Re: [REVIEWv2 PATCH 02/13] vb2: fix handling of data_offset and v4l2_plane.reserved[]

[Hans Verkuil <hansverk@cisco.com>]

On 04/11/2014 03:48 PM, Tomasz Stanislawski wrote:
> On 04/11/2014 03:03 PM, Hans Verkuil wrote:
>> On 04/11/2014 02:48 PM, Tomasz Stanislawski wrote:
>>> On 04/07/2014 03:11 PM, Hans Verkuil wrote:
>>>> From: Hans Verkuil <hans.verkuil@cisco.com>
>>>>
>>>> The videobuf2-core did not zero the 'planes' array in __qbuf_userptr()
>>>> and __qbuf_dmabuf(). That's now memset to 0. Without this the reserved
>>>> array in struct v4l2_plane would be non-zero, causing v4l2-compliance
>>>> errors.
>>>>
>>>> More serious is the fact that data_offset was not handled correctly:
>>>>
>>>> - for capture devices it was never zeroed, which meant that it was
>>>>   uninitialized. Unless the driver sets it it was a completely random
>>>>   number. With the memset above this is now fixed.
>>>>
>>>> - __qbuf_dmabuf had a completely incorrect length check that included
>>>>   data_offset.
>>>
>>> Hi Hans,
>>>
>>> I may understand it wrongly but IMO allowing non-zero data offset
>>> simplifies buffer sharing using dmabuf.
>>> I remember a problem that occurred when someone wanted to use
>>> a single dmabuf with multiplanar API.
>>>
>>> For example, MFC shares a buffer with DRM. Assume that DRM device
>>> forces the whole image to be located in one dmabuf.
>>>
>>> The MFC uses multiplanar API therefore application must use
>>> the same dmabuf to describe luma and chroma planes.
>>>
>>> It is intuitive to use the same dmabuf for both planes and
>>> data_offset=0 for luma plane and data_offset = luma_size
>>> for chroma offset.
>>>
>>> The check:
>>>
>>>> -		if (planes[plane].length < planes[plane].data_offset +
>>>> -		    q->plane_sizes[plane]) {
>>>
>>> assured that the logical plane does not overflow the dmabuf.
>>>
>>> Am I wrong?
>>
>> Yes :-)
>>
>> For video capture the data_offset field is set by the *driver*, not the
>> application. In practice data_offset is the size of a header that is in
>> front of the actual image.
>>
>> You cannot use data_offset for the purpose you describe. To do that a new
>> offset field would have to be added (user_offset?). I'm not opposed to
>> that, I think it is a valid use-case for both dmabuf and userptr and
>> even mmap in combination with CREATE_BUFS.
> 
> Ok. What do you think about allowing the user to set this field?
> The driver would be allowed to adjust it.
> This is a slight change of semantics. As long as application
> was forced to treat data_offset as a reserved field (I mean zeroing)
> then this change would be backward compatible.

It's not quite that easy. Suppose the driver adds a 1024 bytes header
before the image (so data_offset is 1024) and the application sets
data_offset to 512. What does that mean? That the driver starts capturing
at start_of_buffer + 512 and then returns 1024+512 as data_offset? Or that
the driver replaces it by 1024 anyway?

Besides, there has never been a requirement that apps set it to 0 for
video capture. They only have to set it for video output. So this doesn't
work anyway.

> 
> Otherwise, a new field could be introduces as you mentioned.
> The name buffer_offset or simply offset might be more
> appropriate than user_offset.
> 
> I can prepare some RFC about this extension.

I would go with a new field.

Regards,

	Hans