I am currently setting up a Proxmox box that has the usual selfhosted stuff (Nextcloud, Jellyfin, etc) and I want all of these services in different containers/VMs. I am planning to start sharing this with family/friends who are not tech savvy, so I want excellent security.
I was thinking of restricting certain services to certain VLANs, and only plugging those VLANs into the CT/VMs that need them.
Currently, each CT/VM has a network interface (for example eth0) which gives them internet access (for updates and whatnot) and an interface that I use for SSH and management (for example eth1). These interfaces are both on different VLANs and I must use Wireguard to get onto the management network.
I am thinking of adding another interface just for “consumption” which my users would get onto via a separate Wireguard server, and they would use this to actually use the services.
I could also add another network just to connect to an internal NFS server to share files between CT/VMs, and this would have its own VLAN and require an additional interface per host that connects to it.
I have lots of other ideas for networks which would require additional interfaces per CT/VM that uses them.
From my experience, using a “VLAN-Aware” bridge and assigning VLANs per interface via the GUI is best practice. However, Proxmox does not support multiple VLANs per interface using this method.
I have an IPv6-only network, so I could theoretically assign multiple IPs per interface. Then I would use Linux VLANs from within the guest OS. However, this is a huge pain and I do not want to do this. And it is less secure because a compromised VM/CT could change its VLAN tag itself.
I am asking if adding many virtual interfaces per CT/VM is good practice, or if there is a better way to separate internal networks. Or maybe I should rethink the whole thing and not use one network per use-case.
I am especially curious about performance impacts of multiple interfaces.
You must log in or # to comment.
I wouldn’t let every VM have an interface into your management network, regardless of how you implement this. Your management network should be segregated with the ability to route to all the other VLANs with an appropriate firewall setup that only allows “related/established” connections back into it.
As for your services, having them on separate VLANs is fine, but it seems like you would benefit from having a reverse proxy to forward things to the appropriate VLAN, to reduce your management overhead.
But in general, having multiple interfaces per VM is fine. There shouldn’t be any performance hit or anything. But remember that if you have a compromised VM, it’ll be on any networks you give it an interface in, so minimizing that is key for security purposes. Ideally it would live in a VLAN that only has Internet access and/or direct access to your reverse proxy.
Having multiple interfaces in each vm can lead to issues with routing if you screw something up.
Like you said I’d expose the services via reverse proxy in the public vlan, and enable ssh access on the firewall only from a jumpbox or the ip of your pc (or maybe the vlan you are in).
Yeah, it can for sure. Definitely worth mentioning. Gotta watch what interface is set as the default router, or you’re bound to have a bad time. That said, the same is true with his originally proposed solution of pushing a trunk port to the VM, so it’s not any worse in that regard.
But yeah, full agreement on the correct solution. Keep it simple.
I agree with this. The only vm I have that has multiple interfaces is an opnsense router vm heavily optimized for kvm to reach 10gb speeds.
One of the interfaces beyond wan and lan is an interface that links to a proxmox services bridge. It’s a proxbridge I gave to a container and is just a gateway in opnsense. It points traffic destined for services directly at the container ip. It keeps the service traffic on the bridge instead of having to hit the physical network.
Agreed, I prefer trunk with native to the vlan for services, each container that the reverse proxy will hit in its own vlan (or multiples for differing sets of services, but I can be excessive).
I’d block any traffic initiated from that vlan to all others, and I’d also only allow the specific ports needed for the services. Then fully open initiated from the general internal vlan.
I use using docker networks but that’s me. They are created for every service and it’s easy to target the gateway. Just make sure DNS is correct for your hostnames.
Lately I’ve been optimizing remote services for reverse proxy passthru. Did you know that it can break streams momentarily and make your proxy work a little harder if your host names don’t match outside and in?
So in other words if you want full passthru of a tcp or udp stream to your server without the proxy breaking it then opening a new stream you would have to make sure the internal network and external network are using the same fqdn for the service you are targeting.
It actually can break passthru via sni if they don’t use the same hostname and cause a slight delay. Kinda matters for things like streaming videos. Especially if you are using a reverse proxy and the service supports quic or http2.
So a reverse proxy entry that simply passes without breaking the stream and resending it might ook like…
Obviously you would need to get the http port working on jellyfin and have ipv6 working with internal DNS in this example.
server { listen 443 ssl; listen [::]:443 ssl; # Listen on IPv6 address server_name jellyfin.example.net; ssl_certificate /path/to/ssl_certificate.crt; ssl_certificate_key /path/to/ssl_certificate.key; location / { proxy_pass https://jellyfin.example.net:8920; # Use FQDN ... } }Full pass through has no advantage when my reverse proxy terminates ssl and internal services are http only right?
Regardless of fqdn nginx has to decrypt and restream anyways.
This would be correct if you are terminating ssl at the proxy and it’s just passing it to http. However, if you can enable SSL on the service it’s possible to take advantage of full passthru if you care about such things.
Ahh nice good to know. For my use case I’d rather not distribute the certificates to all my services.
When I was experimenting with this it didn’t seem like you had to distribute the cert to the service itself. As long as the internal service was an https port. The certificate management was still happening on the proxy.
The trick was more getting the host names right and targeting the proxy for the hostname resolution.
Either way IP addresses are much easier but it is nice to observe a stream being completely passed through. I’m sure it takes a load off the proxy and stabilizes connections.
