Re: [PATCH v7 1/2] qom: new object to associate device to numa node

[Jonathan Cameron via <qemu-devel@nongnu.org>]

On Thu, 29 Feb 2024 13:00:27 +0000
Ankit Agrawal <ankita@nvidia.com> wrote:

> >> >>> Jonathan, you pointed out interface design issues in your review of v2.>  
> >> >> Are you fully satisfied with the interface in v3?
> >> >>
> >> >> Yes. I'm fine with the interface in this version (though it's v7, so I'm lost
> >> >> on v2 vs v3!)  
> >> >
> >> > Looks like I can't count to 7!
> >> >
> >> > With NUMA capitalized in the doc comment, QAPI schema
> >> > Acked-by: Markus Armbruster <armbru@redhat.com>
> >> >
> >> > Thanks!  
> >>
> >> Thanks! Will fix that in the next version.  
> >
> > The following is really me arguing with myself, so can probably be
> > ignored, but maybe it will spark an idea from someone else!
> >
> > One trivial tweak that might make our life easier if anyone adds
> > support in the future for the other device handle type might be to go
> > with simply dev rather than pci-dev.
> >
> > There is a sticky corner though if a device is a PCI device
> > and in ACPI DSDT so maybe we are better off adding acpi-dev
> > to take either pci-dev or acpi-dev?  
> 
> That use case does complicate the situation. Do you of any such
> use case for generic initiator?

In physical systems yes - in QEMU not yet, though it's a quirk
of the available ids to get to the ACPI devices (which oddly
are PCI devices :()

> 
> As for your suggestion of using acpi-dev as the arg to take both
> pci-dev and acpi-dev.. Would that mean sending a pure pci device
> (not the corner case you mentioned) through the acpi-dev argument
> as well? Not sure if that would appropriate.

Ah, looking up my description is unclear. I meant two optional parameters
pci-dev or acpi-dev - which one was supplied would indicate the type
of handle to be used.

> 
> > Annoyingly for generic ports, (I'm reusing this infrastructure here)
> > the kernel code currently only deals with the ACPI form (for CXL host
> > bridges).  Given I point that at the bus of a PXB_CXL it is both
> > a PCI device, and the only handle we have for getting to the
> > Root Bridge ACPI handle.  
> 
> So IIUC, you need to pass a PCI device to the generic port object, but use
> that to reach the ACPI handle and build the Generic port affinity structure
> for an ACPI device?

Yes.  Slightly shortcut is that the UID is the bus number for all the
relevant devices so I can abuse that.  QEMU doesn't keep track of
the ACPI handles directly so this is the current cleanest solution.

> 
> > So I think I've argued myself around to thinking we need to extend
> > the interface with another optional parameter if we ever do support
> > the ACPI handle for generic initiators :(
> >
> > Jonatha