Re: [PATCH V3 1/6] dt-bindings: nvmem: layouts: add U-Boot environment variables layout

[Miquel Raynal <miquel.raynal@bootlin.com>]

Hi Rob,

robh@kernel.org wrote on Mon, 15 Jan 2024 11:09:03 -0600:

> On Thu, Jan 04, 2024 at 10:10:13AM +0100, Rafał Miłecki wrote:
> > On 4.01.2024 08:58, Miquel Raynal wrote:
> > > robh@kernel.org wrote on Wed, 3 Jan 2024 17:11:29 -0700:
> > > > On Thu, Dec 21, 2023 at 06:34:16PM +0100, Rafał Miłecki wrote:
> > > > > From: Rafał Miłecki <rafal@milecki.pl>
> > > > > 
> > > > > U-Boot env data is a way of storing firmware variables. It's a format
> > > > > that can be used of top of various storage devices. Its binding should
> > > > > be an NVMEM layout instead of a standalone device.
> > > > > 
> > > > > This patch adds layout binding which allows using it on top of MTD NVMEM
> > > > > device as well as any other. At the same time it deprecates the old
> > > > > combined binding.
> > > > 
> > > > I don't understand the issue. From a DT perspective, there isn't. A
> > > > partition is not a device, but is describing the layout of storage
> > > > already.
> > > 
> > > Actually I think what Rafał wants to do goes in the right direction but
> > > I also understand from a binding perspective it may be a little
> > > confusing, even more if we consider "NVMEM" a Linux specific concept.
> > > 
> > > There is today a "u-boot env" NVMEM *device* description which
> > > almost sits at the same level as eg. an eeprom device. We cannot
> > > compare "an eeprom device" and "a u-boot environment" of course. But
> > > that's truly what is currently described.
> > > 
> > > * Current situation
> > > 
> > > 	Flash device -> U-Boot env layout -> NVMEM cells
> 
> Isn't it?:
> 
> Flash device -> fixed-partitions -> U-Boot env layout -> NVMEM cells
> 
> > > 
> > > * Improved situation
> > > 
> > > 	Any storage device -> NVMEM -> U-Boot env layout -> NVMEM cells
> 
> Why is this better? We don't need a container to say 'this is NVMEM 
> stuff' or 'this is MTD stuff'. 'U-Boot env layout' can tell us 'this is 
> NVMEM stuff' or whatever the kernel decides in the future.

Yes, I also want the U-boot env layout to tell us "this is nvmem
stuff". But that's not the case today. Today, it says "this is NVMEM
stuff on top of mtd stuff". This was a mistake in the first place, but
this compatible is heavily tight to mtd and cannot work on anything
else. And correcting this is IMO the right direction.

> > > The latter is of course the most relevant description as we expect
> > > storage devices to expose a storage-agnostic interface (NVMEM in
> > > this case) which can then be parsed (by NVMEM layouts) in a storage
> > > agnostic way.
> > > 
> > > In the current case, the current U-Boot env binding tells people to
> > > declare the env layout on top of a flash device (only). The current
> > > description also expects a partition node which is typical to flash
> > > devices. Whereas what we should have described in the first place is a
> > > layout that applies on any kind of NVMEM device.
> > > 
> > > Bonus point: We've been working the last couple years on clarifying
> > > bindings, especially with mtd partitions (with the partitions{}
> > > container) and NVMEM layouts (with the nvmem-layout{} container).
> > > The switch proposed in this patch makes use of the latter, of course.
> > 
> > Thanks Miquèl for filling bits I missed in commit description. Despite
> > years in Linux/DT I still struggle with more complex designs
> > documentation.
> > 
> > 
> > As per Rob's comment I think I see his point and a possible design
> > confusion. If you look from a pure DT perspective then "partitions" and
> > "nvmem-layout" serve a very similar purpose. They describe device's data
> > content structure. For fixed structures we have very similar
> > "fixed-partitions" and "fixed-cells".
> > 
> > If we were to design those bindings today I'm wondering if we couldn't
> > have s/partitions/layout/ and s/nvmem-layout/layout/.
> 
> Why!? It is just a name, and we can't get rid of the old names. We don't 
> need 2 names.

We need 2 names because we are not capturing the same concepts here?

> > Rob: other than having different bindings for MTD vs. NVMEM layouts I
> > think they overall design makes sense. A single device may have content
> > structurized on more than 1 level:
> > 1. You may have fixed layout at top level (multiple partitions)
> > 2. Single partitions may have their own layouts (like U-Boot env data)
> 
> Sure. Partitions is for 1 and Layouts is for 2.
> 
> > Maybe ideally above should look more like:
> > 
> > flash@0 {
> > 	compatible = "<flash-compatible>";
> > 
> > 	layout {
> > 		compatible = "fixed-layout";
> 
> Why does 'partitions' and 'fixed-partitions' not work here?

They do, and that's actually what we use. This example just illustrates
another proposal from Rafal. No panic :)

> 
> > 		#address-cells = <1>;
> > 		#size-cells = <1>;
> > 
> > 		partition@0 {
> > 			reg = <0x0 0x40000>;
> > 			label = "u-boot";
> > 		};
> > 
> > 		partition@40000 {
> > 			reg = <0x40000 0x10000>;
> > 			label = "u-boot-env";
> > 
> > 			layout {
> > 				compatible = "u-boot,env-layout";
> > 			};
> > 		};
> > 
> > 		partition@50000 {
> > 			reg = <0x50000 0x100000>;
> > 			label = "u-boot";
> > 		};
> > 	};
> > };
> > 
> > but I can clearly see a use for nested "layout"s. As I said maybe we
> > just shouldn't be so open in calling those MTD or NVMEM devices as that
> > is kind of Linux specific.
> 
> The overall structure should be agnostic to the subsystem. Specific 
> compatibles like 'u-boot,env' can be tied to a subsystem.
> 
> Maybe some things need to be both MTD and NVMEM. MTD to operate on the 
> opague region and NVMEM to access the contents.
> 
> 
> > I'm not sure if we should try renaming "nvmem-layout" to "layout" or
> > "partitions" in similar way at this point.
> 
> You can't rename. It's an ABI though maybe the whole "nvmem-layout" is 
> new enough we can. It's looking like it was a mistake to accept any of 
> this.

I don't think so.

A partition and a layout are not the same concept, as acknowledged
above. We need both, and we need both because we can encapsulate
both as well:

flash { partitions { partA@x { layout { cell@Y } } } }

Renaming nvmem-layout to layout can be done if you want, I don't mind,
but I don't see the point in doing that.

Thanks,
Miquèl