Hi,

> I forgot about this. I hate to rain on your parade Stefan, but
> if you made one global worklist that will complete the discussion.

here we go. Changes since the last version:

-One global worklist. Still using  __LINK_STATE_LINKWATCH_PENDING to
know of a pending event fast
-unsigned char operstate instead of short. If there was no alignment,
this would have reduced the size of struct net_device by one byte ;-)
-removed usage if in-kernel-IFF_RUNNING as a mirror. Useless if we want
to broadcast complete operstate via netlink
-Map only NETDEV_OPER_UP and NETDEV_OPER_UNKNOWN to IFF_RUNNING. I have
kept UNKNOWN as a compatibility kludge for the majority of drivers that
cannot determine any operstate yet
-Use dev_hold()/dev_put()

While doing tests with a hacked vlan driver that creates
NETDEV_OPER_LOWERDOWN/_UP events I found that I get a "No buffer space
available" in ip monitor if the event list is longer than about 20
entries. This can be worked around with setsockopt on SO_RCVBUF, but
does anyone have a clue why netlink events are that expensive?

Cheers, Stefan