Hi Hans,
On 13/01/2016 07:58, Hans Verkuil wrote:
Hi Dave,
On 01/12/2016 09:38 PM, Dave Stevenson wrote:
Hi All.
Apologies for what feels like such a newbie question, but I've failed to
find useful information elsewhere.
I'm one of the ex-Broadcom developers who is still supporting Raspberry
Pi, although I'm not employed by Pi Foundation or Trading.
My aim is to open up that platform by exposing the CSI2 receiver block
(and eventually parts of the ISP) via V4L2. The first use case would be
for the Toshiba TC358743 HDMI to CSI2 converter, but it should be
applicable to any of the other device drivers too.
Sadly it probably won't be upstreamable as it will require the GPU to do
most of the register poking to avoid potential IP issues (Broadcom not
having released the docs for the relevant hardware blocks). In that
regard it will be fairly similar to the existing V4L2 driver for the Pi
camera.
There is now the driver for the TC358743 in mainline, but my stumbling
block is finding a useful example of how to actually use it. The commit
text by Mats Randgaard says it was "tested on our hardware and all the
implemented features works as expected", but I don't know what that
hardware was or how it was used.
It's for Cisco video conferencing equipment, but it basically boils down to
capturing HDMI input over a CSI2 bus. I believe it's on an omap4.
I know Philip Zabel also developed for the TC358743. Philip, do you have a
git tree available that shows how it is used?
An example tree would be great :-)
The media controller API seems to be part of the answer, but that seems
to be a large overhead for an application to have to connect together
multiple sub-devices when it is only interested in images out the back.
The MC is only needed if you have hardware that allows for complex and/or
dynamic internal video routing. For a standard linear video pipeline it
is not needed.
Things may become more complex if I can integrate in the ISP or expose
multiple CSI rx instances, but initially I can keep it simple.
Is there something that sets up default connections that I'm missing?
Somewhere within device tree?
Typically the bridge driver (i.e. the platform driver that sets up the
pipeline and creates the video devices) will use the device tree to find
the v4l2-subdevice(s) it has to load and hooks them into the pipeline.
drivers/media/platform/am437x/am437x-vpfe.c looks to be a decent example
of that.
Perfect. Bedtime reading :-)
Of course, if you have more complex pipelines, then you need to support
the MC.
I have looked at the OMAP4 ISS driver as a vaguely similar device, but
that seemingly covers the image processing pipe only, not hooking in to
the sensor drivers.
Sensor drivers are hooked in in function iss_register_entities(), see the
section for "/* Register external entities */".
OK, stuff squirreled away in iss_platform_data. I'll take a look in more
detail later.
I've also got a slight challenge in that ideally I want the GPU to
allocate the memory, and ARM map that memory (we already have a service
to do that), but I can't see how that would fit in with the the existing
videobuf modes. Any thoughts on how I might be able to support that? The
existing V4L2 driver ends up doing a full copy of every buffer from GPU
memory to ARM, which isn't great for performance.
There may be an option to use contiguous memory and get the GPU to map
that, but it's more involved as I don't believe the supporting code is
on the Pi branch.
The proper way to do this is that the GPU can export buffers as a DMABUF file
descriptor, then import them in V4L2 (V4L2_MEMORY_DMABUF). The videobuf2
v4l2 framework will handle all the details for you, so it is trivial on
the v4l2 side.
If the GPU doesn't support dmabuf, then I'm not sure if you can do this
without horrible hacks.
It doesn't directly support dmabuf. :-( The GPU has full access to RAM,
but doesn't have an IOMMU so buffers have to be contiguous.
I'll have a look into how complex importing the code to map contiguous
ARM memory into GPU space is. I suspect it can be made to work, but
remember it took a long time to get working cleanly at Brcm, and that
was on Android with ION buffers controlled from userspace. Going from
ION to dmabuf shouldn't be too bad, but it will need a bit of work to
sort it from the kernel.
I can always go for the painful copy to start with, and improve the
buffer handling later.
Thanks for the advice.
Dave
Regards,
Hans
Any help much appreciated.
Thanks.
Dave
PS If those involved in the TC358743 driver are reading, a couple of
quick emails over the possibility of bringing the audio in over CSI2
rather than I2S would be appreciated. I can split out the relevant CSI2
ID stream, but have no idea how I would then feed that through the
kernel to appear via ALSA.
--
To unsubscribe from this list: send the line "unsubscribe linux-media" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-media" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
