Hacker News Reader

Comment by avidiax | Hacker News Reader

> Fact is you'd run into exactly the same problems as with IPv6.

If you treat IPv4 addresses as a routable prefix (same as today), then the internet core routers don't change at all.

Only the edge equipment would need to be IPv4+ aware. And even that awareness could be quite gradual, since you would have NAT to fall back on when receiving an IPv4 classic packet at the network. It can even be customer deployed. Add an IPv4+ box on the network, assign it the DMZ address, and have it hand out public IPV4+ addresses and NAT them to the local IPv4 private subnet.

IPv6 seems to be a standard that suffered from re-design by committee. Lots of good ideas were incorporated, but it resulted in a stack that had only complicated backwards compatibility. It has taken the scale of mobile carriers to finally make IPv6 more appealing in some cases than IPv4+NAT, but I think we are still a long way from any ISP being able to disable IPv4 support.

View on HN

> Only the edge equipment would need to be IPv4+ aware.

"Only"? That's still the networking stack of every desktop, laptop, phone, printer, room presentation device, IoT thing-y. Also every firewall device. Then recompile every application to use the new data structures with more bits for addresses.

And let's not forget you have to update all the DNS code because A records are hardcoded to 32-bits, so you need a new record type, and a mechanism to deal with getting both long and short addresses in the reply (e.g., Happy Eyeballs). Then how do you deal with a service that only has a "IPv4+" address but application code that is only IPv4-plain?

Basically all the code and infrastructure that needed to be updated and deployed for IPv6 would have to be done for IPv4+.

But the desktop/laptop/phone/printer was the EASIEST thing to change in that 30 year history. And it would have been the easiest thing to demand a change req from a company for.

Yes: but the process would have been exactly the same whether for a hypothetical IPv4+ or the IPng/IPv6 that was decided on; pushing new code to every last corner of the IP universe.

How could it have been otherwise given the original network structures were all of fixed lengths of 32 bits?

The new code would have been vastly simpler. IPv6 is second system syndrome personified.

What we needed was the equivalent of ASCII->UTF8.

As someone with non-ascii and non-latin-1 characters in my surname, I can tell you that the ascii->utf8 migration still hasn’t finished.

Just a few weeks ago I ordered something from JBL US and somehow on the UPS sticker an "Á" became a caret "^"

shrug

Most of the world is a circus.

If you hand UTF-8 that actually uses anything added by utf-8 to something that can only render ASCII, the text will be garbled. People can read garbled text ok if it’s a few missing accented characters in a western language, but it’s no good for Japanese or Arabic.

In networking terms, this is like a protocol which can reach ipv4 hosts only but loses packets to the ipv4+ hosts randomly depending on what it passes through. Who would adopt a networking technology that fails randomly?

If we have IPv4 address 1.2.3.4, and the hypothetical IPv4+ adds 1.2.3.4.1.2.3.4 (or longer), how would a IPv4-only router handle 1.2.3.4.1.2.3.4? If an IPv4-only host or application gets a DNS response with 1.2.3.4.1.2.3.4, how is it supposed to use it?

As I see it, the transition mechanism for some IPv4+ that 'only' has longer addresses is exactly the same as for IPv6: new code paths that use new data structures, with a gradual rollout with tech refreshes and code updates where hosts slowly go from IPv4-only to IPv4-and-IPv4+ at different rates in different organizations.

If you think it's somehow different, can you explain how it is so? What proposal available (especially when IPng was being decided on in the 1990s) would have allowed for a transition that is different than the one described above (gradual, uncoördinated rollout)?

* https://datatracker.ietf.org/doc/html/rfc1726

* https://datatracker.ietf.org/doc/html/rfc1752

The proposal is that IPv4+ would be interpretable as an IPv4 packet. Either the IP header is extended, or we add another protocol layer for the IPv4+ bits (IPv4+ is another envelope for the user payload).

DNS is like today: A and AAAA records for IPv4 and IPv4+ respectively.

Core routers do not need to know about IPv4+, and might never know.

The transition is similar to 6to4. The edge router does translation to allow IPv4+ hosts to connect to IPv4 hosts. IPv4 hosts are unable to connect to IPv4+ directly (only via NAT). So it has the similar problem to IPv6 that you ideally want all servers to have a full IPv4 address.

What you don't have is a completely parallel addressing system, requirements to upgrade all routers (only edge routers for 4+ networks), requirements to have your ISP cooperate (they can just give you an IPv4 and you handle IPv4+ with your own router), and no need that the clients have two stacks operating at once.

It's essentially a better NAT, one where the clients behind other NATs can directly connect, and where the NAT gradually disappears completely.

And in 30 years, all of that has basically already happened and afoption is still absymal.

v6 has nearly 3 billion users. How is that abysmal?

We've never done something like the v4->v6 migration before, on this sort of scale. It's not clear what the par time for something like this is. Maybe 30 years is a normal amount of time for it to take?

HTTP->HTTPS was this kind of scale, and it was smooth because they changed as little as possible while also being very careful about default behaviors.

3 billion people sorta use ipv6, but not really, cause almost all of those also rely on ipv4 and no host can really go ipv6-only. Meanwhile, many sites are HTTPS-only.

And because it's a layer 7 thing, so it only required updating the server and client software, not the OS... and only the client and server endpoints and not the routers in between... and because we only have two browser vendors who between them can push the ecosystem around, and maybe half a dozen relevant web server daemons.

Layer 3 of the Internet is the one that requires support in all software and on all routers in the network path, and those are run by millions of people in hundreds of countries with no central entity that can force them to do anything.

HTTP->HTTPS is only similar in terms of number of users, not in terms of the deployment itself. The network effects for IP are much stronger than for HTTP.

They don't "sorta" use v6, they're properly using it, and you can certainly go v6-only. I'm posting from a machine with no v4. Also, if you want to go there: HTTPS was released before IPv6, and yet still no browser is HTTPS only, despite how much easier it is to deploy it.

I know they aren't very comparable in a technical way, but look at the mindset. IPv6 included decisions that knowingly made it more different from v4 than strictly needed, cause they wanted it to be perfect day 1. If they did HTTPS like this, it'd be tied to HTTP/2.

Most browsers now discourage plain HTTP with a warning. Any customer-facing server basically needs to use HTTPS now. And you're rare if you actually have no ipv4, not even via a tunnel.

HTTP has the leeway to do that because they have an easier technical job deploying updates.

If they only got one shot at changing HTTP, do you think they would have tied TLS to HTTP/2 or given up on HTTP/2 altogether?

The compromised "ipv4+" idea a bunch of people keep asking for wouldn't require changing the spec down the road. ISPs would just need to clean up their routes later, and SLAAC could still exist as an optional (rather than default) feature for anyone inclined to enable later. Btw, IPv6 spec was only finalized in 2017, wasn't exactly one-shot.

I don't know if HTTP's job is easier. Maybe on the client side, since there were never that many browsers, but you have load-balancers, CDNs, servers, etc. HTTP/2 adoption is still dragging out because of how many random things don't support it. Might be a big reason why gRPC isn't so popular too.

> HTTP->HTTPS was this kind of scale, and it was smooth because they changed as little as possible while also being very careful about default behaviors.

HTTP->HTTPS is not equivalent in any way. The payload in HTTP and HTTPS are exactly the same; HTTPS simply adds a wrapper (e.g., stunnel can be used with an HTTP-only web server). Further HTTP(S) is only on the end points, and specifically in the application layer: your OS, switch, firewall, CPE, ISP router(s), etc, all can be left alone.

If you're not running a web browser or web server (i.e., FTP, SMTP, DNS, database) then there are zero changes that need to be made to any code on a system. This is not true for changing the number of bits the addressing space: every piece of code that calls socket(), bind(), connect(), etc, has to be touched.

Whereas the primary purpose of IPng was to expand the address space, which means your OS, switch, firewall, CPE, ISP router(s), etc, all have to be modified to handle more address bits in the Layer 3 protocol data unit.

Plus stuff at the application layer like DNS (since A records are 32-bit only, you need an entire new network type): entire new library functions had to be created (e.g., gethostbyname() replaced by getaddrinfo()).

I hear people say the IETF/IP Wizards of the 1990s should have "just" picked an IPng that was a larger address space, but don't explain how IPv4 and hypothetical IPv4+ would actually work. Instead of 1.1.1.1, a packet comes in with 1.1.1.1.1.1.1.1: how would a non-IPv4+ router know what to do with that? How would non-updated routers and firewalls be able to handle longer addresses? How would non-updated DNS code be able to handle new record types with >32 bits?

HTTP->HTTPS looks easy in hindsight, but there were plenty of ways it could have gone wrong. They took the path of least resistance, unlike ipv6. I know they're different layers ofc.

To answer the last question, routers would need IPv4+ support, just like ipv6 which already happened. The key is it's much easier for users to switch after. No dual stack, you get the same address, routes, DNS, and middleboxes like NAT initially. ISPs can't hand out longer addrs like /40 until things like DNS are upgraded in-place to support that, but again those are pretty invisible changes throughout the stack.

> To answer the last question, routers would need IPv4+ support, just like ipv6 which already happened.

So exactly like IPv6: you need to roll out new code everywhere.

> The key is it's much easier for users to switch after. No dual stack, you get the same address, routes, DNS, and middleboxes like NAT initially. ISPs can't hand out longer addrs like /40 until things like DNS are upgraded in-place to support that, but again those are pretty invisible changes throughout the stack.

So exactly like IPv6: you need to roll out new code everywhere.

Would organization have rolled out in IPv4+ any differently than IPv6? Some early, some later, some questioning the need at all. It's the exact same coördination / herding cats problem.

It's a simple toggle on vs asking orgs to redo their entire network. In both cases you need routers and network stacks to support the new packet format, but that isn't the hard part of ipv6, we already got there and people still aren't switching.

Sorry, I'm still not seeing how a IPv4+ would be any less complicated (or as simple) as IPv6. In either case you would still have to:

* roll out new code everywhere

* enable the protocol on your routers

* get address block(s) assigned to you

* put those blocks into BGP

* enable the protocol on middleware boxes

* have translation boxes for new-protocol hosts talk to old-protocol-only hosts

* enable the protocol on end hosts

And just because you do it, does not mean anyone else would do in the same timeframe (or ever). You're back in the chicken-and-egg of whether servers/services do it first ("where are the clients?"), or end-devices ("where are the services?").

Everything you listed was already done for ipv6 or is trivial to enable, but people still aren't switching, because of all the things you didn't list.

What did they not list?

Redo all your addresses and routes, reconfigure or replace NAT and DHCP, reconfigure firewall, change your DNS entries at minimum. If it's a home or small business and you don't want to fight the defaults, you go from NAT to NATless.

No, routers would have to be fixed anyway, because even if you put extra bits into extension header we have 30 years of experience that routers and ISPs will regularly fuck around with those extra bits - it's related to why we have TLS GREASE option.

Application rework would be exactly the same as with v6, because the issue was not with v6 but with BSD Sockets API exposing low-level details to userland.

> Only the edge equipment would need to be IPv4+ aware. And even that awareness could be quite gradual, since you would have NAT to fall back on when receiving an IPv4 classic packet at the network. It can even be customer deployed. Add an IPv4+ box on the network, assign it the DMZ address, and have it hand out public IPV4+ addresses and NAT them to the local IPv4 private subnet.

Congratulations, you’ve re-invented CGNAT, with none of the benefits, and the additional hassle of it being an entirely new protocol!

No. No “extra bits” on an IPv4 address would have ever worked. NAT itself is a bug. Suggesting that as an intentional design is disingenuous.

I have not "reinvented CGNAT". It is hierarchal public addressing similar to IPv4 and IPv6.

The edge router has an IPv4+ subnet (either a classic v4 address, or part of a v4+ address). It maintains an L2 routing table with ARP+, and routes IPv4+ packets to the endpoint without translation. Private subnetting and NAT is only needed to support legacy IPv4 clients.

CGNAT pools IPv4 public addresses and has an expanded key for each connection, and translates either 4 to 6 or into a private IPv4 subnet. My proposal needs no pooling and only requires translation if the remote host is IPv4 classic and the edge router is not assigned a full IPv4+/24.

Not just the edge router. Every router between the ISP edge and the destination edge.

And since the goal is “backwards-compatability”, you’d always need to poll, because a “legacy” IPv4 client would also be unable to send packets to the IPv4+ destination. Or receive packets with an IPv4+ source address.

And it would be an absolute nightmare to maintain. CGNAT + a quasi backwards-compatible protocol where the backwards-compatability wouldn’t work in practice.

So you would have exactly the same problem as IPv6. I can say the same about v4 and v6 today. You could just turn off IPv4 on the internet, and we’d only need to do translation on the edge for the legacy clients that would still use IPv4. You can even put IPv4 addresses in IPv6 packets!

I think you've actually reinvented 6to4, or something morally very close to it.

Each v4 address has a corresponding /48 of IPv6 tunnelled to it. The router with that IP receives the tunnelled v6 packets, extracts them and routes them on natively to the end host. This is something that v6 already does, so you don't need to make posts complaining about how dumb they were for not doing it.

That's quite true, but in this counterfactual, IPv4+ doesn't pretend that 6to4 is just a transition mechanism to an all-IPv6 future. That is, IPv4+ is as-if 6to4 was the default, preferred, or only mechanism, and core routers were never demanded to upgrade.

It's an edge based solution similar to NAT, but directly addressable. And given that it extends IPv4, I think it would have been much more "marketable" than IPv6 was.

But again, this is all counterfactual. The IETF standardized IPv6, and 30 years on it's still unclear that we will deprecate IPv4 anytime soon.

But we do want a v6-only future, right? We don't want to be running both protocols forever, which is what you'd be asking for.

Reply on news.ycombinator.com