Re: open sockets preventing unregister_netdevice from completing in linux-next (next-20120724)

From: Eric Dumazet <redacted>
Date: Thu, 26 Jul 2012 19:15:28 +0200

On Wed, 2012-07-25 at 15:17 -0700, David Miller wrote:

quoted

From: Eric Dumazet <redacted>
Date: Wed, 25 Jul 2012 16:38:48 +0200

quoted

Yes, we miss what was done with rt_cache_flush() : find all cached
routes and release all dev references...

We can fix this with a two-pronged approach:

1) Walk the FIB info nexthops and invalidate.

2) Entries not cached in the FIB info nexthops go into a
   per-netns list which is scanned as well.

I'll try to work on this if nobody beats me to it.

With your latest patch, I can "rmmod tg3" while sockets are active.

Not sure we need all this now ?

(the trick is probably in fib_semantics.c, when you changed
dst_release() to dst_free())

That's not the problem.  Not all routes are cached in the FIB nexthop.

Any route that doesn't resolve to a FIB info (255.255.255.255, etc.)
or uses special features (tclassid, etc.) doesn't get cached.

Therefore if a socket gets that kind of route, and then becomes
inactive, we can hold onto the device references forever.  The entity
with the route has to call dst_ops->check() to see that the route has
become invalid, but if the application is inactive, that will never
happen.

We have to put such non-cached routes into a special list or table of
some kind, so that we can zap a netdevice if it wants to go down or
unload before the final dst_release() happens.