Re: [PATCH v4 0/6] media: v4l2: Add extended fmt and buffer ioctls

[Helen Koike <helen.koike@collabora.com>]

On 7/20/20 8:57 AM, Tomasz Figa wrote:
> Hi Helen,
> 
> On Fri, Jul 17, 2020 at 1:55 PM Helen Koike <helen.koike@collabora.com> wrote:
>>
>> Hi,
>>
>> I'm sorry for taking too long to submit v4.
>>
>> It is not perfect, not all v4l2-compliance tests passes, but I'd like a review,
>> specially on the API and potential problems, so I can focus on improving implementation
>> and maybe drop the RFC tag for next version.
>>
>> Follow below what changed in v4 and some items I'd like to discuss:
>>
>>
>> * Ioctl to replace v4l2_pix_format
>> ---------------------------------------------------------------------------------
>> During last media summit, we agreed to create ioctls that replace the v4l2_pix_format
>> struct and leave the other structs in the v4l2_format union alone.
>> Thus I refactored the code to receive struct v4l2_ext_pix_format, and I renamed the
>> ioctls, so now we have:
>>
>> int ioctl(int fd, VIDIOC_G_EXT_FMT, struct v4l2_ext_pix_format *argp);
>> int ioctl(int fd, VIDIOC_S_EXT_FMT, struct v4l2_ext_pix_format *argp);
>> int ioctl(int fd, VIDIOC_TRY_EXT_FMT, struct v4l2_ext_pix_format *argp);
>>
>> The only valid types are V4L2_BUF_TYPE_VIDEO_CAPTURE and V4L2_BUF_TYPE_VIDEO_OUTPUT,
>> all the other types are invalid with this API.
>>
>>
>> * Modifiers
>> ---------------------------------------------------------------------------------
>> I understand that unifying DRM and V4L2 pixel formats is not possible, but I'd like
>> to unify the modifiers [1].
>>
>> [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/uapi/drm/drm_fourcc.h#n290
>>
>> Should we use the DRM modifiers directly in the V4L2 API?
>> Or should we move this header to a common place and change the prefix? (which requires
>> us to sync with DRM community).
>> Or should we create a v4l2 header, defining V4L2_ prefixed macros mapping to DRM_
>> macros?
>>
>> For now, patch 1/6 includes drm/drm_fourcc.h and it is using DRM_FORMAT_MOD_*
>>
>> As discussed before, It would be nice to have documentation describing DRM fourcc
>> equivalents (I'm not sure if someone started this already), listing the number of
>> planes per format.
>>
>> We should also document which pixelformats are valid for the EXT_API, since multiplanar
>> and tile versions like V4L2_PIX_FMT_NV12MT_16X16 (which seems equivalent to
>> DRM_FORMAT_MOD_SAMSUNG_16_16_TILE, and could have a more generic name) should be
>> replaced by a modifier.
>>
>> Regarding flags [2] field in struct v4l2_pix_format_mplane [3]:
>> The only defined flag is V4L2_PIX_FMT_FLAG_PREMUL_ALPHA, and it is only used by vsp1 driver.
>> Which I believe could be replaced by a modifier, to avoid another field that changes
>> pixel formats, so I removed it from the EXT API (we can always add it back later with
>> the reserved fields).
>>
>> [2] https://linuxtv.org/downloads/v4l-dvb-apis-new/userspace-api/v4l/pixfmt-reserved.html#format-flags
>> [3] https://linuxtv.org/downloads/v4l-dvb-apis-new/userspace-api/v4l/pixfmt-v4l2-mplane.html?highlight=v4l2_pix_format_mplane#c.v4l2_pix_format_mplane
>>
>> We also discussed to add a new ENUM_FMT_EXT ioctl to return all pixelformats + modifiers
>> combinations. I still didn't add it in this version, but I don't think it affects
>> what is in this RFC and it can be added later.
>>
>>
>> * Buffers/Plane offset
>> ---------------------------------------------------------------------------------
>>
>> My understanding is that inside a memory buffer we can have multiple planes in random
>> offsets.
>> I was comparing with the DRM API [4], where it can have the same dmabuf for multiple
>> planes in different offsets, and I started to think we could simplify our API, so
>> I took the liberty to do some more changes, please review struct v4l2_ext_plane in
>> this RFC.
>>
>> I removed the data_offset, since it is unused (See Laurent's RFC repurposing this
>> field [5]). And comparing to the DRM API, it seems to me we only need a single offset
>> field.
>>
>> We could also check about overlapping planes in a memory buffer, but this is complicated
>> if we use the same memory buffer with different v4l2_ext_buffer objects. We can also leave
>> to the driver to check situations that may cause HW errors.
>>
>> [4] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/uapi/drm/drm_mode.h#n489
>> [5] https://patchwork.linuxtv.org/patch/29177/
>>
>>
>> * Multistream Channels
>> ---------------------------------------------------------------------------------
>> During last media summit, we discussed about adding a channel number to the API to
>> support multistreams. i.e, to have multiple queues through a single video node.
>>
>> Use cases:
>>
>>     - Blitters: can take multiple streams as input, which would require multiple OUTPUT queues.
>>
>>     As Nicolas was explaining me:
>>     "The blitters comes with a lot of variation between hardware. Most blitters at
>>     least support 3 frames buffer. 2 inputs and one output. The second input is usually
>>     optional, as the output buffer data is not always overwritten (e.g. SRC_OVER
>>     blend or 1 input). Some of them have additional solid color or pattern that can
>>     be used too. Advanced blitters will have composition feature, and may support more
>>     input buffers to reduce the added latency that would be normally done through cascading
>>     the operations. Note that each input can have different size and different cropping
>>     region. Many blitters can scale and render to a sub-region of the CAPTURE buffer."
>>
>>     - Multis-calers: can produce multiple streams, which would require multiple CAPTURE queues.
>>
>>     As Nicolas was explaining me:
>>     "This type of HW (or soft IP) is commonly found on HW used to produce internet
>>     streams for fragmented and scalable protocols (HLS, DASH).  Basically they are
>>     used to transform one stream into multiple sized streams prior from being encoded."
>>
>> Modeling as channels allows the API to have synchronized Start/Stop between queues,
>> and also avoid the complexity of using the Media API in a topology with multiple video
>> nodes, which complicates userspace.
>>
>> This requires adding a new channel id in ioctls for formats (G_FMT/S_FMT/TRY_FMT), and
>> also for buffers (QBUF/DBUF).
>> We also need a mechanism to enumerate channels and their properties.
>> Since we don't have a clear view how this would work, for now I'm leaving reserved bits
>> in the structs, so we can add them later.
>>
>>
>> * Timecode
>> ---------------------------------------------------------------------------------
>> During last media summit, we discussed to return the v4l2_timecode field to the API,
>> since Nicolas mentioned that, even if it is not used by any upstreamed driver, it
>> is used by out-of-tree drivers.
>>
>> I've been discussing with Nicolas about this, and we can avoid adding too many metadata
>> to the buffer struct by using the Read-Only Request API [6] for retrieving more information
>> when required, similar to HDR.
>>
>> The RO Request API has the ability to read a control using a request that has already
>> completed, the control value lives as long as the request object. If it's not read
>> (or if there was no request), the data is simply ignored/discard.
>>
>> Since no upstream driver uses the timecode field, there are no conversions that need
>> to be done.
>>
>> [6] https://patchwork.kernel.org/cover/11635927/
>>
>>
>> * Other changes (and some questions) in this version:
>> ---------------------------------------------------------------------------------
>> - Added reserved fields to struct
>>
>> - The only difference between previously proposed VIDIOC_EXT_EXPBUF and VIDIOC_EXPBUF,
>> was that with VIDIOC_EXT_EXPBUF we can export multiple planes at once. I think we
>> can add this later, so I removed it from this RFC to simplify it.
>>
>> - v4l2_buffer [7] has a memory field (enum v4l2_memory [8]). We kept this field in
>> struct v4l2_ext_buffer, buf I was wondering if this shouldn't be in struct v4l2_ext_plane
>> instead.
>>
>> [7] https://linuxtv.org/downloads/v4l-dvb-apis-new/userspace-api/v4l/buffer.html?highlight=v4l2_buffer#c.v4l2_buffer
>> [8] https://linuxtv.org/downloads/v4l-dvb-apis-new/userspace-api/v4l/buffer.html?highlight=v4l2_memory#c.v4l2_memory
>>
>> - In struct v4l2_ext_pix_format, we have:
>>
>>         struct v4l2_plane_ext_pix_format plane_fmt[VIDEO_MAX_PLANES];
>>
>> The number of planes can be deducted from plane_fmt[i].sizeimage != 0, so I removed
>> the num_planes field. Please let me know if we can't use sizeimage for this.
>> In DRM, we know the number of planes from drm_mode_fb_cmd2 by the number of handle
>> args passed which are not 0.
>> This also avoids num_planes to be bigger then VIDEO_MAX_PLANES.
>>
>> - Added flags field to struct v4l2_ext_create_buffers
>>
>>
>> * Fixed bugs here and there
>> ---------------------------------------------------------------------------------
>> I fixed some bugs found with v4l2-compliance (not all of them yet),
>> through script v4l-utils/contrib/test/test-media.
>>
>> I adapted what Boris did for v4l-utils in previous version to this version:
>> https://gitlab.collabora.com/koike/v4l-utils/-/tree/ext-api/wip
>>
>> Boris' questions regarding DMABUF in last version still holds [9].
>>
>> [9] https://patchwork.linuxtv.org/project/linux-media/cover/20191008091119.7294-1-boris.brezillon@collabora.com/
>>
>>
>> Please, let me know your feedback,
>> Helen
>>
>>
>> Boris Brezillon (5):
>>   media: v4l2: Extend pixel formats to unify single/multi-planar
>>     handling (and more)
>>   media: videobuf2: Expose helpers to implement the _ext_fmt and
>>     _ext_buf hooks
>>   media: mediabus: Add helpers to convert a ext_pix format to/from a
>>     mbus_fmt
>>   media: vivid: Convert the capture and output drivers to
>>     EXT_FMT/EXT_BUF
>>   media: vimc: Implement the ext_fmt and ext_buf hooks
>>
>> Hans Verkuil (1):
>>   media: v4l2: Add extended buffer operations
>>
>>  .../media/common/videobuf2/videobuf2-core.c   |   2 +
>>  .../media/common/videobuf2/videobuf2-v4l2.c   | 549 +++++-----
>>  .../media/test-drivers/vimc/vimc-capture.c    |  61 +-
>>  drivers/media/test-drivers/vimc/vimc-common.c |   6 +-
>>  drivers/media/test-drivers/vimc/vimc-common.h |   2 +-
>>  drivers/media/test-drivers/vivid/vivid-core.c |  70 +-
>>  .../test-drivers/vivid/vivid-touch-cap.c      |  26 +-
>>  .../test-drivers/vivid/vivid-touch-cap.h      |   3 +-
>>  .../media/test-drivers/vivid/vivid-vid-cap.c  | 169 +---
>>  .../media/test-drivers/vivid/vivid-vid-cap.h  |  15 +-
>>  .../media/test-drivers/vivid/vivid-vid-out.c  | 193 ++--
>>  .../media/test-drivers/vivid/vivid-vid-out.h  |  15 +-
>>  drivers/media/v4l2-core/v4l2-dev.c            |  50 +-
>>  drivers/media/v4l2-core/v4l2-ioctl.c          | 934 ++++++++++++++++--
>>  include/media/v4l2-ioctl.h                    |  60 ++
>>  include/media/v4l2-mediabus.h                 |  42 +
>>  include/media/videobuf2-core.h                |   6 +-
>>  include/media/videobuf2-v4l2.h                |  21 +-
>>  include/uapi/linux/videodev2.h                | 144 +++
>>  19 files changed, 1650 insertions(+), 718 deletions(-)
> 
> I don't see any documentation being added by this series. I think it's
> been a long standing issue with this series and makes it difficult to
> review the UAPI itself, in separation from the kernel implementation
> details, which is especially important for any non-kernel developers
> willing to provide feedback.

Agreed, I'll re-spin a new version with docs.

Regards,
Helen

> 
> Best regards,
> Tomasz
>