[pve-devel] [PATCH qemu-server 1/1] create_vm: assume HA state 'started' when live-restoring guests

[Michael Köppl]

On Fri Nov 21, 2025 at 11:05 AM CET, Thomas Lamprecht wrote:
> Am 21.11.25 um 10:45 schrieb Fiona Ebner:
>> Am 20.11.25 um 5:34 PM schrieb Michael Köppl:
>>> To avoid shutting down the VM when performing a live-restore, consider
>>> live-restore=1 to translate to a HA state of 'started', similar to
>>> start=1.
>>>
>>> Reported-by: Fiona Ebner <f.ebner at proxmox.com>
>>> Signed-off-by: Michael Köppl <m.koeppl at proxmox.com>
>>> ---
>>> Used defined-or since in pve-manager, we only ever set one of the two
>>> values and in any other case I'd give precedence to the 'start'
>>> parameter if it is explicitly set.
>> 
>> This could be part of the commit message. But from looking at the code,
>> it seems like the actual behavior if start=0 and live-restore=1 is still
>> doing a live restore, so the HA state should still be started in that case.
>> 
>>>
>>>  src/PVE/API2/Qemu.pm | 3 ++-
>>>  1 file changed, 2 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/src/PVE/API2/Qemu.pm b/src/PVE/API2/Qemu.pm
>>> index c580bf63..33427ca7 100644
>>> --- a/src/PVE/API2/Qemu.pm
>>> +++ b/src/PVE/API2/Qemu.pm
>>> @@ -1237,7 +1237,8 @@ __PACKAGE__->register_method({
>>>          my $bwlimit = extract_param($param, 'bwlimit');
>>>          my $force = extract_param($param, 'force');
>>>          my $pool = extract_param($param, 'pool');
>>> -        my $start_after_create = extract_param($param, 'start');
>>> +        my $start_after_create = extract_param($param, 'start')
>>> +            // extract_param($param, 'live-restore');
>> 
>> It's already started during create, so using the $start_after_create
>> variable seems like a slight misfit. Why not just also check for
>> $live_restore when setting the HA state when the resource is added?
>> 
>
> We lock the config on restore, or?
> If so, couldn't we handle this in the HA stack and do not shutdown if a
> restore lock is present in the config?

Thought a bit about this for my approach for v2. I implemented v2 like
@Fiona suggested ($start_after_create || $live_restore), but also added
a check that only sets the status of an existing resource instead of
trying to add one when restoring over an existing VM that already is a
HA resource. I'm not sure I understand why this would instead be handled
in the HA stack depending on the existence of the lock, though. Setting
the state correctly should already be enough, no? Or am I missing
something?