[libvirt] [PATCH] domain_conf: Include the correct console alias

[Michal Privoznik]

On 28.06.2013 17:01, Daniel P. Berrange wrote:
> On Fri, Jun 28, 2013 at 04:45:08PM +0200, Michal Privoznik wrote:
>> On 28.06.2013 15:55, Daniel P. Berrange wrote:
>>> On Fri, Jun 28, 2013 at 03:31:17PM +0200, Michal Privoznik wrote:
>>>> For some crazy backward compatibility, a console can by just an alias to
>>>> a serial device. This is detected in the XML formating function which
>>>> takes the values to format from corresponding serial device. Including
>>>> the device alias. This results in wrong alias being written into the XML
>>>> definition:
>>>>
>>>>     <console type='pty' tty='/dev/pts/5'>
>>>>       ...
>>>>       <alias name='serial0'/>
>>>>     </console>
>>>>
>>>> While holding the correct alias still in the memory, it doesn't matter.
>>>> However, it starts to matter as soon as libvirtd is restarted and the
>>>> (incorrect) alias is read from status file.
> 
> I don't actually see this problem at all.
> 
> Starting with a guest containing
> 
>     <serial type='pty'>
>       <target port='0'/>
>     </serial>
>     <console type='pty'>
>       <target type='serial' port='0'/>
>     </console>
> 
> After virsh start, it contains
> 
>     <serial type='pty'>
>       <source path='/dev/pts/7'/>
>       <target port='0'/>
>       <alias name='serial0'/>
>     </serial>
>     <console type='pty' tty='/dev/pts/7'>
>       <source path='/dev/pts/7'/>
>       <target type='serial' port='0'/>
>       <alias name='serial0'/>
>     </console>
> 
> 
> The XML in $XDG_RUNTIME_DIR/libvirt/qemu/run/vm1.xml contains
> exactly the same.
> 
> If i now kill libvirtd and restart it, the XML is still
> reported correctly.
> 
> How do you reproduce the bug ?

The problem is, the internal array (vm->def->consoles) doesn't contain
just 'consoleX' but 'serialX' too after libvirtd restarts. I've noticed
this while testing my chardev hotplug patches. If user wants to hotplug
a console, I need to iterate over vm->def->consoles to find the next
free X to generate the alias. I am using qemuDomainDeviceAliasIndex for
that. The function takes alias prefix as one of its arguments. However,
if the prefix doesn't match, an error is returned. And since the aliases
vary over libvirtd restarts I think that's the problem. Of course, I can
workaround it in my patches, but c'mon we want the libvirt source code
to be hack-less, don't we? :)

> 
>>>
>>> This isn't any more correct that your previous patch. For the
>>> dummy <console> elements, the alias *must* be copied from /
>>> identical to the corresponding <serial> device config.
>>>
>>>
>>> Daniel
>>>
>>
>> So what are you saying is, it's the <serial> that has the wrong alias
>> assigned? The output should be like this then?
>>
>>     <serial type='pty'>
>>       <source path='/dev/pts/5'/>
>>       <target type='isa-serial' port='0'/>
>>       <alias name='console0'/>
>>       <!-- originaly was serial0 -->
>>     </serial>
>>
>>     <console type='pty' tty='/dev/pts/5'>
>>       <source path='/dev/pts/5'/>
>>       <target type='serial' port='0'/>
>>       <alias name='console0'/>
>>     </console>
> 
> No, the alias name must match that of the device id= parameter in the
> QEMU command line which is 'serial0' for a serial port.
> 
> 
> Daniel
>