The processing mechanism for blocking IO processes in Linux drivers

We need to consider how to respond when the driver cannot satisfy the request immediately. Normally, the calling process does not care about the status of the driver, so we need to handle it accordingly in the driver. A common approach is to block requests to the process.

Blocking I/O means that if the required resources cannot be obtained when performing device operations, the current process will be suspended until operable conditions are met before the operation is performed. The suspended process goes to sleep and is removed from the scheduler's run queue until the waiting conditions are met. On the contrary, non-blocking I/O does not suspend the process when the device operation cannot be performed, but chooses to give up or continue polling until the operation can be performed.

“

Waiting queue is a classic mechanism for handling blocking I/O.

”

1. Blocking the I/O processing flow in the driver

In summary, the blocking I/O processing flow includes 4 parts. First, initialize the waiting queue list, which stores the processes that need to be blocked. Then initialize a waiting queue and add the process that currently needs to be blocked to the waiting queue linked list. Then set the process to be interruptible and block the process from sleeping. Finally, when certain conditions are met, that is, resources are available, the processes in the waiting queue are awakened.
The following figure describes the blocking I/O processing flow:

2. Initialize the waiting queue list

Before initializing the waiting queue linked list, we first need to define a wait_queue_head_t variable. For example, in the ISP driver of RK3399, the data structure struct rkisp1_stream contains wait_queue_head_t done;. Perform initialization operations by calling init_waitqueue_head(&stream->done);.

void rkisp1_stream_init(struct rkisp1_device *dev, u32 id)
{
 struct rkisp1_stream *stream = &dev->stream[id];

 memset(stream, 0, sizeof(*stream));
 stream->id = id;
 stream->ispdev = dev;

 INIT_LIST_HEAD(&stream->buf_queue);
 init_waitqueue_head(&stream->done);
 spin_lock_init(&stream->vbq_lock);
 ...
}

wait_queue_head_tThe prototype of the variable is __wait_queue_head, as shown below:

struct __wait_queue_head {
 spinlock_t  lock;
 struct list_head task_list;
};

init_waitqueue_head()The function that is actually executed is __init_waitqueue_head(), and its function definition is as follows:

void __init_waitqueue_head(wait_queue_head_t *q, const char *name, struct lock_class_key *key)
{
 spin_lock_init(&q->lock);
 lockdep_set_class_and_name(&q->lock, key, name);
 INIT_LIST_HEAD(&q->task_list);
}

3. 等待队列处理

调用DECLARE_WAITQUEUE(wait, current)将当前进程初始化为等待队列。注意，这里的等待队列和等待队列链表头可不是一个东东。

#define DECLARE_WAITQUEUE(name, tsk)     \
 wait_queue_t name = __WAITQUEUE_INITIALIZER(name, tsk)

等待队列的定义如下：

struct __wait_queue {
 unsigned int  flags;
 void   *private;
 wait_queue_func_t func;
 struct list_head task_list;
};

等待队列和等待队列链表头是通过add_wait_queue()结合到一起的。

init_waitqueue_head(&delay_wait);
add_wait_queue(&delay_wait, &wait);

以上代码是将等待队列进程加入到等待队列链表中：

static inline void __add_wait_queue(wait_queue_head_t *head, wait_queue_t *new)
{
 list_add(&new->task_list, &head->task_list);
}

4. 阻塞进程处理

阻塞进程处理包括两部分内容，首先设置进程的睡眠状态，包括TASK_INTERRUPTIBLE和TASK_UNINTERRUPTIBLE两种。前者用于可中断睡眠，后者用于不可中断睡眠。然后，将当前进程退出调度器让出CPU的使用权。

set_current_state(TASK_INTERRUPTIBLE);
schedule();

5. 唤醒处理

唤醒处理通常位于中断处理函数或某些动作成功执行之后，特定条件满足时，唤醒通过阻塞队列睡眠的进程。例如：

void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
       int success, int behind)
{
 if (!bitmap)
  return;
 if (behind) {
  if (atomic_dec_and_test(&bitmap->behind_writes))
   wake_up(&bitmap->behind_wait);
  pr_debug("dec write-behind count %d/%lu\n",
    atomic_read(&bitmap->behind_writes),
    bitmap->mddev->bitmap_info.max_write_behind);
 }
...
}

The above is the detailed content of The processing mechanism for blocking IO processes in Linux drivers. For more information, please follow other related articles on the PHP Chinese website!