Hi, this is a second round of RFC patches to move forward on discussion about proper kernel support for the TI DS90UB9xx serializer/deserializer chipsets with I2C address translation. RFCv2 is a major improvement over RFCv1, with several parts rewritten from scratch. I2C address translationis now in its own file, there's decent remote GPIO support, the deser driver is much closer to being complete, I added a minimal serializer driver and, last but not least, forwarding one video stream works. My long-term goal is to be able to model different camera modules [or display or other modules] similarly to beaglebone capes or rpi hats, up to a model where: 1. there can be different camera modules being designed over time 2. there can be different base boards being designed over time 3. there is a standard interconnection between them (mechanical, electrical, communication bus) 4. camera modules and base boards are designed and sold independently (thanks to point 3) 5. a camera module can be removed, and a different model connected, at runtime while the other modules are streaming To allow the required flexibility I introduced the idea of an I2C alias pool that must be defined in device tree. It is the responsibility of the DT writer to fill the pool with addresses that are otherwise unused on the local bus. The pool could not be filled automatically because there might be conflicting chips on the local bus that are unknown to the software, or that are just connected later. Addresses from the pool are assigned to remote I2C slaves at runtime, when they are added on the remote bus. The technical details of how address translation works are documented in patch 2. The big beast is hotplugging. Unfortunately this does not seem to be doable "the right way" at the moment for at least two reasons. First, a v4l2 media graph cannot be modified while the pipe is streaming, and AFAIK this will not be possible soon. Second, because handling hotplug of devicetree-based peripherals would require runtime DT overlay insertion and removal, which is a slowly progressing work, but again not ready currently. To overcome at least some of these limitations I found a compromise. The model that I would consider "the correct one" looks like this: <-- base board --> <------- remote camera module -------> .---------------------. .-----. .-----. | SER | | CPU |----| DES |========|----------.----------| `-----' `-----' FPD | GPIO ctl | I2C adap |----+----+----. Link 3 `---------------------' | | | cable |||| remote I2C slaves remote GPIO pins Here the deserializer (DES) is always present and connected, so it can be probed vie DT during boot. The serializer (SER) is instantiated at runtime when a link is established on the FPD-Link cable and the model detected. An I2C adapter is created under the SER, and all the remote I2C slaves are then instantiated under it. But this would require to stop and modify the v4l2 pipe, including the cameras still connected, just because one of them has been removed (or a cable has gone faulty). The comprimise I took looks like this: .------------------. .-----. .-----. | |========| SER | | CPU |----| DES .----------| `-----' `-----' | | I2C adap |----+----+----. `------------------' | | | remote I2C slaves In this case the I2C adapter (representing the "remote" bus) is instantiated under the DES, and is always present. This stil allows proper hotplugging of the SER, and userspace can still add/remove remote I2C slaves. But it makes it possible to instantiate a sensor and leave it always instantiated, so that the v4l2 pipe is never modified and "believes" the sensor is always there. Of course this opens other issues, and requires non-standard wachanisms to start/stop the sensor and handle missing frames when it is disconnected. My prototype design works thanks to the above structure, a somewhat tweaked sensor driver and a bit of help from userspace. Finally, remote GPIOs. .------------------. .-----. .-----. | |========| SER | | CPU |----| DES .----------| `-----' `-----' | | GPIO ctl | |||| `------------------' remote GPIO pins Similar to the I2C adapter, I chose to instantiate on the DES the GPIO controllers to control the remote GPIOs, even if it looks like it should be under the serializers. This decision allows to have the remote GPIOs described in DT, and always available, so that e.g. the always-instantiated sensor driver DT node can say 'reset-gpios = <&deser 1 2 0>'. That's all, see the patches for the details. References: [0] Vladimir Zapolskiy proposal on other TI chips: https://www.spinics.net/lists/linux-gpio/msg33291.html [1] Kieran Bingham's patches covering Maxim chips: https://www.spinics.net/lists/linux-media/msg142367.html [RFCv1] https://lore.kernel.org/linux-media/20190108223953.9969-1-luca@xxxxxxxxxxxxxxxx/ Luca Ceresoli (6): i2c: core: let adapters be notified of client attach/detach i2c: add I2C Address Translator (ATR) support media: dt-bindings: add DS90UB954-Q1 video deserializer media: dt-bindings: add DS90UB953-Q1 video serializer media: ds90ub954: new driver for TI DS90UB954-Q1 video deserializer media: ds90ub953: new driver for TI DS90UB953-Q1 video serializer .../bindings/media/i2c/ti,ds90ub953-q1.txt | 42 + .../bindings/media/i2c/ti,ds90ub954-q1.txt | 194 ++ drivers/i2c/Kconfig | 9 + drivers/i2c/Makefile | 1 + drivers/i2c/i2c-atr.c | 557 ++++++ drivers/i2c/i2c-core-base.c | 16 + drivers/media/i2c/Kconfig | 24 + drivers/media/i2c/Makefile | 3 + drivers/media/i2c/ds90ub953.c | 354 ++++ drivers/media/i2c/ds90ub954.c | 1575 +++++++++++++++++ include/dt-bindings/media/ds90ub953.h | 16 + include/linux/i2c-atr.h | 82 + include/linux/i2c.h | 16 + 13 files changed, 2889 insertions(+) create mode 100644 Documentation/devicetree/bindings/media/i2c/ti,ds90ub953-q1.txt create mode 100644 Documentation/devicetree/bindings/media/i2c/ti,ds90ub954-q1.txt create mode 100644 drivers/i2c/i2c-atr.c create mode 100644 drivers/media/i2c/ds90ub953.c create mode 100644 drivers/media/i2c/ds90ub954.c create mode 100644 include/dt-bindings/media/ds90ub953.h create mode 100644 include/linux/i2c-atr.h -- 2.17.1