linux模块编程（三）——线程的约会completion [转]

http://blog.csdn.net/qb_2008/article/details/6837262

 

 上节中我们已经掌握了创建大量内核线程的能力，可惜线程之间还缺乏配合。要知道学习ITC(inter thread communication)，和学习IPC(inter process communication)一样，不是件简单的事情。本节就暂且解释一种最简单的线程同步手段—completion。

        打开include/linux/completion.h，你就会看到completion使用的全部API。这里简单介绍一下。

 

struct completion{
    unsigned int done;
    wait_queue_head_t wait;
};

void init_completion(struct completion *x);
void wait_for_completion(struct completion *x);
void wait_for_completion_interruptible(struct completion *x);
void wait_for_completion_killable(struct completion *x);
unsigned long wait_for_completion_timeout(struct completion *x,
    unsigned long timeout);
unsigned long wait_for_completion_interruptible_timeout(struct completion *x,
    unsigned long timeout);
bool try_wait_for_completion(struct completion *x);
bool completion_done(struct completion *x);
void complete(struct completion *x);
void complete_all(struct completion *x);

struct completion{
	unsigned int done;
	wait_queue_head_t wait;
};

void init_completion(struct completion *x);
void wait_for_completion(struct completion *x);
void wait_for_completion_interruptible(struct completion *x);
void wait_for_completion_killable(struct completion *x);
unsigned long wait_for_completion_timeout(struct completion *x, 
	unsigned long timeout);
unsigned long wait_for_completion_interruptible_timeout(struct completion *x, 
	unsigned long timeout);
bool try_wait_for_completion(struct completion *x);
bool completion_done(struct completion *x);
void complete(struct completion *x);
void complete_all(struct completion *x);

       首先是struct completion的结构，由一个计数值和一个等待队列组成。我们就大致明白，completion是类似于信号量的东西，用completion.done来表示资源是否可用，获取不到的线程会阻塞在completion.wait的等待队列上，直到其它线程释放completion。这样理解在实现上不错，但我认为completion不是与具体的资源绑定，而是单纯作为一种线程间同步的机制，它在概念上要比信号量清晰得多。以后会逐渐看到，线程间事件的同步大多靠completion，而资源临界区的保护大多靠信号量。所以说，completion是一种线程间的约会。

           init_completion初始化completion结构。初此之外，linux当然还有在定义变量时初始化的方法，都在completion.h中。

      wait_for_completion等待在completion上。如果加了interruptible，就表示线程等待可被外部发来的信号打断；如果加了killable，就表示线程只可被kill信号打断；如果加了timeout，表示等待超出一定时间会自动结束等待，timeout的单位是系统所用的时间片jiffies(多为1ms)。

      try_wait_for_completion则是非阻塞地获取completion。它相当于wait_for_completion_timeout调用中的timeout值为0。

      completion_done检查是否有线程阻塞在completion上。但这个API并不准确，它只是检查completion.done是否为0，为0则认为有线程阻塞。这个API并不会去检查实际的等待队列，所以用时要注意。

      complete唤醒阻塞在completion上的首个线程。

      complete_all唤醒阻塞在completion上的所有线程。它的实现手法很粗糙，把completion.done的值设为UINT_MAX/2，自然所有等待的线程都醒了。所以如果complete_all之后还要使用这个completion，就要把它重新初始化。

 

      好，completion介绍完毕，下面就来设计我们的模块吧。

      我们模拟5个周期性线程的运行。每个周期性线程period_thread的周期各不相同，但都以秒为单位，有各自的completion变量。period_thread每个周期运行一次，然后等待在自己的completion变量上。为了唤醒period_thread，我们使用一个watchdog_thread来模拟时钟，每隔1s watchdog_thread就会检查哪个period_thread下一周期是否到来，并用相应的completion唤醒线程。

      下面就动手实现吧。

1、把上节建立的kthread子目录，复制为新的completion子目录。

 

2、修改hello.c，使其内容如下。

#include <linux/init.h>
#include <linux/module.h>
#include <linux/kthread.h>
#include <linux/completion.h>

MODULE_LICENSE("Dual BSD/GPL");

#define PERIOD_THREAD_NUM 5

staticint periods[PERIOD_THREAD_NUM] =
    { 1, 2, 4, 8, 16 };

staticstruct task_struct *period_tsks[PERIOD_THREAD_NUM];

staticstruct task_struct watchdog_tsk;

staticstruct completion wakeups[PERIOD_THREAD_NUM];


staticint period_thread(void *data)
{
    int k = (int)data;
    int count = -1;

    do{
        printk("thread%d: period=%ds, count=%d\n", k, periods[k], ++count);
        wait_for_completion(&wakeups[k]);
    }while(!kthread_should_stop());
    return count;
}

staticint watchdog_thread(void *data)
{
    int k;
    int count = 0;

    do{
        msleep(1000);
        count++;
        for(k=0; k<PERIOD_THREAD_NUM; k++){
            if (count%periods[k] == 0)
                complete(&wakeups[k]);
        }
    }while(!kthread_should_stop());
    return count;
}

staticint hello_init(void)
{
    int k;

    printk(KERN_INFO "Hello, world!\n");

    for(k=0; k<PERIOD_THREAD_NUM; k++){
        init_completion(&wakeups[k]);
    }

    watchdog_tsk = kthread_run(watchdog_thread, NULL, "watchdog_thread");

    if(IS_ERR(watchdog_tsk)){
        printk(KERN_INFO "create watchdog_thread failed!\n");
        return 1;
    }

    for(k=0; k<PERIOD_THREAD_NUM; k++){
        period_tsks[k] = kthread_run(period_thread, (void*)k, "period_thread%d", k);
        if(IS_ERR(period_tsks[k]))
            printk(KERN_INFO "create period_thread%d failed!\n", k);
    }
    return 0;
}

staticvoid hello_exit(void)
{
    int k;
    int count[5], watchdog_count;

    printk(KERN_INFO "Hello, exit!\n");
    for(k=0; k<PERIOD_THREAD_NUM]; k++){
        count[k] = 0;
        if(!IS_ERR(period_tsks[k]))
            count[k] = kthread_stop(period_tsks[k]);
    }
    watchdog_count = 0;
    if(!IS_ERR(watchdog_tsk))
        watchdog_count = kthread_stop(watchdog_tsk);

    printk("running total time: %ds\n", watchdog_count);
    for(k=0; k<PERIOD_THREAD_NUM; k++)
        printk("thread%d: period %d, running %d times\n", k, periods[k], count[k]);
}

module_init(hello_init);
module_exit(hello_exit);

#include <linux/init.h>
#include <linux/module.h>
#include <linux/kthread.h>
#include <linux/completion.h>

MODULE_LICENSE("Dual BSD/GPL");

#define PERIOD_THREAD_NUM 5

static int periods[PERIOD_THREAD_NUM] = 
	{ 1, 2, 4, 8, 16 };

static struct task_struct *period_tsks[PERIOD_THREAD_NUM];

static struct task_struct watchdog_tsk;

static struct completion wakeups[PERIOD_THREAD_NUM];


static int period_thread(void *data)
{
	int k = (int)data;
	int count = -1;

	do{
		printk("thread%d: period=%ds, count=%d\n", k, periods[k], ++count);
		wait_for_completion(&wakeups[k]);
	}while(!kthread_should_stop());
	return count;
}

static int watchdog_thread(void *data)
{
	int k;
	int count = 0;
	
	do{
		msleep(1000);
		count++;
		for(k=0; k<PERIOD_THREAD_NUM; k++){
			if (count%periods[k] == 0)
				complete(&wakeups[k]);
		}
	}while(!kthread_should_stop());
	return count;
}

static int hello_init(void)
{
	int k;

	printk(KERN_INFO "Hello, world!\n");

	for(k=0; k<PERIOD_THREAD_NUM; k++){
		init_completion(&wakeups[k]);
	}

	watchdog_tsk = kthread_run(watchdog_thread, NULL, "watchdog_thread");

	if(IS_ERR(watchdog_tsk)){
		printk(KERN_INFO "create watchdog_thread failed!\n");
		return 1;
	}

	for(k=0; k<PERIOD_THREAD_NUM; k++){
		period_tsks[k] = kthread_run(period_thread, (void*)k, "period_thread%d", k);
		if(IS_ERR(period_tsks[k]))
			printk(KERN_INFO "create period_thread%d failed!\n", k);
	}
	return 0;
}

static void hello_exit(void)
{
	int k;
	int count[5], watchdog_count;

	printk(KERN_INFO "Hello, exit!\n");
	for(k=0; k<PERIOD_THREAD_NUM]; k++){
		count[k] = 0;
		if(!IS_ERR(period_tsks[k]))
			count[k] = kthread_stop(period_tsks[k]);
	}
	watchdog_count = 0;
	if(!IS_ERR(watchdog_tsk))
		watchdog_count = kthread_stop(watchdog_tsk);

	printk("running total time: %ds\n", watchdog_count);
	for(k=0; k<PERIOD_THREAD_NUM; k++)
		printk("thread%d: period %d, running %d times\n", k, periods[k], count[k]);
}

module_init(hello_init);
module_exit(hello_exit);

3、编译运行模块，步骤参照前例。为保持模块的简洁性，我们仍然使用了kthread_stop结束线程，这种方法虽然简单，但在卸载模块时等待时间太长，而且这个时间会随线程个数和周期的增长而增长。

 

4、使用统一的exit_flag标志来表示结束请求，hello_exit发送completion信号给所有的周期线程，最后调用kthread_stop来回收线程返回值。这样所有的周期线程都是在被唤醒后看到exit_flag，自动结束，卸载模块时间大大缩短。下面是改进过后的hello.c，之前的那个姑且叫做hello-v1.c好了。

#include <linux/init.h>
#include <linux/module.h>
#include <linux/kthread.h>
#include <linux/completion.h>

MODULE_LICENSE("Dual BSD/GPL");

#define PERIOD_THREAD_NUM 5

staticint periods[PERIOD_THREAD_NUM] =
    { 1, 2, 4, 8, 16 };

staticstruct task_struct *period_tsks[PERIOD_THREAD_NUM];

staticstruct task_struct watchdog_tsk;

staticstruct completion wakeups[PERIOD_THREAD_NUM];

staticint exit_flag = 0;

staticint period_thread(void *data)
{
    int k = (int)data;
    int count = -1;

    do{
        printk("thread%d: period=%ds, count=%d\n", k, periods[k], ++count);
        wait_for_completion(&wakeups[k]);
    }while(!exit_flag);
    return count;
}

staticint watchdog_thread(void *data)
{
    int k;
    int count = 0;

    do{
        msleep(1000);
        count++;
        for(k=0; k<PERIOD_THREAD_NUM; k++){
            if (count%periods[k] == 0)
                complete(&wakeups[k]);
        }
    }while(!exit_flag);
    return count;
}

staticint hello_init(void)
{
    int k;

    printk(KERN_INFO "Hello, world!\n");

    for(k=0; k<PERIOD_THREAD_NUM; k++){
        init_completion(&wakeups[k]);
    }

    watchdog_tsk = kthread_run(watchdog_thread, NULL, "watchdog_thread");

    if(IS_ERR(watchdog_tsk)){
        printk(KERN_INFO "create watchdog_thread failed!\n");
        return 1;
    }

    for(k=0; k<PERIOD_THREAD_NUM; k++){
        period_tsks[k] = kthread_run(period_thread, (void*)k, "period_thread%d", k);
        if(IS_ERR(period_tsks[k]))
            printk(KERN_INFO "create period_thread%d failed!\n", k);
    }
    return 0;
}

staticvoid hello_exit(void)
{
    int k;
    int count[5], watchdog_count;

    printk(KERN_INFO "Hello, exit!\n");
    exit_flag = 1;
    for(k=0; k<PERIOD_THREAD_NUM]; k++)
        complete_all(&wakeups[k]);

    for(k=0; k<PERIOD_THREAD_NUM]; k++){
        count[k] = 0;
        if(!IS_ERR(period_tsks[k]))
            count[k] = kthread_stop(period_tsks[k]);
    }
    watchdog_count = 0;
    if(!IS_ERR(watchdog_tsk))
        watchdog_count = kthread_stop(watchdog_tsk);

    printk("running total time: %ds\n", watchdog_count);
    for(k=0; k<PERIOD_THREAD_NUM; k++)
        printk("thread%d: period %d, running %d times\n", k, periods[k], count[k]);
}

module_init(hello_init);
module_exit(hello_exit);

#include <linux/init.h>
#include <linux/module.h>
#include <linux/kthread.h>
#include <linux/completion.h>

MODULE_LICENSE("Dual BSD/GPL");

#define PERIOD_THREAD_NUM 5

static int periods[PERIOD_THREAD_NUM] = 
	{ 1, 2, 4, 8, 16 };

static struct task_struct *period_tsks[PERIOD_THREAD_NUM];

static struct task_struct watchdog_tsk;

static struct completion wakeups[PERIOD_THREAD_NUM];

static int exit_flag = 0;

static int period_thread(void *data)
{
	int k = (int)data;
	int count = -1;

	do{
		printk("thread%d: period=%ds, count=%d\n", k, periods[k], ++count);
		wait_for_completion(&wakeups[k]);
	}while(!exit_flag);
	return count;
}

static int watchdog_thread(void *data)
{
	int k;
	int count = 0;
	
	do{
		msleep(1000);
		count++;
		for(k=0; k<PERIOD_THREAD_NUM; k++){
			if (count%periods[k] == 0)
				complete(&wakeups[k]);
		}
	}while(!exit_flag);
	return count;
}

static int hello_init(void)
{
	int k;

	printk(KERN_INFO "Hello, world!\n");

	for(k=0; k<PERIOD_THREAD_NUM; k++){
		init_completion(&wakeups[k]);
	}

	watchdog_tsk = kthread_run(watchdog_thread, NULL, "watchdog_thread");

	if(IS_ERR(watchdog_tsk)){
		printk(KERN_INFO "create watchdog_thread failed!\n");
		return 1;
	}

	for(k=0; k<PERIOD_THREAD_NUM; k++){
		period_tsks[k] = kthread_run(period_thread, (void*)k, "period_thread%d", k);
		if(IS_ERR(period_tsks[k]))
			printk(KERN_INFO "create period_thread%d failed!\n", k);
	}
	return 0;
}

static void hello_exit(void)
{
	int k;
	int count[5], watchdog_count;

	printk(KERN_INFO "Hello, exit!\n");
	exit_flag = 1;
	for(k=0; k<PERIOD_THREAD_NUM]; k++)
		complete_all(&wakeups[k]);

	for(k=0; k<PERIOD_THREAD_NUM]; k++){
		count[k] = 0;
		if(!IS_ERR(period_tsks[k]))
			count[k] = kthread_stop(period_tsks[k]);
	}
	watchdog_count = 0;
	if(!IS_ERR(watchdog_tsk))
		watchdog_count = kthread_stop(watchdog_tsk);

	printk("running total time: %ds\n", watchdog_count);
	for(k=0; k<PERIOD_THREAD_NUM; k++)
		printk("thread%d: period %d, running %d times\n", k, periods[k], count[k]);
}

module_init(hello_init);
module_exit(hello_exit);

5、编译运行改进过后的模块。可以看到模块卸载时间大大减少，不会超过1s。

 

 

      经过本节，我们学会了一种内核线程间同步的机制—completion。线程们已经开始注意相互配合，以完成复杂的工作。相信它们会越来越聪明的。

 

     

附注：

     completion的实现在kernel/sched.c中。这里的每个API都较短，实现也较为简单。completion背后的实现机制其实是等待队列。等待队列的实现会涉及到较多的调度问题，这里先简单略过。