Re: Using virtio-vhost-user or vhost-pci

[Stefan Hajnoczi]

On Tue, Oct 13, 2020 at 02:14:22AM +0300, Nikos Dragazis wrote:
> On 12/10/20 10:22 μ.μ., Cosmin Chenaru wrote:
> > I am currently running multiple VMs, connected in between by the DPDK 
> > vhost-switch. A VM can start, reboot, shutdown, so much of this is dynamic 
> > and the vhost-switch handles all of these. So these VMs are some sort of 
> > "endpoints" (I could not find a better naming).
> > 
> > The code which runs on the VM endpoints is somehow tied to the vhost-switch 
> > code, and if I change something on the VM which breaks the compatibility, I 
> > need to recompile the vhost-switch and restart. The problem is that most of 
> > the time I forget to update the vhost-switch, and I run into other problems.
> > 
> > If I could use a VM as a vhost-switch instead of the DPDK app, then I hope 
> > that in my endpoint code which runs on the VM, I can add functionality to 
> > make it also run as a switch, and forward the packets between the other VMs 
> > like the current DPDK switch does. Doing so would allow me to catch this 
> > out-of-sync between the VM endpoint code and the switch code at compile 
> > time, since they will be part of the same app.
> > 
> > This would be a two-phase process. First to run the DPDK vhost-switch 
> > inside a guest VM, and then to move the tx-rx part into my app.
> > 
> > Both Qemu and the DPDK app use "vhost-user". I was happy to see that I can 
> > start Qemu in vhost-user server mode:
> > 
> >     <interface type='vhostuser'>
> >       <mac address='52:54:00:9c:3a:e3'/>
> >       <source type='unix' path='/home/cosmin/vsocket.server' mode='server'/>
> >       <model type='virtio'/>
> >       <driver queues='2'>
> >         <host mrg_rxbuf='on'/>
> >       </driver>
> >       <address type='pci' domain='0x0000' bus='0x00' slot='0x04' 
> > function='0x0'/>
> >     </interface>
> > 
> > This would translate to these Qemu arguments:
> > 
> > -chardev socket,id=charnet1,path=/home/cosmin/vsocket.server,server -netdev 
> > type=vhost-user,id=hostnet1,chardev=charnet1,queues=2 -device 
> > virtio-net-pci,mrg_rxbuf=on,mq=on,vectors=6,netdev=hostnet1,id=net1,mac=52:54:00:9c:3a:e3,bus=pci.0,addr=0x4
> > 
> > But at this point Qemu will not boot the VM until there is a vhost-user 
> > client connecting to Qemu. I even tried adding the "nowait" argument, but 
> > Qemu still waits. This will not work in my case, as the endpoint VMs could 
> > start and stop at any time, and I don't even know how many network 
> > interfaces the endpoint VMs will have.

The "server" mode simply creates a listen socket instead of connecting.
It does not mean that QEMU acts as the vhost-user device backend. QEMU
is still the frontend. The UNIX domain socket "client" and "server"
relationship is independent of the vhost-user protocol frontend
(previously called "master") and device backend (previously called
"slave") relationship.

> > 
> > I then found the virtio-vhost-user transport protocol [2], and was thinking 
> > that I could at least start the packet-switching VM, and then let the DPDK 
> > app inside it do the forwarding of the packets. But from what I understand, 
> > this creates a single network interface inside the VM on which the DPDK app 
> > can bind. The limitation here is that if another VM wants to connect to the 
> > packet-switching VM, I need to manually add another virtio-vhost-user-pci 
> > device (and a new vhost-user.sock) before this packet-switching VM starts, 
> > so this is not dynamic.

Yes, each switch port needs its own virtio-vhost-user device because it
is the partner to a VM's virtio-net device.

It is possible to write a guest application that:

1. Opens multiple virtio-vhost-user devices and handles the connection
   lifecycle so that devices may be disconnected some of the time.

and/or

2. Reacts to virtio-vhost-user hotplug (i.e. udev/uevents) to
   dynamically add/remove ports.

> > The second approach for me would be to use vhost-pci [3], which I could not 
> > fully understand how it works, but I think it presents a network interface 
> > to the guest kernel after another VM connects to the first one.

This also requires multiple vhost-pci device instances if you want
multiple switch ports.

> > I realize I made a big story and probably don't make too much sense, but 
> > one more thing. The ideal solution for me would be a combination of the 
> > vhost-user socket and the vhost-pci socket. The Qemu will start the VM and 
> > the socket will wait in the background for vhost-user connections. When a 
> > new connection is established, Qemu should create a hot-plugable PCI 
> > network card and either the guest kernel or the DPDK app inside to handle 
> > the vhost-user messages.

vhost-pci and virtio-vhost-user don't present a network card to the
switch. Instead the switch acts as the device emulator for the
virtio-net NICs that other VMs are using. This has performance
advantages because no data copy and extra packet queuing is necessary
when there is just 1 NIC instead of 2 point-to-point NICs.

Stefan

signature.asc
Description: PGP signature