Re: slowly or no configured connection usage

[Dan Williams <dcbw redhat com>]

On Mon, 2015-08-03 at 03:00 +0200, Ingo Flaschberger wrote:
> Hi,
>
> I use nm 0.9.10 with debian jessie at a raspberry pi.
>
> Problem 1:
> If I configure a single connection with auth ipv4, auto ipv6, auto 
> connect and delete all other connections it takes severall minutes after 
> reboot till the connection is used.
> Also a new "eth0" connection is always generated.
> Connection name is "internet", config is attached.

I think this one is probably due to the kernel-assigned IPv6LL address
that the device gets when something in the boot process brings the
interface IFF_UP.

> Problem 2:
> If I configure a single connection with static ipv4, link-local ipv6, 
> auto connect and delete all other connections after reboot the 
> connections is never used.
> Only a new generated "eth0" connection is used.
> Connection name is "emergency", config is attached.

This is due to the same reason.

> Digging into code - I wonder why a new connection is generated and why 
> nm_utils_match_connection / check_possible_match / check_ip6_method  and 
> check_ip4_method behave such strange?

When NM starts up, it will read the existing interface configuration and
attempt to match it to a stored connection.  If it doesn't match, NM
tries very hard not to touch the interface, because something else
configured it and NM won't blow that away.

In case #2, I'll bet that the generated connection "eth0" was
ipv4.method=disabled, ipv6.method=link-local.  The connection you
created was ipv4.method=manual.  On bootup, the interface has only an
IPv6 link-local address, but not the static address from your
connection.  Therefore, the runtime config of the interface (ipv6ll
only) doesn't match any stored connection, and the existing
configuration wins.

We realize that treating an interface with only an IPv6LL address isn't
very useful though, especially since the kernel will assign an LL
address whenever the interface is brought IFF_UP, which often happens
during boot for various reasons.  Therefore, NM 1.0.4 and later have
switched to ignoring interface configurations that have only an IPv6LL
address, and will use another configured connection instead.  We could
potentially backport that to 0.9.10, or at least make the patch
available to distros.

Dan