[内核]Linux workqueue

转自：http://blog.chinaunix.net/uid-24148050-id-296982.html

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：

• 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. 而workqueue是通过内核进程实现的，就没有上述限制的，而且工作队列函数可以休眠。
• tasklet始终运行在被初始提交的同一处理器上，workqueue不一定
• tasklet不能确定延时时间（即使很短），workqueue可以设定延迟时间

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

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 struct workqueue_struct *create_workqueue(const char *name);
 5 void destroy_workqueue(struct workqueue_struct *queue);
 6 INIT_WORK(_work, _func);
 7 INIT_DELAYED_WORK(_work, _func);
 8 int queue_work(struct workqueue_struct *wq, struct work_struct *work);
 9 int queue_delayed_work(struct workqueue_struct *wq,struct delayed_work *dwork, unsigned long delay);
10 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
11             struct delayed_work *dwork, unsigned long delay);
12 int cancel_work_sync(struct work_struct *work);
13 int cancel_delayed_work_sync(struct delayed_work *dwork);
14 void flush_workqueue(struct workqueue_struct *wq);

Workqueue编程接口

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

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          for_each_possible_cpu(cpu) {
11              cwq = init_cpu_workqueue(wq, cpu);
12              err = create_workqueue_thread(cwq, cpu);
13              start_workqueue_thread(cwq, cpu);
14          }
15     }

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     const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
 5     struct task_struct *p;
 6     p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
 7     if (IS_ERR(p))
 8         return PTR_ERR(p);
 9     cwq->thread = p;
10     return 0;
11 }

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

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

 1  int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
 2 {
 3     int ret = 0;
 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 /* Preempt must be disabled. */
13 static void __queue_work(struct cpu_workqueue_struct *cwq, struct work_struct *work)
14 {
15     unsigned long flags;
16     spin_lock_irqsave(&cwq->lock, flags);
17     insert_work(cwq, work, 1);
18     spin_unlock_irqrestore(&cwq->lock, flags);
19 }
20 static void insert_work(struct cpu_workqueue_struct *cwq,
21                 struct work_struct *work, int tail)
22 {
23     set_wq_data(work, cwq);
24     /*
25      * Ensure that we get the right work->data if we see the
26      * result of list_add() below, see try_to_grab_pending().
27      */
28     smp_wmb();
29     if (tail)
30         list_add_tail(&work->entry, &cwq->worklist);
31     else
32         list_add(&work->entry, &cwq->worklist);
33     wake_up(&cwq->more_work);
34 }

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 int schedule_delayed_work(struct delayed_work *dwork,unsigned long delay)
 6 {
 7     timer_stats_timer_set_start_info(&dwork->timer);
 8     return queue_delayed_work(keventd_wq, dwork, delay);
 9 }
10 void flush_scheduled_work(void)
11 {
12     flush_workqueue(keventd_wq);
13 }