question: CSI on imx8mq with (any) CSI2 camera / experience with mx6s_capture?

Martin Kepplinger martin.kepplinger at
Tue Dec 22 10:55:34 EST 2020

On 22.12.20 16:35, Laurent Pinchart wrote:
> Hi Martin,
> On Tue, Dec 22, 2020 at 03:06:28PM +0100, Martin Kepplinger wrote:
>> On 10.12.20 17:04, Laurent Pinchart wrote:
>>> On Thu, Dec 10, 2020 at 02:12:30PM +0100, Martin Kepplinger wrote:
>>>> On 10.12.20 13:17, Laurent Pinchart wrote:
>>>>> On Thu, Dec 10, 2020 at 11:24:00AM +0200, Laurent Pinchart wrote:
>>>>>> On Thu, Dec 10, 2020 at 09:17:48AM +0100, Martin Kepplinger wrote:
>>>>>>> hi,
>>>>>>> TL;DR: did you use the NXP "mx6s_capture" csi bridge driver with other
>>>>>>> cameras?
>>>>>> I've recently worked on camera support for i.MX8MM (whose camera IP
>>>>>> cores are, if not identical, very similar to the i.MX8MQ's). The least I
>>>>>> can say is that it was painful :-(
>>>>>> I'm using an MT9M114 sensor, which can produce RAW8, RAW10 and YUV and
>>>>>> has a CSI-2 interface. My first use case is to capture RAW10, which
>>>>>> isn't supported by the mx6s_capture driver.
>>>> so did you successfully use the NXP mx6s_capture driver with that sensor
>>>> too?
>>> I haven't tried. The mx6s_capture driver and the mainline driver share a
>>> similar architecture, as they came from the same code base. They have
>>> diverged, with the  mainline driver receiving bug fixes and new
>>> features, and my large no-yet-upstreamed patch series adds lots of fixes
>>> that are required for the mt9m114 driver to work with the driver. For
>>> that reason, I've considered that porting the mt9m114 driver to the NXP
>>> mx6s_capture, and fixing the same issues in the mx6s_capture driver than
>>> were present in the mainline driver, would require lots of time for
>>> likely very little gain.
>>> That being said, I've compared the two drivers, and I haven't seen
>>> anything in mx6s_capture that would address the specific issues I'm
>>> facing. I've ordered yesterday an i.MX8MM EVK with an OV5640 camera
>>> module, and I will test those with the mainline driver. If they don't
>>> work, and assuming the NXP kernel driver works on that platform, I'll
>>> have two code bases to compare.
>>>>>>> I try to use a CSI2 camera (hi846 I'm writing a driver for) on imx8mq:
>>>>>>> Using NXP's CSI bridge driver
>>>>>>> as well as the CSI driver itself:
>>>>>>> works fine when using the ov5640 camera with this driver:
>>>>>>> (I realize there is a CSI bridge driver in staging, but that need more
>>>>>>> work to be actually used. Of course after this the goal is to fix and
>>>>>>> use it; and mainline a CSI phy driver too.)
>>>>>> I have lots of patches for this driver, which I've developed on i.MX7D
>>>>>> for a separate project. I'd like to mainline them, but this is blocked
>>>>>> by one last issue that I haven't been able to solve yet. In a nutshell,
>>>>>> the CSI writes two consecutive frames in each buffer, overflowing the
>>>>>> allocated memory. The registers that control the buffer size seem to be
>>>>>> programmed correctly as far as I can tell. I've reported this issue to
>>>>>> NXP but haven't received any feedback yet.
>>>> that's different from where I am. So you don't get any interrupt (EOF or
>>>> other) when only *one* frame is written into one buffer?
>>> I don't get any FB1_DMA_DONE or FB2_DMA_DONE interrupt at least. I
>>> haven't checked other interrupts. I thus get half of the frame rate. If
>>> I configure the CSI_CSIIMAG_PARA register with half of the actual image
>>> height, buffers then contain a single image and the frame rate goes to
>>> the expected value. That gives me a workaround, but without
>>> understanding the root cause, I'm worried that just halving the height
>>> would introduce other breakages.
>> Hi Laurent,
>> Thanks a lot, exchanging things this way alone is much appreciated. I'll
>> give you detailed feedback on the staging drivers later - I tried
>> running them, they probe but still a (gstreamer) stream would not start.
>> But most all of your addition make a lot of sense.
>> Just so you know, to emphasize that once more: I'm still on the nxp
>> drivers (yes, they are very similar) just because the ov56 cam works
>> with them over here.
> I've recently bought an i.MX8MM EVK, with the OV5647 camera module, and
> when I try to capture images, with the NXP kernel, I just don't get any.
> *sigh* I'll have to debug that.

oh you have it already. very good.

I guess the nxp kernel directly *should* work, but I use a kernel based 
on v5.9.X and include the drivers like so:

>> The main problem I have is that, as soon as I try
>> to use a different pixelformat than V4L2_PIX_FMT_YUYV that basically
>> nothing works anymore. userspace stops to "be interested" pretty soon.
>> If I do (wrongly) use that pixelformat, I do get (only) one frame (dma)
>> completed, the usual way.
>> And when I break the ov56 by using a wrong pixelformat in mx6s_capture
>> (nxp bridge driver), I get equally get only one frame (dma) completed there.
>> How can I make this thing use a different pixelformat (by setting the
>> media bus format in the sensor driver)?
> All the components have to support the format you want to capture. This
> includes the CSI driver, the MIPI_CSI driver, and the OV5647 driver. The
> pixel format only matters in the CSI driver, for the other two it's the
> media bus format that matters. To capture raw bayer data, you'll first
> need to extend the OV5640 driver to support the corresponding media bus
> format (and of course configure the sensor to output that format), then
> the MIPI_CSI driver to support that media bus format as well, and
> finally the CSI driver.

ok, the sensor I work with outputs raw bayer only, and that sounds like 
I should pay more attention to the mipi csi driver.

that links to my experiments I occasionally update:

>> thanks again for sharing your code early. have a nice rest of the year,


More information about the linux-arm-kernel mailing list