core: avahi on Linux uses incorrect address for P-t-P interface

core: avahi on Linux uses incorrect address for P-t-P interface

Hi,

I sent a less knowledgeable question about avahi-daemon and
point-to-point links a few days ago,
http://lists.freedesktop.org/archives/avahi/2011-January/001969.html.
When I didn't get a response to this, I decided to build avahi from
source and step through it and see how it builds its list of interfaces
and their addresses.

This is in iface-linux.c, netlink_callback(). It looks for a RTM_NEWADDR
message, then extracts the payload of type IFA_ADDRESS.

Short story: I think it should be using the payload of type IFA_LOCAL,
not the payload of type IFA_ADDRESS.

In the VM where I was running these experiments, there are 3 interfaces
-- lo, eth0 and tun0. I printed out the IFA_ADDRESS and IFA_LOCAL for
all 3 of these; for lo and eth0 these are the same address; for tun0
(IFF_POINTOPOINT), IFA_ADDRESS is the remote end and IFA_LOCAL is the
local end.

I'm no expert on Linux rtnetlink or these IFA fields, but quoting
/usr/include/linux/if_addr.h:

/*
 * Important comment:
 * IFA_ADDRESS is prefix address, rather than local interface address.
 * It makes no difference for normally configured broadcast
 * interfaces,
 * but for point-to-point IFA_ADDRESS is DESTINATION address,
 * local address is supplied in IFA_LOCAL attribute.
 */

See also this stackoverflow question/answer:
http://stackoverflow.com/questions/4678637/what-is-difference-between-ifa-local-and-ifa-address-in-rtnetlink-linux

Does anyone know why avahi is looking for IFA_ADDRESS here, and whether
there's any reason not to use IFA_LOCAL instead?

Assuming there's not a specific reason to use IFA_ADDRESS here, I
propose the patch attached at the end of this message, which works for
me. (The bug this fixes is described in the earlier message linked above
-- avahi chooses the wrong address to associate with the P-t-P
interface, and if you enable avahi's reflector, avahi tries to call
sendmsg() using that as the source address and the kernel always,
correctly, fails the sendmsg() call with EINVAL.)

(And when/why this changed -- I was able to use avahi over P-t-P
interfaces on Linux several years ago; I don't know what avahi version I
was using at the time.)

thanks,

Matt