[libvirt] [PATCH 1/1] Add vlantag parameter for openvswitch ports

[Laine Stump]

On 08/10/2012 11:21 AM, Kyle Mestery (kmestery) wrote:
> On Aug 9, 2012, at 3:14 PM, Laine Stump wrote:
>> On 08/08/2012 03:47 PM, Kyle Mestery wrote:
>>> Add the ability to specify a "vlantag" parameter for bridge
>>> networks with a virtualport type of openvswitch. This
>>> allows for specifying the port is on a single VLAN, and
>>> should receive untagged traffic (e.g. vlantag=10), or that
>>> a port should receive tagged traffic and is a trunk port
>>> (e.g. vlantag=10,11).
>> Ha! We must have a telepathic link!
>>
> Well, we're at least thinking in the same general direction. :)
>
>> I'm just about to add vlan support for SR-IOV VFs in PCI passthrough
>> (aka "hostdev") and macvtap passthrough modes (the only thing that's
>> kept me from action is indecision about where to put the vlan tag in the
>> xml), and since at least one of those modes doesn't use <virtualport> at
>> all, I was planning to add a <vlan tag='x'/> element at the toplevel of
>> interface. There would also be a similar element at the toplevel of
>> <network>, and another within <portgroup>.
>>
>> Something like this:
>>
>> <interface type='hostdev'>
>>  <mac address='52:54:00:11:22:33'/>
>>  <vlan tag='42'/>
>>  ...
>> </interface>
>>
>> In this case, there is no <network> involved. Here's a case that uses a
>> network:
>>
>>
>> <interface type='network'>
>>  <mac address='52:54:00:11:22:33'/>
>>  <source network='net42'/>
>>  ...
>> </interface>
>>
>> <network>
>>  <name>net42</name>
>>  <vlan tag='42'/>
>>  <forward mode='passthrough'>
>>    <interface dev='eth5'/>
>>    <interface dev='eth6'/>
>>    <interface dev='eth7'/>
>>  </forward>
>> </network>
>>
>> In this second case, when the interface was brought up, it would
>> allocate a physical device from the pool, and inherit the vlan tag,
>> which would be set in the SR-IOV card's hardware as it's assigned to the
>> guest. Note that hostdev passthrough would behave similarly (except
>> forward mode would be 'hosted')
>>
> Those cases both look good. I think the formatting works just fine for
> virtualport type=openvswitch as well, something like this:
>
> Single VLAN (no trunk):
> <interface type='bridge'>
>  <mac address='52:54:00:30:23:a6'/>
>  <source bridge='data-br'/>
>  <vlan tag='70'/>
>  <virtualport type='openvswitch'> 
>    <parameters interfaceid='cdbbbc31-b7fe-16ca-a715-cc7cc76e18b2'>
>  </virtualport>                    
>  <model type='virtio'/>                  
>  <address type='pci' domain='0x0000' bus='0x00' slot='0x03' function='0x0'/>
> </interface>
>
> Single VLAN (trunk):
> <interface type='bridge'>
>  <mac address='52:54:00:30:23:a6'/>
>  <source bridge='data-br'/>
>  <vlan tag='70'/ trunk=yes>
>  <virtualport type='openvswitch'> 
>    <parameters interfaceid='cdbbbc31-b7fe-16ca-a715-cc7cc76e18b2'>
>  </virtualport>                    
>  <model type='virtio'/>                  
>  <address type='pci' domain='0x0000' bus='0x00' slot='0x03' function='0x0'/>
> </interface>
>
> Multiple VLANs (trunk):
> <interface type='bridge'>
>  <mac address='52:54:00:30:23:a6'/>
>  <source bridge='data-br'/>
>  <vlan trunk='yes'>
>   <tag id='70'>
>   <tag id='71'>
>  </vlan>
>  <virtualport type='openvswitch'> 
>    <parameters interfaceid='cdbbbc31-b7fe-16ca-a715-cc7cc76e18b2'>
>  </virtualport>                    
>  <model type='virtio'/>                  
>  <address type='pci' domain='0x0000' bus='0x00' slot='0x03' function='0x0'/>
> </interface>
>
>
>> Finally, there may be a network with multiple vlans. portgroups could be
>> used for that:
>>
>> <interface type='network'>
>>  <mac address='52:54:00:11:22:33'/>
>>  <source network='physnet' portgroup='production'/>
>>  ...
>> </interface>
>>
>> <network>
>>  <name>net42</name>
>>  <forward mode='passthrough'>
>>    <interface dev='eth5'/>
>>    <interface dev='eth6'/>
>>    <interface dev='eth7'/>
>>  </forward>
>>  <portgroup name='production' default='yes'>
>>    <vlan tag='42'/>
>>    <bandwidth>
>>      ..
>>    </bandwidth>
>>  </portgroup>
>>  <portgroup name='test'>
>>    <vlan tag='666'/>
>>    <bandwidth>
>>      ..
>>    </bandwidth>
>>  </portgroup>
>> </network>
>>
>> Selecting a different portgroup would put you on a difference vlan of
>> the same switch. We would need to make a decision about which would take
>> precedence if the vlan tag was given in multiple places - should the
>> domain's request be honored, in order to allow it to made individual
>> modifications to the general config in the <network>? Or should the
>> network, as an entity of higher authority, be allowed to override what
>> the domain asks for?
>>
> I think this is the most elegant solution, and dovetails nicely with the fact that
> virtualport type='openvswitch' ports already support port-profiles. I think in general,
> my feeling here is the domain XML should override the network element. This is
> similar to an interface override of a port-profile on a Cisco switch, for example, where
> you define configuration in a port-profile, but allow configuration directly on the
> interface itself to override the port-profile configuration.

Okay. Thanks - it's good to know how the rest of the world works. Best
to mirror that as much as possible.

>
>> I've never used vlan trunk groups, but I'm guessing whatever information
>> you needed could be added as attributes to the <vlan> element. BTW, in
>> general we don't like to have multiple pieces of data in a single
>> element. We would prefer that they be in separate elements. So maybe if
>> you wanted a single vlan tag, you could do it as above, but when you
>> wanted a trunk group, you could do something like:
>>
>>   <vlan trunk='yes'>
>>     <tag id='23'/>
>>     <tag id='24'/>
>>     <tag id='25'/>
>>   </vlan>
>>
>> (you might allow omitting the "trunk='yes'") Or possibly put them all at
>> the top level:
>>
>>   <vlan tag='23'/>
>>   <vlan tag='24'/>
>>   <vlan tag='25'/>
>>
>> and if you wanted a vlan trunk with only a single vlan in it:
>>
>>   <vlan tag='23' trunk='yes'/>
>>
>> (actually, having multiple toplevel elements could get messy if we
>> started talking about merging the elements from
>> network/portgroup/interface together, so maybe that's not such a good idea).
>>
>> *Still another* possibility would be to put the vlan element as
>> described above inside <virtualport>, then allow <virtualport> with no
>> type to contain a <vlan> element. This would require a change to my
>> recent <virtualport> merging patches, but they haven't been pushed yet.
>> I'm not convinced I like that option, though.
>>
>> I think this needs a day or so to stew...
> So I guess the question is, are you going to role this up into your patches? If so, you
> can likely make use of some pieces of the patches I posted. Let me know, and I can
> also work with you on this as well.

How about this - I'll try to produce patches today that 1) get the
<vlan> element into <interface>, <network>, and <portgroup> as a
virNetDevVlan object, and 2) merge all of those properly so that you end
up with an optional virNetDevVlanPtr vlan in the actual netdef,
accessible via the new function virDomainNetGetActualVlan(net).

I'm thinking virNetDevVlan will just be something like this:

  struct _virNetDevVlan {
      bool trunk; /* true if this is a trunk */
      unsigned int *tag; /* pointer to array of tags */
  }

I guess this will need to be defined in src/util, since it needs to be
passed into virnetdevopenvswitch.c and virnetdevvportprofile.c


I'll submit those patches by themselves, and leave the patches that
actually use the vlan info until that's done. That way you'll have what
you need to rework your patches as soon as possible.