On 11.11.2024 21:07, Stewart Hildebrand wrote:
> On 10/28/24 14:41, Roger Pau Monné wrote:
>> I see previous code also catered for VFs being added without the PF
>> being present, so I assume there was some need for this.
>
> This is exactly the source of the confusion on my part. In the removal
> path, the consensus seems to be that removing a PF with VFs still
> present indicates an error. Then shouldn't the opposite also be true?
> Adding a VF without a PF should also indicate an error.
>
> I see the PF-adding logic was added in 942a6f1376d8 ("x86/PCI-MSI:
> properly determine VF BAR values"). Searching the mailing list archives
> didn't reveal much about it [0].
>
> [0]
> https://lore.kernel.org/xen-devel/[email protected]/
>
> The only time I've observed this path being taken is by manually
> calling PHYSDEVOP_pci_device_add. I've resorted to calling
> PHYSDEVOP_pci_device_{remove,add} from userspace in order to test this,
> because the Linux kernel doesn't behave this way.
The goal was to avoid returning an error when we don't strictly need to.
With the overall adjustment ...
> I can't think of a good rationale for catering to VFs being added
> without a PF, so I'll turn it into an error for the next rev.
... changing this may indeed result in better overall consistency.
Jan
