On Sat, Aug 27, 2022 at 04:48:50PM +0200, Andrew Lunn wrote: > On Fri, Aug 26, 2022 at 09:49:40AM +0200, Oleksij Rempel wrote: > > On Fri, Aug 26, 2022 at 12:27:51AM +0200, Andrew Lunn wrote: > > > > + ieee802.3-pairs: > > > > + $ref: /schemas/types.yaml#/definitions/int8-array > > > > + description: Array of number of twisted-pairs capable to deliver power. > > > > + Since not all circuits are able to support all pair variants, the array of > > > > + supported variants should be specified. > > > > + Note - single twisted-pair PSE is formally know as PoDL PSE. > > > > + items: > > > > + enum: [1, 2, 4] > > > > > > It is not clear to me what you are describing here. It looks like the > > > number of pairs? That does not seem like a hardware property. The > > > controller itself should be able to tell you how many pairs it can > > > feed. > > > > > > A hardware property would be which pairs of the socket are connected > > > to a PSE and so can be used to deliver power. > > > > Good point, this will be needed as well. But not right now. > > That is another point. You are adding properties which no driver > actually uses. That is unusual. > > I think i would rename your current driver to regulator. That is all > it is, and it only needs one property, the regulator itself. Its yaml > description should only have the regulator, and nothing else. > > When other drivers start to be added, we can think about each property > they add, and decided if they are generic, or specific to a > driver/board. Generic properties we can add to one shared .yaml file, > device/board specific properties get added to that drivers .yaml > > > > But i'm not sure how > > > that would be useful to know. I suppose a controller capable of > > > powering 4 pair, but connected to a socket only wired to supply 2, can > > > then disable 2 pairs? > > > > Not only. Here are following reasons: > > - not all boards use a controller in form of IC. Some boards are the > > controller. So, there is no other place to describe, what kind of > > controller this board is. For example - currently there are no known > > ICs to support PoDL (ieee802.3-pairs == 1), early adopters are > > implementing it by using MOSFETs coupled with ADCs and some extra > > logic on CPU side: > > https://www.ti.com/lit/an/snla395/snla395.pdf > > - not all ICs provide a way for advanced communication (I2C, SPI, MDIO). > > Some of them will provide only bootstrapping and some pin status > > feedback: > > https://www.analog.com/media/en/technical-documentation/data-sheets/4279fa.pdf > > - Even if we are able to communicate with the IC, there are still board > > specific limitations. > > I expect each of these will provide some sort of driver. It could be > board specific, or it could be MCU specific if the same MCU is used > multiple times and each implementation looks the same to Linux. I > suppose there could even be a library which implements SCCP via a > bit-banging GPIO line, and it has a binding for the two GPIO? > > And if there is no communication at all with it, you cannot represent > it in Linux, so you don't need to worry about it. > > Each driver should come with its own .yaml file, and we should review > it, and decided are the properties common or not. > > > I hope we can agree that some property is need to tell what kind of PSE > > specification is used by this node. > > > > The next challenge is to name it. We have following options: > > 1. PoE, PoE+, PoE++, 4PPoE, PoDL > > 2. 802.3af, 802.3at, 802.bt, 802.3bu, 802.3cg > > 3. Physical property of this specifications > > > > Option 1 is mostly using marketing names, except of PoDL. This names are > > not used in the ieee 802.3-2018 specification. Systematic research of > > this marketing names would give following results: > > - PoE is about delivering power over two twisted pairs and is related to > > 802.3af and 802.3at specs. > > - PoE+ is about delivering power over two twisted pairs and is related > > only to 802.3at. > > - PoE++ is the same as 4PPoE or power over four twisted pairs and is related > > to 802.3bt. > > - PoDL is related to 802.3bu and 802.3cg. Which is power over one > > twisted pair > > > > All of this names combine different properties: number of twisted pairs > > used to deliver power, maximal supported power by the system and > > recommendation for digital interface to communicate with the PSE > > controller (MDIO registers). Since system I currently use do not follow > > all of this recommendations, it is needed to describe them separately. > > > > Option 2 is interesting only for archaeological investigation. Final > > snapshots of 802.3 specification do not provide mapping of extensions to > > actual parts of the spec. I assume, no software developer will be able > > to properly set the devicetree property by using specification extension > > names. > > > > Option 3 provide exact physical property of implementation by using same > > wording provided by the 802.3-2018 spec. This option is easy to verify > > by reviewing the board schematics and it is easy to understand without > > doing historical analysis of 802.3 spec. > > I would go for option 3. We want well defined concepts, and > specifications provide that. > > > > > + > > > > + ieee802.3-pse-type: > > > > + $ref: /schemas/types.yaml#/definitions/uint8 > > > > + minimum: 1 > > > > + maximum: 2 > > > > + description: PSE Type. Describes classification- and class-capabilities. > > > > + Not compatible with PoDL PSE Type. > > > > + Type 1 - provides a Class 0, 1, 2, or 3 signature during Physical Layer > > > > + classification. > > > > + Type 2 - provides a Class 4 signature during Physical Layer > > > > + classification, understands 2-Event classification, and is capable of > > > > + Data Link Layer classification. > > > > > > Again, the controller should know what class it can support. Why do we > > > need to specify it? What could make sense is we want to limit the > > > controller to a specific type? > > > > If we are using existing controller - yes. But this binding is designed for the > > system where no special PSE IC is used. > > I would expect a discreet implementation to also have a driver. The > specific discreet implementation should have a compatible, and a yaml > file describing whatever properties it needs. And since the driver is > specific to the discreet implementation, it should know what it can do > in terms of PSE Type, etc. > > If the same discrete implementation is used on multiple boards, and > there are board specific limitations, then we need properties to limit > what it can do. Maybe those limits are then described in the shared > .yaml file, since limits like this probably are generic. > > In general, i expect we will end up with two classes of properties: > > Hardware controls: I2C bus address, SPI address, gpios for bit banging > SCCP, GPIOs for turning power on/off and sensing etc. > > Board specific limitations: Max class, Max current, max Type etc. Ok, so current plan is: - rename this driver and binding to pse-regulator - i will integrate the "ieee802.3-pairs" property, since this driver need to know which field it need to fill in the ethtool response (PSE vs PoDL PSE) - compatible will be "ieee802.3-pse-regulator" Correct? Regards, Oleksij -- Pengutronix e.K. | | Steuerwalder Str. 21 | http://www.pengutronix.de/ | 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 | Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |