[libvirt] [PATCH 0/3] Emit event on lease attach/detach

[Michal Privoznik]

On 4/16/19 1:39 AM, Cole Robinson wrote:
> On 4/5/19 3:57 AM, Michal Privoznik wrote:
>> Unfortunately, we can't emit VIR_DOMAIN_EVENT_ID_DEVICE_ADDED or
>> VIR_DOMAIN_EVENT_ID_DEVICE_REMOVED event because that carries device
>> alias within itself and leases don't have one.
>>
> 
> Hmm. I understand that <leases> aren't really devices so it doesn't have
> any direct reason to give it an info structure and assign it an alias.
> But that's really an argument for why they shouldn't have been  a device
> to begin with. I presume we did it that way to take advantage of the
> existing hotplug APIs. But then adding a special purpose event seems
> like going in the opposite direction.
> 
> How invasive is it to add an 'info' bit to lease devices, allocate
> aliases with the rest of qemu devices, maybe add an ADDRESS_TYPE_LEASE
> or something that is just an explicit no op for the ADDRESS consumers.
> My guess is it's not too bad but there could be dragons lurking
> 
> There's already places internally where this would simplify things: it's
> annoying that device iterator code already needs to play games like
> 'check the address of all devices, oh except leases and graphics which
> don't have info, oh and hostdev info is a pointer for some reason'

Yeah, it's annoying, but I remember discussion from when I was 
introducing user aliases. I wanted to just have some dummy element that 
would not mean anything to libvirt but users could set it and then match 
devices using it as a mark. It was rejected because it doesn't map onto 
anything in qemu. Well, nor do leases, nor would their address.

I don't have a preference here really. But if we wanted to invent 
addresees for leases then I guess we would have to do it for already 
running domains too (when daemon restarts).

I guess we need somebody else's input here. Dan/Peter?

Michal