[pve-devel] [PATCH qemu-server 1/1] fix 1734: clone VM: if deactivation fails demote error to warning

[Friedrich Weber]

On 06/03/2024 13:40, Fiona Ebner wrote:
> Am 06.03.24 um 11:47 schrieb Hannes Duerr:
>> @@ -3820,7 +3821,13 @@ __PACKAGE__->register_method({
>>  
>>  		if ($target) {
>>  		    # always deactivate volumes - avoid lvm LVs to be active on several nodes
>> -		    PVE::Storage::deactivate_volumes($storecfg, $vollist, $snapname) if !$running;
>> +		    eval {
>> +			PVE::Storage::deactivate_volumes($storecfg, $vollist, $snapname) if !$running;
>> +		    };
>> +		    my $err = $@;
>> +		    if ($err) {
>> +			log_warn("$err\n");
>> +		    }
>>  		    PVE::Storage::deactivate_volumes($storecfg, $newvollist);
> 
> We might also want to catch errors here. Otherwise, the whole clone
> operation (which might've taken hours) can still fail just because of a
> deactivation error. But when failing here, we shouldn't move the config
> file (or the LV can get active on multiple nodes more easily).

I think succeeding but not moving the config file when deactivating
$newvollist fails sounds like it could lead to unexpected behavior.
Right now, when running `qm clone 101 [...] --target node2` on node1
succeeds, one can be sure there will be an VM 101 on node2. But if we
cannot deactivate $newvollist and thus don't move the config file, the
command succeeds but VM 101 instead exists on node1 (correct me if I'm
wrong), which may be confusing e.g. if the clone is automated.

To avoid that, I'd lean towards keeping the behavior of failing the task
if deactivating $newvollist fails. After all, at least in case of LVM
not being able to deactivate because the device is in use, we just
created $newvollist so hopefully nobody else should be accessing it.