svn commit: r297326 - head/sys/nfs

[Ian Lepore]

On Sat, 2016-04-02 at 23:39 +0800, Julian Elischer wrote:
> On 28/03/2016 7:16 AM, Ian Lepore wrote:
> >    The theory here is that since we can only have one default
> > route, the one
> >    most likely to be correct for mounting the rootfs is the one
> > that was
> >    delivered along with the rootpath option.
> 
> We can have multiple default routes with weights, and we can have 
> different default routes in different fibs.
> since interfaces can have fibs assigned to them you need to install 
> the new default if it is useable in the fib of the interface.
> This is complicated and most companies that use these features have 
> special code to do it..

Does any of that apply during bootpc_init(), before rootfs is even
mounted?  The dhcp server just delivers a router address, and on a
multihomed machine you can end up with different addresses coming from
different servers (which may be insane as an actual working config, I
dunno).  In testing what I found was that only the first address
delivered ever actually got installed and all subsequent ones resulted
in an (ignored) EEXIST error.

If what you say applies to the bootp milieu, I'm going to need help
making it right.  I picked up that PR when I stumbled across it because
I'm somewhat familiar with the bootp routines from earlier bugfixes,
but the PR really addressed two issues, and on the default route vs.
proxy-arp parts of it I'm out of my depth.

I suspect in the real world there probably aren't too many
configurations that require a gateway to access the nfs server
providing a machine's rootfs.  It seems fair to assume that if a
gateway is required to reach that machine, then the dhcp server that
told you about that nfs server will also be the one to give you a
useful gateway address for reaching it.  I would expect more complex
routing stuff to be handled by the rc scripts.

-- Ian