[pve-devel] [PATCH guest-common/qemu-server/manager/docs v3 0/4] implement experimental vgpu live migration

[Dominik Csapak]

On 5/31/24 13:14, Fiona Ebner wrote:
> Am 19.04.24 um 14:45 schrieb Dominik Csapak:
>> and some useful cleanups
>>
>> Resending even there was not much feedback, because i worked in some
>> minor fixes/changes in the meantime.
>>
>> A user tested the previous patch series and only found one issue with
>> the ui, see the comments on bug #5175
>>
>> https://bugzilla.proxmox.com/show_bug.cgi?id=5175
>>
>>
>> This is implemented for mapped resources. This requires driver and
>> hardware support, but aside from nvidia vgpus there don't seem to be
>> many drivers (if any) that do support that.
>>
>> qemu already supports that for vfio-pci devices, so nothing to be
>> done there besides actively enabling it.
>>
>> Since we currently can't properly test it here and very much depends on
>> hardware/driver support, mark it as experimental everywhere (docs/api/gui).
>> (though i tested the live-migration part manually here by using
>> "exec:cat > /tmp/test" for the migration target, and "exec: cat
>> /tmp/test" as the 'incoming' parameter for a new vm start, which worked ;) )
>>
>> i opted for marking them migratable at the mapping level, but we could
>> theoretically also put it in the hostpciX config instead.
>> (though imho it fits better in the cluster-wide resource mapping config)
>>
>> also the naming/texts could probably be improved, but i think
>> 'live-migration-capable' is very descriptive and i didn't want to
>> use an overly short name for it (which can be confusing, see the
>> 'shared' flag for storages)
>>
>> guest-common 2/4 semi-breaks pve-manager without pve-manager 1/5
>> and qemu-server without 3/10
>> (would always fail for devices capable of mediated devices)
>>
> 
> And guest-common 4/4 fully breaks old qemu-server because it removes the
> find_on_current_node() functions.
> 

true, though we could leave it in, and mark it with a TODO to remove in 9.x ?
but i'll leave that to the packaging people to decide which variant is less
work/hassle ;)

>> guest-common 1,2; qemu-server 1-6; pve-manager 1,2
>> are preparations/cleanups mostly and could be applied independently
>>