Re: [PATCH v3 14/30] vfio/pci: re-introduce CONFIG_VFIO_PCI_ZDEV

[Cornelia Huck]

On Mon, Feb 07 2022, Matthew Rosato <mjrosato@xxxxxxxxxxxxx> wrote:

> On 2/7/22 3:35 AM, Cornelia Huck wrote:
>> On Fri, Feb 04 2022, Matthew Rosato <mjrosato@xxxxxxxxxxxxx> wrote:
>> 
>>> This was previously removed as unnecessary; while that was true, subsequent
>>> changes will make KVM an additional required component for vfio-pci-zdev.
>>> Let's re-introduce CONFIG_VFIO_PCI_ZDEV as now there is actually a reason
>>> to say 'n' for it (when not planning to CONFIG_KVM).
>> 
>> Hm... can the file be split into parts that depend on KVM and parts that
>> don't? Does anybody ever use vfio-pci on a non-kvm s390 system?
>> 
>
> It is possible to split out most of the prior CLP/ vfio capability work 
> (but it would not be a totally clean split, zpci_group_cap for example 
> would need to have an inline ifdef since it references a KVM structure) 
> -- I suspect we'll see more of that in the future.
> I'm not totally sure if there's value in the information being provided 
> today -- this CLP information was all added specifically with 
> userspace->guest delivery in mind.  And to answer your other question, 
> I'm not directly aware of non-kvm vfio-pci usage on s390 today; but that 
> doesn't mean there isn't any or won't be in the future of course.  With 
> this series, you could CONFIG_KVM=n + CONFIG_VFIO_PCI=y|m and you'll get 
> the standard vfio-pci support but never any vfio-pci-zdev extension.

Yes. Remind me again: if you do standard vfio-pci without the extensions
grabbing some card-specific information and making them available to the
guest, you get a working setup, it just always looks like a specific
card, right?

>
> If we wanted to provide everything we can where KVM isn't strictly 
> required, then let's look at what a split would look like:
>
> With or without KVM:
> zcpi_base_cap
> zpci_group_cap (with an inline ifdef for KVM [1])
> zpci_util_cap
> zpci_pfip_cap
> vfio_pci_info_zdev_add_caps	
> vfio_pci_zdev_open (ifdef, just return when !KVM  [1])
> vfio_pci_zdev_release (ifdef, just return when !KVM [1])
>
> KVM only:
> vfio_pci_zdev_feat_interp
> vfio_pci_zdev_feat_aif
> vfio_pci_zdev_feat_ioat
> vfio_pci_zdev_group_notifier
>
> I suppose such a split has the benefit of flexibility / 
> future-proofing...  should a non-kvm use case arrive in the future for 
> s390 and we find we need some s390-specific handling, we're still 
> building vfio-pci-zdev into vfio-pci by default and can just extend that.
>
> [1] In this case I would propose renaming CONFIG_VFIO_PCI_ZDEV as we 
> would once again always be building some part of vfio-pci-zdev with 
> vfio-pci on s390 -- maybe something like CONFIG_VFIO_PCI_ZDEV_KVM (wow 
> that's a mouthful) and then use this setting to check "KVM" in my above 
> split.  Since this setting will imply PCI, VFIO_PCI and KVM, we can then 
> s/CONFIG_VFIO_PCI_ZDEV/CONFIG_VFIO_PCI_ZDEV_KVM/ for the rest of the 
> series (to continue covering cases like we build KVM but not pci, or not 
> vfio-pci)
>
> How does that sound?

Complex :)

I'm not really sure whether it's worth the hassle on an odd chance that
we may want it for a !KVM usecase in the future (that goes beyond the
"base" vfio-pci support.) OTOH, it would be cleaner. I'm a bit torn on
this one.