Re: [PATCH v7 0/5] Add support for MAX96714/F and MAX96717/F GMSL2 ser/des

[Tomi Valkeinen <tomi.valkeinen@ideasonboard.com>]

Hi,

On 30/04/2024 16:19, Julien Massot wrote:
> Change since v6:
>    - Remove mention of C-PHY for MAX96717, this serializer is D-PHY only
>    - Remove bus-type requirement for MAX96717
>    - Minor changes requested by Sakari
>    - Workaround a MAX96717 issue, which occurs when stopping
>      the CSI source before stopping the MAX96717 CSI receiver.
> 
> Power management is not included in this patchset. The GMSL link is
> not always resuming when the deserializer is suspended without
> suspending the serializer.
> 
> Change since v5:
>   - Reverse fallback logic: max9671{4,7} can fallback to max9671{4,7}F
>   - use const instead of enum for max9671{4,7}f compatible as suggested
> 
> Change since v4:
>   - Add support for MAX96717 and MAX96714 and use them as a fallback for
>     MAX96717F and MAX96714F respectively
>   - The drivers are now compatible with MAX96717 and MAX96714 since no change in
>     the logic is needed
>   - Reference 'i2c-gate' instead of 'i2c-controller' in the bindings
> 
> Change since v3:
> - bindings
>    - Renamed bindings to drop the 'f' suffix
>    - Add bus type to MAX96717 and remove from MAX9674
>    - Add lane-polarities to both bindings
> 
> - drivers
>    - Address changes requested by Sakari in v3
>    - use v4l2_subdev_s_stream_helper for MAX96714
>    - do not init regmap twice in the MAX96714 driver
>    - Fix compilations on 32 bits platforms
> 
> Change since v2:
> - Convert drivers to use CCI helpers
> - Use generic node name
> - Use 'powerdown' as gpio name instead of 'enable'
> - Add pattern generator support for MAX96714
> 
> These patches add support for Maxim MAX96714F deserializer and
> MAX96717F serializer.
> 
> MAX96714F has one GMSL2 input port and one CSI2 4 lanes output port,
> MAX96717F has one CSI2 input port and one GMSL2 output port.
> 
> The drivers support the tunnel mode where all the
> CSI2 traffic coming from an imager is replicated through the deserializer
> output port.
> 
> Both MAX96714F and MAX96717F are limited to a 3Gbps forward link rate
> leaving a maximum of 2.6Gbps for the video payload.

(I see this mail turned out to be a collection of thoughts rather than 
clear questions... Bear with me =))

I know I'm very late to this party, and perhaps these topics have 
already been discussed, but as I will most likely be doing some GMSL 
work in the future I wanted to ask these questions. My main 
questions/concerns are related to the i2c and the representation of the 
links in the DT.

First, I know these particular devices are one input, one output 
serializer and deserializer, so there's not much to do wrt. i2c 
translation/gating. But even here I wonder how does one support a case 
where a single local i2c bus would have two deserializer devices (with 
different i2c addresses), connected to two identical camera modules?

Controlling the deserializers would work fine, but as the serializers 
and the remote peripherals (sensor) would answer to identical i2c 
addresses, it would conflict and not work.

If I understand the HW docs right, a way (maybe there are others?) to 
handle this would be:
- deser probes, but keeps the link disabled by default
- deser reads the initial serializer i2c address from the DT, but also a 
new address which we want the serializer to have (which doesn't conflict 
with the other serializer)
- deser enables the link and immediately (how to be sure the other deser 
driver doesn't do this at the same time?) sends a write to the 
serializer's DEV_ADDR, changing the serializer's i2c address.
- deser can now add the serializer linux i2c device, so that the 
serializer can probe
- the serializer should prevent any remote i2c transactions until it has 
written the SRC_A/B and DST_A/B registers, to get translation for the 
remote peripherals (or maybe the deser driver should do this part too).

Am I on the right track with the above?

Now, maybe having such a HW config, two deserializers on a single i2c 
bus, doesn't happen in real life, but this issue comes up with 
multi-port deserializers. And while those deserializers are different 
devices than what's added in this series, the serializers used may be 
the same as here. This means the serializer drivers and DT bindings 
should be such that multi-port deserializers can be supported.

As I said, I'm late (and new) to this party, and struggling to consume 
and understand all the related specs and drivers, so I hope you can give 
some insight into how all this might be implemented in the future =).

Have you looked at the FPD-Link drivers (ds90ub9xx)? The i2c management 
is a bit different with those (with my current understanding, a bit 
saner...), but I wonder if similar style would help here, or if the 
i2c-atr could be utilized. It would be nice (but I guess not really 
mandatory in any way) to have similar style in DT bindings for all 
ser-des solutions.

To summarize the i2c management on both FPD-Link and GMSL (if I have 
understood it right):

In FPD-Link the deserializer does it all: it has registers for the 
serializer i2c aliases, and for i2c address translation (per port). So 
when the deser probes, it can program suitable i2c addresses (based on 
data from DT), which will be the addresses visible on the main i2c bus, 
and thus there are never any conflicts.

In addition to that, the drivers utilize i2c-atr, which means that new 
linux i2c busses are created for each serializer. E.g. the deser might 
be, say, on i2c bus 4, and also the serializers, via their i2c aliases, 
would be accessible bus 4. When the serializer drivers probe they will 
create new i2c busses with i2c-atr. So with a 4 port deserializer we 
might get i2c busses 5, 6, 7 and 8. The linux i2c devices for remote 
peripherals (sensors mainly) would be created on these busses with their 
real i2c addresses. When a sensor driver does an i2c write to its 
device, the i2c-atr will catch the write, change the address according 
to the translation table, and do an actual write on the i2c bus 4. This 
would result in the deser HW to catch this write, switch the address 
back to the "real" one, and send it to the appropriate serializer, which 
would then send the i2c transaction on its i2c bus.

In GMSL the deser just forwards everything it sees on the i2c bus, if a 
port is enabled. The deser has no other support related to i2c. The 
serializers have DEV_ADDR register which can be used to change the 
address the serializers respond to, and the serializers also have i2c 
translation for two remote peripherals.

But if the i2c translation is used, it would mean that, say, the sensor 
driver would need to use the "virtual" address, not the real one to 
communicate with the sensor device, which doesn't sound right...

You have used i2c-gate for both the deser and the ser. I don't have 
experience with i2c-gate, but how can we manage the serializer i2c 
address and the i2c address translation with it?

One difference with the FPD-Link and this series' DT bindings is that I 
have a "links" node in the deser, instead of just adding the serializers 
under an i2c node. In FPD-Link case this allowed me to better represent 
the hardware and the configuration needed.

So... Perhaps my bottom line question is: do we need something similar 
to what the FPD-Link uses (links, i2c-atr) to fully support GMSL 
devices? And also, if we merge the DT bindings in this series, will we 
have gone into a corner wrt. how we can manage the i2c?

  Tomi