Re: [PATCH] ipv6: Fixed source specific default route handling.

From: Matthias Schiffer
Date: Mon Jun 22 2015 - 13:51:28 EST

On 06/22/2015 07:58 AM, Steven Barth wrote:
> On 22.06.2015 00:35, Matthias Schiffer wrote:
>> Could you explain in detail what you mean with "If you want specific SA,
>> add same route with higher metric and/or (more) specific src match."?
>> Routes aren't bound to specific addresses except via the "src" attribute
>> (which is called prefsrc in the kernel), which is exactly what it not
>> working. I can't control the chosen source address at all when
>> source-specific routes are involved.
> Except that prefsrc and src are two different beasts and usually ip route from transates to
> RTA_SRC instead of RTA_PREFSOURCE when used with a prefix length.
> Try adding two routes to the same destination with the same metric but different source values with PREFSRC (e.g. IPv4) and then
> try doing the same with SRC (e.g. IPv6). The former will fail but the latter will succeed.

Ah sorry, I didn't know that "src" and "prefsrc" were distinct concepts.
I meant to refer to "src" whenever I wrote "prefsrc". What are the
precise semantics of the "src" attribute? Any RFC I can read, or is this
a Linux-specific concept?

> was the original draft for source-address dependent routing IIRC so might be a good read.

Thanks for the link, that helps a bit.

>> Even though the source-specific route has a higher metric than the
>> generic one, the source-specific one shadows the generic route.
> (was a bit ago since I read into this so please correct me if I am wrong)
> IIRC this is intentional since longest-prefix-match beats metric here
> and the source-address match counts to being more-specific here. See also above difference between PREFSRC and SRC.

Ah, that would explain the metric issue. I looks like the source of my
confusion is that for source-specific routes *all* addresses are in the
candidate set, not only the addresses of the outgoing interface (which
makes sense as ip6_route_get_saddr() is called with a NULL rt6_info in
the source-specific case).

I'm not sure if this can be fixed in a sane way (as there seems to be a
dependency cycle: source address should depend on outgoing interface,
which depends on the chosen route, which depends on the source address),
but it leads to highly unintuitive source address selection :(

Markus suggested in the commit message not to call ip6_route_output at
all before the source address has been selected. Wouldn't this make it
impossible to choose the source address depending on the outgoing
interface in the non-source-specific case as well?

> Cheers,
> Steven

Thanks for the explanation,

Attachment: signature.asc
Description: OpenPGP digital signature