[libvirt] [PATCH 0/5] Interface pools and passthrough mode

[Laine Stump]

On 11/29/2011 02:53 PM, Daniel P. Berrange wrote:
> On Tue, Nov 29, 2011 at 03:46:13PM +0000, Shradha Shah wrote:
>> Interface Pools and Passthrough mode:
>>
>> Current Method:
>> The passthrough mode uses a macvtap a direct connection to connect each guest to the network. The physical interface to be used is picked from among those listed in<interface>  sub elements of the<forward>  element.
>>
>> The current specification for<forward>  extends to allow 0 or more<interface>  sub-elements:
>> Example:
>> <forward mode='passthrough' dev='eth10'/>
>> <interface dev='eth10'/>
>> <interface dev='eth12'/>
>> <interface dev='eth18'/>
>> <interface dev='eth20'/>
>> </forward>
>>
>> However with an ethernet card with 64 VF's or more, the above method gets tedious on the system.
>>
>> On the other hand, just parameterizing a string (eth%d) is inadequate, eg, when there are multiple non-contiguous ranges.


Heh. You've gone through some of the same thought process I went through 
when I wrote the original code :-)


>>
>> Proposed Method:
>> The 5 patches provided along with this introductory e-mail
>>
>> i) Introduce a structure to store the state of all the virtual functions attached to each physical function
>> ii) Find a free virtual function given the physical function.
>>
>> The forward specification will hence only mention the physical function as the interface sub element:
>> Example:
>> <forward mode='passthrough' dev='eth2'/>
>> <interface dev='eth2'/>
>> </forward>
> I can see what you mean about it being tedious to construct the config with
> all 64 (or more) VF's listed, but this proposed change has a couple of issues
>
> First of all, the change you describe would be a semantic change in the
> network XML, which would break compatibility with previous releases of
> libvirt. Since we consider XML to be poart of our long term ABI stability
> guarantee I don't think we can do the change.
>
> Ignoring the ABI issue, I'm concerned that as we get PFs with an increasingly
> large number of VFs, we may well *not* want to associate all VFs with a single
> virtual network definition. eg, we might wna to put 32 VFs in one network and
> 32 VFs in another network.  Or if we have 2 PFs, we might want to interleave
> VFs from several PFs across virtual networks. If all we can do is list the
> PF in the XML, we loose significant flexibility in how VFs are assigned.

My first concern too when I saw the patch was the semantic change (but 
also the loss of flexibility), which is obviously a no-go. It's a 
convenient capability to have though, so it would be nice to get it in 
somehow. What if we allowed including all the VFs associated with a PF 
by adding an extra attribute?  e.g.:

<interface dev='eth10' type='sriov'/>

(or whatever is more appropriate in place of "sriov"). Or possibly a 
different element type could be used:

<pf dev='eth10'/>

(didn't want to spend time thinking of a better name than "pf"...).

At the time the network is created, this would cause libvirt to get the 
list of all VFs for the given PF and put them into the pool. This could 
be used instead of, or in combination with, the existing <interface 
dev='eth1'/> form. Thus the existing semantics would be preserved, the 
flexibility of specifying individual devices would be retained, and the 
desired convenience of adding all VFs of a PF with a single line would 
be added.