[libvirt] [PATCH v2 3/3] network: Taint networks that are using hook script

Michal Privoznik mprivozn at redhat.com
Mon Feb 10 17:07:31 UTC 2014


On 10.02.2014 15:52, Michal Privoznik wrote:
> On 08.02.2014 11:51, Laine Stump wrote:
>> On 02/07/2014 10:52 PM, Antoni Segura Puimedon wrote:
>>>
>>> ----- Original Message -----
>>>> From: "Laine Stump" <laine at laine.org>
>>>> To: libvir-list at redhat.com
>>>> Cc: "Michal Privoznik" <mprivozn at redhat.com>
>>>> Sent: Friday, February 7, 2014 1:17:10 PM
>>>> Subject: Re: [libvirt] [PATCH v2 3/3] network: Taint networks that
>>>> are using hook script
>>>>
>>>> On 02/05/2014 12:11 PM, Michal Privoznik wrote:
>>>>> Basically, the idea is copied from domain code, where tainting
>>>>> exists for a while. Currently, only one taint reason exists -
>>>>> VIR_NETWORK_TAINT_HOOK to mark those networks which caused invoking
>>>>> of hook script.
>>>> What's missing here is that the network status XML doesn't include a
>>>> <taint> element.
>>>>
>>>> Also, I think if a network is tainted, and domain that connects to that
>>>> network should be tainted as well.
>>>>
>>>> Of course what would make this more useful would be if would could
>>>> determine when a hook script actually *did* something for a particular
>>>> network/interface (since presumably people are usually going to write
>>>> their network hook scripts to only take action for particular networks
>>>> and/or domains, not for *all* networks). I don't know that there's a
>>>> way
>>>> to do that without either 1) having a different hook script for each
>>>> network, or 2) trusting the hook script to return some sort of status
>>>> indicating whether or not it did anything. Obviously (2) is not a good
>>>> idea, but we may want to think about (1) in the future (for qemu and
>>>> lxc
>>>> hook scripts as well) - instead of just looking for
>>>> /etc/libvirt/hook/network, we could first look for
>>>> /etc/libvirt/hook/network.${netname} and exec that instead if found (or
>>>> in addition). But I think that can be deferred until later.
>>> Actually I kind of like the option (2). I think it could make a lot
>>> of sense
>>> that the hook would be able to add an attribute to the network
>>> definition
>>> xml, e.g. <bandwidth hooked="1"> so that libvirt would know that that
>>> part
>>> has been taken care of by the hook. Of course, it might be a bad idea
>>> for
>>> libvirt to blindly accept any kind of modification, but something
>>> like what
>>> I propose does not seem eminently dangerous.
>>
>> The reason I don't like option (2) is that it requires trusting the hook
>> to leave its mark if it modifies anything, and that's exactly why we
>> want to taint the networks that call a hook - because we don't/can't
>> trust the hook.
>>
>> I wonder if there might be some way to allow a hook to add information
>> to the network's xml in some well-defined location, though. This
>> information would not be used/trusted by libvirt at all, but would only
>> be there, for example, so that a later "stop/unplug" hook could retrieve
>> it, rather than being required to keep its state externally.
>>
>
> Well, we may make the hook script to return the network xml that libvirt
> will parse and startup. For example:
>
> 1) network with <bandwidth/> is about to start. The network XML is
> passed to the script.
>
> 2) The script sees <network> ... <bandwidth/> ... </network> and do all
> the tc magic. Then it produces the same XML minus <bandwidth/>
>
> 3) Libvirt parses the <network> ... </network> without the bandwidth
> knowing that the script has taken care of it. If it doesn't we may error
> out because <bandwidth/> is not supported yet (assuming the right type
> of network for this little example). The whole network startup process
> would be aborted then.

On a second thought, this is not such a clever idea. Users are not 
expected to have knowledge about libvirt internals. So for example, if 
they were to set <bandwidth/> themselves, they may had used some 
iptables rules to mark packets. So their hook script removes 
<bandwidth/> from the XML definition. But they leave nwfilter in, which 
will interfere with their iptables rules. Or even the less obvious one: 
<ip address=''/> with range (in which case libvirt inserts some iptables 
rules too).

Long story short, each configuration knob in the network xml may affect 
others and since it is viewed as internal implementation, it is subject 
to change. Hence, updating the libvirt may lead to breaking users' hook 
scripts.

Michal




More information about the libvir-list mailing list