Re: Outbound interface

[Dennis Ferguson <dennis.c.ferguson%gmail.com@localhost>]

On 20 Apr, 2012, at 06:46 , Mouse wrote:

>> Is it entirely up to the routing to determine which interface/wire to
>> send out a packet on?  [...]
> 
>> I must admit that I would have guessed that bind():ing would have
>> some kind of bearing on outbound interface, [...]
> 
> See RFC 1122 section 3.3.4.2.  You are seeing a dissonance between the
> strong-ES model your guessing expects and the partially-strong
> partially-weak model NetBSD actually implements.

Note, though, that even if NetBSD did implement a strong-ES model it still
wouldn't send the packet out the bge0 interface; it would drop it instead.
You need to route on the destination address and if the only route you
have to the destination is through bge1 then either the packet is leaving
the box through bge1 or it is going no where.  I think the former is generally
more useful behavior.  Implementing a "perfect" strong-ES model requires a
separate routing table per interface, something which NetBSD doesn't do (but
which would be a quite useful configuration option to have available, for this
and for other uses).

I would also object to the notion that a local address on an interface "names"
that interface in any useful way.  While I admit this is arguable, and is
slightly less clear for ethernet interfaces because of additional constraints
related to the use of ARP, in general the only practical function of a local
address on an interface is to provide a default for local address selection
procedures, in particular one which maximizes the probability that the recipient
of the packets you send will know a route back to you.  Because this is all the
local interface address really does it is not unreasonable to configure the
same local address on many interfaces, to configure local addresses not
associated with any interface (generally implemented by adding them to the
loopback) or to operate interfaces with no local address configured (which
you might do if the fallback default local address selection provided a
useful result).  A local address may hence "name" more than one interface,
or no interface, and an interface is not guaranteed to have a local address
"name", which really suggests that the local address doesn't "name" anything
in a useful way.  An interface name is actually "bge0", or some software
equivalent (like the if_index), since that does unambiguously name something.

Given this, I would argue that all bind()ing a local address to a socket does 
for
outbound packets is to override the kernel's most-likely-to-succeed selection of
a local address with one of your own choosing.  It alters nothing else, the
outbound routing to a destination still is what it is.  You are permitted to
do this if it is useful to you, but you are on your own if you do since you
are in effect overriding what would generally be expected to be the "best"
local address to pick with something else.

Dennis Ferguson