Re: [RFC PATCH 0/1] dt-bindings: arm: Add scmi_devid paramter for

From: Stefano Stabellini
Date: Fri Mar 04 2022 - 19:45:38 EST


On Fri, 4 Mar 2022, Sudeep Holla wrote:
> (sorry for the delay, had to move my email setup and some mails were
> stuck in outbox and I missed to notice)
>
> On Thu, Feb 24, 2022 at 03:34:01PM -0800, Stefano Stabellini wrote:
> > On Thu, 24 Feb 2022, Stefano Stabellini wrote:
> > > On Thu, 24 Feb 2022, Sudeep Holla wrote:
> > > > On Tue, Feb 22, 2022 at 05:15:49PM +0000, Oleksii Moisieiev wrote:
> > > > > Hi Sudeep,
> > > > >
> > > > > On Tue, Feb 22, 2022 at 04:06:37PM +0000, Sudeep Holla wrote:
> > > > > > Hi Oleksii,
> > > > > >
> > > > > > My initial feedback on this. And thanks Cristian for making it so easy as
> > > > > > you have covered most of the things in depth(which I might have not done
> > > > > > myself that well)
> > > > > >
> > > > > > On Tue, Feb 22, 2022 at 11:00:03AM +0000, Cristian Marussi wrote:
> > > > > > > On Mon, Feb 21, 2022 at 05:26:46PM +0000, Oleksii Moisieiev wrote:
> > > > > > > > Introducing new parameter called scmi_devid to the device-tree bindings.
> > > > > > > > This parameter should be set for the device nodes, which has
> > > > > > > > clocks/power-domains/resets working through SCMI.
> > > > > >
> > > > > > I prefer you had given more details on your usage model here instead of
> > > > > > pointing to the other Xen thread as it helps for someone without much
> > > > > > background on Xen or your use-case to review this.
> > > > > >
> > > > > Let me describe the process in few words:
> > > > > We implemented a new feature, called SCI-mediator in Xen.
> > > > > The proposed implementation allows Guests to communicate with the Firmware using SCMI
> > > > > protocol with SMC as a transport. Other implementation are also
> > > > > possible, such as SCMI-Mailbox, SCPI-mailbox etc.
> > > > >
> > > > > In this feature Xen is the Trusted Agent, which receives the following
> > > > > information in Xen device-tree:
> > > > > 1) All channels should be described, each channel defined as
> > > > > arm,scmi-shmem node;
> > > > > 2) Scmi node arm,scmi-smc with protocols description;
> > > >
> > > > Sounds good so far.
> > > >
> > > > > 3) scmi-devid should be set in nodes, which works through SCMI.
> > > > >
> > > >
> > > > Why is this needed for Guest OS, you need not populate this if Guest OS
> > > > is not required to use it, right ? If it is needed just by Xen hypervisor,
> > > > lets talk about that and why it is bad idea to mix that with general
> > > > SCMI bindings.
> > >
> > > I'll try to help Oleksii by answering here, I hope I am not off the mark
> > > :-)
> > >
> > > I think Sudeep is right, scmi-devid is not needed by the guest OS.
> > >
> > > The host device tree is a more interesting discussion. As the host
> > > device tree is meant to be generic and not tied to a specific version of
> > > Linux, it should fully describe the SCMI interface available. If the
> > > device tree is provided to a Trusted Agent, then it should also have the
> > > scmi-devid information, right?
> > >
> > >
> > > > > On start Xen inits itself as trusted agent and requests agent
> > > > > configuration by using BASE_DISCOVER_AGENT message. This message is sent
> > > > > to each configured channel to get agent_id
> > > > >
> > > > > On Domain creation stage Xen will do the following steps:
> > > > > 1) Assign channel to the Guest and map channel address to the Domain
> > > > > address. For the Domain this address should be the same;
> > > > > 2) Generate arm,scmi-shmem and arm,scmi-smc nodes if needed for Guest
> > > > > device-tree (the device-tree which should be passed to the Guest);
> > > > > 3) Process devices, which are passed through to this Guest and set
> > > > > BASE_SET_DEVICE_PERMISSIONS for the scmi-devid, received from the
> > > > > device-node;
> > > > >
> > > >
> > > > I am confused here. So the Xen knows which devices are assigned to each
> > > > Guest OS but doesn't know device ID for them, but relies on the device
> > > > tree node ?
> > >
> > > Which devices go to which guest OS is a user-provided configuration. For
> > > instance, a user can say: "assing /amba/ethernet@ff0e0000 to dom1". This
> > > is normal and not related to SCMI: when a user configures a static
> > > partitioning system, they decide which resources belong to which domain.
> > >
> > > So Xen is told that /amba/ethernet@ff0e0000 is supposed to go to dom1.
> > > Xen proceeds to map memory and interrupts corresponding to
> > > /amba/ethernet@ff0e0000 to dom1. So far so good. What about SCMI?
> > >
> > > In Oleksii's design, Xen is going to assign one of the available SCMI
> > > channels to dom1 and restrict its permission to only
> > > /amba/ethernet@ff0e0000. To do that, Xen needs to know the scmi-devid of
> > > /amba/ethernet@ff0e0000. As far as I can tell there is nothing
> > > Xen-specific in this activitity, that's why I asked Oleksii to reach out
> > > to the upstream device tree community to improve the generic bindings
> > > for everyone's benefits.
> >
> > Let's leave Linux and Xen aside for the moment. What are other possible
> > Trusted Agents? (Maybe TF-A?) How do they get the scmi-devid? It looks
> > like it was supposed to come from device tree but nobody got around to
> > adding it to the binding because it is not used by Linux?
>
> I do agree we need this info and probably device tree is the way. But what
> I disagree here is that it needs to part of existing SCMI bindings which are
> for the SCMI users only and not for one that may provide the interface(SCMI
> platform/server/arbitrator/passthrough/..whatever). You can have bindings for
> them as part of system device tree initiative and can be merged back to Linux
> if that happens. Or we may even take the whole devicetree bindings out of
> the Linux one day (when all the stars are aligned :) )

I would love to hear Rob's opinion on what I am about to write next on
the topic of whether the binding should be under linux.git.

I am not sure if the policy is that only device tree bindings actively
used by Linux are present under
linux.git/Documentation/devicetree/bindings/. There are a tons of other
projects using device tree and without a central point for keeping these
bindings the specification will shatter. Given that Linux prefers to
keep the bindings under linux.git, then the logic conclusion is that
linux.git/Documentation/devicetree/bindings/ should also hold bindings
not actively used by Linux right at the moment. Especially bindings that
could be used by Linux in the future. Otherwise we risk a new binding
being used by U-boot, Xen, Zephyr and others then Linux introduces an
incompatible version of it. Nobody would win in that situation.


> > After all, we are currently using in Xen a property called
> > "linux,pci-domain". We might as well have Linux in the future use a
> > property called "xen,scmi-devid" to even things out :-)
>
> Sure or we may add a generic one in the future as mentioned in the other
> email for reasons mentioned there.

[...]

> The fact that we don't need this to be part of SCMI OSPM user bindings,
> it is not addressed and can be considered as a gap.
>
> + The reason I want to keep it xen specific at the moment as we had some
> plan to extended the device-id usage in the spec which hasn't progressed
> a bit(I must admit that before you ask), and this addition should not be
> obstruct that future development. If we align with what we define xen
> specific as part of $subject work, we can always define generic binding
> in the future and slowly make the other obsolete over the time.

Keep in mind that device tree is supposed to be backward compatible,
with or without a "xen," prefix. The process of updating the binding and
making the older binding obsolete won't be any different whether the
property is called "super-official-device-id" or "xen,scmi-devid".

(Also, it is not Xen specific, but as we have enstablished, it is for any
Trusted Agents.)

Why not review the bindings now also considering a future Linux use of
it?