Weekly Edition
Return to the Kernel page
| |||||||||||||
Video4Linux2 part 6b: Streaming I/O
With the read() and write() methods, each video frame is copied between user and kernel space as part of the I/O operation. When streaming I/O is being used, instead, this copying does not happen; instead, the application and the driver exchange pointers to buffers. These buffers will be mapped into the application's address space, making it possible to perform zero-copy frame I/O. There are two different types of streaming I/O buffers:
Note that drivers are not required to support streaming I/O, and, if they do support streaming, they do not have to handle both buffer types. A driver which is more flexible will support more applications; in practice, it seems that most applications are written to use memory-mapped buffers. It is not possible to use both types of buffer simultaneously. We will now delve into the numerous grungy details involved in supporting streaming I/O. Any Video4Linux2 driver writer will need to understand this API; it is worth noting, however, that there is a higher-level API which can help in the writing of streaming drivers. That layer (called video-buf) can make life easier when the underlying device can support scatter/gather I/O. The video-buf API will be discussed in a future installment. Drivers which support streaming I/O should inform the application of that fact by setting the V4L2_CAP_STREAMING flag in their vidioc_querycap() method. Note that there is no way to describe which buffer types are supported; that comes later.
The v4l2_buffer structureWhen streaming I/O is active, frames are passed between the application and the driver in the form of struct v4l2_buffer. This structure is a complicated beast which will take a while to describe. A good starting point is to note that there are three fundamental states that a buffer can be in:
These states, and the operations which cause transitions between them, come together as shown in the diagram below:
The actual v4l2_buffer structure looks like this:
struct v4l2_buffer
{
__u32 index;
enum v4l2_buf_type type;
__u32 bytesused;
__u32 flags;
enum v4l2_field field;
struct timeval timestamp;
struct v4l2_timecode timecode;
__u32 sequence;
/* memory location */
enum v4l2_memory memory;
union {
__u32 offset;
unsigned long userptr;
} m;
__u32 length;
__u32 input;
__u32 reserved;
};
The index field is a sequence number identifying the buffer; it is only used with memory-mapped buffers. Like other objects which can be enumerated in the V4L2 interface, memory-mapped buffers start with index 0 and go up sequentially from there. The type field describes the type of the buffer, usually V4L2_BUF_TYPE_VIDEO_CAPTURE or V4L2_BUF_TYPE_VIDEO_OUTPUT. The size of the buffer is given by length, which is in bytes. The size of the image data contained within the buffer is found in bytesused; obviously bytesused <= length. For capture devices, the driver will set bytesused; for output devices the application must set this field. field describes which field of an image is stored in the buffer; fields were discussed in part 5a of this series. The timestamp field, for input devices, tells when the frame was captured. For output devices, the driver should not send the frame out before the time found in this field; a timestamp of zero means "as soon as possible." The driver will set timestamp to the time that the first byte of the frame was transferred to the device - or as close to that time as it can get. timecode can be used to hold a timecode value, useful for video editing applications; see this table for details on timecodes. The driver maintains a incrementing count of frames passing through the device; it stores the current sequence number in sequence as each frame is transferred. For input devices, the application can watch this field to detect dropped frames. memory tells whether the buffer is memory-mapped or user-space. For memory-mapped buffers, m.offset describes where the buffer is to be found. The specification describes it as "the offset of the buffer from the start of the device memory," but the truth of the matter is that it is simply a magic cookie that the application can pass to mmap() to specify which buffer is being mapped. For user-space buffers, instead, m.userptr is the user-space address of the buffer. The input field can be used to quickly switch between inputs on a capture device - assuming the device supports quick switching between frames. The reserved field should be set to zero. Finally, there are several flags defined:
Buffer setupOnce a streaming application has performed its basic setup, it will turn to the task of organizing its I/O buffers. The first step is to establish a set of buffers with the VIDIOC_REQBUFS ioctl(), which is turned by V4L2 into a call to the driver's vidioc_reqbufs() method:
int (*vidioc_reqbufs) (struct file *file, void *private_data,
struct v4l2_requestbuffers *req);
Everything of interest will be in the v4l2_requestbuffers structure, which looks like this:
struct v4l2_requestbuffers
{
__u32 count;
enum v4l2_buf_type type;
enum v4l2_memory memory;
__u32 reserved[2];
};
The type field describes the type of I/O to be done; it will usually be either V4L2_BUF_TYPE_VIDEO_CAPTURE for a video acquisition device or V4L2_BUF_TYPE_VIDEO_OUTPUT for an output device. There are other types, but they are beyond the scope of this article. If the application wants to use memory-mapped buffers, it will set memory to V4L2_MEMORY_MMAP and count to the number of buffers it wants to use. If the driver does not support memory-mapped buffers, it should return -EINVAL. Otherwise, it should allocate the requested buffers internally and return zero. On return, the application will expect the buffers to exist, so any part of the task which could fail (memory allocation, for example) should be done at this stage. Note that the driver is not required to allocate exactly the requested number of buffers. In many cases there is a minimum number of buffers which makes sense; if the application requests fewer than the minimum, it may actually get more buffers than it asked for. In your editor's experience, for example, the mplayer application will request two buffers, which makes it susceptible to overruns (and thus lost frames) if things slow down in user space. By enforcing a higher minimum buffer count (adjustable with a module parameter), the cafe_ccic driver is able to make the streaming I/O path a little more robust. The count field should be set to the number of buffers actually allocated before the method returns. Setting count to zero is a way for the application to request that all existing buffers be released. In this case, the driver must stop any DMA operations before freeing the buffers or terrible things could happen. It is also not possible to free buffers if they are current mapped into user space. If, instead, user-space buffers are to be used, the only fields which matter are the buffer type and a value of V4L2_MEMORY_USERPTR in the memory field. The application need not specify the number of buffers that it intends to use; since the allocation will be happening in user space, the driver need not care. If the driver supports user-space buffers, it need only note that the application will be using this feature and return zero; otherwise the usual -EINVAL return is called for. The VIDIOC_REQBUFS command is the only way for an application to discover which types of streaming I/O buffer are supported by a given driver.
Mapping buffers into user spaceIf user-space buffers are being used, the driver will not see any more buffer-related calls until the application starts putting buffers on the incoming queue. Memory-mapped buffers require more setup, though. The application will typically step through each allocated buffer and map it into its address space. The first stop is the VIDIOC_QUERYBUF command, which becomes a call to the driver's vidioc_querybuf() method:
int (*vidioc_querybuf)(struct file *file, void *private_data,
struct v4l2_buffer *buf);
On entry to this method, the only fields of buf which will be set are type (which should be checked against the type specified when the buffers were allocated) and index, which identifies the specific buffer. The driver should make sure that index makes sense and fill in the rest of the fields in buf. Typically drivers store an array of v4l2_buffer structures internally, so the core of a vidioc_querybuf() method is just a structure assignment. The only way for an application to access memory-mapped buffers is to map them into their address space, so a vidioc_querybuf() call will typically be followed by a call to the driver's mmap() method - this method, remember, is stored in the fops field of the video_device structure associated with this device. How the driver handles mmap() will depend on just how the buffers are set up in the kernel. If the buffer can be mapped up front with remap_pfn_range() or remap_vmalloc_range(), that should be done at this time. For buffers in kernel space, pages can also be mapped individually at page-fault time by setting up a nopage() method in the usual way. A good discussion of handling mmap() can be found in Linux Device Drivers for those who need it. When mmap() is called, the VMA structure passed in should have the address of one of your buffers in the vm_pgoff field - right-shifted by PAGE_SHIFT, of course. It should, in particular, be the offset value that your driver returned in response to a VIDIOC_QUERYBUF call. Please iterate through your list of buffers and be sure that the incoming address matches one of them; video drivers should not be a means by which hostile programs can map arbitrary regions of memory. The offset value you provide can be almost anything, incidentally. Some drivers just return (index<<PAGE_SHIFT), meaning that the incoming vm_pgoff field should just be the buffer index. The one thing you should not do is store the actual kernel-space address of the buffer in offset; leaking kernel addresses into user space is never a good idea. When user space maps a buffer, the driver should set the V4L2_BUF_FLAG_MAPPED flag in the associated v4l2_buffer structure. It must also set up open() and close() VMA operations so that it can track the number of processes which have the buffer mapped. As long as this buffer remains mapped somewhere, it cannot be released back to the kernel. If the mapping count of one or more buffers drops to zero, the driver should also stop any in-progress I/O, as there will be no process which can make use of it.
Streaming I/OSo far we have looked at a lot of setup without the transfer of a single frame. We're getting closer, but there is one more step which must happen first. When the application obtains buffers with VIDIOC_REQBUFS, those buffers are all in the user-space state; if they are user-space buffers, they do not really even exist yet. Before the application can start streaming I/O, it must put at least one buffer into the driver's incoming queue; for an output device, of course, those buffers should also be filled with valid frame data. To enqueue a buffer, the application will issue a VIDIOC_QBUF ioctl(), which the V4L2 maps into a call to the driver's vidioc_qbuf() method:
int (*vidioc_qbuf) (struct file *file, void *private_data,
struct v4l2_buffer *buf);
For memory-mapped buffers, once again, only the type and index fields of buf are valid. The driver can just perform the obvious checks (type and index make sense, the buffer is not already on one of the driver's queues, the buffer is mapped, etc.), put the buffer on its incoming queue (setting the V4L2_BUF_FLAG_QUEUED flag), and return. User-space buffers can be more complicated at this point, because the driver will have never seen this buffer before. When using this method, applications are allowed to pass a different address every time they enqueue a buffer, so the driver can do no setup ahead of time. If your driver is bouncing frames through a kernel-space buffer, it need only make a note of the user-space address provided by the application. If you are trying to DMA the data directly into user-space, however, life is significantly more challenging. To ship data directly into user space, the driver must first fault in all of the pages of the buffer and lock them into place; get_user_pages() is the tool to use for this job. Note that this function can perform significant amounts of memory allocation and disk I/O - it could block for a long time. You will need to take care to ensure that important driver functions do not stall while get_user_pages(), which can block for long enough for many video frames to go by, does its thing. Then there is the matter of telling the device to transfer image data to (or from) the user-space buffer. This buffer will not be contiguous in physical memory - it will, instead, be broken up into a large number of separate 4096-byte pages (on most architectures). Clearly, the device will have to be able to do scatter/gather DMA operations. If the device transfers full video frames at once, it will need to accept a scatterlist which holds a great many pages; a VGA-resolution image in a 16-bit format requires 150 pages. As the image size grows, so will the size of the scatterlist. The V4L2 specification says:
If required by the hardware the driver swaps memory pages within
physical memory to create a continuous area of memory. This happens
transparently to the application in the virtual memory subsystem of
the kernel.
Your editor, however, is unwilling to recommend that driver writers attempt this kind of deep virtual memory trickery. A more promising approach could be to require user-space buffers to be located in hugetlb pages, but no drivers do that now. If your device transfers images in smaller pieces (a USB camera, for example), direct DMA to user space may be easier to set up. In any case, when faced with the challenges of supporting direct I/O to user-space buffers, the driver writer should (1) be sure that it is worth the trouble, given that applications tend to expect to use memory-mapped buffers anyway, and (2) make use of the video-buf layer, which can handle some of the pain for you. Once streaming I/O starts, the driver will grab buffers from its incoming queue, have the device perform the requested transfer, then move the buffer to the outgoing queue. The buffer flags should be adjusted accordingly when this transition happens; fields like the sequence number and time stamp should also be filled in at this time. Eventually the application will want to claim buffers in the outgoing queue, returning them to the user-space state. That is the job of VIDIOC_DQBUF, which becomes a call to:
int (*vidioc_dqbuf) (struct file *file, void *private_data,
struct v4l2_buffer *buf);
Here, the driver will remove the first buffer from the outgoing queue, storing the relevant information in *buf. Normally, if the outgoing queue is empty, this call should block until a buffer becomes available. V4L2 drivers are expected to handle non-blocking I/O, though, so if the video device has been opened with O_NONBLOCK, the driver should return -EAGAIN in the empty-queue case. Needless to say, this requirement also implies that the driver must support poll() for streaming I/O. The only remaining step is to actually tell the device to start performing streaming I/O. The Video4Linux2 driver methods for this task are:
int (*vidioc_streamon) (struct file *file, void *private_data,
enum v4l2_buf_type type);
int (*vidioc_streamoff)(struct file *file, void *private_data,
enum v4l2_buf_type type);
The call to vidioc_streamon() should start the device after checking that type makes sense. The driver can, if need be, require that a certain number of buffers be in the incoming queue before streaming can be started. When the application is done it should generate a call to vidioc_streamoff(), which must stop the device. The driver should also remove all buffers from both the incoming and outgoing queues, leaving them all in the user-space state. Of course, the driver must be prepared for the application to simply close the device without stopping streaming first. (Log in to post comments)
|
|||||||||||||
![[Buffer states]](/images/ns/kernel/v4l2_buffers.png)