工作队列(workqueue)

项目需要，在驱动模块里用内核计时器timer_list实现了一个状态机。
郁闷的是，运行时总报错“Scheduling while atomic”，网上搜了一下：

"Scheduling while atomic" indicates that you've tried to sleep somewhere that you shouldn't - like within a spinlock-protected critical section or an interrupt handler.

改进程序，在计时器里使用了workqueue，搞定问题。顺便把workqueue的实现代码总结了一下


一、workqueue简介

workqueue与tasklet类似，都是允许内核代码请求某个函数在将来的时间被调用（抄《ldd3》上的）
每个workqueue就是一个内核进程。

workqueue与tasklet的区别：
   1.tasklet是通过软中断实现的，在软中断上下文中运行，tasklet代码必须是原子的
     workqueue是通过内核进程实现的，就没有上述限制的，最爽的是，工作队列函数可以休眠
        
     PS: 我的驱动模块就是印在计时器中调用了可休眠函数，所以出现了cheduling while atomic告警
         内核计时器也是通过软中断实现的

   2.tasklet始终运行在被初始提交的同一处理器上，workqueue不一定
   3.tasklet不能确定延时时间（即使很短），workqueue可以设定延迟时间


二、workqueue的API

 
workqueue的API自2.6.20后发生了变化

1. #include <linux/workqueue.h>
   
2. struct workqueue_struct;
   
3. struct work_struct;
   
4. 
   
5. struct workqueue_struct *create_workqueue(const char *name);
   
6. void destroy_workqueue(struct workqueue_struct *queue);
   
7. 
   
8. INIT_WORK(_work, _func);
   
9. INIT_DELAYED_WORK(_work, _func);
   
10. 
    
11. int queue_work(struct workqueue_struct *wq, struct work_struct *work);
    
12. int queue_delayed_work(struct workqueue_struct *wq,struct delayed_work *dwork, unsigned long delay);
    
13. 
    
14. int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
    
15.             struct delayed_work *dwork, unsigned long delay);
    
16. 
    
17. int cancel_work_sync(struct work_struct *work);
    
18. int cancel_delayed_work_sync(struct delayed_work *dwork);
    
19. 
    
20. void flush_workqueue(struct workqueue_struct *wq);

Workqueue编程接口

序号	接口函数	说明
1	create_workqueue	用于创建一个workqueue队列，为系统中的每个CPU都创建一个内核线程。输入参数： @name：workqueue的名称
2	create_singlethread_workqueue	用于创建workqueue，只创建一个内核线程。输入参数： @name：workqueue名称
3	destroy_workqueue	释放workqueue队列。输入参数： @ workqueue_struct：需要释放的workqueue队列指针
4	schedule_work	调度执行一个具体的任务，执行的任务将会被挂入Linux系统提供的workqueue——keventd_wq输入参数： @ work_struct：具体任务对象指针
5	schedule_delayed_work	延迟一定时间去执行一个具体的任务，功能与schedule_work类似，多了一个延迟时间，输入参数： @work_struct：具体任务对象指针 @delay：延迟时间
6	queue_work	调度执行一个指定workqueue中的任务。输入参数： @ workqueue_struct：指定的workqueue指针 @work_struct：具体任务对象指针
7	queue_delayed_work	延迟调度执行一个指定workqueue中的任务，功能与queue_work类似，输入参数多了一个delay。

下面实例是不指定delay时间的workqueue
(代码基于2.6.24)

1. struct my_work_stuct{
   
2.     int test;
   
3.     struct work_stuct save;
   
4. };
   
5. 
   
6. struct my_work_stuct test_work;
   
7. struct workqueue_struct *test_workqueue;
   
8. 
   
9. void do_save(struct work_struct *p_work)
   
10. {
    
11.     struct my_work_struct *p_test_work = container_of(p_work, struct my_work_stuct, save);
    
12.     printk("%d\n",p_test_work->test);
    
13. }
    
14.   
    
15. void test_init()
    
16. {
    
17.     test_workqueue = create_workqueue("test_workqueue");
    
18.     if (!test_workqueue)
    
19.         panic("Failed to create test_workqueue\n");
    
20. 
    
21.     INIT_WORK(&(test_work.save), do_save);
    
22. 
    
23.     queue_work(test_workqueue, &(test_work.save));
24. }
    
25. 
    
26. void test_destory(void)
    
27. {
    
28.     if(test_workqueue)
    
29.         destroy_workqueue(test_workqueue);
    
30. }

三、workqueue的实现


工作队列workqueue不是通过软中断实现的，它是通过内核进程实现的



首先，创建一个workqueue，实际上就是建立一个内核进程

1. create_workqueue("tap_workqueue")
   
2. --> __create_workqueue(“tap_workqueue”, 0, 0)
   
3. --> __create_workqueue_key((name), (singlethread), (freezeable), NULL, NULL){
   
4.          wq = kzalloc(sizeof(*wq), GFP_KERNEL);
   
5.          wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
   
6.          wq->name = name;
   
7.          wq->singlethread = singlethread;
   
8.          wq->freezeable = freezeable;
   
9.          INIT_LIST_HEAD(&wq->list);
   
10. 
    
11.          for_each_possible_cpu(cpu) {
    
12.              cwq = init_cpu_workqueue(wq, cpu);
    
13.              err = create_workqueue_thread(cwq, cpu);
    
14.              start_workqueue_thread(cwq, cpu);
    
15.          }
    
16.     }

create_workqueue_thread 建立了一个内核进程 worker_thread（linux_2_6_24/kernel/workqueue.c）

1. create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
   
2. {
   
3.     struct workqueue_struct *wq = cwq->wq;
   
4. 
   
5.     const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
   
6.     struct task_struct *p;
   
7. 
   
8.     p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
   
9. 
   
10.     if (IS_ERR(p))
    
11.         return PTR_ERR(p);
    
12. 
    
13.     cwq->thread = p;
    
14. 
    
15.     return 0;
    
16. }

内核进程worker_thread做的事情很简单，死循环而已，不停的执行workqueue上的work_list
（linux_2_6_24/kernel/workqueue.c）

1. int worker_thread (void *__cwq)
   
2. {
   
3.     struct cpu_workqueue_struct *cwq = __cwq;
   
4.     /*下面定义等待队列项*/
   
5.     DEFINE_WAIT(wait);
   
6. 
   
7.     /*下面freezeable一般为0*/
   
8.     if (cwq->wq->freezeable)
   
9.         set_freezable();
   
10. 
    
11.     /*提高优先级别*/
    
12.     set_user_nice(current, -5);
    
13. 
    
14.     for (;;) {
    
15.         /*在cwq->more_work上等待, 若有人调用queue_work,该函数将调用wake_up(&cwq->more_work) 激活本进程*/
    
16.         prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
    
17. 
    
18.         /*work队列空则切换出去*/
    
19.         if (!freezing(current) && !kthread_should_stop() && list_empty(&cwq->worklist))
    
20.             schedule();
    
21. 
    
22. 
    
23.         /*切换回来则结束等待 说明有人唤醒cwq->more_work上的等待 有work需要处理*/
    
24.         finish_wait(&cwq->more_work, &wait);
    
25. 
    
26.         /*下面空，因为没有定义电源管理*/
    
27.         try_to_freeze();
    
28. 
    
29.         if (kthread_should_stop())
    
30.             break;
    
31. 
    
32.         /*run_workqueue依次处理工作队列上所有的work*/
    
33.         run_workqueue(cwq);
    
34.     }
    
35.     return 0;
    
36. }
    
37. 
    
38. 
    
39. /*run_workqueue依次处理工作队列上所有的work*/
    
40. static void run_workqueue(struct cpu_workqueue_struct *cwq)
    
41. {
    
42.     spin_lock_irq(&cwq->lock);
    
43.     cwq->run_depth++;
    
44.     if (cwq->run_depth > 3) {
    
45.         /* morton gets to eat his hat */
    
46.         printk("%s: recursion depth exceeded: %d\n",
    
47.             __FUNCTION__, cwq->run_depth);
    
48.         dump_stack();
    
49.     }
    
50. 
    
51.     while (!list_empty(&cwq->worklist)) {
    
52.         struct work_struct *work = list_entry(cwq->worklist.next,
    
53.                         struct work_struct, entry);
    
54.         work_func_t f = work->func;
    
55. #ifdef CONFIG_LOCKDEP
    
56.         /*
    
57.          * It is permissible to free the struct work_struct
    
58.          * from inside the function that is called from it,
    
59.          * this we need to take into account for lockdep too.
    
60.          * To avoid bogus "held lock freed" warnings as well
    
61.          * as problems when looking into work->lockdep_map,
    
62.          * make a copy and use that here.
    
63.          */
    
64.         struct lockdep_map lockdep_map = work->lockdep_map;
    
65. #endif
    
66. 
    
67.         cwq->current_work = work;
    
68.         list_del_init(cwq->worklist.next);
    
69.         spin_unlock_irq(&cwq->lock);
    
70. 
    
71.         BUG_ON(get_wq_data(work) != cwq);
    
72.         work_clear_pending(work);
    
73.         lock_acquire(&cwq->wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
    
74.         lock_acquire(&lockdep_map, 0, 0, 0, 2, _THIS_IP_);
    
75. 
    
76.         f(work); /*执行work项中的func*/
    
77.        
    
78.         lock_release(&lockdep_map, 1, _THIS_IP_);
    
79.         lock_release(&cwq->wq->lockdep_map, 1, _THIS_IP_);
    
80. 
    
81.         if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
    
82.             printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
    
83.                     "%s/0x%08x/%d\n",
    
84.                     current->comm, preempt_count(),
    
85.                         task_pid_nr(current));
    
86.             printk(KERN_ERR " last function: ");
    
87.             print_symbol("%s\n", (unsigned long)f);
    
88.             debug_show_held_locks(current);
    
89.             dump_stack();
    
90.         }
    
91. 
    
92.         spin_lock_irq(&cwq->lock);
    
93.         cwq->current_work = NULL;
    
94.     }
    
95.     cwq->run_depth--;
    
96.     spin_unlock_irq(&cwq->lock);
    
97. }

将一个work加入到指定workqueue的work_list中（文件linux_2_6_24 /kernel /workqueue .c）

  int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)

1. {
   
2.     int ret = 0;
   
3. 
   
4.     if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
   
5.         BUG_ON(!list_empty(&work->entry));
   
6.         __queue_work(wq_per_cpu(wq, get_cpu()), work);
   
7.         put_cpu();
   
8.         ret = 1;
   
9.     }
   
10.     return ret;
    
11. }
    
12. 
    
13. 
    
14. /* Preempt must be disabled. */
    
15. static void __queue_work(struct cpu_workqueue_struct *cwq, struct work_struct *work)
    
16. {
    
17.     unsigned long flags;
    
18. 
    
19.     spin_lock_irqsave(&cwq->lock, flags);
    
20.     insert_work(cwq, work, 1);
    
21.     spin_unlock_irqrestore(&cwq->lock, flags);
    
22. }
    
23. 
    
24. static void insert_work(struct cpu_workqueue_struct *cwq,
    
25.                 struct work_struct *work, int tail)
    
26. {
    
27.     set_wq_data(work, cwq);
    
28.     /*
    
29.      * Ensure that we get the right work->data if we see the
    
30.      * result of list_add() below, see try_to_grab_pending().
    
31.      */
    
32.     smp_wmb();
    
33.     if (tail)
    
34.         list_add_tail(&work->entry, &cwq->worklist);
    
35.     else
    
36.         list_add(&work->entry, &cwq->worklist);
    
37.     wake_up(&cwq->more_work);
    
38. }

四、共享队列

其实内核有自己的一个workqueue，叫keventd_wq，这个工作队列也叫做“共享队列”。
do_basic_setup --> init_workqueues --> create_workqueue("events");

若驱动模块使用的workqueue功能很简单的话，可以使用“共享队列”，不用自己再建一个队列
使用共享队列，有这样一套API

1. int schedule_work(struct work_struct *work)
   
2. {
   
3.     queue_work(keventd_wq, work);
   
4. }
   
5. 
   
6. int schedule_delayed_work(struct delayed_work *dwork,unsigned long delay)
   
7. {
   
8.     timer_stats_timer_set_start_info(&dwork->timer);
   
9.     return queue_delayed_work(keventd_wq, dwork, delay);
   
10. }
    
11. 
    
12. void flush_scheduled_work(void)
    
13. {
    
14.     flush_workqueue(keventd_wq);
    
15. }
16. 转自http://blog.chinaunix.net/uid-24148050-id-296982.html