacfa0 IM	527	assert(nr_cpus <= MAX_NR_CPUS);
f37a291c	528	assert((int)nr_cpus >= 0);
ddcacfa0	529
58d7e993 IM	530	/*
	531	* We dont want to block the signals - that would cause
	532	* child tasks to inherit that and Ctrl-C would not work.
	533	* What we want is for Ctrl-C to work in the exec()-ed
	534	* task, but being ignored by perf stat itself:
	535	*/
f7b7c26e	536	atexit(sig_atexit);
58d7e993 IM	537	signal(SIGINT, skip_signal);
	538	signal(SIGALRM, skip_signal);
	539	signal(SIGABRT, skip_signal);
	540
42202dd5 IM	541	status = 0;
	542	for (run_idx = 0; run_idx < run_count; run_idx++) {
	543	if (run_count != 1 && verbose)
3d632595	544	fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
42202dd5 IM	545	status = run_perf_stat(argc, argv);
	546	}
	547
	548	print_stat(argc, argv);
	549
	550	return status;
ddcacfa0	551	}