Re: [PATCH v2 3/3] media: pci: intel: ivsc: Add acquire/release API for ivsc

[Hans de Goede]

Hi,

On 2/20/23 09:57, Wu, Wentong wrote:
> 
> 
>> -----Original Message-----
>> From: Sakari Ailus <sakari.ailus@xxxxxxxxxxxxxxx>
>> Sent: Monday, February 20, 2023 4:37 PM
>>
>> Hi Wentong,
>>
>> On Mon, Feb 20, 2023 at 03:50:55AM +0000, Wu, Wentong wrote:
>>>
>>>
>>>> -----Original Message-----
>>>> From: Laurent Pinchart <laurent.pinchart@xxxxxxxxxxxxxxxx>
>>>> Sent: Friday, February 17, 2023 7:19 PM
>>>>
>>>> Hi Wentong,
>>>>
>>>> On Fri, Feb 17, 2023 at 06:10:22AM +0000, Wu, Wentong wrote:
>>>>> On Sent: Wednesday, February 15, 2023 5:46 PM, Laurent Pinchart wrote:
>>>>>> On Wed, Feb 15, 2023 at 10:03:29AM +0100, Hans de Goede wrote:
>>>>>>> On 2/14/23 17:06, Sakari Ailus wrote:
>>>>>>>> On Mon, Feb 13, 2023 at 10:23:47AM +0800, Wentong Wu wrote:
>>>>>>>>> IVSC directly connects to camera sensor on source side, and
>>>>>>>>> on output side it not only connects ISH via I2C, but also
>>>>>>>>> exposes MIPI
>>>>>>>>> CSI-2 interface to output camera sensor data. IVSC can use
>>>>>>>>> the camera sensor data to do AI algorithm, and send the results to
>> ISH.
>>>>>>>>> On the other end, IVSC can share camera sensor to host by
>>>>>>>>> routing the raw camera sensor data to the exposed MIPI
>>>>>>>>> CSI-2 interface. But they can not work at the same time, so
>>>>>>>>> software APIs are defined to sync the ownership.
>>>>>>>>>
>>>>>>>>> This commit defines the interfaces between IVSC and camera
>>>>>>>>> sensor driver in include/linux/ivsc.h. The camera driver
>>>>>>>>> controls ownership of the CSI-2 link and sensor with the
>>>>>>>>> acquire/release APIs. When acquiring camera, lane number
>>>>>>>>> and link freq are also required by IVSC frame router.
>>>>>>>>
>>>>>>>> The more I learn about this system, the more I'm inclined to
>>>>>>>> think this functionality should be exposed as a V4L2 sub-device.
>>>>>>>> IVSC doesn't really do anything to the data (as long as it
>>>>>>>> directs it towards the CSI-2 receiver in the SoC), but it is
>>>>>>>> definitely part of the image pipeline.
>>>>>>>
>>>>>>> Yes I happened to discuss this exact same thing with Laurent
>>>>>>> at FOSDEM and we also came to the conclusion that the IVSC
>>>>>>> chip should be modeled as a V4L2 sub-device.
>>>>>>
>>>>>> Agreed.
>>>>>
>>>>> Thanks for your quick review and conclusion, I'm fresh to media
>>>>> sub-system, is there any convention that I should follow to
>>>>> upstream this kind of v4l2 sub-device driver so that not too much
>>>>> back and forth?
>>>>
>>>> This is a tentative proposal, as I'm not very familiar with the
>>>> hardware
>>>> architecture:
>>>>
>>>> - The subdev should have two pads, a sink pad connected to the camera
>>>>   sensor, and a source pad connected to the CSI-2 receiver in the IPU.
>>>>
>>>> - As for any new driver, the subdev driver should use the active state
>>>>   managed by the v4l2-subdev core. This requires calling
>>>>   v4l2_subdev_init_finalize() at probe time. See
>>>>   https://git.linuxtv.org/media_tree.git/commit/?id=a2514b9a634a for an
>>>>   example of a subdev driver converted to this mechanism.
>>>>
>>>> - As we're talking about CSI-2, the subdev driver should use the streams
>>>>   API that was recently merged, and in particular support the
>>>>   .get_frame_desc(), .set_routing(), .enable_streams() and
>>>>   .disable_streams() operations.
>>>>
>>>> - I don't see a need to support V4L2 controls in the subdev driver, but
>>>>   I may be missing something.
>>>>
>>>> - The driver should be validated with v4l2-compliance, part of
>>>>   v4l-utils.
>>>>
>>>
>>> Thanks for the detail, but I have one more question, during probe of
>>> sensor(v4l2-sudev) driver, it will configure camera sensor connected
>>> to IVSC via I2C, but before that it should acquire camera sensor's
>>> ownership from IVSC, how v4l2 framework guarantee this?
>>
>> Please wrap the lines at 74 characters or so when replying.
> 
> Thanks, I will.
> 
>>
>> Do you mean accessing the sensor via I²C also requires acquiring the sensor
>> from IVSC?
> 
> Yes

Hmm, that is going to be a problem since  we really want to have
independent driver for the 2 which are not aware of each other.

I think that maybe we can model this part of the ivsc functionality
as an i2c-mux. But then we will need to somehow change the parent
of the i2c device for the sensor to the output of this mux ...

I guess this means adding some code (some hack likely) to
drivers/platform/x86/intel/int3472/discrete.c which' proe function 
is already guaranteed to run before the sensor's i2c-client gets
instantiated, because of the ACPI _DEP on the INT3472 ACPI device
in the DSDT.

This is going to be a tricky problem to tackle.

Regards,

Hans