1An introduction to the videobuf layer
2Jonathan Corbet <corbet@lwn.net>
3Current as of 2.6.33
4
5The videobuf layer functions as a sort of glue layer between a V4L2 driver
6and user space.  It handles the allocation and management of buffers for
7the storage of video frames.  There is a set of functions which can be used
8to implement many of the standard POSIX I/O system calls, including read(),
9poll(), and, happily, mmap().  Another set of functions can be used to
10implement the bulk of the V4L2 ioctl() calls related to streaming I/O,
11including buffer allocation, queueing and dequeueing, and streaming
12control.  Using videobuf imposes a few design decisions on the driver
13author, but the payback comes in the form of reduced code in the driver and
14a consistent implementation of the V4L2 user-space API.
15
16Buffer types
17
18Not all video devices use the same kind of buffers.  In fact, there are (at
19least) three common variations:
20
21 - Buffers which are scattered in both the physical and (kernel) virtual
22   address spaces.  (Almost) all user-space buffers are like this, but it
23   makes great sense to allocate kernel-space buffers this way as well when
24   it is possible.  Unfortunately, it is not always possible; working with
25   this kind of buffer normally requires hardware which can do
26   scatter/gather DMA operations.
27
28 - Buffers which are physically scattered, but which are virtually
29   contiguous; buffers allocated with vmalloc(), in other words.  These
30   buffers are just as hard to use for DMA operations, but they can be
31   useful in situations where DMA is not available but virtually-contiguous
32   buffers are convenient.
33
34 - Buffers which are physically contiguous.  Allocation of this kind of
35   buffer can be unreliable on fragmented systems, but simpler DMA
36   controllers cannot deal with anything else.
37
38Videobuf can work with all three types of buffers, but the driver author
39must pick one at the outset and design the driver around that decision.
40
41[It's worth noting that there's a fourth kind of buffer: "overlay" buffers
42which are located within the system's video memory.  The overlay
43functionality is considered to be deprecated for most use, but it still
44shows up occasionally in system-on-chip drivers where the performance
45benefits merit the use of this technique.  Overlay buffers can be handled
46as a form of scattered buffer, but there are very few implementations in
47the kernel and a description of this technique is currently beyond the
48scope of this document.]
49
50Data structures, callbacks, and initialization
51
52Depending on which type of buffers are being used, the driver should
53include one of the following files:
54
55    <media/videobuf-dma-sg.h>		/* Physically scattered */
56    <media/videobuf-vmalloc.h>		/* vmalloc() buffers	*/
57    <media/videobuf-dma-contig.h>	/* Physically contiguous */
58
59The driver's data structure describing a V4L2 device should include a
60struct videobuf_queue instance for the management of the buffer queue,
61along with a list_head for the queue of available buffers.  There will also
62need to be an interrupt-safe spinlock which is used to protect (at least)
63the queue.
64
65The next step is to write four simple callbacks to help videobuf deal with
66the management of buffers:
67
68    struct videobuf_queue_ops {
69	int (*buf_setup)(struct videobuf_queue *q,
70			 unsigned int *count, unsigned int *size);
71	int (*buf_prepare)(struct videobuf_queue *q,
72			   struct videobuf_buffer *vb,
73			   enum v4l2_field field);
74	void (*buf_queue)(struct videobuf_queue *q,
75			  struct videobuf_buffer *vb);
76	void (*buf_release)(struct videobuf_queue *q,
77			    struct videobuf_buffer *vb);
78    };
79
80buf_setup() is called early in the I/O process, when streaming is being
81initiated; its purpose is to tell videobuf about the I/O stream.  The count
82parameter will be a suggested number of buffers to use; the driver should
83check it for rationality and adjust it if need be.  As a practical rule, a
84minimum of two buffers are needed for proper streaming, and there is
85usually a maximum (which cannot exceed 32) which makes sense for each
86device.  The size parameter should be set to the expected (maximum) size
87for each frame of data.
88
89Each buffer (in the form of a struct videobuf_buffer pointer) will be
90passed to buf_prepare(), which should set the buffer's size, width, height,
91and field fields properly.  If the buffer's state field is
92VIDEOBUF_NEEDS_INIT, the driver should pass it to:
93
94    int videobuf_iolock(struct videobuf_queue* q, struct videobuf_buffer *vb,
95			struct v4l2_framebuffer *fbuf);
96
97Among other things, this call will usually allocate memory for the buffer.
98Finally, the buf_prepare() function should set the buffer's state to
99VIDEOBUF_PREPARED.
100
101When a buffer is queued for I/O, it is passed to buf_queue(), which should
102put it onto the driver's list of available buffers and set its state to
103VIDEOBUF_QUEUED.  Note that this function is called with the queue spinlock
104held; if it tries to acquire it as well things will come to a screeching
105halt.  Yes, this is the voice of experience.  Note also that videobuf may
106wait on the first buffer in the queue; placing other buffers in front of it
107could again gum up the works.  So use list_add_tail() to enqueue buffers.
108
109Finally, buf_release() is called when a buffer is no longer intended to be
110used.  The driver should ensure that there is no I/O active on the buffer,
111then pass it to the appropriate free routine(s):
112
113    /* Scatter/gather drivers */
114    int videobuf_dma_unmap(struct videobuf_queue *q,
115			   struct videobuf_dmabuf *dma);
116    int videobuf_dma_free(struct videobuf_dmabuf *dma);
117
118    /* vmalloc drivers */
119    void videobuf_vmalloc_free (struct videobuf_buffer *buf);
120
121    /* Contiguous drivers */
122    void videobuf_dma_contig_free(struct videobuf_queue *q,
123				  struct videobuf_buffer *buf);
124
125One way to ensure that a buffer is no longer under I/O is to pass it to:
126
127    int videobuf_waiton(struct videobuf_buffer *vb, int non_blocking, int intr);
128
129Here, vb is the buffer, non_blocking indicates whether non-blocking I/O
130should be used (it should be zero in the buf_release() case), and intr
131controls whether an interruptible wait is used.
132
133File operations
134
135At this point, much of the work is done; much of the rest is slipping
136videobuf calls into the implementation of the other driver callbacks.  The
137first step is in the open() function, which must initialize the
138videobuf queue.  The function to use depends on the type of buffer used:
139
140    void videobuf_queue_sg_init(struct videobuf_queue *q,
141				struct videobuf_queue_ops *ops,
142				struct device *dev,
143				spinlock_t *irqlock,
144				enum v4l2_buf_type type,
145				enum v4l2_field field,
146				unsigned int msize,
147				void *priv);
148
149    void videobuf_queue_vmalloc_init(struct videobuf_queue *q,
150				struct videobuf_queue_ops *ops,
151				struct device *dev,
152				spinlock_t *irqlock,
153				enum v4l2_buf_type type,
154				enum v4l2_field field,
155				unsigned int msize,
156				void *priv);
157
158    void videobuf_queue_dma_contig_init(struct videobuf_queue *q,
159				       struct videobuf_queue_ops *ops,
160				       struct device *dev,
161				       spinlock_t *irqlock,
162				       enum v4l2_buf_type type,
163				       enum v4l2_field field,
164				       unsigned int msize,
165				       void *priv);
166
167In each case, the parameters are the same: q is the queue structure for the
168device, ops is the set of callbacks as described above, dev is the device
169structure for this video device, irqlock is an interrupt-safe spinlock to
170protect access to the data structures, type is the buffer type used by the
171device (cameras will use V4L2_BUF_TYPE_VIDEO_CAPTURE, for example), field
172describes which field is being captured (often V4L2_FIELD_NONE for
173progressive devices), msize is the size of any containing structure used
174around struct videobuf_buffer, and priv is a private data pointer which
175shows up in the priv_data field of struct videobuf_queue.  Note that these
176are void functions which, evidently, are immune to failure.
177
178V4L2 capture drivers can be written to support either of two APIs: the
179read() system call and the rather more complicated streaming mechanism.  As
180a general rule, it is necessary to support both to ensure that all
181applications have a chance of working with the device.  Videobuf makes it
182easy to do that with the same code.  To implement read(), the driver need
183only make a call to one of:
184
185    ssize_t videobuf_read_one(struct videobuf_queue *q,
186			      char __user *data, size_t count,
187			      loff_t *ppos, int nonblocking);
188
189    ssize_t videobuf_read_stream(struct videobuf_queue *q,
190				 char __user *data, size_t count,
191				 loff_t *ppos, int vbihack, int nonblocking);
192
193Either one of these functions will read frame data into data, returning the
194amount actually read; the difference is that videobuf_read_one() will only
195read a single frame, while videobuf_read_stream() will read multiple frames
196if they are needed to satisfy the count requested by the application.  A
197typical driver read() implementation will start the capture engine, call
198one of the above functions, then stop the engine before returning (though a
199smarter implementation might leave the engine running for a little while in
200anticipation of another read() call happening in the near future).
201
202The poll() function can usually be implemented with a direct call to:
203
204    unsigned int videobuf_poll_stream(struct file *file,
205				      struct videobuf_queue *q,
206				      poll_table *wait);
207
208Note that the actual wait queue eventually used will be the one associated
209with the first available buffer.
210
211When streaming I/O is done to kernel-space buffers, the driver must support
212the mmap() system call to enable user space to access the data.  In many
213V4L2 drivers, the often-complex mmap() implementation simplifies to a
214single call to:
215
216    int videobuf_mmap_mapper(struct videobuf_queue *q,
217			     struct vm_area_struct *vma);
218
219Everything else is handled by the videobuf code.
220
221The release() function requires two separate videobuf calls:
222
223    void videobuf_stop(struct videobuf_queue *q);
224    int videobuf_mmap_free(struct videobuf_queue *q);
225
226The call to videobuf_stop() terminates any I/O in progress - though it is
227still up to the driver to stop the capture engine.  The call to
228videobuf_mmap_free() will ensure that all buffers have been unmapped; if
229so, they will all be passed to the buf_release() callback.  If buffers
230remain mapped, videobuf_mmap_free() returns an error code instead.  The
231purpose is clearly to cause the closing of the file descriptor to fail if
232buffers are still mapped, but every driver in the 2.6.32 kernel cheerfully
233ignores its return value.
234
235ioctl() operations
236
237The V4L2 API includes a very long list of driver callbacks to respond to
238the many ioctl() commands made available to user space.  A number of these
239- those associated with streaming I/O - turn almost directly into videobuf
240calls.  The relevant helper functions are:
241
242    int videobuf_reqbufs(struct videobuf_queue *q,
243			 struct v4l2_requestbuffers *req);
244    int videobuf_querybuf(struct videobuf_queue *q, struct v4l2_buffer *b);
245    int videobuf_qbuf(struct videobuf_queue *q, struct v4l2_buffer *b);
246    int videobuf_dqbuf(struct videobuf_queue *q, struct v4l2_buffer *b,
247		       int nonblocking);
248    int videobuf_streamon(struct videobuf_queue *q);
249    int videobuf_streamoff(struct videobuf_queue *q);
250
251So, for example, a VIDIOC_REQBUFS call turns into a call to the driver's
252vidioc_reqbufs() callback which, in turn, usually only needs to locate the
253proper struct videobuf_queue pointer and pass it to videobuf_reqbufs().
254These support functions can replace a great deal of buffer management
255boilerplate in a lot of V4L2 drivers.
256
257The vidioc_streamon() and vidioc_streamoff() functions will be a bit more
258complex, of course, since they will also need to deal with starting and
259stopping the capture engine.
260
261Buffer allocation
262
263Thus far, we have talked about buffers, but have not looked at how they are
264allocated.  The scatter/gather case is the most complex on this front.  For
265allocation, the driver can leave buffer allocation entirely up to the
266videobuf layer; in this case, buffers will be allocated as anonymous
267user-space pages and will be very scattered indeed.  If the application is
268using user-space buffers, no allocation is needed; the videobuf layer will
269take care of calling get_user_pages() and filling in the scatterlist array.
270
271If the driver needs to do its own memory allocation, it should be done in
272the vidioc_reqbufs() function, *after* calling videobuf_reqbufs().  The
273first step is a call to:
274
275    struct videobuf_dmabuf *videobuf_to_dma(struct videobuf_buffer *buf);
276
277The returned videobuf_dmabuf structure (defined in
278<media/videobuf-dma-sg.h>) includes a couple of relevant fields:
279
280    struct scatterlist  *sglist;
281    int                 sglen;
282
283The driver must allocate an appropriately-sized scatterlist array and
284populate it with pointers to the pieces of the allocated buffer; sglen
285should be set to the length of the array.
286
287Drivers using the vmalloc() method need not (and cannot) concern themselves
288with buffer allocation at all; videobuf will handle those details.  The
289same is normally true of contiguous-DMA drivers as well; videobuf will
290allocate the buffers (with dma_alloc_coherent()) when it sees fit.  That
291means that these drivers may be trying to do high-order allocations at any
292time, an operation which is not always guaranteed to work.  Some drivers
293play tricks by allocating DMA space at system boot time; videobuf does not
294currently play well with those drivers.
295
296As of 2.6.31, contiguous-DMA drivers can work with a user-supplied buffer,
297as long as that buffer is physically contiguous.  Normal user-space
298allocations will not meet that criterion, but buffers obtained from other
299kernel drivers, or those contained within huge pages, will work with these
300drivers.
301
302Filling the buffers
303
304The final part of a videobuf implementation has no direct callback - it's
305the portion of the code which actually puts frame data into the buffers,
306usually in response to interrupts from the device.  For all types of
307drivers, this process works approximately as follows:
308
309 - Obtain the next available buffer and make sure that somebody is actually
310   waiting for it.
311
312 - Get a pointer to the memory and put video data there.
313
314 - Mark the buffer as done and wake up the process waiting for it.
315
316Step (1) above is done by looking at the driver-managed list_head structure
317- the one which is filled in the buf_queue() callback.  Because starting
318the engine and enqueueing buffers are done in separate steps, it's possible
319for the engine to be running without any buffers available - in the
320vmalloc() case especially.  So the driver should be prepared for the list
321to be empty.  It is equally possible that nobody is yet interested in the
322buffer; the driver should not remove it from the list or fill it until a
323process is waiting on it.  That test can be done by examining the buffer's
324done field (a wait_queue_head_t structure) with waitqueue_active().
325
326A buffer's state should be set to VIDEOBUF_ACTIVE before being mapped for
327DMA; that ensures that the videobuf layer will not try to do anything with
328it while the device is transferring data.
329
330For scatter/gather drivers, the needed memory pointers will be found in the
331scatterlist structure described above.  Drivers using the vmalloc() method
332can get a memory pointer with:
333
334    void *videobuf_to_vmalloc(struct videobuf_buffer *buf);
335
336For contiguous DMA drivers, the function to use is:
337
338    dma_addr_t videobuf_to_dma_contig(struct videobuf_buffer *buf);
339
340The contiguous DMA API goes out of its way to hide the kernel-space address
341of the DMA buffer from drivers.
342
343The final step is to set the size field of the relevant videobuf_buffer
344structure to the actual size of the captured image, set state to
345VIDEOBUF_DONE, then call wake_up() on the done queue.  At this point, the
346buffer is owned by the videobuf layer and the driver should not touch it
347again.
348
349Developers who are interested in more information can go into the relevant
350header files; there are a few low-level functions declared there which have
351not been talked about here.  Also worthwhile is the vivi driver
352(drivers/media/platform/vivi.c), which is maintained as an example of how V4L2
353drivers should be written.  Vivi only uses the vmalloc() API, but it's good
354enough to get started with.  Note also that all of these calls are exported
355GPL-only, so they will not be available to non-GPL kernel modules.
356