[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 "util/event.h"
#include "util/debug.h"
#include "util/header.h"
#include "util/cpumap.h"
#include "util/thread.h"

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

static struct perf_event_attr default_attrs[] = {

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

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

};

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

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

static int			*fd[MAX_NR_CPUS][MAX_COUNTERS];

static int			event_scaled[MAX_COUNTERS];

static volatile int done = 0;

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

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

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

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

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

	return sqrt(variance_mean);
}

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

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

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

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

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

	if (system_wide) {
		unsigned int cpu;

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

	return ncreated;
}

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

	return 0;
}

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

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

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

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

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

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

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

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

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

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

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

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

	if (!system_wide)
		nr_cpus = 1;

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

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

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

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

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

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

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

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

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

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

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

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

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

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

	t1 = rdclock();

	update_stats(&walltime_nsecs_stats, t1 - t0);

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

	return WEXITSTATUS(status);
}

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

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

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

	fprintf(stderr, " %14.6f  %-24s", msecs, event_name(counter));

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

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

	fprintf(stderr, " %14.0f  %-24s", avg, event_name(counter));

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

		if (total)
			ratio = avg / total;

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

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

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

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

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

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

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

	if (scaled == -1) {
		fprintf(stderr, " %14s  %-24s\n",
			"<not counted>", event_name(counter));
		return;
	}

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

	print_noise(counter, avg);

	if (scaled) {
		double avg_enabled, avg_running;

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

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

	fprintf(stderr, "\n");
}

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

	fflush(stdout);

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

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

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

	fprintf(stderr, "\n");
	fprintf(stderr, " %14.9f  seconds time elapsed",
			avg_stats(&walltime_nsecs_stats)/1e9);
	if (run_count > 1) {
		fprintf(stderr, "   ( +- %7.3f%% )",
				100*stddev_stats(&walltime_nsecs_stats) /
				avg_stats(&walltime_nsecs_stats));
	}
	fprintf(stderr, "\n\n");
}

static volatile int signr = -1;

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

	signr = signo;
}

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

	if (signr == -1)
		return;

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

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

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

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

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

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

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

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

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

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

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

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

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

	return status;
}
Commit	Line	Data
ddcacfa0	1	/*
bf9e1876 IM	2	* builtin-stat.c
	3	*
	4	* Builtin stat command: Give a precise performance counters summary
	5	* overview about any workload, CPU or specific PID.
	6	*
	7	* Sample output:
ddcacfa0	8
bf9e1876 IM	9	$ perf stat ~/hackbench 10
bf9e1876 IM	10	Time: 0.104
ddcacfa0	11
bf9e1876	12	Performance counter stats for '/home/mingo/hackbench':
ddcacfa0	13
bf9e1876 IM	14	1255.538611 task clock ticks # 10.143 CPU utilization factor
	15	54011 context switches # 0.043 M/sec
	16	385 CPU migrations # 0.000 M/sec
	17	17755 pagefaults # 0.014 M/sec
	18	3808323185 CPU cycles # 3033.219 M/sec
	19	1575111190 instructions # 1254.530 M/sec
	20	17367895 cache references # 13.833 M/sec
	21	7674421 cache misses # 6.112 M/sec
ddcacfa0	22
bf9e1876	23	Wall-clock time elapsed: 123.786620 msecs
ddcacfa0	24
5242519b IM	25	*
	26	* Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
	27	*
	28	* Improvements and fixes by:
	29	*
	30	* Arjan van de Ven <arjan@linux.intel.com>
	31	* Yanmin Zhang <yanmin.zhang@intel.com>
	32	* Wu Fengguang <fengguang.wu@intel.com>
	33	* Mike Galbraith <efault@gmx.de>
	34	* Paul Mackerras <paulus@samba.org>
6e750a8f	35	* Jaswinder Singh Rajput <jaswinder@kernel.org>
5242519b IM	36	*
5242519b IM	37	* Released under the GPL v2. (and only v2, not any later version)
ddcacfa0 IM	38	*/
ddcacfa0 IM	39
1a482f38	40	#include "perf.h"
16f762a2	41	#include "builtin.h"
148be2c1	42	#include "util/util.h"
5242519b IM	43	#include "util/parse-options.h"
5242519b IM	44	#include "util/parse-events.h"
8f28827a FW	45	#include "util/event.h"
8f28827a FW	46	#include "util/debug.h"
60666c63	47	#include "util/header.h"
a12b51c4	48	#include "util/cpumap.h"
d6d901c2	49	#include "util/thread.h"
ddcacfa0	50
ddcacfa0	51	#include <sys/prctl.h>
42202dd5	52	#include <math.h>
16c8a109	53
cdd6c482	54	static struct perf_event_attr default_attrs[] = {
ddcacfa0	55
56aab464 IM	56	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
	57	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
	58	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
	59	{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
f4dbfa8f	60
56aab464 IM	61	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
56aab464 IM	62	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
56aab464 IM	63	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
56aab464 IM	64	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
dd86e72a IM	65	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES },
dd86e72a IM	66	{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES },
f4dbfa8f	67
ddcacfa0	68	};
5242519b	69
c0555642	70	static bool system_wide = false;
f37a291c	71	static unsigned int nr_cpus = 0;
3d632595	72	static int run_idx = 0;
ddcacfa0	73
3d632595	74	static int run_count = 1;
2e6cdf99	75	static bool no_inherit = false;
c0555642	76	static bool scale = true;
933da83a	77	static pid_t target_pid = -1;
d6d901c2 ZY	78	static pid_t target_tid = -1;
	79	static pid_t *all_tids = NULL;
	80	static int thread_num = 0;
933da83a	81	static pid_t child_pid = -1;
c0555642	82	static bool null_run = false;
ddcacfa0	83
d6d901c2	84	static int *fd[MAX_NR_CPUS][MAX_COUNTERS];
42202dd5	85
63d40deb	86	static int event_scaled[MAX_COUNTERS];
3d632595	87
60666c63 LW	88	static volatile int done = 0;
60666c63 LW	89
506d4bc8 PZ	90	struct stats
506d4bc8 PZ	91	{
8a02631a	92	double n, mean, M2;
506d4bc8	93	};
42202dd5	94
9e9772c4 PZ	95	static void update_stats(struct stats *stats, u64 val)
9e9772c4 PZ	96	{
8a02631a	97	double delta;
9e9772c4	98
8a02631a PZ	99	stats->n++;
	100	delta = val - stats->mean;
	101	stats->mean += delta / stats->n;
	102	stats->M2 += delta*(val - stats->mean);
9e9772c4 PZ	103	}
9e9772c4 PZ	104
506d4bc8 PZ	105	static double avg_stats(struct stats *stats)
506d4bc8 PZ	106	{
8a02631a	107	return stats->mean;
506d4bc8	108	}
42202dd5	109
506d4bc8	110	/*
63d40deb PZ	111	* http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
63d40deb PZ	112	*
8a02631a PZ	113	* (\Sum n_i^2) - ((\Sum n_i)^2)/n
	114	* s^2 = -------------------------------
	115	* n - 1
63d40deb PZ	116	*
	117	* http://en.wikipedia.org/wiki/Stddev
	118	*
	119	* The std dev of the mean is related to the std dev by:
	120	*
	121	* s
	122	* s_mean = -------
	123	* sqrt(n)
	124	*
506d4bc8 PZ	125	*/
	126	static double stddev_stats(struct stats *stats)
	127	{
8a02631a PZ	128	double variance = stats->M2 / (stats->n - 1);
8a02631a PZ	129	double variance_mean = variance / stats->n;
42202dd5	130
63d40deb	131	return sqrt(variance_mean);
506d4bc8	132	}
42202dd5	133
506d4bc8	134	struct stats event_res_stats[MAX_COUNTERS][3];
506d4bc8 PZ	135	struct stats runtime_nsecs_stats;
	136	struct stats walltime_nsecs_stats;
	137	struct stats runtime_cycles_stats;
11018201	138	struct stats runtime_branches_stats;
be1ac0d8	139
b9ebdcc0 JSR	140	#define MATCH_EVENT(t, c, counter) \
	141	(attrs[counter].type == PERF_TYPE_##t && \
	142	attrs[counter].config == PERF_COUNT_##c)
	143
cca03c0a	144	#define ERR_PERF_OPEN \
cdd6c482	145	"Error: counter %d, sys_perf_event_open() syscall returned with %d (%s)\n"
cca03c0a	146
084ab9f8	147	static int create_perf_stat_counter(int counter)
ddcacfa0	148	{
cdd6c482	149	struct perf_event_attr *attr = attrs + counter;
d6d901c2	150	int thread;
084ab9f8	151	int ncreated = 0;
16c8a109	152
ddcacfa0	153	if (scale)
a21ca2ca IM	154	attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED \|
a21ca2ca IM	155	PERF_FORMAT_TOTAL_TIME_RUNNING;
ddcacfa0 IM	156
ddcacfa0 IM	157	if (system_wide) {
f37a291c IM	158	unsigned int cpu;
f37a291c IM	159
cca03c0a	160	for (cpu = 0; cpu < nr_cpus; cpu++) {
d6d901c2 ZY	161	fd[cpu][counter][0] = sys_perf_event_open(attr,
d6d901c2 ZY	162	-1, cpumap[cpu], -1, 0);
084ab9f8 ACM	163	if (fd[cpu][counter][0] < 0)
	164	pr_debug(ERR_PERF_OPEN, counter,
	165	fd[cpu][counter][0], strerror(errno));
	166	else
	167	++ncreated;
ddcacfa0 IM	168	}
ddcacfa0 IM	169	} else {
2e6cdf99 SE	170	attr->inherit = !no_inherit;
2e6cdf99 SE	171	if (target_pid == -1 && target_tid == -1) {
6be2850e ZY	172	attr->disabled = 1;
	173	attr->enable_on_exec = 1;
	174	}
d6d901c2 ZY	175	for (thread = 0; thread < thread_num; thread++) {
	176	fd[0][counter][thread] = sys_perf_event_open(attr,
	177	all_tids[thread], -1, -1, 0);
084ab9f8 ACM	178	if (fd[0][counter][thread] < 0)
	179	pr_debug(ERR_PERF_OPEN, counter,
	180	fd[0][counter][thread],
	181	strerror(errno));
	182	else
	183	++ncreated;
d6d901c2	184	}
ddcacfa0	185	}
084ab9f8 ACM	186
084ab9f8 ACM	187	return ncreated;
ddcacfa0 IM	188	}
ddcacfa0 IM	189
c04f5e5d IM	190	/*
	191	* Does the counter have nsecs as a unit?
	192	*/
	193	static inline int nsec_counter(int counter)
	194	{
b9ebdcc0 JSR	195	if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) \|\|
b9ebdcc0 JSR	196	MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
c04f5e5d IM	197	return 1;
	198
	199	return 0;
	200	}
	201
	202	/*
2996f5dd	203	* Read out the results of a single counter:
c04f5e5d	204	*/
2996f5dd	205	static void read_counter(int counter)
c04f5e5d	206	{
849abde9	207	u64 count[3], single_count[3];
f37a291c IM	208	unsigned int cpu;
f37a291c IM	209	size_t res, nv;
c04f5e5d	210	int scaled;
d6d901c2	211	int i, thread;
c04f5e5d IM	212
c04f5e5d IM	213	count[0] = count[1] = count[2] = 0;
2996f5dd	214
c04f5e5d	215	nv = scale ? 3 : 1;
cca03c0a	216	for (cpu = 0; cpu < nr_cpus; cpu++) {
d6d901c2 ZY	217	for (thread = 0; thread < thread_num; thread++) {
	218	if (fd[cpu][counter][thread] < 0)
	219	continue;
	220
	221	res = read(fd[cpu][counter][thread],
	222	single_count, nv * sizeof(u64));
	223	assert(res == nv * sizeof(u64));
	224
	225	close(fd[cpu][counter][thread]);
	226	fd[cpu][counter][thread] = -1;
	227
	228	count[0] += single_count[0];
	229	if (scale) {
	230	count[1] += single_count[1];
	231	count[2] += single_count[2];
	232	}
c04f5e5d IM	233	}
	234	}
	235
	236	scaled = 0;
	237	if (scale) {
	238	if (count[2] == 0) {
9e9772c4	239	event_scaled[counter] = -1;
2996f5dd	240	count[0] = 0;
c04f5e5d IM	241	return;
c04f5e5d IM	242	}
2996f5dd	243
c04f5e5d	244	if (count[2] < count[1]) {
9e9772c4	245	event_scaled[counter] = 1;
c04f5e5d IM	246	count[0] = (unsigned long long)
	247	((double)count[0] * count[1] / count[2] + 0.5);
	248	}
	249	}
9e9772c4 PZ	250
	251	for (i = 0; i < 3; i++)
	252	update_stats(&event_res_stats[counter][i], count[i]);
	253
	254	if (verbose) {
	255	fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
	256	count[0], count[1], count[2]);
	257	}
	258
be1ac0d8 IM	259	/*
	260	* Save the full runtime - to allow normalization during printout:
	261	*/
b9ebdcc0	262	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
9e9772c4	263	update_stats(&runtime_nsecs_stats, count[0]);
b9ebdcc0	264	if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
9e9772c4	265	update_stats(&runtime_cycles_stats, count[0]);
11018201 AB	266	if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter))
11018201 AB	267	update_stats(&runtime_branches_stats, count[0]);
2996f5dd IM	268	}
2996f5dd IM	269
f37a291c	270	static int run_perf_stat(int argc __used, const char **argv)
42202dd5 IM	271	{
	272	unsigned long long t0, t1;
	273	int status = 0;
084ab9f8	274	int counter, ncreated = 0;
051ae7f7	275	int child_ready_pipe[2], go_pipe[2];
6be2850e	276	const bool forks = (argc > 0);
051ae7f7	277	char buf;
42202dd5 IM	278
	279	if (!system_wide)
	280	nr_cpus = 1;
	281
60666c63	282	if (forks && (pipe(child_ready_pipe) < 0 \|\| pipe(go_pipe) < 0)) {
051ae7f7 PM	283	perror("failed to create pipes");
	284	exit(1);
	285	}
	286
60666c63	287	if (forks) {
6be2850e	288	if ((child_pid = fork()) < 0)
60666c63 LW	289	perror("failed to fork");
60666c63 LW	290
6be2850e	291	if (!child_pid) {
60666c63 LW	292	close(child_ready_pipe[0]);
	293	close(go_pipe[1]);
	294	fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
	295
	296	/*
	297	* Do a dummy execvp to get the PLT entry resolved,
	298	* so we avoid the resolver overhead on the real
	299	* execvp call.
	300	*/
	301	execvp("", (char **)argv);
	302
	303	/*
	304	* Tell the parent we're ready to go
	305	*/
	306	close(child_ready_pipe[1]);
	307
	308	/*
	309	* Wait until the parent tells us to go.
	310	*/
	311	if (read(go_pipe[0], &buf, 1) == -1)
	312	perror("unable to read pipe");
	313
	314	execvp(argv[0], (char **)argv);
	315
	316	perror(argv[0]);
	317	exit(-1);
	318	}
051ae7f7	319
d6d901c2 ZY	320	if (target_tid == -1 && target_pid == -1 && !system_wide)
	321	all_tids[0] = child_pid;
	322
051ae7f7	323	/*
60666c63	324	* Wait for the child to be ready to exec.
051ae7f7 PM	325	*/
051ae7f7 PM	326	close(child_ready_pipe[1]);
60666c63 LW	327	close(go_pipe[0]);
60666c63 LW	328	if (read(child_ready_pipe[0], &buf, 1) == -1)
a92bef0f	329	perror("unable to read pipe");
60666c63	330	close(child_ready_pipe[0]);
051ae7f7 PM	331	}
051ae7f7 PM	332
42202dd5	333	for (counter = 0; counter < nr_counters; counter++)
084ab9f8 ACM	334	ncreated += create_perf_stat_counter(counter);
	335
	336	if (ncreated == 0) {
	337	pr_err("No permission to collect %sstats.\n"
	338	"Consider tweaking /proc/sys/kernel/perf_event_paranoid.\n",
	339	system_wide ? "system-wide " : "");
	340	if (child_pid != -1)
	341	kill(child_pid, SIGTERM);
	342	return -1;
	343	}
42202dd5 IM	344
	345	/*
	346	* Enable counters and exec the command:
	347	*/
	348	t0 = rdclock();
42202dd5	349
60666c63 LW	350	if (forks) {
	351	close(go_pipe[1]);
	352	wait(&status);
	353	} else {
6be2850e	354	while(!done) sleep(1);
60666c63	355	}
42202dd5	356
42202dd5 IM	357	t1 = rdclock();
42202dd5 IM	358
9e9772c4	359	update_stats(&walltime_nsecs_stats, t1 - t0);
42202dd5 IM	360
	361	for (counter = 0; counter < nr_counters; counter++)
	362	read_counter(counter);
	363
	364	return WEXITSTATUS(status);
	365	}
	366
849abde9	367	static void print_noise(int counter, double avg)
42202dd5	368	{
849abde9 PZ	369	if (run_count == 1)
	370	return;
	371
	372	fprintf(stderr, " ( +- %7.3f%% )",
	373	100 * stddev_stats(&event_res_stats[counter][0]) / avg);
42202dd5 IM	374	}
42202dd5 IM	375
849abde9	376	static void nsec_printout(int counter, double avg)
44175b6f	377	{
506d4bc8	378	double msecs = avg / 1e6;
44175b6f	379
6e750a8f	380	fprintf(stderr, " %14.6f %-24s", msecs, event_name(counter));
44175b6f	381
b9ebdcc0	382	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
506d4bc8 PZ	383	fprintf(stderr, " # %10.3f CPUs ",
506d4bc8 PZ	384	avg / avg_stats(&walltime_nsecs_stats));
44175b6f IM	385	}
	386	}
	387
849abde9	388	static void abs_printout(int counter, double avg)
44175b6f	389	{
c7f7fea3 IM	390	double total, ratio = 0.0;
c7f7fea3 IM	391
506d4bc8	392	fprintf(stderr, " %14.0f %-24s", avg, event_name(counter));
44175b6f	393
506d4bc8	394	if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
c7f7fea3 IM	395	total = avg_stats(&runtime_cycles_stats);
	396
	397	if (total)
	398	ratio = avg / total;
	399
	400	fprintf(stderr, " # %10.3f IPC ", ratio);
7255fe2a LDM	401	} else if (MATCH_EVENT(HARDWARE, HW_BRANCH_MISSES, counter) &&
7255fe2a LDM	402	runtime_branches_stats.n != 0) {
11018201 AB	403	total = avg_stats(&runtime_branches_stats);
	404
	405	if (total)
	406	ratio = avg * 100 / total;
	407
dd86e72a	408	fprintf(stderr, " # %10.3f %% ", ratio);
11018201	409
7255fe2a	410	} else if (runtime_nsecs_stats.n != 0) {
c7f7fea3 IM	411	total = avg_stats(&runtime_nsecs_stats);
	412
	413	if (total)
	414	ratio = 1000.0 * avg / total;
	415
	416	fprintf(stderr, " # %10.3f M/sec", ratio);
44175b6f	417	}
44175b6f IM	418	}
44175b6f IM	419
2996f5dd IM	420	/*
	421	* Print out the results of a single counter:
	422	*/
	423	static void print_counter(int counter)
	424	{
849abde9	425	double avg = avg_stats(&event_res_stats[counter][0]);
63d40deb	426	int scaled = event_scaled[counter];
2996f5dd	427
2996f5dd	428	if (scaled == -1) {
6e750a8f	429	fprintf(stderr, " %14s %-24s\n",
2996f5dd IM	430	"<not counted>", event_name(counter));
	431	return;
	432	}
c04f5e5d	433
44175b6f	434	if (nsec_counter(counter))
849abde9	435	nsec_printout(counter, avg);
44175b6f	436	else
849abde9 PZ	437	abs_printout(counter, avg);
	438
	439	print_noise(counter, avg);
506d4bc8 PZ	440
	441	if (scaled) {
	442	double avg_enabled, avg_running;
	443
	444	avg_enabled = avg_stats(&event_res_stats[counter][1]);
	445	avg_running = avg_stats(&event_res_stats[counter][2]);
d7c29318	446
210ad39f	447	fprintf(stderr, " (scaled from %.2f%%)",
506d4bc8 PZ	448	100 * avg_running / avg_enabled);
506d4bc8 PZ	449	}
44175b6f	450
c04f5e5d IM	451	fprintf(stderr, "\n");
	452	}
	453
42202dd5 IM	454	static void print_stat(int argc, const char **argv)
	455	{
	456	int i, counter;
	457
ddcacfa0 IM	458	fflush(stdout);
	459
	460	fprintf(stderr, "\n");
60666c63	461	fprintf(stderr, " Performance counter stats for ");
d6d901c2	462	if(target_pid == -1 && target_tid == -1) {
60666c63 LW	463	fprintf(stderr, "\'%s", argv[0]);
	464	for (i = 1; i < argc; i++)
	465	fprintf(stderr, " %s", argv[i]);
d6d901c2 ZY	466	} else if (target_pid != -1)
	467	fprintf(stderr, "process id \'%d", target_pid);
	468	else
	469	fprintf(stderr, "thread id \'%d", target_tid);
44db76c8	470
42202dd5 IM	471	fprintf(stderr, "\'");
	472	if (run_count > 1)
	473	fprintf(stderr, " (%d runs)", run_count);
	474	fprintf(stderr, ":\n\n");
2996f5dd	475
c04f5e5d IM	476	for (counter = 0; counter < nr_counters; counter++)
c04f5e5d IM	477	print_counter(counter);
ddcacfa0	478
ddcacfa0	479	fprintf(stderr, "\n");
566747e6	480	fprintf(stderr, " %14.9f seconds time elapsed",
506d4bc8	481	avg_stats(&walltime_nsecs_stats)/1e9);
566747e6 IM	482	if (run_count > 1) {
566747e6 IM	483	fprintf(stderr, " ( +- %7.3f%% )",
506d4bc8 PZ	484	100*stddev_stats(&walltime_nsecs_stats) /
506d4bc8 PZ	485	avg_stats(&walltime_nsecs_stats));
566747e6 IM	486	}
566747e6 IM	487	fprintf(stderr, "\n\n");
ddcacfa0 IM	488	}
ddcacfa0 IM	489
f7b7c26e PZ	490	static volatile int signr = -1;
f7b7c26e PZ	491
5242519b	492	static void skip_signal(int signo)
ddcacfa0	493	{
6be2850e	494	if(child_pid == -1)
60666c63 LW	495	done = 1;
60666c63 LW	496
f7b7c26e PZ	497	signr = signo;
	498	}
	499
	500	static void sig_atexit(void)
	501	{
933da83a CW	502	if (child_pid != -1)
	503	kill(child_pid, SIGTERM);
	504
f7b7c26e PZ	505	if (signr == -1)
	506	return;
	507
	508	signal(signr, SIG_DFL);
	509	kill(getpid(), signr);
5242519b IM	510	}
	511
	512	static const char * const stat_usage[] = {
60666c63	513	"perf stat [<options>] [<command>]",
5242519b IM	514	NULL
	515	};
	516
5242519b IM	517	static const struct option options[] = {
5242519b IM	518	OPT_CALLBACK('e', "event", NULL, "event",
86847b62 TG	519	"event selector. use 'perf list' to list available events",
86847b62 TG	520	parse_events),
2e6cdf99 SE	521	OPT_BOOLEAN('i', "no-inherit", &no_inherit,
2e6cdf99 SE	522	"child tasks do not inherit counters"),
5242519b	523	OPT_INTEGER('p', "pid", &target_pid,
d6d901c2 ZY	524	"stat events on existing process id"),
	525	OPT_INTEGER('t', "tid", &target_tid,
	526	"stat events on existing thread id"),
5242519b	527	OPT_BOOLEAN('a', "all-cpus", &system_wide,
3d632595	528	"system-wide collection from all CPUs"),
b26bc5a7	529	OPT_BOOLEAN('c', "scale", &scale,
3d632595	530	"scale/normalize counters"),
c0555642	531	OPT_INCR('v', "verbose", &verbose,
743ee1f8	532	"be more verbose (show counter open errors, etc)"),
42202dd5 IM	533	OPT_INTEGER('r', "repeat", &run_count,
42202dd5 IM	534	"repeat command and print average + stddev (max: 100)"),
0cfb7a13 IM	535	OPT_BOOLEAN('n', "null", &null_run,
0cfb7a13 IM	536	"null run - dont start any counters"),
5242519b IM	537	OPT_END()
	538	};
	539
f37a291c	540	int cmd_stat(int argc, const char *argv, const char prefix __used)
5242519b	541	{
42202dd5	542	int status;
d6d901c2	543	int i,j;
42202dd5	544
a0541234 AB	545	argc = parse_options(argc, argv, options, stat_usage,
a0541234 AB	546	PARSE_OPT_STOP_AT_NON_OPTION);
d6d901c2	547	if (!argc && target_pid == -1 && target_tid == -1)
5242519b	548	usage_with_options(stat_usage, options);
9e9772c4	549	if (run_count <= 0)
42202dd5	550	usage_with_options(stat_usage, options);
ddcacfa0	551
c3043569 JSR	552	/* Set attrs and nr_counters if no event is selected and !null_run */
	553	if (!null_run && !nr_counters) {
	554	memcpy(attrs, default_attrs, sizeof(default_attrs));
	555	nr_counters = ARRAY_SIZE(default_attrs);
	556	}
ddcacfa0	557
a12b51c4 PM	558	if (system_wide)
	559	nr_cpus = read_cpu_map();
	560	else
	561	nr_cpus = 1;
ddcacfa0	562
d6d901c2 ZY	563	if (target_pid != -1) {
	564	target_tid = target_pid;
	565	thread_num = find_all_tid(target_pid, &all_tids);
	566	if (thread_num <= 0) {
	567	fprintf(stderr, "Can't find all threads of pid %d\n",
	568	target_pid);
	569	usage_with_options(stat_usage, options);
	570	}
	571	} else {
	572	all_tids=malloc(sizeof(pid_t));
	573	if (!all_tids)
	574	return -ENOMEM;
	575
	576	all_tids[0] = target_tid;
	577	thread_num = 1;
	578	}
	579
	580	for (i = 0; i < MAX_NR_CPUS; i++) {
	581	for (j = 0; j < MAX_COUNTERS; j++) {
	582	fd[i][j] = malloc(sizeof(int)*thread_num);
	583	if (!fd[i][j])
	584	return -ENOMEM;
	585	}
	586	}
	587
58d7e993 IM	588	/*
	589	* We dont want to block the signals - that would cause
	590	* child tasks to inherit that and Ctrl-C would not work.
	591	* What we want is for Ctrl-C to work in the exec()-ed
	592	* task, but being ignored by perf stat itself:
	593	*/
f7b7c26e	594	atexit(sig_atexit);
58d7e993 IM	595	signal(SIGINT, skip_signal);
	596	signal(SIGALRM, skip_signal);
	597	signal(SIGABRT, skip_signal);
	598
42202dd5 IM	599	status = 0;
	600	for (run_idx = 0; run_idx < run_count; run_idx++) {
	601	if (run_count != 1 && verbose)
3d632595	602	fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
42202dd5 IM	603	status = run_perf_stat(argc, argv);
	604	}
	605
084ab9f8 ACM	606	if (status != -1)
084ab9f8 ACM	607	print_stat(argc, argv);
42202dd5 IM	608
42202dd5 IM	609	return status;
ddcacfa0	610	}