Daniel P. Berrangé <[email protected]> writes:

> On Sat, Oct 04, 2025 at 01:33:48PM -0400, Michael S. Tsirkin wrote:
>> On Mon, Sep 29, 2025 at 11:07:06AM +0100, Daniel P. Berrangé wrote:
>> > > Well that's because e.g. kvmtest actually depends on pci-testdev.
>> > > IOW it's actually supported.
>> > 
>> > This again just sounds like a downstream 'support' rationalization.
>> > I'm still not seeing a compelling reason why the vhost user generic
>> > device should be disabled by default in upstream, especially if we
>> > mark it as an experimental device with an x- prefix. 
>> 
>> We can do that. I am still somewhat puzzled by whether making
>> it unsupported/experimental addresses the actual need, which
>> seems to be to expose it to end users?
>
> My interpretation is that simply having the device exist by default
> in QEMU builds is the minimum bar. If we declare it supported, then
> that is a "nice to have"  guarantee for long term.
>
>> Once something is used in the field, we can't take it back
>> whether we added x- to the name or not.
>> 
>> What are your thoughts if it's not marked as experimental?
>
> In general my view is that we can't protect against user foolishness.
> If they provide inappropriate configuration options to this device
> and get broken behaviour, so be it.  If they file bugs against QEMU
> our assistance will be very minimal - they get to do the debugging.
>
> On our side as maintainers, the important question is whether exposing
> this device ties our hands for future code development.
>
> eg would it unacceptably limit our ability to refactor things in future,
> while keeping compat for this exposed device ?

The arguments it exposes aren't really going to change. All it does is
allow you simulate what the boilerplate sets in stone.

> If it would be an undue burden, then it would be worth marking it as
> experimental to give us the freedom to make incompatible changes.

IMHO no I don't think it would be a burden.

>
> If exposing the device has minimal burden anticipated on future work,
> then no need to mark it experimental
>
> With regards,
> Daniel

-- 
Alex Bennée
Virtualisation Tech Lead @ Linaro

Reply via email to