Comment # 8 on bug 1093877 from Per Jessen

(In reply to Neil Brown from comment #7)

> Are you using the same initramfs file for all your clones? 

No, I build a separate one on each clone.  Years ago it was possible to have
just one for all, there was logic in the linuxrc code to pick up the MAC
address of the active interface, from <somewhere>.  I'm not sure that works
anymore. 

> This doesn't explain the difference between "good" and "bad" though.  Having
> looked more deeply, I'm surprised they aren't the other way around.
> 
> The hostname that is used in the client-id is determined inside the kernel. 
> The first time any NFS request is initiated, the current hostname is copied
> and that used ever more, even if the hostname is changed again.
> So if there is a mount attempt before dhcp completes, the value from
> /etc/hostname will be used.  If the first attempt is after dhcp completes,
> the value it sets will be used.

Right, that sounds plausible, although why would there be a mount attempt
before dhcp has completed?  

> In the case where the root filesystem is mounted using a hostname, you
> presumably need to have the network up before a DNS lookup can succeed. 

I don't know why one uses ip address, the other hostname. Maybe I need to take
a closer look at that.

> By the time it succeeds, the new hostname should be set.  If, however, the
> mount is performed with an IP address, that could start trying before the
> network is up.
> So I do wonder if you got those backwards.  Though maybe the host name
> 'rainbox' is in /etc/hosts.

No, nothing in /etc/hosts.  Well, I might have mixed up those initrds, but it's
easily reproduced. 

> I think the short answer is that dracut needs to make sure dhcp completes
> before it even tries to mount an NFS root filesystem.  I'm surprised it
> doesn't already,

Me too, that is really odd.