Hacker News Reader

Comment by zinconfire | Hacker News Reader

The energy that is being spent convincing millions of nodes to talk to 127/8 might be better spent convincing those nodes to talk IPv6.

View on HN

There’s a lot less effort required to update an existing IPv4 system to accept a few more ranges than to entirely rearchitect a network to work with IPv6.

Plus there’s a lot of energy in this world to do both without one taking away effort from the other.

But as to the actual proposition, I would rather the 240/4 space was changed to RFC1918 private space, or at least half of it.

Really?

I for one am not looking forward to all the frankenbugs when new stuff on those IPs talk to old stuff (middle and endboxes)...

The great part about IPv6 in that regard is that you are either capable or not, not mostly compatible as an older client would be with this "ipv4+" thingy..

IPv6 implementors are inconsistent in what they support. Eg Android and it not supporting DHCPv6 [1].

Also NAT in IPv6 is not as widely supported as it is in IPv4 (and yes NAT can and is used for reasons other than security, such as policy based routing and routers with multiple disparate WANs like Fibre and Cellular).

Plus there will always be someone somewhere with a middlebox, be it IPv4 or IPv6.

1. https://lostintransit.se/2020/05/22/its-2020-and-androids-ip...

NAT is not security. [1]

Why would you want to do NAT in IPv6? You can do all the things you mentioned with IPv6 without NAT.

1. https://samy.pl/slipstream/

I never said NAT was security.

No, the things I mentioned cannot be done easily with IPv6. You cannot easily switch a router from using one WAN to another (with different IPs) without requiring IPv6 clients having to obtain new addresses via SLAAC (and hence downtime), or use NAT (which is what IPv4 does, but not well supported in IPv6).

You cannot easily have 2-3 levels of networks where router3 is attached to the lan of router2 which is attached to the lan of router1, which has internet access. This is because DHCP-PD is not well supported (see Comcast) or the prefix delegated is insufficiently sized (a /56 minimum should be standard but some will give a /60 or worse, just a /64!). This is bread and butter to do in IPv4.

Isn't it possible to have IPs from both WANs assigned to devices on your network? In that case you don't even have to switch.

Problem is, the device doesn't have to know, which address is the working one - e.g. which prefix is working on the upstream. It might break connections etc. A solution of course could be to have NAT66 and I really mean address translation not masquerading/ port translation here - as such, much easier to do as it is a 1:1 mapping or to get provider independent space or become a LIR and have two providers announce your space with respective BGP configuration for failover. It is a solvable problem but isn't much guidance and experience with this at smaller companies that might not want to run BGP or go through hoops to get PI space etc.

Exactly this.

And you are right BGP is a non-starter for smaller sites, especially where cellular is involved. Often small site businesses just need a backup internet connection to access the web and emails, which doesn’t need the overhead of BGP. Also there is a non-trivial cost to using BGP (equipment, expertise, and ISP will charge more for the privilege) which for smaller business doesn’t make economic sense.

You sound like you know what you're talking about, which makes me a bit confused because you appear to be misinformed here.

Your router will advertise new routes immediately when it's given a new WAN address, there's no downtime incurred explicitly from IPv6 in that scenario: local connections will still route until the timeout anyway.

SLAAC issues (and can issue) more than a single IP per device, leading to interesting bugs like the ones Sam Bowne talks about in windows.[0]

Changing a WAN interface _does_ incur downtime on ipv4 too, unless you're assigning virtual NICs and then failing over: which IPv6 would support in identical ways.

[0]: https://youtu.be/YOglp0m35D8?t=1362

Yes sorry I should have been clearer, it’s the connections from clients having to timeout and re-establish upon IP change that is the issue. Specifically where those packets might be going over a SD-WAN type solution where a WAN change is non-disruptive to SD-WAN destine traffic. It relies on clients having NAT addresses and using those addresses for Internet destination addresses (not just private space).

Another issue with SLAAC/lack of NAT on the LAN clients is you can’t do policy based routing on the router, eg I’ve got a fast fibre but it’s costly to use, I’ve also got a cheap and cheerful cable/DSL line that I use for backups and FTP. I want to route all requests to my backup service (which has a domain that resolves to a public IPv6 address) via the cable/DSL and everything else via Fibre. Easy to do on IPv4.

But, nothing you said is inherent to NAT, you should read up a bit more on iBPF for the latter and the former works exactly the same way on v6 as v4. Static routing is not really a thing unless you’re a really bad network admin- for sure your ISP wont be doing static routing to your home.

Can you talk about this some more? Also do you mean eBPF or iBPF or iBGP? If it's BGP that's just too complicated/expensive for a lot of the small businesses I deal with.

Policy routing works exactly the same on IPv6 as on IPv4. People misusing NAT or internal addresses to handle policy routing are the problem.

I've no idea why SLAAC would prevent you from doing policy routing - the prefix is deterministic.

In IPv4 I presume you guessed who each host is by some sort of DHCP registration. For IPv6 you would use either NDP, or perhaps even way more secure SEND enhancement with PKI instead. Of course static assignment still just works, including in NDP/SEND. (e.g. by MAC address or per link)

RIPng/OSPFv3/IS-IS/EIGRP can also tell other hosts which route you prefer packets to return by with better granularity and dynamically. This works even better with v6 since global addresses are always available.

The above doesn’t work if you are dealing with client devices that don’t support routing protocols (Ie phones, laptops, tablets, CCTV systems, PCs etc - everything except routers and servers) and especially with BYOD devices.

Also perhaps I missed something in your reply, but policy based routing - the definition I have of it - encompasses more than routing different subnets via different WANs. Ie routing based on destination port, IP DSCP flags, WAN utilisation etc rather than just by source.

And as I’ve mentioned elsewhere, the cost and complexity of running routing protocols is a non-starter for many small and/or low margin businesses. Think retail or fast food or fuel etc.

> You cannot easily switch a router from using one WAN to another.

You cannot easily do that with NAT either.

Every single one of your TCP sessions will die, and likely every session going over UDP as well. Unless you're talking some special protocol specifically designed to handle clients changing IPs.

Your machine may not notice, but the machines responding to your packets sure will.

Yes you can if you are using a SD-WAN solution running on top of those WANs.

Finally someone who get's it.

First - you can't actually get static IPv6 despite the claims that there is plenty of it as a home user in the US. This static IPv6 for everyone is basically a total lie.

It's super simple with IPv4 to setup an internal network with known hosts / DHCP lease reservations if needed / dns etc, and have all that stay totally unchanged as the WAN link bounces around. If you are supporting a bunch of users (who yes, will configure a printer using the printers IP etc) then this makes like simple. You can do this at home too easily etc.

There is a weird myopia in the IPv6 world about use cases for IP. It's like they just ignored what people need and want, and went a totally different direction.

We need IPv4+ with another octet and be done with this :)

I'm serious, anyone interested in trying to push this. Keep everything the same except with another octet at the front?

For your local DNS names, etc, you could use the link-local IPv6 address.

For some reason these change endlessly on my network.

RFC4941 or similar privacy extension stuff might explain this.

Let's keep the IPv4 header, but start counting packets from byte -2 instead of 0.

i need ipv6 nat because isp gives me only 1 ip.

At this point Google's refusal to implement DHCPv6 is well documented and is being brought up in a lot of arguments. Easy solution: if you need DHCPv6 you don't deploy Android devices and move on.

It’s only easy if you don’t have to support Android devices. Often it’s not possible to do so.

That's not how it works. For BYOD or home use, the device is always more important than the network. You can't just tell somebody "lol, Android not supported". Most of the phones out there are Androids, and not everyone prefers iOS even without the extra cost attached to an Apple phone.

DHCPv6 isn't mandatory to use IPv6, and SLAAC works fine in those cases (and neither are mandatory actually, can't be). This isn't related to IPv6 overall.

Android devices are not optional in some parts of world with low iPhone adaption.

In my previous company of 250+ people only 2 had iphones.

NAT is the worst thing that could have happened to Internet. Without NAT people would have been able to host their own blogs, pages, photos decades ago from their home. We could have had far more decentralized applications than we have today and would have prevented this walled garden mess of Internet.

The ability to address any system on the Internet with it's unique IP is what original Internet was supposed to be. NAT was temporary solution to address the issue of limited IPs but now that is single protocol that is preventing the migration to IPv6

I don't know why this is being down voted. The entire modern bullshit Internet exists because of NAT. We can't host stuff ourselves and have to use some central system. Since nothing is free, the provider has to inject more and more ads everywhere.

I bet the actual cost of NAT when all things are considered is well into the trillions of dollars.

Any attempt to keep IPv4 alive is not a net win. It's another few trillion down the drain.

Because it implies that it is the single reason that people don't host their own sites anymore, when in reality people don't because it is generally too difficult, expensive, unreliable and redundant to do so.

Almost nobody wants to host their own social media or messenger site and if you did your friends wouldn't use it too.

Besides, anyone who currently wants to do so can get a cheap VPS for dollars a month.

There is no telling what would happen. It is only difficult, expensive, unreliable because no one does it?

I think there is proof of the opposite in how people host giant torrents without much issues, some buy seed boxes, their friends use it too.

I wouldn’t put CGNAT and general NAT in the same basket.

My ISP and every ISP I’ve used since the 90s has given me a public IP, only exception is my cellular links which are CGNAT’d. I’ve got no problems hosting stuff with my current connection, while my LAN devices are on RFC1918 address space.

I guess it depends on where you live here in India only one ISP among several that I have used over 15 years gives public IP.

Agreed. NAT may or may not have made the walled gardens, but it sure killed their competition.

The worst part is that now it's not just baked into the network layer, it's baked into security assumptions as well. What a mess.

> Without NAT people would have been able to host their own blogs, pages, photos decades ago from their home.

Reliability would have been horrendous. Stuff would disappear forever much more than it does now. It would have been slow because of limited upstream.

It would have also killed the bandwidth of people not hosting themselves if they were on DOCSIS? ( cable modems ).

Still people are generally not NATed on a home/business connection. NAT is not the problem.

Around here (USA) ISPs still give publically routables IPs to customers, and the customer is responsible for running their own NAT if they want to have multiple devices on their network. Most routers support port forwarding, which allows those households to have publically routable TCP or UDP ports on any device within the network. In theory UPNP allows devices to set this up seemlessly.

It has become less common with the rise of clouds. But it used to be common for multiplayer games to be done by one player setting up their instance as a server.

*Some/most ISPs. Some smaller ISPs use CGNAT, and I believe most cellular providers also use CGNAT for IPv4.

Not implementing DHCPv6 is not "inconsistent" it's from the start a thing you should NOT rely on existing.

And for what it’s worth, linux has already supported 240/4 as unicast for many years already.

The IPv4 systems that need to be updated - the hosts - will not be. Not in a year, not in ten years. This plan is a non starter.

v4 should, at this point, basically exist as fallback prefixes for the devices that will hang around the next 30 Y without proper v6 support etc.

Routing in the small is not hard. I am absolutely baffled that people are claiming BGP is some kind of super hard thing to deal with. BGP has a ton of problems and lots of ugly duct tape (rpki, etc.) that is just done poorly but it isn’t hard.

> BGP has a ton of problems and lots of ugly duct tape (rpki, etc.) that is just done poorly but it isn’t hard.

For a lot of folks, for internal purposes, OSPF would also work pretty well (or IS-IS).

also very true

It’s probably easier to add some more addresses than to switch to IPv6 but I hope you agree that switching to IPv6 ultimately is inevitable. And then you are comparing the optional costs of adding that small number of addresses against the switch costs that you are going to incur anyway.

IPv6 has been around since the mid-late 90s, it’s an incredibly slow migration as a lot of things were just not thought through. Yes the switch will eventually happen, but we’re already talking decades, which is a lot of time to unlock a few IPv4 subnets while we wait.

The alternative take is that it’s an incredibly slow migration because we keep unlocking a few IPv4 subnets so we can wait a little longer.

> There’s a lot less effort required to update an existing IPv4 system to accept a few more ranges than to entirely rearchitect a network to work with IPv6.

I'm not so sure, actually.

It's still a case of literally billions of devices that just plain need to be replaced.

Not to mention millions(?) of multimillion dollar routers that need to be scrapped, even though they handle IPv4 and IPv6 in silicon just fine.

> But as to the actual proposition, I would rather the 240/4 space was changed to RFC1918 private space, or at least half of it.

Nobody's stopping you. It'll never be usable on the Internet anyway.

I am a bit sad that what I felt was the proposal of most immediate benefit to the internet was not the one that got all the attention. That one is here:

https://datatracker.ietf.org/doc/draft-schoen-intarea-unicas...

There is some merit to having more IPv4 space e.g. for new players trying to start a business or for IPv6 transition technologies, where you need at least some IPv4 space e.g. for the NAT64 gateway. This should be considered in the application process for the space, if the unicast extension will be widely supported technically.

The real question here is, whether Microsoft will patch Windows to support the unicast extension. More specifically, which versions and whether it will include Windows XP/ Windows Server 2003/ 7/ Windows Server 2008 and various Windows CE versions. I would say, there are so many WinXP/WS2003 machines in industrial networks still, that not having a patch enabling support might be a major blocker. In some cases, you just cannot replace the operating system without also replacing the machine it is driving.

Yes, more or cheaper IPv4 space will prolong the pain of having two protocols one of which is the default but has inherent problems in terms of scalability (IPv4).

There might also be some damage from changing 0/8 and 127/8 to be global unicast and carving out 0.0.0.0/32 and 127.0/16 for their current usage. Also the lower address as alternative broadcast to unicast switchover could be "fun" in some networks. The reality is, ransomware and other attacks are quite common and creating new doors into legacy networks with less than ideal maintenance through changing rules should be evaluated. I don't think the typical water treatment plant, hospital, pharmacy or an aluminum smelter follow even major RFCs closely.

I know, I have discovered (and fixed) the 1/8 blackhole problem in a subsidiary at my previous job. I would need to really check 0/8 and 127/8 there to verify they are ready for the unicast extension as well. That company had an IT department and the IT worked reasonably well there but there were some very specific constraints such as Windows NT driving some industrial machines. This wasn't the common case, but it wasn't the only "special" host on the network. In such places, such changes perhaps don't need to be implemented, but you never know, when the inconsistency will bite you.

Reply on news.ycombinator.com