Hacker News Reader

Comment by kazinator | Hacker News Reader

A sane extension to IPv4 using 64 bit addresses, and not disturbing anything unrelated to address space exhaustion: not replacing ARP with the ICMP-based neighbor bullshit (so that you now need an IP address to get an IP address) and so on.

https://en.wikipedia.org/wiki/Second-system_effect

Even the use of colons in the textual IP address notation is moronic. The colon is already used in a notation for port numbers: 1.2.3.4:567. Under IPv6, you have to use square brackets around the address if it is followed by colon-delimited port number. Ugh!

A 64 bit extension to IP can just use a four-part dotted address, where the parts are 16 bit decimals. Or perhaps an 8 part dotted address. A range of zeros can be two consecutive dots: 10..1 would be 10.0.0.0.0.0.0.1.

View on HN

I quite like the clarity of using link-local addresses within IPv6 addressing, along with scoped multicast. Replacing the functional of ARP from within the protocol (as opposed to ARP itself, which operates as a sibling to IPv4 with its own Ethertype) is quite elegant. DHCPv6 operating without the hack of using destination 255.255.255.255 and source 0.0.0.0 addresses, and thus without requiring special OS level privileges, also feels like a nice simplification.

Upgrading to 64-bit address space would also require a rip and replace of all carrier routers, customer CPEs, and re-writing of software.

We might as well take advantage of a compatibility breaking upgrade to fit in all of the other upgrades.

> We might as well take advantage of a compatibility breaking upgrade to fit in all of the other upgrades.

These are the shoals that so many projects have foundered upon.

... and that's how you end up with current situation

Replacing inaddr_t with inaddr6_t is trivial, and does not require any significant effort at all.

Replacing everything else is hard. My house is on IPv4 and will be on IPv4 for foreseeable future, because all of my setup (filtering, metering, monitoring, fixed IP assignment) is completely unusable in the new IPv6 world with privacy extensions where addresses are not only random, but change constantly.

You do know of RFC 7217 address assignment resulting in stable addresses as long as you don't move into another network? It's even the default on some Linux distributions.

If you want predicatable addressing, you can always forego privacy extensions and use EUI64-style interface identifiers.

Yes, I know about RFC 7217. But I also know that default network setup in recent Linuxes, Android and iOS all uses privacy extensions. And even if they can be disabled, I want to keep them -- they are there for a reason, I don't want to give every single internet website a permanent, unique identifier.

This is not an insolvable problem. I can write another daemon to keep the history of IP<->Mac associations, change my accounting scripts to look the matching MAC, change the firewall to filter based on mac address, and so on -- but it is a lot of work, and no benefits.

The NAT is not a problem at my home network, generic openwrt router has enough CPU power for my bandwidth. Direct addressing for each device is useless, as I would still have to list every port and every device I want to expose -- even if NAT goes away, the firewall with default-deny incoming policy stays.

This, IMHO, is the biggest problem with IPv6 adoption. If this was only "recompile the problem and change IP address regex", then we would be done ages ago. Instead, it just got strictly worse for most small network owners.

How does that solve the incentive problem? (everyone wants to keep supporting IPv4 so it continues to work fine and there's no incentive to upgrade)

They should have just extended IPv4 similar to how UTF-8 works.

Which would still require everyone to buy new routers.

People bought several generations of new routers since work on IPv6 began in the early 1990's. If that work had been sanely scoped, we'd all be using IPv6 by now.

People bought new routers, for example, because of faster ethernet and fibre. Faster switching, more ports. Bigger routing tables. At the last hop, also due to Wi-Fi and rounds of newer Wi-Fi.

A smart upgrade to IP could have been hammered out by a team of half a dozen skilled engineers well before Y2K rolled in.

We already have the best upgrade that's actually possible given the design of IPv4 -- the fundamental problem is a combination of the pigeonhole principle (which means there's no way to allow unmodified v4-only hosts to talk to v6 hosts) and the fact that there's nobody in a position to force people to use the necessary modifications.

As a result, we end up with several generations of hardware that support everything necessary but where the people who wrote the config for them are too lazy to actually turn it on.

So much this. NDP is a horrible mess.

Reply on news.ycombinator.com