Re: [PATCH v7 2/2] hw/acpi: Implement the SRAT GI affinity structure

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

On Tue, 5 Mar 2024 08:17:18 +0000
Ankit Agrawal <ankita@nvidia.com> wrote:

> >>> Please add a test.  tests/qtest/bios-tables-test.c
> >>> + relevant table dumps.  
> >>
> >> Here I need to add a test that creates a vfio-pci device and numa
> >> nodes and link using the acpi-generic-initiator object. One thing
> >> here is that the -device vfio-pci needs a host=<bdf> argument. I
> >> probably cannot provide the device bdf from my local setup. So
> >> I am not sure how can I add this test to tests/qtest/bios-tables-test.c.
> >> FYI, the following is a sample args we use for the
> >> acpi-generic-initiator object.
> >>
> >>         -numa node,nodeid=2
> >>         -device vfio-pci-nohotplug,bus=pcie.0,addr=04.0,rombar=0,id=dev0 \
> >>         -object acpi-generic-initiator,id=gi0,pci-dev=dev0,node=2 \
> >>
> >> Moreover based on a quick grep, I don't see any other test that
> >> have -device vfio-pci argument.
> >>
> >> Jonathan, Alex, do you know how we may add tests that is dependent
> >> on the vfio-pci device?  
> >
> > There are none.
> >
> > This would require a host device always available for passthrough and
> > there is no simple solution for this problem. Such tests would need to
> > run in a nested environment under avocado: a pc/virt machine with an
> > igb device and use the PF and/or VFs to check device assignment in a
> > nested guests.
> >
> > PPC just introduced new tests to check nested guest support on two
> > different HV implementations. If you have time, please take a look
> > at tests/avocado/ppc_hv_tests.py for the framework.
> >
> > I will try to propose a new test when I am done with the reviews,
> > not before 9.0 soft freeze though.  
> 
> Thanks for the information. As part of this patch, I'll leave out
> this test change then.

For BIOS table purposes it can be any PCI device. I've been testing
this with a virtio-net-pci but something like virtio-rng-pci will
do fine.  The table contents doesn't care if it's vfio or not.

I can spin a test as part of the follow up Generic Port series that
incorporates both and pushes the limits of the hmat code in general.
Current tests are too tame ;)

Given I don't think we have clarification from ACPI spec side on
the many to one mapping you are using, I'd just use a 1-1 in any
test.


Jonathan