drivers/cpuidle/profiler.c

/*
 * Copyright (c) 2018 Park Bumgyu, Samsung Electronics Co., Ltd <bumgyu.park@samsung.com>
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License version 2 as
 * published by the Free Software Foundation.
 *
 * CPUIDLE profiler for Exynos
 */

#include <linux/device.h>
#include <linux/kobject.h>
#include <linux/cpuidle.h>
#include <linux/slab.h>
#include <soc/samsung/exynos_perf_cpuidle.h>

/* whether profiling has started */
static bool profile_started;

/*
 * Represents statistic of idle state.
 * All idle states are mapped 1:1 with cpuidle_stats.
 */
struct cpuidle_stats {
	/* time to enter idle state */
	ktime_t			idle_entry_time;

	/* number of times an idle state is entered */
	unsigned int		entry_count;

	/* number of times the entry into idle state is canceled */
	unsigned int		cancel_count;

	/* time in idle state */
	unsigned long long	time;
};

/* description length of idle state */
#define DESC_LEN	32

/*
 * Manages idle state where cpu enters individually. One cpu_idle_state
 * structure manages a idle state for each cpu to enter, and the number
 * of structure is determined by cpuidle driver.
 */
struct cpu_idle_state {
	/* description of idle state */
	char 			desc[DESC_LEN];

	/* idle state statstics for each cpu */
	struct cpuidle_stats	stats[NR_CPUS];
};

/* cpu idle state list and length of cpu idle state list */
static struct cpu_idle_state *cpu_idle_state;
static int cpu_idle_state_count;

/*
 * Manages idle state in which multiple cpus unit enter. Each idle state
 * has one group_idle_state structure.
 */
struct group_idle_state {
	/* idle state id, it must be unique */
	int			id;

	/* description of idle state */
	char 			desc[DESC_LEN];

	/* idle state statstics */
	struct cpuidle_stats	stats;
};

/*
 * To easily manage group_idle_state dynamically, manage the list as an
 * list. Currently, the maximum number of group idle states supported is 5,
 * which is unlikely to exceed the number of states empirically.
 */
#define MAX_GROUP_IDLE_STATE	5

/* group idle state list and length of group idle state list */
static struct group_idle_state * group_idle_state[MAX_GROUP_IDLE_STATE];
static int group_idle_state_count;

/************************************************************************
 *                              Profiling                               *
 ************************************************************************/
static void idle_enter(struct cpuidle_stats *stats)
{
	stats->idle_entry_time = ktime_get();
	stats->entry_count++;
}

static void idle_exit(struct cpuidle_stats *stats, int cancel)
{
	s64 diff;

	/*
	 * If profiler is started with cpu already in idle state,
	 * idle_entry_time is 0 because entry event is not recorded.
	 * From the start of the profile to cpu wakeup is the idle time,
	 * but ignore this because it is complex to handle it and the
	 * time is not large.
	 */
	if (!stats->idle_entry_time)
		return;

	if (cancel) {
		stats->cancel_count++;
		return;
	}

	diff = ktime_to_us(ktime_sub(ktime_get(), stats->idle_entry_time));
	stats->time += diff;

	stats->idle_entry_time = 0;
}

/*
 * cpuidle_profile_cpu_idle_enter/cpuidle_profile_cpu_idle_exit
 * : profilie for cpu idle state
 */
void cpuidle_profile_cpu_idle_enter(int cpu, int index)
{
	if (!profile_started)
		return;

	idle_enter(&cpu_idle_state[index].stats[cpu]);
	exynos_perf_cpu_idle_enter(cpu, index);
}

void cpuidle_profile_cpu_idle_exit(int cpu, int index, int cancel)
{
	if (!profile_started)
		return;

	exynos_perf_cpu_idle_exit(cpu, index, cancel);
	idle_exit(&cpu_idle_state[index].stats[cpu], cancel);
}

/*
 * cpuidle_profile_group_idle_enter/cpuidle_profile_group_idle_exit
 * : profilie for group idle state
 */
void cpuidle_profile_group_idle_enter(int id)
{
	int i;

	if (!profile_started)
		return;

	for (i = 0; i < group_idle_state_count; i++)
		if (group_idle_state[i]->id == id)
			break;

	idle_enter(&group_idle_state[i]->stats);
}

void cpuidle_profile_group_idle_exit(int id, int cancel)
{
	int i;

	if (!profile_started)
		return;

	for (i = 0; i < group_idle_state_count; i++)
		if (group_idle_state[i]->id == id)
			break;

	idle_exit(&group_idle_state[i]->stats, cancel);
}

/************************************************************************
 *                          Profile start/stop                          *
 ************************************************************************/
/* totoal profiling time */
static s64 profile_time;

/* start time of profile */
static ktime_t profile_start_time;

/* idle ip */
static int idle_ip_stats[4][32];
extern char *idle_ip_names[4][32];

static void clear_stats(struct cpuidle_stats *stats)
{
	if (!stats)
		return;

	stats->idle_entry_time = 0;

	stats->entry_count = 0;
	stats->cancel_count = 0;
	stats->time = 0;
}

static void reset_profile(void)
{
	int cpu, i;

	profile_start_time = 0;

	for (i = 0; i < cpu_idle_state_count; i++)
		for_each_possible_cpu(cpu)
			clear_stats(&cpu_idle_state[i].stats[cpu]);

	for (i = 0; i < group_idle_state_count; i++)
		clear_stats(&group_idle_state[i]->stats);

	memset(idle_ip_stats, 0, sizeof(idle_ip_stats));
}

static void do_nothing(void *unused)
{
}

static void cpuidle_profile_start(void)
{
	if (profile_started) {
		pr_err("cpuidle profile is ongoing\n");
		return;
	}

	exynos_perf_cpuidle_start();

	reset_profile();
	profile_start_time = ktime_get();

	profile_started = 1;

	preempt_disable();
	/* wakeup all cpus to start profile */
	smp_call_function(do_nothing, NULL, 1);
	preempt_enable();

	pr_info("cpuidle profile start\n");
}

static void cpuidle_profile_stop(void)
{
	if (!profile_started) {
		pr_err("CPUIDLE profile does not start yet\n");
		return;
	}

	exynos_perf_cpuidle_stop();

	pr_info("cpuidle profile stop\n");

	preempt_disable();
	/* wakeup all cpus to stop profile */
	smp_call_function(do_nothing, NULL, 1);
	preempt_enable();

	profile_started = 0;

	profile_time = ktime_to_us(ktime_sub(ktime_get(), profile_start_time));
}

/************************************************************************
 *                               IDLE IP                                *
 ************************************************************************/
void cpuidle_profile_idle_ip(int index, unsigned int idle_ip)
{
	int i;

	/*
	 * Return if profile is not started
	 */
	if (!profile_started)
		return;

	for (i = 0; i < 32; i++) {
		/*
		 * If bit of idle_ip has 1, IP corresponding to its bit
		 * is not idle.
		 */
		if (idle_ip & (1 << i))
			idle_ip_stats[index][i]++;
	}
}

/************************************************************************
 *                              Show result                             *
 ************************************************************************/
static int calculate_percent(s64 residency)
{
	if (!residency)
		return 0;

	residency *= 100;
	do_div(residency, profile_time);

	return residency;
}

static unsigned long long cpu_idle_time(int cpu)
{
	unsigned long long idle_time = 0;
	int i;

	for (i = 0; i < cpu_idle_state_count; i++)
		idle_time += cpu_idle_state[i].stats[cpu].time;

	return idle_time;
}

static int cpu_idle_ratio(int cpu)
{
	return calculate_percent(cpu_idle_time(cpu));
}

static ssize_t show_result(struct kobject *kobj,
				     struct kobj_attribute *attr,
				     char *buf)
{
	int ret = 0;
	int cpu, i, bit;

	if (profile_started) {
		ret += snprintf(buf + ret, PAGE_SIZE - ret,
				"CPUIDLE profile is ongoing\n");
		return ret;
	}

	ret += snprintf(buf + ret, PAGE_SIZE - ret,
		"#############################################################\n");
	ret += snprintf(buf + ret, PAGE_SIZE - ret,
		"Profiling Time : %lluus\n", profile_time);

	ret += snprintf(buf + ret, PAGE_SIZE - ret,
		"\n");

	ret += snprintf(buf + ret, PAGE_SIZE - ret,
		"[total idle ratio]\n");
	ret += snprintf(buf + ret, PAGE_SIZE - ret,
		"#cpu      #time    #ratio\n");
	for_each_possible_cpu(cpu)
		ret += snprintf(buf + ret, PAGE_SIZE - ret,
			"cpu%d %10lluus   %3u%%\n",
			cpu, cpu_idle_time(cpu), cpu_idle_ratio(cpu));

	ret += snprintf(buf + ret, PAGE_SIZE - ret,
		"\n");

	/*
	 * Example of cpu idle state profile result.
	 * Below is an example from the quad core architecture. The number of
	 * rows depends on the number of cpu.
	 *
	 * [state : {desc}]
	 * #cpu   #entry   #cancel      #time    #ratio
	 * cpu0     985        8      8808916us    87%
	 * cpu1     340        2      8311318us    82%
	 * cpu2     270        7      8744801us    87%
	 * cpu3     330        2      9001329us    89%
	 */
	for (i = 0; i < cpu_idle_state_count; i++) {
		ret += snprintf(buf + ret, PAGE_SIZE - ret,
			"[state : %s]\n", cpu_idle_state[i].desc);
		ret += snprintf(buf + ret, PAGE_SIZE - ret,
			"#cpu   #entry   #cancel      #time    #ratio\n");
		for_each_possible_cpu(cpu) {
			struct cpuidle_stats *stats = &cpu_idle_state[i].stats[cpu];

			ret += snprintf(buf + ret, PAGE_SIZE - ret,
				"cpu%d   %5u    %5u   %10lluus   %3u%%\n",
				cpu,
				stats->entry_count,
				stats->cancel_count,
				stats->time,
				calculate_percent(stats->time));
		}

		ret += snprintf(buf + ret, PAGE_SIZE - ret,
				"\n");
	}

	/*
	 * Example of group idle state profile result.
	 * The number of results depends on the number of group idle state.
	 *
	 * [state : {desc}]
	 * #entry   #cancel      #time    #ratio
	 *   52        1       4296397us    42%
	 *
	 * [state : {desc}]
	 * #entry   #cancel      #time    #ratio
	 *    20        0      2230528us    22%
	 */
	for (i = 0; i < group_idle_state_count; i++) {
		struct cpuidle_stats *stats = &group_idle_state[i]->stats;

		ret += snprintf(buf + ret, PAGE_SIZE - ret,
			"[state : %s]\n", group_idle_state[i]->desc);
		ret += snprintf(buf + ret, PAGE_SIZE - ret,
			"#entry   #cancel      #time    #ratio\n");
		ret += snprintf(buf + ret, PAGE_SIZE - ret,
			"%5u    %5u   %10lluus   %3u%%\n",
			stats->entry_count,
			stats->cancel_count,
			stats->time,
			calculate_percent(stats->time));

		ret += snprintf(buf + ret, PAGE_SIZE - ret,
				"\n");
	}

	ret += snprintf(buf + ret, PAGE_SIZE - ret, "[IDLE-IP statistics]\n");
	for (i = 0; i < 4; i++) {
		for (bit = 0; bit < 32; bit++) {
			if (!idle_ip_stats[i][bit])
				continue;

			ret += snprintf(buf + ret, PAGE_SIZE - ret,
				"busy IP : %s(count = %d)\n",
				idle_ip_names[i][bit], idle_ip_stats[i][bit]);
		}
	}

	ret += snprintf(buf + ret, PAGE_SIZE - ret,
		"#############################################################\n");

	return ret;
}

/*********************************************************************
 *                          Sysfs interface                          *
 *********************************************************************/
static ssize_t show_cpuidle_profile(struct kobject *kobj,
				     struct kobj_attribute *attr,
				     char *buf)
{
	int ret = 0;

	if (profile_started)
		ret += snprintf(buf + ret, PAGE_SIZE - ret,
				"CPUIDLE profile is ongoing\n");
	else
		ret = show_result(kobj, attr, buf);

	return ret;
}

static ssize_t store_cpuidle_profile(struct kobject *kobj,
				      struct kobj_attribute *attr,
				      const char *buf, size_t count)
{
	int input;

	if (!sscanf(buf, "%1d", &input))
		return -EINVAL;

	if (!!input)
		cpuidle_profile_start();
	else
		cpuidle_profile_stop();

	return count;
}

static struct kobj_attribute cpuidle_profile_attr =
	__ATTR(profile, 0644, show_cpuidle_profile, store_cpuidle_profile);

static struct attribute *cpuidle_profile_attrs[] = {
	&cpuidle_profile_attr.attr,
	NULL,
};

static const struct attribute_group cpuidle_profile_group = {
	.attrs = cpuidle_profile_attrs,
};

/*********************************************************************
 *                   Initialize cpuidle profiler                     *
 *********************************************************************/
void __init
cpuidle_profile_cpu_idle_register(struct cpuidle_driver *drv)
{
	struct cpu_idle_state *state;
	int state_count = drv->state_count;
	int i;

	state = kzalloc(sizeof(struct cpu_idle_state) * state_count,
							GFP_KERNEL);
	if (!state) {
		pr_err("%s: Failed to allocate memory\n", __func__);
		return;
	}

	for (i = 0; i < state_count; i++)
		strncpy(state[i].desc, drv->states[i].desc, DESC_LEN - 1);

	cpu_idle_state = state;
	cpu_idle_state_count = state_count;

	exynos_perf_cpu_idle_register(drv);
}

void __init
cpuidle_profile_group_idle_register(int id, const char *name)
{
	struct group_idle_state *state;

	state = kzalloc(sizeof(struct group_idle_state), GFP_KERNEL);
	if (!state) {
		pr_err("%s: Failed to allocate memory\n", __func__);
		return;
	}

	state->id = id;
	strncpy(state->desc, name, DESC_LEN - 1);

	group_idle_state[group_idle_state_count] = state;
	group_idle_state_count++;
}

static int __init cpuidle_profile_init(void)
{
	struct class *class;
	struct device *dev;
	int ret = 0;

	class = class_create(THIS_MODULE, "cpuidle");
	dev = device_create(class, NULL, 0, NULL, "cpuidle_profiler");

	ret = sysfs_create_group(&dev->kobj, &cpuidle_profile_group);
	if (ret)
		pr_err("%s: failed to create sysfs group", __func__);

	return ret;
}
late_initcall(cpuidle_profile_init);