Support multiple IP addresses per interface

[jorgensigvardsson]

The setup:

• interface vlan1
  • 192.168.2.200
  • 169.254.29.54 (link local address

mdnsd sends out a query for its own services, presumably to avoid conflicts. The queries have the source address 169.254.29.54:5353 and destination 224.0.0.251:5353. mdnsd somehow responds to its own query, and it says it has the IP-address 192.168.2.200. The response however has the source address 169.254.29.54:5353.

The log files are filled with messages like these:

Jan  1 00:10:56 ptwos-coronet-61-9f-20 mdnsd[1424]: vlan1: conlicting name detected ptwos-coronet-61-9f-20-167._ssh._tcp.local. for type 16, reloading config ...
Jan  1 00:10:56 ptwos-coronet-61-9f-20 mdnsd[1424]: vlan1: conlicting name detected ptwos-coronet-61-9f-20-168._ssh._tcp.local. for type 33, reloading config ...
Jan  1 00:10:56 ptwos-coronet-61-9f-20 mdnsd[1424]: vlan1: conlicting name detected ptwos-coronet-61-9f-20-168.local. for type 1, reloading config ...
Jan  1 00:10:56 ptwos-coronet-61-9f-20 mdnsd[1424]: vlan1: conlicting name detected ptwos-coronet-61-9f-20-168._ssh._tcp.local. for type 16, reloading config ...
Jan  1 00:10:57 ptwos-coronet-61-9f-20 mdnsd[1424]: vlan1: conlicting name detected ptwos-coronet-61-9f-20-169._ssh._tcp.local. for type 33, reloading config ...
Jan  1 00:10:57 ptwos-coronet-61-9f-20 mdnsd[1424]: vlan1: conlicting name detected ptwos-coronet-61-9f-20-169.local. for type 1, reloading config ...
Jan  1 00:10:57 ptwos-coronet-61-9f-20 mdnsd[1424]: vlan1: conlicting name detected ptwos-coronet-61-9f-20-169._ssh._tcp.local. for type 16, reloading config ...
Jan  1 00:10:58 ptwos-coronet-61-9f-20 mdnsd[1424]: vlan1: conlicting name detected ptwos-coronet-61-9f-20-170._ssh._tcp.local. for type 33, reloading config ...
Jan  1 00:10:58 ptwos-coronet-61-9f-20 mdnsd[1424]: vlan1: conlicting name detected ptwos-coronet-61-9f-20-170.local. for type 1, reloading config ...
Jan  1 00:10:58 ptwos-coronet-61-9f-20 mdnsd[1424]: vlan1: conlicting name detected ptwos-coronet-61-9f-20-170._ssh._tcp.local. for type 16, reloading config ...

If I drop either address from the interface, it has no problems.

[jorgensigvardsson]

This is really mind boggling. Tried the same setup on my PC, and mdnsd could handle multiple addresses just fine.

It's as if the multicast packets are bouncing around in my network. 🫣😵‍💫

[jorgensigvardsson]

Or rather, the mdns packets are sent to the network from the interface, as well back onto itself! But only on my embedded device. It's as if IP_MULTICAST_LOOP is enabled and mdnsd does not recognize the sender address (its own address).

Right now I feel like I'm rubber ducking myself on your issue page, sorry about that. You might have a clue what's going on. Either way I think I have a lead to what I should focus my debugging efforts on on Monday. 😬

[troglobit]

I have no idea what could be going on here. The hostname in your logs rings familiar though, knowing it's a switch-based embedded platform, are you sure the frames are not looped back from the switchcore? You can test the theory by running tcpdump on the lower-level port interfaces or inspecting the RMON counters for multicast in/out.

[jorgensigvardsson]

Thanks for the tip! It will be fun figuring this out!

[jorgensigvardsson]

So, now that it's not Friday and my brain is a bit less fried, it seems like my issue has been fixed in 0.12.

[troglobit]

So, now that it's not Friday and my brain is a bit less fried, it seems like my issue has been fixed in 0.12.

He he, OK 😏

[jorgensigvardsson]

I think I may have to reopen this issue. I think in fact there is a bug in mdnsd with respect to multiple IP addresses per interface. Here's my reasoning:

1. At startup all interfaces are enumerated using getifaddrs() in iface_init()
   1. Any interface X is iterated for each address on the interface
   2. The first address for interface X is recorded, all other addresses are in effect ignored
2. Found addresses in step 1 are attached to the iface->mdns instance (one address per interface only)
3. In mdnsd_in() answers are processed at the end. The source address is checked against the address for the current interface's mdns instance. Because only one of the interface's addresses are recorded as "ours", any packet received for any of the other addresses on the interface, the result is a conflict, causing mdnsd to reconfigure itself
4. After reconfiguration, the entire probe and conflict loop is started again

It's odd that the source address of the packet going out from mdnsd is the first address of the interface, but the bound address of the multicast socket seem to be the last address of the interface. I can't really comment on this behavior, or if it's a proper behavior of the system.

One solution to this particular problem would be keeping a list of all interface addresses and compare against the list in 3. Not sure if that's a proper solution though if my system is behaving badly.

Do you have any thoughts about this?

[jorgensigvardsson]

(I've tested and debugged this on 1.12 by the way)

[jorgensigvardsson]

I think that what I'm saying is that:

1. When mdnsd starts up, it picks one address as "this is us"
2. bind() picks one address from the interface and uses that as sender address

These addresses don't always match, and that causes mdnsd to think there are conflicts when it sees its own packets.

[troglobit]

Hmm, yeah that makes sense. We could do a better job of that.

Unfortunately I don't have the time (or energy) at the moment to help out. Maybe someone else who has been active recently in the project is interested?

[jorgensigvardsson]

I'm actually trying to fix it right now, but with limited success. Limitations set by too much caffeine. 🥴

I thought I had a solution for it, but I fail to grok the mark and sweep stuff. In my eyes, it has no effect, but I'm sure I've missed something. The mark and sweep algorithm is now also a different beast, as the changes I'm working with has changed the interface:address ratio from 1:1 to 1:n.

I'll give it a go again tomorrow once this caffeine overload has simmered...

[iam-TJ]

Came across this issue whilst checking if I can leverage libmdnsd to build a transparent translating proxy a-la squid.

I run IPv6-only networks, multi-homed hosts, and for IPv6 multiple addresses of different scopes (link-local, site-local (ULA), and global) are commonly attached to each interface and I've seen similar issues to this with both avahi and systemd-resolved that due to either 1) advertising only a sub-set of IP addresses or 2) unpredictable ordering of the IP addresses (global > ULA > link-local) causes hard-to-diagnose client failures of various types.

Avahi has an optional nice way to prevent getting confused by itself or multiple mDNS instances all publishing:

[publish]
add-service-cookie=yes

This adds org.freedesktop.Avahi.cookie= to the TXT record where the value is unique to the avahi process. A similar approach for mdnsd might be appropriate.

However, for me, I require the responder to publish all addresses belonging to an interface but with an option to tell it which scope(s) to publish or to set an ordering of scopes, types (static, dynamic (DHCP or RA) or addresses.

My use-case is to proxy and transparently translate multi-function printer resource records from a VLAN where they are isolated and be able to:

1. optionally translate from IPv4-only devices to IPv6 address(es) of the host
2. optionally translate the Port=
3. optionally drop some records. E.g. in a CUPS environment where CUPS shares the printers the proxy drops the _ipp._tcp/_printer._tcp records but allows _scanner._tcp such that clients print via the CUPS printserver but can directly discover and use the scanner.