libvirt users - Mar 2019 - Re: KVM-Docker-Networking using TAP and MACVLAN

On 3/13/19 2:26 PM, Martin Kletzander wrote:
> IIUC, you are using the tap0 device, but it is not plugged anywhere.
> By that I
> mean there is one end that you created and passed through into the VM,
> but there
> is no other end of that.  I can think of some complicated ways how to
> do what
> you are trying to, but hopefully the above explanation will move you
> forward and
> you'll figure out something better than what I'm thinking about
right
> now.  What
> usually helps me is to think of a way this would be done with hardware
> and
> replicate that as most of the technology is modelled after HW anyway.  Or
> someone else will have a better idea.
>
> Before sending it I just thought, wouldn't it be possible to just have
> a veth
> pair instead of the tap device?  one end would go to the VM and the
> other one
> would be used for the containers' macvtaps...
What I am trying to achieve is the most performant way to connect a set
of containers to the KVM while having proper isolation. As the Linux
bridge does not support port isolation I started with a 'bridge'
networking and MACVLAN using a VLAN for each container, but this comes
at the cost of bridging and the VLAN trunk on the KVM side. The simplest
(and hopefully therefore most performant) solution I could come up with
was using a 'virtio' NIC in the KVM, with 'direct' connection in
'vepa'
mode to 'some other end' on the host, TAP in its simplest form, which
Docker then uses for its MACVLAN network.


I am not quite sure if I understood you correctly with the 'other end'.
With the given configuration I would expect that one end of the TAP
device is connected to the NIC in the KVM (and it actually is, it has an
IP address assigned in the KVM and is serving the web configurator) and
the other end is connected to the MACVLAN network of Docker. If this is
not how TAP works, how do I then provide a 'simple virtual NIC' which
has one end in the KVM itself and the other on the host (without using
bridging or alike). I always thought then when using 'bridge' network
libvirt does exactly that, it creates a TAP device on the host and
assigns it to a bridge.


According to the man page I have to specify both interfaces when
creating the 'vdev' device, but how would I do that on the host with one
end being in the KVM?



br Lars

On Wed, Mar 13, 2019 at 11:40:51PM +0100, Lars Lindstrom
wrote:
>
>On 3/13/19 2:26 PM, Martin Kletzander wrote:
>> IIUC, you are using the tap0 device, but it is not plugged anywhere.
>> By that I
>> mean there is one end that you created and passed through into the VM,
>> but there
>> is no other end of that.  I can think of some complicated ways how to
>> do what
>> you are trying to, but hopefully the above explanation will move you
>> forward and
>> you'll figure out something better than what I'm thinking about
right
>> now.  What
>> usually helps me is to think of a way this would be done with hardware
>> and
>> replicate that as most of the technology is modelled after HW anyway. 
Or
>> someone else will have a better idea.
>>
>> Before sending it I just thought, wouldn't it be possible to just
have
>> a veth
>> pair instead of the tap device?  one end would go to the VM and the
>> other one
>> would be used for the containers' macvtaps...
>
>What I am trying to achieve is the most performant way to connect a set
>of containers to the KVM while having proper isolation. As the Linux
>bridge does not support port isolation I started with a 'bridge'
>networking and MACVLAN using a VLAN for each container, but this comes
>at the cost of bridging and the VLAN trunk on the KVM side. The simplest
>(and hopefully therefore most performant) solution I could come up with
>was using a 'virtio' NIC in the KVM, with 'direct'
connection in 'vepa'
>mode to 'some other end' on the host, TAP in its simplest form,
which
>Docker then uses for its MACVLAN network.
>
I hope I'm not misunderstanding something, but the way I understand it is
the
following:

TAP device is 1 device, a virtual ethernet card or an emulated network card.
Usually (VPNs and such) some user process binds to the device, can read data
which would normally go out on the wire and can write data that will look like
they came on the wire from the outside.  This is "one end of the
device", the
software is there instead of a wire connected to it.  The "other end"
shows up
in the OS as a network device that it can use to getting an IP address, sending
other packets, etc.

I think we both understand that, I just want to make sure we are on the same
page and also to have something to reference with my highly technical term (like
"other end") =)

Then when you create the VM and give it a VEPA access to the tap device, it
essentially just "duplicates" that device (the part that's in the
OS) and
whatever the VM tries to send will be sent out over the wire.

When you create the containers with MACVLAN (macvtap) it does similar thing, it
creates virtual interface over the tap device and whatever the container is
trying to send ends up being sent over the wire.

That is why I think this cannot work for you (unless you have a program that
binds to the tap device and hairpins all the packets, but that's not what
you
want I think).
>
>I am not quite sure if I understood you correctly with the 'other
end'.
Yeah, sorry for my knowledge on how this is actually called.  I found out,
however, the "other end" is whatever _binds_ to the device.
>With the given configuration I would expect that one end of the TAP
>device is connected to the NIC in the KVM (and it actually is, it has an
>IP address assigned in the KVM and is serving the web configurator) and
>the other end is connected to the MACVLAN network of Docker. If this is
From how I understand it you are plugging the same end to all of them.
>not how TAP works, how do I then provide a 'simple virtual NIC'
which
>has one end in the KVM itself and the other on the host (without using
This is what veth pair does, IIUC.  I imagine it as a patch cable.
>bridging or alike). I always thought then when using 'bridge'
network
>libvirt does exactly that, it creates a TAP device on the host and
>assigns it to a bridge.
>
Yes, but it is slightly more complicated.  There is a device that is connected
to the bridge which is on the host, so that the host can communicate with it.
However, when you start a VM, new device is created that is plugged in the
bridge and then passed to the VM.  That is done for each VM (and talking about
the default network).
>
>According to the man page I have to specify both interfaces when
>creating the 'vdev' device, but how would I do that on the host with
one
>end being in the KVM?
>
I cannot try this right now, but I would try something like this:

ip link add dev veth-vm type veth peer name veth-cont

and then put veth-vm in the VM (type='direct' would work, but I can
imagine
type='ethernet' might be even faster) and start the containers with
macvtap
using veth-cont.

I am in no position to even estimate what the performance is, neither compare it
to anything.  I would, however, imagine this could be pretty low-overhead
solution.

Let me know if that works or if I made any sense, I'd love to hear that.

Have a nice day,
Martin

On 3/14/19 4:06 PM, Martin Kletzander wrote:
> I cannot try this right now, but I would try something like this:
>
> ip link add dev veth-vm type veth peer name veth-cont
>
> and then put veth-vm in the VM (type='direct' would work, but I can
> imagine
> type='ethernet' might be even faster) and start the containers with
> macvtap
> using veth-cont.

I really appreciate your effort. The problem is in fact that I
misunderstood the manual. I thought that VETH requires two existing
devices, but in fact it creates two devices being connected (this
virtual link exists purely in the kernel, so performance should be fine).

As there is still no connectivity I reduced the setup to the bare
minimum, I kickstarted the server, installed QEMU and defined two
Debian-based KVM with a 'direct' device in 'vepa' mode and a
VETH
between them. When started, both KVM create a MACVTAP assigned to each
end of the VETH. Both KVM can ping each other (without any additional
route configured) and the VETH is LOWER_UP and UP even when no KVM is
running.

I then replaced one Debian-based KVM with pfSense and both KVM can still
ping each other.

I then created a MACVLAN docker network with the VETH as parent and
replaced the second Debian-based KVM with a Docker container using this
network. I am not quite sure why - but there is connectivity now. There
must have been some configuration issue on the server that was resolved
with kickstarting it.

Using type ETHERNET and the VETH as target did not work, "Unable to
create tap device veth1: Invalid argument". I then removed the VETH and
kept the ETHERNET configuration setting, which caused a TAP device
(according to error message) being created when the KVM is started. To
my confusion, this TAP device can be actually used as a parent for the
Docker MACVLAN network while still having connectivity! The downside is
that there is now an order dependency; the Docker network is unusable
when the KVM is shut down. In addition, when the KVM is shut down while
the container is still running the container must be restarted to get
networking going again.

So, there are two possible ways to achieve connectivity now
(VETH/VEPA/PASSTHROUGH and ETHERNET/TAP). Unfortunately there is a
'but'. I then added another container and even though the KVM device has
been configured VEPA, the containers are still able to contact each
other, so there is no isolation. I assume this is because just the 'KVM
end' of the VETH is in VEPA mode, whereas the 'Docker end' of the
VETH
is in BRIDGE mode. Unfortunately for the ETHERNET/TAP way no mode can be
configured in the KVM domain (I would assume because no MACVTAP is
involved).

The problem is that I cannot seem to figure out how to configure the
Docker network to use VEPA mode.


br Lars