This is not correct, there are many ways to do this without resorting to NAT hacks like uPnP.

For example, ISPs hand out entire /64 per customer (at a minimum, almost all will give you larger subnets if you ask). This allows you to have a separate IP address per application, which can be allocated from an unfirewalled pool.

There's also the fact that almost all major OSs now come with sane firewall defaults, also allowing applications to punch holes via APIs. (You could also have done this with IPv4 if it had a larger address space.)

If you spend some time thinking about how to solve these problems with IPv6, you will find that there are lots of ways to do it that are simpler and far more secure.

None of the solutions you suggest exist today, to my knowledge. Am I wrong?

For example, a single IP address per application would be fine, but there is no protocol that allows me to allocate addresses per application on a central home router such that I can then identify them to permit them individually.

And applications punching holes via APIs does exist in uPnP and NAT-PMP, but isn't that the same thing as what you dismissed above as a hack?

You are not wrong. There is no consistent way to do this today -- most of it is in debate, and there is no standard.

The most reliable way is to rely on your system firewall to prompt you to open up a port when an application binds to it (which can be accomplished with the default firewall software on a modern system.) Again, this is technically possible with IPv4 too, it's just impractical due to address space limitations.

> there is no protocol that allows me to allocate addresses per application on a central home router

Correct. But the whole point it to take the router out of the loop for this, and only use it as a coarse firewall. With IPv6 very device gets a separate /64, so it can simply allocate from its own pool of 10^19 addresses.

> And applications punching holes via APIs does exist in uPnP and NAT-PMP, but isn't that the same thing as what you dismissed above as a hack?

Not the same. With IPv6, you do not need to coordinate across multiple NAT devices (or even worse if you're double/triple NATted as in some countries.) You're also not limited by the maximum NAT table sizes (again amplified in the double/triple NATted situations.) Since each system has 10^19 addresses (preallocated out of a global pool of 10^38 addresses), these can now be treated as system resources and managed intelligently by the kernel.

> The most reliable way is to rely on your system firewall to prompt you to open up a port when an application binds to it...But the whole point it to take the router out of the loop for this, and only use it as a coarse firewall.

I don't trust, and doubt I'll ever trust, the firewall on my Smart TV, even if it had one. That's because I don't trust that it'll ever receive timely security updates. So I must use [the firewall on] a central router.

Then trust your router's firewall. And if you don't trust your router's firewall build your own router (or install OpenWRT on store-bought router).

Which puts us back at square one with upnp to open up holes on the router.

> And applications punching holes via APIs does exist in uPnP and NAT-PMP, but isn't that the same thing as what you dismissed above as a hack?

I think applications communicating with firewalling is not a hack, but applications communicating with routing is a hack. E.g. I have two friends who share a house (and therefore an internet connection) and I can play a game with one or other of them (via uPnP) but not both at the same time. With IPv6 + uPnP (assuming that's supported), they'd still have the same firewalling they have now, but I'd be able to play a game with both of them. No?

Sorry, I've only just seen this. I agree with you on this point, yes. But I think you have a pretty rare and specific case. The more general cry of "IPv6 fixes NAT!" incorrectly expects much more than this, I think.

I agree with you and wish there was a public address/private address pair system, so that connections origionate from the private address (which has NAT-like semantics) while servers intending to be public listen on the public address. Local systems could still connect to the private address, so server applications could still listen there by default but would not listen on the public address unless explicitly configured to do so. Not having this local/wide address distinction makes it more likely that firewalls will simply block all traffic until, as you say, similar systems to those used with NAT are implemented.

Another way to fix this problem is a personal VPN that only establishes outgoing connections automatically by default but can proxy incoming connections when explicitly requested. This is the same type of thing as used with NAT, except without the expectation that your ISP is the one providing it. Considering how obnoxious ISPs are these days in terms of filtering and such, this would be a major improvement. Since I am currently doing this (minus the proxying so far which I haven't needed but could easily add) for $15/year with a IPv4 address and virtual host that is way overprovisioned for what it needs, it seems likely that this could be done on a large scale for IPv6 at a quite low cost.

> Not having this local/wide address distinction makes it more likely that firewalls will simply block all traffic until, as you say, similar systems to those used with NAT are implemented.

Yes, and that would be the sensible thing to do.

This "public/private" distinction doesn't even work: What if you have two LANs that aren't bridged? Possibly because they are connected through a VPN? Or maybe one Ethernet and one WiFi? Can you communicate between them? Or only on "public addresses"? What if the VPN is between two companies, say, who don't trust each other? And why would you make it depend on the "address type" anyhow, why not just have the application drop connections that come from outside the LAN?

Blocking all inbound connections by default is no real problem as long as you have a protocol for telling the firewall which addresses/ports to open to the public - the problem with the "port forwarding" that you need with NAT is that ports collide and that you create asymmetric address translations or bottlenecks and all that crap, a simple protocol that tells the firewall to open a specific port or address would be trivial to implement and to use.

Also, IPv6 has link-local addresses, which are essentially "private addresses for a LAN", so you can actually set up services that are reachable only on a given LAN (and nowhere else) - but that really only makes sense for local LAN management (neighbour discovery and such), not for application protocols, those really should just be reachable from everywhere, as far as the local API is concerned, and access control should then happen based on source addresses, possibly implemented as a stateful packetfilter at the network boundary, as you otherwise severely limit which network architectures can be used.

I just noticed your reply but thanks, I agree that source filtering at the OS level would accomplish the same thing, be more flexible, and be more consistent with existing practice.

I think the main additional assumption I have to (part of) what you describe is that applications (on a general purpose system) cannot be trusted to behave reasonably on the network, however users need insecure local networking to work. The public/private disctinction would work the way I imagined it, but only because I was assuming a limited form of source address filtering: essentially one bit on every interface (including VPNs) for "is this local" (I also assume separate addresses per interface). The second address is so that so wildcard listening only listens on local addresses, and explicit public addresses need to be specified (or otherwise enabled) by the user before a server will listen on it. I didn't realize that just doing source address filtering as you suggest could accomplish the same thing with one address in a much simpler way.

There's an upside to this, yes ingress will still be blocked by default (good!) but with uPnP or something along those lines, we now open a single port and it's a 1:1 connection. No longer are we limited to having one device in the household holding that forwarded port and all traffic hitting the external IP going to a single internal machine.

This would allow two users to have the same port opened. This makes it easier to communicate with others on the internet.

> Firewalling P2P traffic effectively remains just as hard with IPv6 as with NAT.

Nope, it's much easier because the addresses are easy to predict. When both parties are behind NAT, neither side knows which address and port they will appear to be coming from to the other side, while without NAT, both parties know exactly what address the other side will see (it's the one the local socket is bound to ...), that makes it relatively easy to establish a connection through a stateful firewall with a simultaneous connect if both sides can somehow tell each other where their respective sockets are.

It's outright trivial with UDP, not quite that easy with TCP: Assume both sides are behind a stateful firewall, both sides simply send a packet from their socket to the respective remote address/port, which will cause each local firewall to establish a connection tracking entry that will allow packets in the respective opposite direction to pass through. There is a race condition that might cause one of the packets to be rejected by the remote firewall (if it arrives before the locally originated packet passed through the firewall), but that doesn't matter: Afterwards, both firewalls have an entry for the connection, and all subsequent packets will pass through just fine. This is impossible with NAT because you cannot know the remote address that the NAT will assign for the connection, so you cannot send a packet to establish the entry in your firewall, and you cannot learn the address because you would need to receive the packet to see the address on it, which you can't as long as the firewall doesn't know about it.

Also, a protocol for opening ports on the stateful firewall would be much simpler without NAT, as the firewall doesn't need to allocate ports or handle conflicts between multiple clients wishing to be forwarded the same port: The address/port allocation authority is with the client, the client('s kernel/network stack) can simply allocate ports on its assigned addresses and then only needs to tell the firewall that connections to a specific address plus port should be allowed in. This also means that, for example, the firewall forgetting those open ports doesn't cause much trouble: The clients simply can tell it again whith ports to open, and they will be reachable again, at the same address, no need to tell peers a new address because a uPnP gateway forgot which ports it had allocated to which client.

And maybe most importantly: P2P traffic between devices on the same LAN becomes much, much simpler. Currently, if you have two participants of some P2P system in one LAN segment, where both establish an open port on the NAT gateway via uPnP, when one wants to connect to the other, there are roughly three possible scenarios: It can try to connect to the public address/port as established by uPnP, just like any remote client would. If you are unlucky, the NAT gateway only does DNAT, in which case the connection won't work because any responses will come from the local address of the peer instead of from the public address of the NAT gateway, so won't be recognized as belonging to the connection. If you are lucky, the NAT gateway does hairpin NAT, in which case the connection works, but all traffic has to go through the NAT gateway in order to rewrite addresses, instead of directly from one machine to the other at the ethernet level, which usually will be much slower than a direct transfer (even more so when there might be multiple people doing the same thing on the same LAN, where the switch might well be able to forward multiple gigabit flows at once, but your DSL NAT gateway most likely will not, not even if they are both in the same case). Or, if you want to avoid both of that, you'd have to somehow detect that you are in the same LAN and then figure out what the address of the peer on the LAN is in order to connect to that address instead (and note that you can't use the local addresses to decide that, even if you know them, as you might well have addresses in the same private address range, but still be connected to completely different LANs).

Without NAT? You simply connect to whatever address the peer announces. If you are on the same LAN, your machine's routing table will automatically route the packets to your local LAN, directly to the other machine, at full LAN speed. If not, they will be forwarded to your local gateway, through the internet, through the remote stateful firewall, and to the remote peer. No difference at all, apart from the speed, possibly, it just works.