Re: route: an issue caused by local and main table's merge

[Guillaume Nault <gnault@redhat.com>]

On Tue, Mar 10, 2020 at 10:19:24AM -0700, Alexander Duyck wrote:
> On Tue, Mar 10, 2020 at 9:01 AM Guillaume Nault <gnault@redhat.com> wrote:
> >
> > On Tue, Mar 10, 2020 at 04:56:32PM +0100, Guillaume Nault wrote:
> > > On Mon, Mar 09, 2020 at 08:53:53AM -0700, Alexander Duyck wrote:
> > > > Also, is it really a valid configuration to have the same address
> > > > configured as both a broadcast and unicast address? I couldn't find
> > > > anything that said it wasn't, but at the same time I haven't found
> > > > anything saying it is an acceptable practice to configure an IP
> > > > address as both a broadcast and unicast destination. Everything I saw
> > > > seemed to imply that a subnet should be at least a /30 to guarantee a
> > > > pair of IPs and support for broadcast addresses with all 1's and 0 for
> > > > the host identifier. As such 192.168.122.1 would never really be a
> > > > valid broadcast address since it implies a /31 subnet mask.
> > > >
> > > RFC 3031 explicitly allows /31 subnets for point to point links.
> > That RFC 3021, sorry :/
> >
> 
> So from what I can tell the configuration as provided doesn't apply to
> RFC 3021. Specifically RFC 3021 calls out that you are not supposed to
> use the { <network-prefix>, -1 } which is what is being done here. In
> addition the prefix is technically a /24 as configured here since a
> prefix length wasn't specified so it defaults to a class C.
> 
Yes, I was just replying on the use of /31 subnets. I agree that this
case is different.

> Looking over the Linux kernel code it normally doesn't add such a
> broadcast if using a /31 address:
> https://elixir.bootlin.com/linux/v5.6-rc5/source/net/ipv4/fib_frontend.c#L1122
> 
Yes, and that's the right thing to do IMHO.

I think the original problem is that the command is accepted when it's
run after "ip rule add from 2.2.2.2". It should continue to be rejected
instead, as the ip-rule command has no action and is not supposed to
interfere in this case.