Re: [PATCH] efi/libstub/fdt: Standardize the names of EFI stub parameters

From: Stefano Stabellini
Date: Thu Sep 10 2015 - 12:13:29 EST


On Thu, 10 Sep 2015, Mark Rutland wrote:
> On Thu, Sep 10, 2015 at 02:52:25PM +0100, Stefano Stabellini wrote:
> > On Thu, 10 Sep 2015, Mark Rutland wrote:
> > > On Thu, Sep 10, 2015 at 12:37:57PM +0100, Stefano Stabellini wrote:
> > > > On Thu, 10 Sep 2015, Mark Rutland wrote:
> > > > > > > Does Xen not talk to EFI itself and/or give the kernel a virtual EFI
> > > > > > > interface?
> > > > > >
> > > > > > Xen talks to EFI itself but the interface provided to dom0 is somewhat
> > > > > > different: there are no BootServices (Xen calls ExitBootServices before
> > > > > > running the kernel), and the RuntimeServices go via hypercalls (see
> > > > > > drivers/xen/efi.c).
> > > > >
> > > > > That's somewhat hideous; a non Xen-aware OS wouild presumably die if
> > > > > trying to use any runtime services the normal way? I'm not keen on
> > > > > describing things that the OS cannot use.
> > > >
> > > > I agree that is somewhat hideous, but a non-Xen aware OS traditionally
> > > > has never been able to even boot as Dom0. On ARM it can, but it still
> > > > wouldn't be very useful (one couldn't use it to start other guests).
> > >
> > > Sure, but it feels odd to provide the usual information in this manner
> > > if it cannot be used. If you require Xen-specific code to make things
> > > work, I would imagine this information could be dciscovered in a
> > > Xen-specific manner.
> >
> > We need ACPI (or Device Tree) to find that Xen is available on the
> > platform, so we cannot use Xen-specific code to get the ACPI tables.
>
> I don't understand. The proposition already involves passing a custom DT
> to the OS, implying that Xen knows how to boot that OS, and how to pass
> it a DTB.
>
> Consider:
>
> A) In the DT-only case, we go:
>
> DT -> Discover Xen -> Xen-specific stuff
>
>
> B) The proposition is that un the ACPI case we go:
>
> DT -> EFI tables -> ACPI tables -> Discover Xen -> Xen-specific stuff -> override EFI/ACPI stuff
> \-----------------------/
> (be really cautious here)

Well, yes. To be pedantic "override" here would just be the different
delivery method for RuntimeServices. I guess it still counts.


> C) When you could go:
>
> DT -> Discover Xen -> Xen-specific stuff -> Xen-specific EFI/ACPI discovery

I take you mean discovering Xen with the usual Xen hypervisor node on
device tree. I think that C) is a good option actually. I like it. Not
sure why we didn't think about this earlier. Is there anything EFI or
ACPI which is needed before Xen support is discovered by
arch/arm64/kernel/setup.c:setup_arch -> xen_early_init()?

If not, we could just go for this. A lot of complexity would go away.


> D) If you want to be generic:
> EFI -> EFI application -> EFI tables -> ACPI tables -> Xen-specific stuff
> \------------------------------------------/
> (virtualize these, provide shims to Dom0, but handle
> everything in Xen itself)

I think that this is good in theory but could turn out to be a lot of
work in practice. We could probably virtualize the RuntimeServices but
the BootServices are troublesome.


>
> E) Partially-generic option:
> EFI -> EFI application -> Xen detected by registered GUID -> Xen-specific EFI bootloader stuff -> OS in Xen-specific configuration
>
>
> > > > In any case this should be separate from the shim ABI discussion.
> > >
> > > I disagree; I think this is very much relevant to the ABI discussion.
> > > That's not to say that I insist on a particular approach, but I think
> > > that they need to be considered together.
> >
> > Let's suppose Xen didn't expose any RuntimeServices at all, would that
> > make it easier to discuss about the EFI stub parameters?
>
> It would simply the protocol specific to Xen, certainly.
>
> > In the grant scheme of things, they are not that important, as Ian
> > wrote what is important is how to pass the RSDP.
>
> Unfortunately we're still going to have to care about this eventually,
> even if for something like kexec. So we still need to spec out the state
> of things if this is going to be truly generic.

Fair enough. My position is that if we restrict this to RuntimeServices,
it might be possible, but I still prefer C).
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/