They clearly haven't talked to a telco or network device vendor, they would've sold them a VRF/EVPN/L3VPN based solution… for a whole bunch of money :)
You can DIY that these days though, plain Linux software stack, with optional hardware offload on some specific things and devices. Basically, you have a traffic distinguisher (VXLAN tunnel, MPLS label, SRv6, heck even GRE tunnel), keep a whole bunch of VRFs (man ip-vrf) around, and have your end services (server side) bind into appropriate VRFs as needed.
Also, yeah, with IPv6 you wouldn't have this problem. Regardless of whether it's GUAs or ULAs.
Also-also, you can do IPv6 on the server side until the NAT (which is in the same place as in the article), and have that NAT be a NAT64 with distinct IPv6 prefixes for each customer.
I like to think this is what we did. It's a simple Linux software stack - Linux, nftables, WireGuard, Go... But the goal was also to make it automatic and easy to use. It's not for my Mom. But you don't need a CCNP either.
The trick is in the automation and not the stack itself.
The key distinction with a L3VPN setup is that the packets are unmodified from and including the IP layer upwards, they're just encapsulated/labelled/tagged (depending on your choice of distinguisher). That encapsulation/… is a stateless operation, but comes at the cost of MTU (which in your case should be a controllable factor since the inner flows don't really hit uncontrolled devices.) Depending on what you're trying to do, the statelessness can be anything from useless to service critical (the latter if you're under some risk of DoS due to excessive state creation). It can also alleviate NAT problems, e.g. SIP and RTP are "annoying" to NAT.
(ed.: To be fair, 1:1 NAT can be almost stateless too, that is if your server side ["Technician"] can be 1:1 mapped into the customer's network, i.e. the other direction. This only works if you have very few devices on "your" side and/or/according to how many IPs you can grab on the customer network.)
The IPv6/NAT64 approach meanwhile is very similar to what you did, it just gets rid of the need to allocate unique IP addresses to devices. The first 96 bits of the IPv6 address become a customer/site ID, the last 32 bit are the unmodified device IPv4 address.
10. is /8 (24 payload bits), 172.16 is /12 (so 22) and 192.168 is /16. Very little need to spend more than 18 bits of space to map every 'usable' private IPv4 address once per customer. Probably also less than 14 bits (16k) of customers to service.
There's more addresses I didn't know about offhand but found when looking up the 'no DHCP server' autoconf IP address range (Link Local IPv4).
The problem with talking to a telco, is you have to talk with not just one but any your customer may use. And if at the customer location there’s multiple routers in between the cameras and that telco router, it’s a shitshow trying to configure anything.
Much easier to drop some router on site that is telco neutral and connect back to your telco neutral dc/hq.
No good when the upstream is some wifi connection provided by the building management, rather than a telco themselves.
May as well pick a single solution that works across all Internet connections and weird setups, be an expert in that, vs having to manage varying network approaches based on telco presence, local network equipment, operating country, etc.
On the later parts, VRF in my scenarios won’t scale.
Need to provide support access to 10k-50k locations all with the same subnet (industry standard equipment where the vendor mandates specific IP addressing, for better or worse). They are always feeding in data into the core too.
I wouldn't be surprised if a lot of the hardware under management (e.g. IP cameras, NVRs, cable modems) lacks support for IPv6, and/or the customer networks that it's resident on don't have working IPv6 transit.
The solution is to run ipv6 on the overlay and have the customer site gateway thing they have to translate it to target ipv4. Conveniently you can do the translation it more or less statefully and very easily because you can just embed the ipv4 addr in ipv6. For example you could grab a /64 prefix, assign 32 bits to customer/gateway id and other 32 bits to target ipv4 addr.
Coincidentally I think that's an overestimation on the number of devices that don't support IPv6. At this point, vendors have to go out of their way to disable IPv6, and they lose out on some government/enterprise tenders that require IPv6 even if they're not running it (yet).
IPv6 is very badly supported at the low end of the market. Cheap webcams, doorbells, etc. And that not counting already old equipment...
If we had a nuclear war, we could start over. But for now, we are stuck. Blame it on Cisco for inventing NAT.
man this sounds so realistic, a system administrator saying "yes! Now we can migrate to the IPv6" after the 99.9 percent of the world population is killed
IPv6 solves the addressing problem, not the reachability problem. Good luck opening ports in the stateful IPv6 firewalls in the scenarios outlined in TFA:
> And that assumes a single NAT. Many sites have a security firewall behind the ISP modem, or a cellular modem in front of it. Double or triple NAT means configuring port forwarding on two or three devices in series, any of which can be reset or replaced independently.
I'm not really seeing a reason why it would be impossible to open firewalls in that scenario. More work, sure, but by no means impossible. In any case TFA says right up front that it is trying to solve the problem of overlapping subnets, which IPv6 solves nicely.
Then you've probably never worked in any serious networked embedded systems space. Getting people to open ports on the firewall and making the firewall configuration palatable to the end customer is like a quarter of what I think about when my team makes new features.
Hole punching actually works most of the time. A lot more often than you might think. But enterprise firewalls usually don't allow it. And some home routers fail when you check all the anti-intrusion options. But it's the same for other VPNs.
In the residential and small-business space, it's pretty rare. You might need to point it out to the network guy. If the customer wants the service, they should be open to it.
The problem isn’t that it doesn’t work (and it does often not work – one “symmetric NAT” in the old/deprecated terminology is enough), it’s that it’s orders of magnitude more complex than it needs to be.
I’ve also never seen it work for TCP in practice, and not everybody should have to roll their own UDP wrapper for their TCP-expecting application.
Hole punching is a thing. Ports are not normally completely blocked. They allow replies, which can be exploited to do make a connection. Obviously this requires an out of band signaling mechanism. Tailscale does this, so does WebRTC, iirc.
Yes, but I don't believe all firewalls support that, especially for TCP, and as you've mentioned, now you also need to maintain a handshaking mechanism.
The complexity makes sense if you need to transport a lot of data peer-to-peer or the lowest possible latency, but if you don't, you might as well use that coordination server (which outbound-only clients are connecting to) for payload communication as well.
> I'm not really seeing a reason why it would be impossible to open firewalls in that scenario.
Cheap ass ISP-managed routers. Got to be lucky for these rubbish bins to even somewhat reliably provide IPv6 connectivity to clients at all, or you run into bullshit like new /64's being assigned every 24 hours, or they may provide IPv6 but not provide any firewall control...
You can have /both/ a ULA and a Globally Routable address. In practice it works just fine. My internal DNS points to the ULA for internal connectivity and my hosts use their global addresses for external connectivity.
Ah, you mean for cases where you want both stable addresses (even if only internal) and globally reachable ones (even if non-constant)?
Yeah, that works, but everything gets much easier if your internal DNS can just support the varying prefix natively, e.g. via integration with the external-facing DHCP or PPPoE or whatever other address configuration protocol you use, since then you can reach everything both locally and globally by name.
It also gets more fragile. If your ISP can't or doesn't issue you a prefix for whatever reason then your entire IPv6 network stops working even internally. This is even more pertinent if, like me, you're on a 4G LTE connection. Verizon has great IPv6 support, when you can get it, and when you can't I'd still prefer to have a stable internal network.
That's the addressing problem, although I have some bad news on that: NAT is used with IPv6 in some places.
The reachability problem is, even with public addresses, sometimes you have to do the same thing to "configure port forwarding" with stateful IPv6 firewalls as with double or triple NAT IPv4.
I recently just changed my default subnet to 10.X.Y.... rolling two random numbers to make it highly unlikely my home subnet through wireguard would conflict with the subnet where I am connecting from.
This works fine for your end. But the issue we are addressing is on the other end, when you don't control the network and need to reach devices. If all customer sites are running rfc-unroutable blocks, you eventually encounter conflicts. And the conflict will likely be with the 2nd one you try.
My (very large) corporate network uses 172.16 and 10. heavily, which has lead me to set my docker/daemon.json default-address-pools to 84.54.64.0/18, as it's very unlikely we need to communicate with any IPs in Uzbekistan.
C) I got teased for it a long time ago by my other nerd friends.
But the US DOD has huge blocks of prefixes that it doesn't do anything with, presumably they use it for internal routing so every device they have could publicly route without NAT..
One of those prefixes is 7.0.0.0/8.
My home network uses that. I have never had an issue with S2S VPNs.
However, there have been a few bits of software (pfsense for example) which have RFC1918 hardcoded in some areas and treat it like a public network and overwriting it means doing the entire network setup manually without the helping hand of the system to build-out a working boilerplate.
In this vein there's also 3 TEST-NETs, all /24 but still useful. I've been known to use TEST-NET 1 for Wireguard: 192.0.2.0/24. The other two are 198.51.100.0/24 and 203.0.113.0/24.
There's also 198.18.0.0/15, Wikipedia says it's "Used for benchmark testing of inter-network communications between two separate subnets"[1]. Use this if you really want to thumb your nose at the RFC police.
Do you run Docker? Because I remember having to VPN out to a client that used that range, and it caused conflicts where our docker containers couldn't reach the client side to fetch data.
I had this happen at home. I'm not convinced it was a good idea to choose default subnets as /20.
It was pretty easy to cause myself problems with Docker compose. Eventually I run out of subnets in the 172.16 range and it happily created subnets in the 192.168. range. Some of them overlapped with subnets on my LAN.
We chose Go as the development language. Go produces statically compiled binaries that include all dependencies. The only external deps are wireguard, nftables, nmap, etc. All easy stuff. So we have no need for Docker. We publish binaries for ARM64 and AMD64. Avoiding Docker has made it much easier to work with.
This is basically what I use tailscale & their magicdns feature for. I manage a few locally hosted jellyfin servers for myself and some family members, and its the same problem. I just added tailscale to them all and now I can basically do ssh parents.jellyfin.ts.net or inlaws.jellyfin.ts.net
I need to implement this type of thing for supporting networks of family members, but without the media server aspect - just computer/networking support. I'm looking for a cheap and reliable device that I can put in each home, to give the Tailscale "foothold". Do you happen to know of any tiny devices? I was thinking there must be something even cheaper than a Raspberry Pi to perform this single function at each location.
An old micro pc from dell/hp/lenovo. They are often cheaper and more capable than Raspberry Pis. You can just put up a random Linux distro and it will work.
If they have an Apple TV, you can just install the app and use it as an exit node. I would check out the devices that are on their network currently, chances are you can use one of those.
The only drawback are routes - they won't work on the same CIDR (I mean the fact that you can say in Tailscale "if you want to reach the 192.168.16.13 device that does not support Tailscale, go through this Tailscale gateway"). For this I had to shift my parents' network to be able to access stuff like the printer, in a network that clashed with another one of mine.
The way we did it, roting is not a problem. Any Netrinos client (Windows, Mac, or Linux, including the free version) can act as a gateway. It assigns a unique overlay IP to devices on the local network that can't run software themselves, like cameras, NAS units, or printers, and handles the NAT translation.
Think of it like a router's DMZ feature, but inverted. Instead of exposing one device to the internet, each device gets a private address that's only reachable inside your mesh network.
In your experience, how often does Tailscale have to resort to an external relay server to traverse? I’ve had that out the kibosh on bandwidth/latency sensitive applications before.
> The gateway device performs 1:1 NAT. Traffic arriving for 100.97.14.3 is destination-translated to 192.168.1.100, and the source is masqueraded to the gateway's own LAN address.
Couldn't you tell the WG devices that 192.168.2.0/24 refers to the 192.168.1.0/24 network at customer A, such that 192.168.2.55 is routed to 192.168.1.55. Same for 192.168.3.0/24 referring to customer B.
I think this is what the article is getting at but I don't see the value in manually assigning an alias to each non-wg device, versus assigning an alias to the entire LAN.
The problem there is you still need to keep track of the subnets. It works for a while, but it's quite complex. NAT is actually easier when you get into hundreds of sites.
Yeah so instead DNAT, use NETMAP on the gateway device to that LAN. (Sorry if I'm abusing the terminology, I only do this stuff like once a year for homelab.)
eg this is what I'm currently using to alias my home network
The suggested solution involves using the CGNAT /10 in conjunction with a VPN, but I've actually seen someone do this, and still have problems with certain end users where their next hop for routing also involves a router with an IPv4 address in the same space, so it's not really bulletproof either. We may as well consider doing other naughty things like co-opting DoD non-routable /8s or the test net in the RFCs you're not supposed to use, because basically anything you pick is going to have problems.
That does not happen here. The CGNAT addresses are in the VPN tunnel. And the tunnel connects private devices end-to-end. The LAN packets never see the Internet. They are inside the WireGuard packets.
I decided to learn IPv6 recently and I'm pleasantly surprised how simple and elegant it is. Truly a joy. Highly recommend, if you've never worked with IPv6 to try it. It's like discovering a bidet.
One step beyond this is the multi-subnetted network on each side. You get the DNAT working, but then suddenly the app gets more complex over time and suddenly you're calling 192.168.2.x, which leads to async routes. Some traffic works, some traffic works one way, and other traffic disappears.
Then you as the client/app manager pull your hair out as the network team tells you everything is working fine.
Shameless plug - this is exactly the same problem that our team had when we had to maintain a bunch of our customer's servers. All of the subnets were same, and we had to jump through hoops just to access those servers - vpns, port forwarding, dynamic dns with vnc - we've tried it all. That is why we developed https://sshreach.me/ - now it's a click of a button.
The initial idea started as a bunch of ssh tunnels. Been doing that for years. But WireGuard seemed a better solution at scale, and more efficient. When I first saw WiteGuard, it blew my mind how elegantly simple it was. I always hated VPNs. Now I seem to have made them my life...
> But the moment two sites share the same address range, you have an ambiguity that IP routing cannot resolve.
Writing PF or nft rules to NAT these hyper-legacy subnets on the local side of the layer3 tunnel is actually super trivial, like 20 seconds of effort to reason about and write in a config manifest.
Like written the article, a device on the customer site is required. At that point you might as well deploy a router that has a supportable software stack and where possible sober IP instead of legacy IP.
.
I have been running IPv6-only networks since 2005 and have been deploying IPv6-only networks since 2009. When I encountered a small implementation gap in my favorite BSD, I wrote and submitted a patch.
Anyone who complained about their favorite open source OS having an IPv6 implementation gap or was using proprietary software (and then also dumb enough to complain about it), should be ashamed of themselves for doing so on any forum with "hacker" in the name. But we all know they aren't ashamed of themselves because the competency crisis is very real and the coddle culture let's such disease fester.
There is no excuse to not deploy at minimum a dual-stack network if not an IPv6-only network. If you deploy an IPv4-only network you are incompetent, you are shitting up the internet for everyone else, and it would be better for all of humanity if you kept any and all enthusiasm you have for computers entirely to yourself (not a single utterance).
I won't name the 2 large telecoms I know, that don't support IPv6 being used by customers - if you get L2VPN, L3VPN, other typical services etc. it will be IPv4-only. Of course you can buy a wave and do whatever you want with it :-)
Support for IPv6 is notoriously bad in residential modems. They can barely run IPv4. In an enterprise, you can do it properly. But here we are stuck with the junk the ISP gave out. Customers don't care. You have to work with what you've got.
You can DIY that these days though, plain Linux software stack, with optional hardware offload on some specific things and devices. Basically, you have a traffic distinguisher (VXLAN tunnel, MPLS label, SRv6, heck even GRE tunnel), keep a whole bunch of VRFs (man ip-vrf) around, and have your end services (server side) bind into appropriate VRFs as needed.
Also, yeah, with IPv6 you wouldn't have this problem. Regardless of whether it's GUAs or ULAs.
Also-also, you can do IPv6 on the server side until the NAT (which is in the same place as in the article), and have that NAT be a NAT64 with distinct IPv6 prefixes for each customer.
(ed.: To be fair, 1:1 NAT can be almost stateless too, that is if your server side ["Technician"] can be 1:1 mapped into the customer's network, i.e. the other direction. This only works if you have very few devices on "your" side and/or/according to how many IPs you can grab on the customer network.)
The IPv6/NAT64 approach meanwhile is very similar to what you did, it just gets rid of the need to allocate unique IP addresses to devices. The first 96 bits of the IPv6 address become a customer/site ID, the last 32 bit are the unmodified device IPv4 address.
There's more addresses I didn't know about offhand but found when looking up the 'no DHCP server' autoconf IP address range (Link Local IPv4).
https://en.wikipedia.org/wiki/IPv4#Special-use_addresses
Much easier to drop some router on site that is telco neutral and connect back to your telco neutral dc/hq.
May as well pick a single solution that works across all Internet connections and weird setups, be an expert in that, vs having to manage varying network approaches based on telco presence, local network equipment, operating country, etc.
(btw, have you actually read past the first 7 words? I'm much more interested what people think about the latter parts.)
Need to provide support access to 10k-50k locations all with the same subnet (industry standard equipment where the vendor mandates specific IP addressing, for better or worse). They are always feeding in data into the core too.
Much easier to just VPN+NAT.
https://en.wikipedia.org/wiki/List_of_IPv6_transition_mechan...
Coincidentally I think that's an overestimation on the number of devices that don't support IPv6. At this point, vendors have to go out of their way to disable IPv6, and they lose out on some government/enterprise tenders that require IPv6 even if they're not running it (yet).
> And that assumes a single NAT. Many sites have a security firewall behind the ISP modem, or a cellular modem in front of it. Double or triple NAT means configuring port forwarding on two or three devices in series, any of which can be reset or replaced independently.
Nor can everyone control whether their connection supports v6, unfortunately.
I’ve also never seen it work for TCP in practice, and not everybody should have to roll their own UDP wrapper for their TCP-expecting application.
See: https://tailscale.com/blog/how-nat-traversal-works
The complexity makes sense if you need to transport a lot of data peer-to-peer or the lowest possible latency, but if you don't, you might as well use that coordination server (which outbound-only clients are connecting to) for payload communication as well.
Cheap ass ISP-managed routers. Got to be lucky for these rubbish bins to even somewhat reliably provide IPv6 connectivity to clients at all, or you run into bullshit like new /64's being assigned every 24 hours, or they may provide IPv6 but not provide any firewall control...
It'd be nice if DNS servers supported this. Save the 64 host bits in the zone and just use whatever 64 prefix bits happen to be issued right now.
Otherwise it makes a strong case for the continued use of "private networks" and the IPv6 ULA mechanism.
Let's please not. Even without inbound reachability, hole punching is significantly easier given globally routeable addresses.
Yeah, that works, but everything gets much easier if your internal DNS can just support the varying prefix natively, e.g. via integration with the external-facing DHCP or PPPoE or whatever other address configuration protocol you use, since then you can reach everything both locally and globally by name.
It also gets more fragile. If your ISP can't or doesn't issue you a prefix for whatever reason then your entire IPv6 network stops working even internally. This is even more pertinent if, like me, you're on a 4G LTE connection. Verizon has great IPv6 support, when you can get it, and when you can't I'd still prefer to have a stable internal network.
Stateful firewalls are very much a thing on v6. Many mobile ISPs don't allow incoming connections by default, for example.
Many CPEs (home routers) also come with a v6 firewall (I'd guess it's probably more common than not?), and not everybody has admin access to theirs.
The reachability problem is, even with public addresses, sometimes you have to do the same thing to "configure port forwarding" with stateful IPv6 firewalls as with double or triple NAT IPv4.
I kinda don't want to share this because:
A) it's a bad idea
B) it means it will be less unique
and
C) I got teased for it a long time ago by my other nerd friends.
But the US DOD has huge blocks of prefixes that it doesn't do anything with, presumably they use it for internal routing so every device they have could publicly route without NAT..
One of those prefixes is 7.0.0.0/8.
My home network uses that. I have never had an issue with S2S VPNs.
However, there have been a few bits of software (pfsense for example) which have RFC1918 hardcoded in some areas and treat it like a public network and overwriting it means doing the entire network setup manually without the helping hand of the system to build-out a working boilerplate.
There's also 198.18.0.0/15, Wikipedia says it's "Used for benchmark testing of inter-network communications between two separate subnets"[1]. Use this if you really want to thumb your nose at the RFC police.
[1] https://en.wikipedia.org/wiki/List_of_reserved_IP_addresses
Docker defaults to 172.16.0.0/16.
It was pretty easy to cause myself problems with Docker compose. Eventually I run out of subnets in the 172.16 range and it happily created subnets in the 192.168. range. Some of them overlapped with subnets on my LAN.
The only drawback are routes - they won't work on the same CIDR (I mean the fact that you can say in Tailscale "if you want to reach the 192.168.16.13 device that does not support Tailscale, go through this Tailscale gateway"). For this I had to shift my parents' network to be able to access stuff like the printer, in a network that clashed with another one of mine.
Think of it like a router's DMZ feature, but inverted. Instead of exposing one device to the internet, each device gets a private address that's only reachable inside your mesh network.
Couldn't you tell the WG devices that 192.168.2.0/24 refers to the 192.168.1.0/24 network at customer A, such that 192.168.2.55 is routed to 192.168.1.55. Same for 192.168.3.0/24 referring to customer B.
I think this is what the article is getting at but I don't see the value in manually assigning an alias to each non-wg device, versus assigning an alias to the entire LAN.
eg this is what I'm currently using to alias my home network
With other wg peers getting a 192.168.150.0/24 entry in the AllowedIPs for this gateway (if needed).Then you as the client/app manager pull your hair out as the network team tells you everything is working fine.
Writing PF or nft rules to NAT these hyper-legacy subnets on the local side of the layer3 tunnel is actually super trivial, like 20 seconds of effort to reason about and write in a config manifest.
Like written the article, a device on the customer site is required. At that point you might as well deploy a router that has a supportable software stack and where possible sober IP instead of legacy IP.
.
I have been running IPv6-only networks since 2005 and have been deploying IPv6-only networks since 2009. When I encountered a small implementation gap in my favorite BSD, I wrote and submitted a patch.
Anyone who complained about their favorite open source OS having an IPv6 implementation gap or was using proprietary software (and then also dumb enough to complain about it), should be ashamed of themselves for doing so on any forum with "hacker" in the name. But we all know they aren't ashamed of themselves because the competency crisis is very real and the coddle culture let's such disease fester.
There is no excuse to not deploy at minimum a dual-stack network if not an IPv6-only network. If you deploy an IPv4-only network you are incompetent, you are shitting up the internet for everyone else, and it would be better for all of humanity if you kept any and all enthusiasm you have for computers entirely to yourself (not a single utterance).