【Camera专题】Qcom-Camera驱动框架浅析(Hal层

1、前言

本文主要研究高通平台Camera驱动和HAL层代码架构，熟悉高通Camera的控制流程。
平台：Qcom-高通平台
Hal版本：【HAL1】
知识点如下：
从HAL层到driver层：研究Camera以下内容
1.打开(open)流程
2.预览(preview)流程
3.拍照(tackPicture)流程

2、Camera软件架构

camera软件架构

由上图可以看出，Android Camera 框架是 client/service 的架构,

• 1.有两个进程:

  client 进程：可以看成是 AP 端,主要包括 JAVA 代码与一些 native c/c++代码;

  service 进 程：”属于服务端,是 native c/c++代码,主要负责和 linux kernel 中的 camera driver 交互,搜集 linuxkernel 中 cameradriver 传上来的数据,并交给显示系统SurfaceFlinger显示。

client 进程与 service 进程通过 Binder 机制通信, client 端通过调用 service 端的接口实现各个具体的功能。

• 2.最下面的是kernel层的驱动，其中按照V4L2架构实现了camera sensor等驱动，向用户空间提供/dev/video0节点，这些设备节点文件，把操作设备的接口暴露给用户空间。

• 3.在往上是HAL层，高通代码实现了对/dev/video0的基本操作，对接了android的camera相关的interface。

2.1 Camera的open流程

2.1.1 Hal层

Android中Camera的调用流程, 基本是 Java -> JNI -> Service -> HAL -> 驱动层。

frameworks/av/services/camera/libcameraservice/device1/CameraHardwareInterface.h

status_t initialize(CameraModule *module) {
···
    rc = module->open(mName.string(), (hw_device_t **)&mDevice);
···
}

这里调用module->open开始调用到HAL层，那调用的是哪个方法呢？
我们继续往下看:

hardware/qcom/camera/QCamera2/HAL/wrapper/QualcommCamera.cpp

static hw_module_methods_t camera_module_methods = {
    open: camera_device_open,
};

实际上是调用了camera_device_open函数，为了对调用流程更加清晰的认识，
我画了一张流程图(画图工具:processon):

open流程

open流程图已经很清晰明了，我们关注一些重点函数：
在HAL层的 module->open(mName.string(), (hw_device_t **)&mDevice)层层调用，最终会调用到函数mm_camera_open(cam_obj);

hardware/qcom/camera/QCamera2/HAL/core/src/QCameraHWI.cpp

QCameraHardwareInterface::QCameraHardwareInterface(int cameraId, int mode)
{
···
/* Open camera stack! */
    mCameraHandle=camera_open(mCameraId, &mem_hooks);
    //Preview
    result = createPreview();
    //Record
    result = createRecord();
    //Snapshot
    result = createSnapshot();
    /* launch jpeg notify thread and raw data proc thread */
    mNotifyTh = new QCameraCmdThread();
    mDataProcTh = new QCameraCmdThread();
···
}

分析：new QCameraHardwareInterface()进行初始化：主要做了以下动作：

• 1.打开camera
• 2.creat preview stream、record stream、snapshot stream
• 3.创建2个线程(jpeg notify thread和raw data proc thread)

hardware/qcom/camera/QCamera2/stack/mm-camera-interface/src/mm_camera.c

int32_t mm_camera_open(mm_camera_obj_t *my_obj)
{
···
    my_obj->ctrl_fd = open(dev_name, O_RDWR | O_NONBLOCK);
···

}

在V4L2框架中，Camera被看做一个视频设备，使用open函数打开这个设备：这里以阻塞模式打开Camera。

1. 用非阻塞模式打开摄像头设备

cameraFd = open("/dev/video0", O_RDWR | O_NONBLOCK);

2. 如果用阻塞模式打开摄像头设备，上述代码变为：

cameraFd = open("/dev/video0", O_RDWR);

ps:关于阻塞模式和非阻塞模式

应用程序能够使用阻塞模式或非阻塞模式打开视频设备，如果使用非阻塞模式调用视频设备，
即使尚未捕获到信息，驱动依旧会把缓存（DQBUFF）里的东西返回给应用程序。

那么，接下来就会调用到Kernel层的代码

2.1.2Kernel层

kernel/drivers/media/platform/msm/camera_v2/msm.c

static struct v4l2_file_operations msm_fops = {
  .owner  = THIS_MODULE,
  .open   = msm_open,
  .poll   = msm_poll,
  .release = msm_close,
  .ioctl   = video_ioctl2,
#ifdef CONFIG_COMPAT
  .compat_ioctl32 = video_ioctl2,
#endif
};

实际上是调用了msm_open这个函数，我们跟进去看：

static int msm_open(struct file *filep)
{
···
   /* !!! only ONE open is allowed !!! */
   if (atomic_cmpxchg(&pvdev->opened, 0, 1))
       return -EBUSY;

   spin_lock_irqsave(&msm_pid_lock, flags);
   msm_pid = get_pid(task_pid(current));
   spin_unlock_irqrestore(&msm_pid_lock, flags);

   /* create event queue */
   rc = v4l2_fh_open(filep);
   if (rc  < 0)
       return rc;

   spin_lock_irqsave(&msm_eventq_lock, flags);
   msm_eventq = filep->private_data;
   spin_unlock_irqrestore(&msm_eventq_lock, flags);

   /* register msm_v4l2_pm_qos_request */
   msm_pm_qos_add_request();
···
}

分析：
通过调用v4l2_fh_open函数打开Camera,该函数会创建event队列等进行一些其他操作。

接下来我们跟着log去看:
camera open log

<3>[   12.526811] msm_camera_power_up type 1
<3>[   12.526818] msm_camera_power_up:1303 gpio set val 33
<3>[   12.528873] msm_camera_power_up index 6
<3>[   12.528885] msm_camera_power_up type 1
<3>[   12.528893] msm_camera_power_up:1303 gpio set val 33
<3>[   12.534954] msm_camera_power_up index 7
<3>[   12.534969] msm_camera_power_up type 1
<3>[   12.534977] msm_camera_power_up:1303 gpio set val 28
<3>[   12.540162] msm_camera_power_up index 8
<3>[   12.540177] msm_camera_power_up type 1
<3>[   ·
<3>[   ·
<3>[   ·
<3>[   12.562753] msm_sensor_match_id: read id: 0x5675 expected id 0x5675:
<3>[   12.562763] ov5675_back probe succeeded
<3>[   12.562771] msm_sensor_driver_create_i2c_v4l_subdev camera I2c probe succeeded
<3>[   12.564930] msm_sensor_driver_create_i2c_v4l_subdev rc 0 session_id 1
<3>[   12.565495] msm_sensor_driver_create_i2c_v4l_subdev:120
<3>[   12.565507] msm_camera_power_down:1455
<3>[   12.565514] msm_camera_power_down index 0

分析：
最终就是调用msm_camera_power_up上电，msm_sensor_match_id识别sensor id，调用ov5675_back probe()探测函数去完成匹配设备和驱动的工作，msm_camera_power_down下电！

到此 我们的open流程就结束了！！！

2.2 Camera的preview流程

2.2.1 Hal层

hardware/qcom/camera/QCamera2/HAL/QCamera2HWI.cpp

int QCamera2HardwareInterface::startPreview()
{
···
    int32_t rc = NO_ERROR;
···
    rc = startChannel(QCAMERA_CH_TYPE_PREVIEW);
···
}

这里调用startChannel(QCAMERA_CH_TYPE_PREVIEW)，开启preview流。
接来下看我画的一张流程图:(Hal层)

Preview流程
关注一些重点函数:
hardware/qcom/camera/QCamera2/HAL/QCameraChannel.cpp

int32_t QCameraChannel::start()
{
···
    mStreams[i]->start();//流程1
···
    rc = m_camOps->start_channel(m_camHandle, m_handle);//流程2
···
}

进入QCameraChannel::start()函数开始执行两个流程，分别是
mStreams[i]->start()和m_camOps->start_channel(m_camHandle, m_handle);

流程1：mStreams[i]->start()

1.通过mProcTh.launch(dataProcRoutine, this)开启新线程
2.执行CAMERA_CMD_TYPE_DO_NEXT_JOB分支，
3.从mDataQ队列中取出数据并放入mDataCB中,等待数据返回到对应的stream回调中去，
4.最后向kernel请求数据；

流程2：m_camOps->start_channel(m_camHandle, m_handle);

通过流程图，我们可以清晰的看到，经过一系列复杂的调用用，
最后在mm_camera_channel.c中
调用mm_channel_start(mm_channel_t *my_obj)函数，

来看mm_channel_start做了什么事情：
hardware/qcom/camera/QCamera2/stack/mm-camera-interface/src/mm_camera_channel.c

int32_t mm_channel_start(mm_channel_t *my_obj)
{
···
    /* 需要发送cb，因此启动线程 */
    /* 初始化superbuf队列 */
    mm_channel_superbuf_queue_init(&my_obj->bundle.superbuf_queue);
    /* 启动cb线程，通过cb调度superbuf中 */
    snprintf(my_obj->cb_thread.threadName, THREAD_NAME_SIZE, "CAM_SuperBuf");
    mm_camera_cmd_thread_launch(&my_obj->cb_thread,
                                    mm_channel_dispatch_super_buf,
                                    (void*)my_obj);
    /* 启动 cmd 线程，作为superbuf接收数据的回调函数*/
    snprintf(my_obj->cmd_thread.threadName, THREAD_NAME_SIZE, "CAM_SuperBufCB");
    mm_camera_cmd_thread_launch(&my_obj->cmd_thread,
                                mm_channel_process_stream_buf,
                                (void*)my_obj);
    /* 为每个strean分配 buf */
    /*allocate buf*/
    rc = mm_stream_fsm_fn(s_objs[i],
                              MM_STREAM_EVT_GET_BUF,
                              NULL,
                              NULL);
    /* reg buf */
    rc = mm_stream_fsm_fn(s_objs[i],
                              MM_STREAM_EVT_REG_BUF,
                              NULL,
                              NULL);
    /* 开启 stream */
    rc = mm_stream_fsm_fn(s_objs[i],
                              MM_STREAM_EVT_START,
                              NULL,
                              NULL);
···
}

过程包括：

• 1.创建cb thread，cmd thread线程以及
• 2.为每个stream分配buf
• 3.开启stream；
  我们继续关注开启stream后的流程：
  rc = mm_stream_fsm_fn(s_objs[i],MM_STREAM_EVT_START,NULL,NULL);
  调用到
  rc = mm_stream_fsm_reg(my_obj, evt, in_val, out_val)
  hardware/qcom/camera/QCamera2/stack/mm-camera-interface/src/mm_camera_stream.c

int32_t mm_stream_fsm_reg(···)
{
···
    case MM_STREAM_EVT_START:
        rc = mm_stream_streamon(my_obj);
···
}

在mm_camera_stream.c中调用mm_stream_streamon(mm_stream_t *my_obj)函数.

向kernel发送v4l2请求，等待数据回调

int32_t mm_stream_streamon(mm_stream_t *my_obj)
{
···
    enum v4l2_buf_type buf_type = V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE;
···
    rc = ioctl(my_obj->fd, VIDIOC_STREAMON, &buf_type);  
···
}

2.2.2 Kernel层

image

kernel/drivers/media/platform/msm/camera_v2/camera/camera.c
通过ioctl的方式，经过层层调用，最后调用到camera_v4l2_streamon();

static int camera_v4l2_streamon(struct file *filep, void *fh,
    enum v4l2_buf_type buf_type)
{
    struct v4l2_event event;
    int rc; 
    struct camera_v4l2_private *sp = fh_to_private(fh);

    rc = vb2_streamon(&sp->vb2_q, buf_type);
    camera_pack_event(filep, MSM_CAMERA_SET_PARM,
        MSM_CAMERA_PRIV_STREAM_ON, -1, &event);

    rc = msm_post_event(&event, MSM_POST_EVT_TIMEOUT);
···
    rc = camera_check_event_status(&event);
    return rc; 
}

分析:通过msm_post_event发生数据请求，等待数据回调。

Preview完整流程图

Preview完整流程图

到此，preview预览流程结束

2.3 Camera的takePicture流程

事实上，tackPicture流程和preview的流程很类似！

以ZSL模式(零延迟模式)为切入点：

2.3.1 Hal层

hardware/qcom/camera/QCamera2/HAL/QCamera2HWI.cpp

int QCamera2HardwareInterface::takePicture()
{
···
    //流程1
    mCameraHandle->ops->start_zsl_snapshot(mCameraHandle->camera_handle,    
        pZSLChannel->getMyHandle());
···
    //流程2
     rc = pZSLChannel->takePicture(numSnapshots);
···
}

进入QCamera2HardwareInterface::takePicture后，会走2个流程：

• 1.mCameraHandle->ops->start_zsl_snapshot(···);

• 2.pZSLChannel->takePicture(numSnapshots);

流程1：

image

经过层层调用，最终会调用到mm_channel_start_zsl_snapshot
hardware/qcom/camera/QCamera2/stack/mm-camera-interface/src/mm_camera_channel.c

int32_t mm_channel_start_zsl_snapshot(mm_channel_t *my_obj)
{
    int32_t rc = 0; 
    mm_camera_cmdcb_t* node = NULL;

    node = (mm_camera_cmdcb_t *)malloc(sizeof(mm_camera_cmdcb_t));
    if (NULL != node) {
        memset(node, 0, sizeof(mm_camera_cmdcb_t));
        node->cmd_type = MM_CAMERA_CMD_TYPE_START_ZSL;

        /* enqueue to cmd thread */
        cam_queue_enq(&(my_obj->cmd_thread.cmd_queue), node);

        /* wake up cmd thread */
        cam_sem_post(&(my_obj->cmd_thread.cmd_sem));
    } else {
        CDBG_ERROR("%s: No memory for mm_camera_node_t", __func__);
        rc = -1;
    }

    return rc;
}

分析:
该函数主要做了2件事情:

• 1 cam_queue_enq(&(my_obj->cmd_thread.cmd_queue), node);入队
• 2 通过cam_sem_post(&(my_obj->cmd_thread.cmd_sem));唤醒cmd线程

这里的node->cmd_type=MM_CAMERA_CMD_TYPE_START_ZSL

hardware/qcom/camera/QCamera2/stack/mm-camera-interface/src/mm_camera_thread.c

static void *mm_camera_cmd_thread(void *data)
{
···
      case MM_CAMERA_CMD_TYPE_START_ZSL:
         cmd_thread->cb(node, cmd_thread->user_data);
···
}

这里cmd_thread->cb是回调函数：
cmd_thread->cb = mm_channel_process_stream_buf，经过层层复杂的回调
最终:
mm_channel_superbuf_skip(ch_obj, &ch_obj->bundle.superbuf_queue);
super_buf = (mm_channel_queue_node_t*)node->data;
将buffer 取出 且释放list中的node，最终将buffer queue给kernel进行下一次填充.

流程2：

image

同样，经过层层调用，最终调用到mm_channel_request_super_buf

hardware/qcom/camera/QCamera2/stack/mm-camera-interface/src/mm_camera_channel.c

int32_t mm_channel_request_super_buf(mm_channel_t *my_obj, uint32_t num_buf_requested)
{
    /* set pending_cnt
     * will trigger dispatching super frames if pending_cnt > 0 */
    /* send cam_sem_post to wake up cmd thread to dispatch super buffer */
    node = (mm_camera_cmdcb_t *)malloc(sizeof(mm_camera_cmdcb_t));
    if (NULL != node) {
        memset(node, 0, sizeof(mm_camera_cmdcb_t));
        node->cmd_type = MM_CAMERA_CMD_TYPE_REQ_DATA_CB;
        node->u.req_buf.num_buf_requested = num_buf_requested;

        /* enqueue to cmd thread */
        cam_queue_enq(&(my_obj->cmd_thread.cmd_queue), node);

        /* wake up cmd thread */
        cam_sem_post(&(my_obj->cmd_thread.cmd_sem));
    } else {
        CDBG_ERROR("%s: No memory for mm_camera_node_t", __func__);
        rc = -1;
    }

    return rc;
}

分析:该函数和流程1一样:

• 1 cam_queue_enq(&(my_obj->cmd_thread.cmd_queue), node);入队
• 2 通过cam_sem_post(&(my_obj->cmd_thread.cmd_sem));唤醒cmd线程

static void *mm_camera_cmd_thread(void *data)
{
···
      case MM_CAMERA_CMD_TYPE_START_ZSL:
      case MM_CAMERA_CMD_TYPE_REQ_DATA_CB:
         cmd_thread->cb(node, cmd_thread->user_data);
···
}

这里和流程1一样，就不再赘述!

2.3.2 Kernel层

int32_t mm_camera_start_zsl_snapshot(mm_camera_obj_t *my_obj)
{
···
    rc = mm_camera_util_s_ctrl(my_obj->ctrl_fd,
             CAM_PRIV_START_ZSL_SNAPSHOT, &value);
···
}

int32_t mm_camera_util_s_ctrl(int32_t fd,  uint32_t id, int32_t *value)
{
···
    rc = ioctl(fd, VIDIOC_S_CTRL, &control);
···
}

kernel/drivers/media/v4l2-core/v4l2-subdev.c

static long subdev_do_ioctl(struct file *file, unsigned int cmd, void *arg)
{
···
    case VIDIOC_S_CTRL:
        return v4l2_s_ctrl(vfh, vfh->ctrl_handler, arg);
···
}

通过ioctl(fd, VIDIOC_S_CTRL, &control)的方式，借助V4L2框架，调用到kernel层，

最终buffer queue给kernel进行下一次填充。

takePicture完整流程图

takePicture完整流程图

Stay hungry，Stay foolish！