Re: [PATCH 0/4] dt-bindings: imx: add nvmem property

[Rob Herring <robh@kernel.org>]

On Sat, Apr 02, 2022 at 01:52:28AM +0000, Peng Fan wrote:
> > Subject: Re: [PATCH 0/4] dt-bindings: imx: add nvmem property
> > 
> > On Thu, Mar 24, 2022 at 12:11:04PM +0100, Uwe Kleine-König wrote:
> > > Hello,
> > >
> > > On Thu, Mar 24, 2022 at 12:20:20PM +0800, Peng Fan (OSS) wrote:
> > > > From: Peng Fan <peng.fan@nxp.com>
> > > >
> > > > To i.MX SoC, there are many variants, such as i.MX8M Plus which
> > > > feature 4 A53, GPU, VPU, SDHC, FLEXCAN, FEC, eQOS and etc.
> > > > But i.MX8M Plus has many parts, one part may not have FLEXCAN, the
> > > > other part may not have eQOS or GPU.
> > > > But we use one device tree to support i.MX8MP including its parts,
> > > > then we need update device tree to mark the disabled IP status
> > "disabled".
> > > >
> > > > In NXP U-Boot, we hardcoded node path and runtime update device tree
> > > > status in U-Boot according to fuse value. But this method is not
> > > > scalable and need encoding all the node paths that needs check.
> > > >
> > > > By introducing nvmem property for each node that needs runtime
> > > > update status property accoridng fuse value, we could use one
> > > > Bootloader code piece to support all i.MX SoCs.
> > > >
> > > > The drawback is we need nvmem property for all the nodes which maybe
> > > > fused out.
> > >
> > > I'd rather not have that in an official binding as the syntax is
> > > orthogonal to status = "..." but the semantic isn't. Also if we want
> > > something like that, I'd rather not want to adapt all bindings, but
> > > would like to see this being generic enough to be described in a
> > > single catch-all binding.
> > >
> > > I also wonder if it would be nicer to abstract that as something like:
> > >
> > > 	/ {
> > > 		fuse-info {
> > > 			compatible = "otp-fuse-info";
> > >
> > > 			flexcan {
> > > 				devices = <&flexcan1>, <&flexcan2>;
> > > 				nvmem-cells = <&flexcan_disabled>;
> > > 				nvmem-cell-names = "disabled";
> > > 			};
> > >
> > > 			m7 {
> > > 				....
> > > 			};
> > > 		};
> > > 	};
> > >
> > > as then the driver evaluating this wouldn't need to iterate over the
> > > whole dtb but just over this node. But I'd still keep this private to
> > > the bootloader and not describe it in the generic binding.
> > 
> > There's been discussions (under the system DT umbrella mostly) about
> > bindings for peripheral enable/disable control/status. Most of the time it is in
> > context of device assignment to secure/non-secure world or partitions in a
> > system (via a partitioning hypervisor).
> > 
> > This feels like the same thing and could use the same binding. But someone
> > has to take into account all the uses and come up with something. One off
> > solutions are a NAK.
> 
> Loop Stefano.
> 
> Per my understanding, system device tree is not a runtime generated device
> tree, in case I am wrong.

I said it was part of 'system DT' discussions, not that you need 'system 
DT'. There's been binding patches on the list from ST for the 'trustzone 
protection controller' if I remember the name right. I think there was 
another proposal too.


> To i.MX, one SoC has many different parts, one kind part may not have
> VPU, another part may not have GPU, another part may be a full feature
> one. We have a device tree for the full feature one, but we not wanna
> introduce other static device tree files for non-full feature parts.
> 
> So we let bootloader to runtime setting status of a device node according
> to fuse info that read out by bootloader at runtime.

Sounds like the same problem for the OS perspective. A device may or may 
not be available to the OS. The reason being because the device is 
assigned to TZ or another core vs. fused off doesn't matter.


> I think my case is different with system device tree, and maybe NXP i.MX
> specific. So I would introduce a vendor compatible node, following Uwe's
> suggestion. We Just need such binding doc and device node in Linux kernel
> tree. The code to scan this node is in U-Boot.

Again, device assignment is a common problem. I'm only going to accept a 
common solution.

Rob