Re: [PATCH 0/4] Fix routing metrics

From: Steffen Klassert <steffen.klassert@secunet.com>
Date: Thu, 9 Feb 2012 13:44:11 +0100

On Wed, Feb 08, 2012 at 03:18:37PM -0500, David Miller wrote:

quoted

From: Steffen Klassert <steffen.klassert@secunet.com>
Date: Wed, 8 Feb 2012 08:30:37 +0100

quoted

On Mon, Feb 06, 2012 at 03:29:16PM -0500, David Miller wrote:

quoted

Thinking about this, it seems overkill to check this on every metric
access.

You have an opportunity to validate metrics when the peer is bound to
the route.

This is because any change to the FIB metrics, is in turn a change
to the FIB, which therefore invalidates the entire routing cache.

So you can be sure that a new route cache entry will be created, and
at that creation time you can ensure that we'll respect the updated
FIB metrics if encessary.

Not sure if I get you right here, but that's what this patchset
does. It invalidates the metrics on the peer by incrementing
peer_genid in rt_cache_invalidate() which is invoked on every FIB
change. Then, on slowpath route lookup it checks in rt_init_metrics()
whether the peer_genid changed. If it changed, it exchanges the
invalidated merics with new ones and copies the informations from
the FIB into it.

If the routing cache is invalided, you'll "see" this updated inetpeer
because every single routing cache entry will get rebuilt.

Hm, I still don't get your point. Could you specify this please?

When a route cache entry is created and the peer_genid does not match
the genid on the inetpeer, fresh inetpeer metrics are allocated  and
then published. After that, the new metrics are in proper state and
ready to use.

Right, which is exactly what you want to happen.

Checking on every metric access is therefore pointless and needless.

The peer_genid only increments when the routing cache is flushed,
therefore every subsequent access to the metrics will go through the
route cache entry creation path first, and therefore that will make
sure fresh inetpeer metrics will be allocated since the peer_genid
does not match.