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>
#include <locale.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 int			nr_cpus				=  0;
static int			run_idx				=  0;

static int			run_count			=  1;
static bool			no_inherit			= false;
static bool			scale				=  true;
static bool			no_aggr				= false;
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 bool			big_num				=  false;
static const char		*cpu_list;


static int			*fd[MAX_NR_CPUS][MAX_COUNTERS];

static int			event_scaled[MAX_COUNTERS];

static struct {
	u64 val;
	u64 ena;
	u64 run;
} cpu_counts[MAX_NR_CPUS][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[MAX_NR_CPUS];
struct stats			runtime_cycles_stats[MAX_NR_CPUS];
struct stats			runtime_branches_stats[MAX_NR_CPUS];
struct stats			walltime_nsecs_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) {
		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:
 * aggregate counts across CPUs in system-wide mode
 */
static void read_counter_aggr(int counter)
{
	u64 count[3], single_count[3];
	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[0], count[0]);
	if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
		update_stats(&runtime_cycles_stats[0], count[0]);
	if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter))
		update_stats(&runtime_branches_stats[0], count[0]);
}

/*
 * Read out the results of a single counter:
 * do not aggregate counts across CPUs in system-wide mode
 */
static void read_counter(int counter)
{
	u64 count[3];
	int cpu;
	size_t res, nv;

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

	nv = scale ? 3 : 1;

	for (cpu = 0; cpu < nr_cpus; cpu++) {

		if (fd[cpu][counter][0] < 0)
			continue;

		res = read(fd[cpu][counter][0], count, nv * sizeof(u64));

		assert(res == nv * sizeof(u64));

		close(fd[cpu][counter][0]);
		fd[cpu][counter][0] = -1;

		if (scale) {
			if (count[2] == 0) {
				count[0] = 0;
			} else if (count[2] < count[1]) {
				count[0] = (unsigned long long)
				((double)count[0] * count[1] / count[2] + 0.5);
			}
		}
		cpu_counts[cpu][counter].val = count[0]; /* scaled count */
		cpu_counts[cpu][counter].ena = count[1];
		cpu_counts[cpu][counter].run = count[2];

		if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
			update_stats(&runtime_nsecs_stats[cpu], count[0]);
		if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
			update_stats(&runtime_cycles_stats[cpu], count[0]);
		if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter))
			update_stats(&runtime_branches_stats[cpu], 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);

	if (no_aggr) {
		for (counter = 0; counter < nr_counters; counter++)
			read_counter(counter);
	} else {
		for (counter = 0; counter < nr_counters; counter++)
			read_counter_aggr(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 cpu, int counter, double avg)
{
	double msecs = avg / 1e6;

	if (no_aggr)
		fprintf(stderr, "CPU%-4d %18.6f  %-24s",
			cpumap[cpu], msecs, event_name(counter));
	else
		fprintf(stderr, " %18.6f  %-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 cpu, int counter, double avg)
{
	double total, ratio = 0.0;
	char cpustr[16] = { '\0', };

	if (no_aggr)
		sprintf(cpustr, "CPU%-4d", cpumap[cpu]);
	else
		cpu = 0;

	if (big_num)
		fprintf(stderr, "%s %'18.0f  %-24s",
			cpustr, avg, event_name(counter));
	else
		fprintf(stderr, "%s %18.0f  %-24s",
			cpustr, avg, event_name(counter));

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

		if (total)
			ratio = avg / total;

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

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

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

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

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

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

/*
 * Print out the results of a single counter:
 * aggregated counts in system-wide mode
 */
static void print_counter_aggr(int counter)
{
	double avg = avg_stats(&event_res_stats[counter][0]);
	int scaled = event_scaled[counter];

	if (scaled == -1) {
		fprintf(stderr, " %18s  %-24s\n",
			"<not counted>", event_name(counter));
		return;
	}

	if (nsec_counter(counter))
		nsec_printout(-1, counter, avg);
	else
		abs_printout(-1, 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");
}

/*
 * Print out the results of a single counter:
 * does not use aggregated count in system-wide
 */
static void print_counter(int counter)
{
	u64 ena, run, val;
	int cpu;

	for (cpu = 0; cpu < nr_cpus; cpu++) {
		val = cpu_counts[cpu][counter].val;
		ena = cpu_counts[cpu][counter].ena;
		run = cpu_counts[cpu][counter].run;
		if (run == 0 || ena == 0) {
			fprintf(stderr, "CPU%-4d %18s  %-24s", cpumap[cpu],
					"<not counted>", event_name(counter));

			fprintf(stderr, "\n");
			continue;
		}

		if (nsec_counter(counter))
			nsec_printout(cpu, counter, val);
		else
			abs_printout(cpu, counter, val);

		print_noise(counter, 1.0);

		if (run != ena) {
			fprintf(stderr, "  (scaled from %.2f%%)",
					100.0 * run / ena);
		}
		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");

	if (no_aggr) {
		for (counter = 0; counter < nr_counters; counter++)
			print_counter(counter);
	} else {
		for (counter = 0; counter < nr_counters; counter++)
			print_counter_aggr(counter);
	}

	fprintf(stderr, "\n");
	fprintf(stderr, " %18.9f  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_BOOLEAN('B', "big-num", &big_num,
		    "print large numbers with thousands\' separators"),
	OPT_STRING('C', "cpu", &cpu_list, "cpu",
		    "list of cpus to monitor in system-wide"),
	OPT_BOOLEAN('A', "no-aggr", &no_aggr,
		    "disable CPU count aggregation"),
	OPT_END()
};

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

	setlocale(LC_ALL, "");

	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);

	/* no_aggr is for system-wide only */
	if (no_aggr && !system_wide)
		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(cpu_list);
	else
		nr_cpus = 1;

	if (nr_cpus < 1)
		usage_with_options(stat_usage, options);

	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;
}