[libvirt] Allowing <interface type="none"> for network interface definition
Hi I am writing a management stack for our platform using libvirt. We have many DomU images which all have PV drivers loaded for network.I want to start these DomU guests with only the PV network device being visible from within the DomU. This is a requirement in the platform and hopefully, I presume there are other people with the same requirement.. But when creating a server, libvirt sends (type ioemu) tag to xend. Which make all of our vm's to have 2 network cards. Using the normal xm command I can send type=none ( vif = [ "mac=00:16:3e:00:a5:57,bridge=eth0,script=vif-bridge,type=none"]) to xend. Which will send qeum-dm "-net none" switch which solves the problem. After doing some investigation, I did some code change in libvirt0.5.0 to support a attribute called "none" with in <interface type="">, that will send xend the tag (type none). Allowing qeum-dm to be configured with "-net none" switch. ;-) So you can define your network tag as <interface type='none'> <!-- This will send qeum-dm -net none switch --> <!-- Other configurations will work as normal --> <mac address='00:16:3e:00:a5:57'/> <source bridge='eth0'/> <target dev='vif47.0'/> </interface> I this the proper way of doing it? :-\ Would this be a general functionality that libvirt community will be interested in (I think this add bit more flexibility to libvirt of the way user want to write their own management stack). If so could I contribute the patch to the mailing list. Looking forward to hearing from you all, I am new to libvirt therefore any comments will be greatly appreciated :-) Thanks Gihan -- Gihan Munasinghe R&D Team Leader XCalibre Communications Ltd. www.flexiscale.com
On Tue, Dec 02, 2008 at 09:24:24PM +0000, Gihan Munasinghe wrote:
Hi
I am writing a management stack for our platform using libvirt. We have many DomU images which all have PV drivers loaded for network.I want to start these DomU guests with only the PV network device being visible from within the DomU.
This is a requirement in the platform and hopefully, I presume there are other people with the same requirement..
But when creating a server, libvirt sends (type ioemu) tag to xend. Which make all of our vm's to have 2 network cards. Using the normal xm command I can send type=none ( vif = [ "mac=00:16:3e:00:a5:57,bridge=eth0,script=vif-bridge,type=none"]) to xend. Which will send qeum-dm "-net none" switch which solves the problem.
After doing some investigation, I did some code change in libvirt0.5.0 to support a attribute called "none" with in <interface type="">, that will send xend the tag (type none). Allowing qeum-dm to be configured with "-net none" switch. ;-) So you can define your network tag as
<interface type='none'> <!-- This will send qeum-dm -net none switch --> <!-- Other configurations will work as normal --> <mac address='00:16:3e:00:a5:57'/> <source bridge='eth0'/> <target dev='vif47.0'/> </interface>
I this the proper way of doing it? :-\
The "type" attribute is used to specify how the network card is connected to the host networking. You're after something that changes what is exposed tto the guest. The XML format for network interfaces already has support for a model element which accomplishes this. eg, for KVM we can select VirtIO network card with: <interface type='bridge'> <model type="virtio"/> <source bridge='br0'/> <mac address='00:16:3e:00:a5:57'/> </interface> This 'model' is not currently used in the Xen driver for libvirt, so we should implement it. A value of 'none' doesn't really make sense. For Xen >= 3.1.0, libvirt will no longer append type=ioemu by default. This lets XenD configure by rtl8139 nic, and PV driver back/front. So I think we just need to allow an explicit choice to override this A couple of choices, either follow XenD's naming <model type="ioemu"/> <model type="netback"/> Or take into account we might want to allow for full range of QEMU nic choices, even if XenD doesn't (yet) expose it <model type="rtl8139"/> <model type="e1000"/> <model type="ne2k_pci"/> <model type="netback"/> I think i tend towards the latter, since its more consistent with naming in QEMU driver. Daniel -- |: Red Hat, Engineering, London -o- http://people.redhat.com/berrange/ :| |: http://libvirt.org -o- http://virt-manager.org -o- http://ovirt.org :| |: http://autobuild.org -o- http://search.cpan.org/~danberr/ :| |: GnuPG: 7D3B9505 -o- F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505 :|
Daniel P. Berrange wrote:
On Tue, Dec 02, 2008 at 09:24:24PM +0000, Gihan Munasinghe wrote:
The "type" attribute is used to specify how the network card is connected to the host networking. You're after something that changes what is exposed tto the guest. The XML format for network interfaces already has support for a model element which accomplishes this. eg, for KVM we can select VirtIO network card with:
<interface type='bridge'> <model type="virtio"/> <source bridge='br0'/> <mac address='00:16:3e:00:a5:57'/> </interface>
Hi Daniel Thanks for the reply by using type=none what I meant was asking qemu-dm not to make any configurations what so ever. so I am asking qemu-dm to not to present a network card to vm at all..
This 'model' is not currently used in the Xen driver for libvirt, so we should implement it. A value of 'none' doesn't really make sense. For Xen >= 3.1.0, libvirt will no longer append type=ioemu by default. This lets XenD configure by rtl8139 nic, and PV driver back/front.
So I think we just need to allow an explicit choice to override this
A couple of choices, either follow XenD's naming
<model type="ioemu"/> <model type="netback"/> Or take into account we might want to allow for full range of QEMU nic choices, even if XenD doesn't (yet) expose it
<model type="rtl8139"/> <model type="e1000"/> <model type="ne2k_pci"/> <model type="netback"/>
I think i tend towards the latter, since its more consistent with naming in QEMU driver.
Daniel
This is some code segments I took out from xend_internal. c if (def->ifname != NULL && !STRPREFIX(def->ifname, "vif")) virBufferVSprintf(buf, "(vifname '%s')", def->ifname); if (def->model != NULL) virBufferVSprintf(buf, "(model '%s')", def->model); So the model tag is already send to xend, and used in xend as well check the following xend code for devuuid in vmConfig['vif_refs']: devinfo = vmConfig['devices'][devuuid][1] dtype = devinfo.get('type', 'ioemu') if dtype != 'ioemu': continue nics += 1 mac = devinfo.get('mac') if mac is None: raise VmError("MAC address not specified or generated.") bridge = devinfo.get('bridge', 'xenbr0') model = devinfo.get('model', 'rtl8139') ret.append("-net") ret.append("nic,vlan=%d,macaddr=%s,model=%s" % (nics, mac, model)) ret.append("-net") ret.append("tap,vlan=%d,ifname=tap%d.%d,bridge=%s" % (nics, self.vm.getDomid(), nics-1, bridge)) if nics == 0: ret.append("-net") ret.append("none") Therefore only way I can achieve -net none switch, is by sending a (type != ioemu ) tag. sending in a ( model ) tag will not give me what I want. So simply I just want to update my xenstore configuration but I am going to ask qemu not to load any network at all.. So as far as qemu is concerned the domain does not have any network.. -net none. That's what I thought of using something like <interface type="none">. Am I missing something here.. Just to let you know, This is what I changed on the level of xend_internal.c if ((hvm) && (xendConfigVersion < 4) && (def->type!=VIR_DOMAIN_NET_TYPE_NONE)) virBufferAddLit(buf, "(type ioemu)"); else{ //Only send type none if the <interface type="none" vm formats> virBufferAddLit(buf, "(type none)"); } I need to send a explicit (type none) or any ( type !ioemu) to get the desired outcome. Thought none would make sense as I don't want any network configured Thanks Gihan
participants (2)
-
Daniel P. Berrange -
Gihan Munasinghe