On Fri, Oct 28, 2011 at 04:15:41PM -0700, Sumit Naiksatam (snaiksat) wrote: > Hi, > > In its current implementation Libvirt makes sure that the network > interfaces that it passes/provision to a VM (for example to qemu[-kvm]) > are already connected to its backend (interfaces/networks) by the time > the VM starts its boot process. In a non virtualized setup it would be > like booting a machine with the Ethernet cable already plugged into a > router/switch port. While in a non virtualized setup you can boot a > machine first (with no physical connection to a router/switch) and later > connect its NIC/s to the switch/router, when you boot a VM via Libvirt > it is not possible to decouple the two actions (VM boot, cable > plug/unplug). > > An example of case where the capability of decoupling the two actions > mentioned above is a requirement in Quantum/NetStack which is the > network service leveraged by OpenStack. The modular design of OpenStack > allows you to: > - provision VMs with NIC/s > - create networks > - create ports on networks > - plug/unplug a VM NIC into/from a given port on a network (at runtime) > > Note that this runtime plug/unplug requirement has nothing to do with > hot plug/unplug of NICs. > The idea is more that of decoupling the provisioning of a VM from the > connection of the VM to the network/s. > This would make it possible to change (at run-time too) the networks the > NIC/s of a given VM are connected to. > > For example, when a VM boots, its interfaces should be in link down > state if the network admin has not connected the VM NIC/s to any > "network" yet. > Even though libvirt already provides a way to change the link state of > an a VM NIC, link state and physical connection are two different things > and should be manageable independently. > > Ideally the configuration syntax should be interface type and hypervisor > type agnostic. > > Let's take QEMU[-kvm] as an example - when Libvirt starts a QEMU VM, it > passes to QEMU a number of file descriptors that map to host backend > interfaces (for example macvtap interfaces). > > In order to introduce this runtime plug/unplug capability, we need a > mechanism that permits to delay the binding between the host macvtap > interfaces and the guest taps (because you cannot know the fd of the > macvtap interfaces before you create them). This means you need a > mechanism that allows you to change such fd/s at runtime: > > - you can close/reset an fd (ie, when you disconnect a VM NIC from its > network) > - you can open/set an fd (ie, when you connect a VM NIC to a network) > > This could probably be a libvirt command that translates to a QEMU > monitor command. > > Can the runtime plug/unplug capability described above be achieved > (cleanly) with another mechanism? > > Is anybody working on implementing something similar? No, but I've long thought about doing this & it is quite straightforward todo really. Ordinarily when we start QEMU we do qemu ... -device e1000,id=nic0,netdev=netdevnic0 \ -netdev user,id=netdevnic0 Todo what you describe we need to be able to: 1. Start QEMU with a NIC, but no netdev 2. Add a netdev to running QEMU. 3. Remove a netdev from a running QEMU 4. Associate a netdev with a NIC in running QEMU We can do 1: $ qemu ... -device e1000,id=nic0 But QEMU prints an annoying warning Warning: nic nic0 has no peer We can do 2 via the monitor: (qemu) netdev_add type=user,id=netdevnic0 We can do 3 via the monitor: (qemu) netdev_del netdevnic0 The problem is 4 - AFAICT we can't connect the existing NIC upto the newly hotplugged netdev, since we can't update the 'netdev' property in the NIC device.

Also if we delete the netdev, we can't clear out the 'netdev' property in the NIC, so its dangling to a netdev that no longer exists.

The latter is fairly harmless, since we can just re-use the name if adding a new backend later. The first problem is a bit of a pain, unless we plug in a 'user' backend on the CLI, and immediately netdev_del it before starting the CPUs. Ideally we'd have some way to set qdev properties for devices so we can associate the NIC with the new netdev. eg when adding a netdev: (qemu) netdev_add type=user,id=netdevnic0 (qemu) set nic0 netdev=netdevnic0 Or removing one (qemu) netdev_add netdevnic0 (qemu) unset nic0 netdev

WRT to libvirt XML config. Normally you specifiy a NIC like <interface type='network'> <mac address='52:54:00:0f:7d:ad'/> <source network='default'/> <model type='virtio'/> </interface> To boot a guest without any netdev backend present, we'd introduce a new network type="none". eg <interface type='none'> <mac address='52:54:00:0f:7d:ad'/> <model type='virtio'/> </interface> The existing API 'virDomainUpdateDevice', can then be used to change the interface config on the fly, adding or removing the netdev by passing in new XML with a different 'type' attribute & <source> element. Finally, when adding & removing the netdev backends to a running guest, we likely want to be able to set the NIC's link carrier, so the guest OS sees that it has lost / gain its network connection & will thus retry DHCP / IPv6 autoconfig. There is already a QEMU montior command 'set_link' for changing the NIC link carrier. A minor problem is that, AFAICT, we can't specify the link carrier state on the command line when specifying the NIC hardware, eg we would want something like this when starting a guest without a netdev back qemu ... -device e1000,id=nic0,link=down

And when adding a netdev we would do (qemu) netdev_add user,id=netdevnic0 (qemu) set nic0 netdev=netdevnic0 (qemu) set_link nic0 up Or when removing a netdev (qemu) set_link nic0 down (qemu) unset nic0 netdev (qemu) netdev_del user,id=netdevnic0

> -----Original Message----- > From: qemu-devel-bounces+benve=cisco.com(a)nongnu.org [mailto:qemu-devel- > bounces+benve=cisco.com(a)nongnu.org] On Behalf Of Markus Armbruster > Sent: Monday, October 31, 2011 7:05 AM > To: Daniel P. Berrange > Cc: libvir-list(a)redhat.com; David Wang (dwang2); Ram Durairaj > (radurair); qemu-devel(a)nongnu.org; Sumit Naiksatam (snaiksat) > Subject: Re: [Qemu-devel] [libvirt] RFC decoupling VM NIC provisioning > fromVM NIC connection to backend networks > > "Daniel P. Berrange" <berrange(a)redhat.com&gt; writes:

> > The problem is 4 - AFAICT we can't connect the existing NIC upto the > newly > > hotplugged netdev, since we can't update the 'netdev' property in the > NIC > > device. > > Yes, that's the missing piece. Can we go ahead and add this missing piece to QEMU? Is there any concern on the QEMU side?

> -----Original Message----- > From: qemu-devel-bounces+benve=cisco.com(a)nongnu.org [mailto:qemu-devel- > bounces+benve=cisco.com(a)nongnu.org] On Behalf Of Daniel P. Berrange > Sent: Monday, October 31, 2011 3:49 AM > To: Sumit Naiksatam (snaiksat) > Cc: libvir-list(a)redhat.com; David Wang (dwang2); Ram Durairaj > (radurair); qemu-devel(a)nongnu.org > Subject: Re: [Qemu-devel] [libvirt] RFC decoupling VM NIC provisioning > from VM NIC connection to backend networks > > On Fri, Oct 28, 2011 at 04:15:41PM -0700, Sumit Naiksatam (snaiksat) > wrote: > > Hi, > > > > In its current implementation Libvirt makes sure that the network > > interfaces that it passes/provision to a VM (for example to qemu[- > kvm]) > > are already connected to its backend (interfaces/networks) by the > time > > the VM starts its boot process. In a non virtualized setup it would > be > > like booting a machine with the Ethernet cable already plugged into a > > router/switch port. While in a non virtualized setup you can boot a > > machine first (with no physical connection to a router/switch) and > later > > connect its NIC/s to the switch/router, when you boot a VM via > Libvirt > > it is not possible to decouple the two actions (VM boot, cable > > plug/unplug). > > > > An example of case where the capability of decoupling the two actions > > mentioned above is a requirement in Quantum/NetStack which is the > > network service leveraged by OpenStack. The modular design of > OpenStack > > allows you to: > > - provision VMs with NIC/s > > - create networks > > - create ports on networks > > - plug/unplug a VM NIC into/from a given port on a network (at > runtime) > > > > Note that this runtime plug/unplug requirement has nothing to do with > > hot plug/unplug of NICs. > > The idea is more that of decoupling the provisioning of a VM from the > > connection of the VM to the network/s. > > This would make it possible to change (at run-time too) the networks > the > > NIC/s of a given VM are connected to. > > > > For example, when a VM boots, its interfaces should be in link down > > state if the network admin has not connected the VM NIC/s to any > > "network" yet. > > Even though libvirt already provides a way to change the link state > of > > an a VM NIC, link state and physical connection are two different > things > > and should be manageable independently. > > > > Ideally the configuration syntax should be interface type and > hypervisor > > type agnostic. > > > > Let's take QEMU[-kvm] as an example - when Libvirt starts a QEMU VM, > it > > passes to QEMU a number of file descriptors that map to host backend > > interfaces (for example macvtap interfaces). > > > > In order to introduce this runtime plug/unplug capability, we need a > > mechanism that permits to delay the binding between the host macvtap > > interfaces and the guest taps (because you cannot know the fd of the > > macvtap interfaces before you create them). This means you need a > > mechanism that allows you to change such fd/s at runtime: > > > > - you can close/reset an fd (ie, when you disconnect a VM NIC from > its > > network) > > - you can open/set an fd (ie, when you connect a VM NIC to a network) > > > > This could probably be a libvirt command that translates to a QEMU > > monitor command. > > > > Can the runtime plug/unplug capability described above be achieved > > (cleanly) with another mechanism? > > > > Is anybody working on implementing something similar? > > No, but I've long thought about doing this & it is quite > straightforward > todo really. Ordinarily when we start QEMU we do > > qemu ... -device e1000,id=nic0,netdev=netdevnic0 \ > -netdev user,id=netdevnic0 > > Todo what you describe we need to be able to: > > 1. Start QEMU with a NIC, but no netdev > 2. Add a netdev to running QEMU. > 3. Remove a netdev from a running QEMU > 4. Associate a netdev with a NIC in running QEMU > > We can do 1: > > $ qemu ... -device e1000,id=nic0 > > But QEMU prints an annoying warning > > Warning: nic nic0 has no peer If we introduce this new functionality, can this warning change? If we change it, would it break any test/script? Actually it is just a warning (not an error). Why do you think it is annoying? (I guess it is supposed to catch misconfigurations) > We can do 2 via the monitor: > > (qemu) netdev_add type=user,id=netdevnic0 > > We can do 3 via the monitor: > > (qemu) netdev_del netdevnic0 > > > The problem is 4 - AFAICT we can't connect the existing NIC upto the > newly > hotplugged netdev, since we can't update the 'netdev' property in the > NIC > device. Also if we delete the netdev, we can't clear out the 'netdev' > property in the NIC, so its dangling to a netdev that no longer exists. > The latter is fairly harmless, since we can just re-use the name if > adding > a new backend later. The first problem is a bit of a pain, unless we > plug > in a 'user' backend on the CLI, and immediately netdev_del it before > starting > the CPUs. Ideally we'd have some way to set qdev properties for devices > so we > can associate the NIC with the new netdev. > > eg when adding a netdev: > > (qemu) netdev_add type=user,id=netdevnic0 > (qemu) set nic0 netdev=netdevnic0 > > Or removing one > > (qemu) netdev_add netdevnic0 > (qemu) unset nic0 netdev > > > WRT to libvirt XML config. Normally you specifiy a NIC like > > <interface type='network'> > <mac address='52:54:00:0f:7d:ad'/> > <source network='default'/> > <model type='virtio'/> > </interface> > > To boot a guest without any netdev backend present, we'd introduce a > new network type="none". eg > > <interface type='none'> > <mac address='52:54:00:0f:7d:ad'/> > <model type='virtio'/> > </interface> > > The existing API 'virDomainUpdateDevice', can then be used to change > the interface config on the fly, adding or removing the netdev by > passing in new XML with a different 'type' attribute & <source> > element. > > Finally, when adding & removing the netdev backends to a running guest, > we likely want to be able to set the NIC's link carrier, so the guest > OS sees that it has lost / gain its network connection & will thus > retry DHCP / IPv6 autoconfig. I assume this is what you meant: - when the NIC does not have a backend netdev configured, its link state is DOWN by default. Right? (*) Would it make sense to provide anyway the possibility of setting the link state in this case too? (ie, ability to force link UP also when there is no backend netdev configured). There may be use cases where this could be needed, but I can't think of any right now. - When you unbind a NIC from a backend netdev (ie, "unset" command above) the NIC's link goes down (unless the above (*) option said otherwise) - When you bind a NIC to a backend netdev (ie, "set" command above), the NIC link state should be copied from the backend netdev link state. For example, when you plug an Ethernet NIC to a switch, the NIC receives link UP only if the switch's port is UP. This means that the NIC link state reflects the backend netdev's link state (ie, the NIC is not always UP regardless of the backend netdev's link state).