Re: [RFC PATCH] x86/xen: Consider Xen PVH support in CONFIG_XEN_PVHVM

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Tue, Feb 24, 2026 at 10:51:35AM +0000, Teddy Astie wrote:
> It's currently possible to build Linux with CONFIG_PVH|CONFIG_XEN_PVHVM
> and no CONFIG_XEN_PVH. That leads to inconsistent kernels that fails with
> "Missing xen PVH initialization" when booting using PVH boot method or
> display various errors and fail to initialize Xen PV drivers when booting
> with PVH-GRUB.
> 
>     platform_pci_unplug: Xen Platform PCI: unrecognised magic value
>     ...
>     # modprobe xen-blkfront
>     modprobe: ERROR: could not insert 'xen_blkfront': No such device
>     # modprobe xen-netfront
>     modprobe: ERROR: could not insert 'xen_netfront': No such device
> 
> When built without CONFIG_XEN_PVH, PVH-specific logic is disabled, hence when
> booting with e.g PVH-OVMF, Linux assumes we are a HVM guest, even when we 
> aren't
> actually one (in the "with HVM emulated devices" sense).
> 
> As it is actually possible to boot Xen PVH without CONFIG_PVH; and that most
> Xen-related logic exist within CONFIG_XEN_PVHVM; consider PVH guests support
> within CONFIG_XEN_PVHVM instead of CONFIG_XEN_PVH.

So the current CONFIG_PVH selection done by CONFIG_XEN_PVH is moot?

> Keep CONFIG_XEN_PVH as a shortcut to enable PVH boot, ACPI support and PVHVM.
> 
> Signed-off-by: Teddy Astie <teddy.astie@xxxxxxxxxx>
> ---
> Cc: Juergen Gross <jgross@xxxxxxxx>
> Cc: Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx>
> Cc: Oleksandr Tyshchenko <oleksandr_tyshchenko@xxxxxxxx>
> 
> A tentative patch, I'm not sure of the way of dealing with the KConfig part,
> keeping CONFIG_XEN_PVH as a shortcut is interesting, but there may be other
> options.
> 
> There are widespreadly used Linux distributions that have a similar 
> configuration
> to this one, thus exhibit this issue i.e fail to boot.

Do you know the underlying cause of not enabling CONFIG_XEN_PVH?  Is
the default set to n on the defconfig?  Or are distros specifically
disabling this option on purpose?

It seems like a step backwards to merge this into some bigger generic
option, we always try to fine-grain as much as possible.

Maybe you could introduce XEN_HVM meta option, that selects both PVHVM
and PVH?

Thanks, Roger.