Re: [patch net-next v2] ipv6: log autoconfiguration failures

[Hannes Frederic Sowa <hannes@stressinduktion.org>]

On Wed, Dec 11, 2013 at 03:54:52PM -0500, David Miller wrote:
> From: Hannes Frederic Sowa <hannes@stressinduktion.org>
> Date: Wed, 11 Dec 2013 20:21:38 +0100
> 
> > On Wed, Dec 11, 2013 at 12:45:14PM +0100, Denys Vlasenko wrote:
> >> If ipv6 auto-configuration does not work, currently it's hard
> >> to track what's going on. This change adds log messages
> >> (at debug level) on every code path where ipv6 autoconf fails.
> >> 
> >> v2: fixed indentation in multi-line log output statements.
> > 
> > Have you seen lots of those problems? Some of those seem like very
> > serious problems and maybe could also deserve a pr_warn or pr_err.
> > 
> > I hope these are one-time errors, so I don't think counters would
> > be helpful.
> 
> I still think that statitics would better serve this issue.
> 
> For one thing, the event would always be counted, whereas with
> pr_debug() someone has to turn on dynamic debugging in order
> to see the message.

I proposed switching some of these to pr_warn/err. pr_debugs get used very
seldomly, I agree.

> You can make them part of the per-inet6_dev MIB, and therefore
> implicitly letting the admin know what interface the events
> occurred on.

If such an error happens I would try to recreate the interface (in case of
virtual ones), thus erasing the statistics.

> I would even prefer an approach involving signalling netlink
> events in such circumstances with more detailed information.

That sounds good. But I would like to hear how often these problems
really occur first. Maybe Denys has some numbers? If that happens a lot
I would like to check if the notifier_chains are correctly setup so that
interfaces correclty notify ipv6 to retry link local address generation.

Greetings,

  Hannes